See what else came in today at byblos.digital →

The feed mix this week was womething like this.

In short, what we received was:

• Fintech / web3? Sell the plumbing.

• AI? Own your data and workflow, rent the model.

• Raising anything? Lead with efficiency.

Continued: Wall Street on chain

Robinhood launched its own blockchain, and Apollo — old-money finance, over a trillion dollars in credit — bought into a crypto lending protocol.

BlackRock plugged a stablecoin (USDe) into Aladdin — its risk software that trillions of dollars run through, so this is crypto getting a real seat at the grown-ups’ table. DTCC — the invisible plumbing behind almost every US stock trade, sitting on ~$114T in custody — set an October date to start doing this onchain. Securitize went public on the NYSE. All in two weeks.

The one to actually watch is the Robinhood’s July 1 event in London, “The World is Flat.” They shipped Robinhood Chain — a fast, cheap layer built on Arbitrum, with no token of its own (fees paid in ETH, a clear “we’re useful, not pumping a coin” move). Tokenized stocks now trade in 120+ countries. And their 7% “Earn” product quietly runs on Morpho — the same DeFi rails the crypto natives use. So an app with ~28M users just put its customers onto DeFi without them noticing.

Institutions came for the boring stuff: 24/7 markets, settlement in seconds not days, yield they can pass on.

The opening is the plumbing institutions now have to buy — credit rails, custody, compliance, price oracles, token issuance. Real demand, real budgets. Building there? Timing’s on your side. Building the 400th retail wallet? Read the room.

And the money’s showing up: real-world assets onchain went from ~$5.4B to ~$19B in about 15 months (+257%). Trackers count it differently and some go higher — so cite your source.

The tiny team is the new flex

Stripe Atlas (they see who’s incorporating) says solo founders are now 63% of new US companies — a record. And a Harvard/INSEAD study of 515 startups found the AI-native ones needed ~40% less capital — and it was a proper controlled study, not a survey, which is why it lands.

Zoom out: revenue per employee (total revenue ÷ headcount) runs ~$300K at old-school SaaS. Lean AI-native startups do $2–4M.

And solo operators clearing $1M+ a year more than doubled from 2023 to 2025.

The anchor is Elias Torres. He sold Drift for $1.2B, watched it swell to thousands of people, and capped his new company (Agency) at 100 — while aiming for $1B in revenue, 80–90% engineers. Building big on a tiny frame.

AI made small teams possible, and also made big teams a liability — more burn, more coordination, more dilution, for output a sharp small team now matches.

Capital efficiency is the price of admission now. Raise less than you think you need, stay small on purpose, and put revenue-per-employee near the top of your deck. “We did this with six people” is becoming more and more common.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

See what else came in today at byblos.digital →

VCs are paying for the layer above the model

Open-source AI was the fastest-moving topic in the newsletters this week. The trigger was GLM-5.2, an open-weight model from the Beijing lab Z.ai. It beats GPT-5.5 on some coding benchmarks and runs at about a sixth of the price. Open-weight models now cost 50–90% less to run than the frontier, and API prices have dropped 90%+ since 2023.

Madrona said that once the model gets this cheap, it stops being the prize. The value moves up to whoever orchestrates it - the routing, the governance, the wrapper that sits on top.

Gradial, an enterprise wrapper, took a $65M Series C from Insight Partners. Baseten raised $1.5B on inference — the plumbing that runs models, not the models themselves. OpenRouter's Fusion doesn't bet on one model at all; it routes prompts across many and beats any single one.

Right now VCs are writing checks to orchestration and inference startups — Gradial, Baseten, OpenRouter — over teams trying to train a cheaper base model. The reason is simple: the cheap base model already exists, and it's open and Chinese. If you're in AI infra, that's the side of the stack the money is on this week.

One caveat: New Capital Partners pointed out it's still not clear who actually keeps the money in this layer. And Databricks admitted one of its "cheap" open training runs really cost about $20M once you counted downtime and mistakes. The thesis is hot.

The smart money already crossed into Chinese open models

The whole China cluster lit up at once this week. GLM-5.2 showed up new. Qwen mentions rose along with those about DeepSeek. The top funds are covering Qwen 7.1× more than the rest of the field. When elite funds are that far ahead of everyone else on a topic, it's usually early.

What lit the fire was policy. On June 12 the US government ordered Anthropic to switch off its top models (Fable 5 and Mythos 5) for all foreign nationals, worldwide, on national-security grounds. Anthropic complied within hours. Overnight, "which model do I build on" turned into a geopolitical question.

Nazaré Ventures' line: "virtue is not a business model" — meaning being the good guys didn't protect Anthropic when the state flipped the switch. Zhipu's founder Tang Jie went the other way and dunked on Elon Musk's timeline, saying Chinese models would close the gap faster than Musk predicted. Both point the same direction: the closed frontier looks less safe to depend on than it did a month ago.

The money's already moving to the open side. Poolside raised at a $12B valuation (Nvidia, Bain, DST). Mistral is sitting around $23B. These are the open-weight bets, and they're getting funded at scale while the closed leaders deal with export controls.

If you build on top of one closed model, that dependency is now a risk an investor can see in your data room. The funds tracking it earliest are the T1s over-covering Qwen and DeepSeek — Social Capital, Madrona, Nazaré, New Capital Partners. If you're in this niche, those are the newsletters to read and the names to pitch.

The stablecoin fight moved from the coin to the money behind it

TradFi walked into stablecoins this week without issuing a single coin. State Street showed up new in the data. Fidelity showed up new. Franklin Templeton jumped a lot. State Street is also being over-covered by the top funds at 3.5× the field — another early-money signal.

Insight Partners said the real question isn't whose logo is on the coin, it's who manages the reserves behind it. State Street and Fidelity both launched money-market funds built to hold stablecoin reserves under the GENIUS Act. Fidelity's launched June 17.

It's not just funds. Avalanche launched a Payments Collective with 28 organizations, including Franklin Templeton, VanEck, Paxos and Anchorage — TradFi institutions plugging into crypto rails directly. On the venture side, the checks went to the rails and the on-ramps, not to new coins: Flutterwave raised a $3.2B Series E led by Ripple, Trace Finance took $32M (CoinFund), El Dorado got $9M (Paradigm).

Meanwhile the field is about to thin out. Europe's MiCA licensing deadline hits July 1, and only 194 of more than 3,000 crypto firms have a license. Binance may get rejected in Greece. Tether's USDT is already getting pulled from major EU exchanges. The unlicensed players are about to be cleared off the board.

Right now the money is going to reserve management, payment rails and compliance infrastructure — State Street, Fidelity, Flutterwave, Trace — not to "launch another stablecoin." If you're building in this space, a real EU license is turning into a moat, and the reserve layer is where the institutional attention sits. The funds to watch here are Insight Partners and Theta Capital.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

Five months ago, VC newsletters were writing about robotics, defense, climate, crypto, AI, and a dozen other categories in roughly equal proportion. This week, across the 600+ newsletters we track at Byblos, the conversation has compressed dramatically.

AI is splitting, but the AI orbit is widening

A year ago, "AI" was one thing VCs wrote about. This week, it's at least five separate active narratives - AI labs, AI agents, the capex bubble debate, AI regulation, AI coding tools - each growing or sustained independently.

Lenny Rachitsky's interview with Fiona Fung from the Claude Code team this week is the artifact form of that shift. It's about how to manage an engineering team when "roles are blurring and agents are everywhere." Compare to how VCs wrote about "the cloud" between 2010 and 2014. The conversation moved from "is this real" to "how do we operate inside it."

What stopped being a story

The other side of compression. Several categories have almost left the VC conversation this month:

• Robotics and humanoids — coverage roughly halved compared to its peak earlier this year

• Quantum computing — barely surfacing

• Climate and energy tech — declining sharply week over week

• Geopolitics and macro framing — fading from posts where it used to anchor every other paragraph

CryptoSlate's piece on Strategy's STRC margin call this week is the inverse of the AI story. It documents a thesis breaking - the corporate Bitcoin treasury trade - without the wider VC ecosystem rallying around it. A year ago, Strategy's first major selloff would have generated industry-wide commentary on whether the treasury thesis is alive, but not this week.

One unexpected exception: space and defense

Almost everything physical-world is cooling. The exception this week: Defense & Aerospace appeared as the sector with the largest jump in fund coverage, and Space & Aerospace cleared the threshold as a breakout theme. Securities' piece on North Korea using crypto and defense tech to build influence is a small example of the wider arc - defense and aerospace are returning as a venture topic after years on the margin.

We're flagging this with caution. The trajectory shows a jump but the underlying coverage is still thin. It could be the start of a real reversal or it could be a one-week artifact.

The self-referential turn

a16z crypto published a piece this week titled "The new comms playbook: show, don't promise." The thesis: builders need to demonstrate substance instead of broadcasting future plans. The irony is unsubtle - a fund whose own content has become more frequent, more self-referential, and more concerned with the meta of investing than with specific portfolio bets, telling builders to focus on substance. The pattern is across the industry. When external markets become harder to call, internal industry meta-content

What this means if you're raising

If you're a founder pitching VCs right now, this matters in two practical ways.

First, the lanes that VCs are actively writing about are the lanes their attention is actually in. If you're in AI agents, infrastructure, or any AI-adjacent category, the conversation is loud and the capital is following. If you're in physical robotics, quantum, climate hardware, or non-AI-connected crypto, the room is quieter than it was 12 months ago.

Second, in a narrowing conversation, specific positioning matters more than category. The funds that publish are watching for builders who occupy a particular lane within a particular narrative - not generalist pitches across an entire category.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

What 700 VC newsletters chose not to cover this week

Funds aren't lazy. Attention just compresses into specific lanes, and major events outside those lanes get filtered out.

Three blind spots from this week.

1. Energy and grid infrastructure

The most-discussed AI-adjacent topic this week was the capex bubble debate. Fourteen funds covering whether all this AI spending makes sense.

Almost none of them wrote about the actual physical infrastructure that capex pays for. Data center buildouts. Grid permitting. Energy contracts. Cooling.

This is where every billion-dollar AI commitment becomes real physical capacity. It's also the bottleneck - US data center permitting takes years, grid queues are full, power costs are climbing. We saw one substantive piece on this across our entire newsletter universe this month.

If you're building in energy or grid software: quiet category, light competition, structurally necessary for AI to play out. Funds will rediscover it eventually, usually 6-12 months after newsletters start covering it.

2. The Series B-C middle

Pre-seed coverage is loud. Mega-rounds at the top get extensive writeups. Everything in between is quiet.

This week Iterative VC sent a "got a weird AI project? we want to fund it" email - friendly pre-seed signal. Same week, a bunch of Series B and C rounds closed at companies you've never heard of, in sales intelligence, vertical SaaS, dev tooling. Almost no coverage.

Newsletters cover headline rounds and feel-good anecdotes. The boring middle - $30-80M raises at $200-500M valuations to build infra for real revenue - falls into silence.

But that's where most actual venture returns come from. When the writing stops covering it, founders raising there lose the narrative cover and the warm-intro flow that comes with "I read about you."

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

On Friday, at 5:21pm, Anthropic got a letter. By that evening, two of the most advanced AI models on the planet were switched off for everyone.

The company that built its entire brand on safety was shut down by the government, citing safety. Fable 5 and Mythos 5 - gone for every customer worldwide, because the order banned access for any foreign national, and Anthropic decided the cleanest way to comply was to pull the plug on all of it.

It seems that the switch that controls your core feature might not be in your hands, your office, or even your country. You rent most of your stack. And some of your landlords live in jurisdictions where you don't get a vote.

This week it became closer to a board-level question, because the order didn't stop at the hardware. The US has restricted the chips behind AI before. This was the first time it reached the models themselves - the thing founders actually build on. And the calmest line in Anthropic's statement was the one that mattered most: all other models were unaffected. The risk was never the vendor. It was depending on one thing you can't replace.

The fix is:

• Abstract the layer. Build your product so you can swap the model (or the rail, or the provider) in a day.

• Keep a second source for anything that fails the 30-day test.

• Stop confusing "it works great" with "it's safe to depend on."

It doesn’t mean avoid US tools, frontier models, or move to a bunker. The best infrastructure available still sits with a handful of providers and there's no clean way around that.

Anthropic, for what it's worth, called the whole thing a misunderstanding and is working to turn the lights back on. Maybe it gets fixed by the time you read this. That's not the part that should keep you up. The part that should is simpler

If you’re a founder, here’s the uncomfortable exercise. Go through your stack and ask one question about each piece: could I rebuild this in 30 days if it disappeared tomorrow? Anything that gets a “no” is a dependancy.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

SAFE, Convertible Notes, or ASA?

Before your first round closes, someone will ask which instrument you're using: a SAFE, a convertible note, or an ASA. All three do the same basic job - they let you raise money now without agreeing on a valuation yet. Instead of arguing over what the company is worth today, you take the cash and agree it converts into equity at your next priced round, usually at a discount or a capped valuation that rewards the investor for being early. Founders use them because they're fast, cheap, and don't force a valuation fight before you have the traction to win it.

That much they share. Where they differ is what you're actually on the hook for in the gap between signing and converting.

A SAFE - short for Simple Agreement for Future Equity, invented by Y Combinator - is the simplest of the three. The investor gives you money now in exchange for the right to shares later, once you raise a priced round. It's not a loan: no interest, no maturity date, no clock. Until that round comes, it sits quietly on the cap table.

A convertible note is a short-term loan that's meant to turn into equity rather than be paid back. Because it's a loan, it charges interest - around 8% is standard in the UK - and it has a maturity date, usually 12 to 24 months out. If you haven't raised a priced round by then, the investor can ask for the money back in cash. It often gets sold as the quick, founder-friendly option, but it's the one that can come due before your next round arrives to cover it. Worth knowing the timing: UK seed-to-Series-A now averages around 29 months. Sign an 18-month note and the repayment date lands roughly a year before the round meant to solve it, which leaves you negotiating from a weak spot.

An ASA - short for Advance Subscription Agreement - is the UK-specific one most founders overlook. The investor pays now for shares that get issued at your next round. Like a SAFE, it's not a loan and carries no interest or maturity, and the outcome at conversion is much the same - but it's compatible with SEIS/EIS tax relief, which makes UK angels meaningfully more likely to invest, at no extra cost to you. The one thing to check is the trigger: an ASA usually converts on a qualifying round, so read that clause or a smaller round might not convert it.

The conversion maths across all three is nearly identical when the cap binds, which is the part founders tend to obsess over and the part that matters least. What actually matters is what you're carrying between signing and converting. For a UK pre-seed raise from UK angels, the ASA is usually the right default - same result as a SAFE, plus the tax relief that helps checks get written. For US or cross-border raises, a SAFE is the cleaner choice. A convertible note makes sense mainly when an investor requires it or it's a bridge from people already set up for them - not as a default.

A high valuation cap feels like winning the negotiation, but it isn't a prize, it's a hurdle. It becomes the number your next round has to beat, and the comp you get measured against if it doesn't. Messari carried a $300M valuation in 2022 and sold for around $10M this year. The mark always comes due eventually, so raise at one your next round can actually clear.

Embarassing bets are the best bets

Boost VC wrote checks into crypto and "sci-fi" in 2013, when admitting you took those meetings was mildly awkward. The returns didn't come from being smart. They came from being willing to sit in the room everyone else found embarrassing. The edge was tolerance for looking a little foolish, which tends to get relabelled as conviction once it works.

The trouble in 2026 is that everyone already agrees with this. "Be contrarian" is the most consensus opinion in venture, and the same people who say it tend to back the identical AI wrapper as everyone else. Real contrarianism feels slightly dumb - the kind of thing your group chat would screenshot.

So it's worth noticing where the less obvious money is actually going. The mega-rounds this month went to physical things rather than apps: Jeff Bezos's industrial-AI startup raised a $12B round, and a German robotics company raised up to $1.4B, led by a stablecoin issuer. Energy and nuclear, dull for years, are suddenly strategic because all this AI has to plug into something. Decentralized AI was a punchline until Friday, when one government switched off a frontier model and handed the "what if no single party could do that" argument a real-world reason to exist. And tokenization keeps crowding into T-bills, the easiest thing to put on-chain, while the harder and larger opportunity sits in things like parametric insurance.

None of these are fashionable, and all of them are getting funded. That's not a coincidence, it's the mechanism. If it feels safe and obvious, you're probably early to being late instead.

In 2013 it was crypto. Right now it's robots, reactors, censorship-resistant compute, and whatever sounds faintly ridiculous this week. The useful skill is genuinely being able to sit in the embarrassing one a little longer than everyone else.

The deals worth knowing

A quick read on the week's biggest moves, and what each one actually tells a founder.

• SpaceX — IPO, ~$75B raised at a $1.77T valuation. The largest IPO in history. The number isn't the point; the liquidity is. Big exits send cash back to LPs, who recommit, which eventually trickles down to early rounds — on a lag. It priced at roughly 92× sales, so read it as froth at the top, not a starting gun.

• Blockworks → Messari, ~$10M. A crypto-data platform worth about $300M in 2022, sold for a sliver of that. Half cautionary tale (the mark comes due), half land grab — whoever owns the data record gets to be the thing everyone checks.

• Prometheus (Bezos) — $12B Series B at a $41B valuation. Industrial AI. The mega-money is moving toward AI that runs factories, not AI that writes captions.

• NEURA Robotics — up to $1.4B Series C, led by Tether. Physical AI, funded by a stablecoin issuer. Crypto money is quietly flowing into robots; atoms are the trade.

• Kalshi — $1.2B Series F. Prediction markets went from "is this even legal" to a billion-dollar round in about 18 months. A category being born in real time.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

Across VC newsletters tracked, one thing became clear: this week reset what investors think AI is actually worth, while regulators in three jurisdictions moved in completely different directions on crypto.

28 unique VCs published. AI dominated the volume. Infrastructure followed. The dominant sentiment across all sources was neutral with a bullish AI undercurrent - funds aren’t celebrating yet, but they’re not running for the exits either.

Top trending names: Pluralis, Quantinuum, tea Protocol.

1. The AI valuation curve is no longer normal — and your raise will be priced against it whether or not it should be.

Anthropic at $965B means every Series A multiple gets benchmarked against a number that doesn’t reflect your stage. If you’re pitching, prepare for partners to anchor on the wrong comps. Have the answer ready: here’s why Anthropic’s revenue trajectory isn’t relevant to ours, and here’s the comp that is.

2. Distribution is the moat. Model performance is the table stake.

Benedict Evans called it “1997 for AI” on Lenny’s podcast this week. The thesis lining up across multiple newsletters: as AI capabilities commoditize, the durable moat shifts from “we have a better algorithm” to “we control access to the user.” If you’re building AI tooling without an embedded distribution channel, you’re competing on a dimension that’s losing value fast.

3. Decentralized AI infrastructure is becoming investable.

Pluralis at $7.6M from USV and CoinFund is small but matters because of who’s writing the check. The decentralized training thesis used to be a crypto sideshow. With EpochAI data showing 20x annual scaling, real VCs are now positioning. For founders in the AI infra layer - particularly those touching crypto rails - this is a window opening.

4. Regulatory arbitrage is real this quarter, and timing matters.

Japan moves forward on stablecoins June 1. UK frames their full crypto regulation for October 2027. US House aggressively investigates prediction markets June 5. Three jurisdictions, three different speeds, three different appetites. If you’re building in crypto-adjacent categories, your incorporation jurisdiction and licensing strategy is now a real moat or a real liability - not a future problem.

5. Public market exits are not back, but the doors are unlocking selectively.

Quantinuum’s IPO on June 3 is the bellwether to watch. If it prices well and trades clean, expect a small wave of deep-tech follow-ons. If it disappoints, the IPO window quietly closes again for everyone but the SpaceX/OpenAI/Anthropic tier. For founders at Series B+, the answer to “should I raise more private or wait for the window” depends on this datapoint in 7 days.

1. The death of the seat: agents replacing SaaS is now consensus

The "AI agents will replace SaaS" thesis went from contrarian to consensus in about six months.

In February, $285 billion in software market cap vanished in a single trading session - ServiceNow, Salesforce, Intuit all down 7-11%. The trigger wasn't a recession or regulation. It was Anthropic releasing Claude Cowork enterprise plugins that let one person automate work previously requiring 5-10 separate SaaS subscriptions. The "SaaSpocalypse" tag stuck.

This week, the Canonical Capital piece "Eyes Are a Tax" articulated where this is heading: the browser-and-dashboard UX is structurally inefficient. The terminal, the place where work actually executes, won. Software companies whose value proposition is "we give you a nice interface to click buttons" are competing against agents that just do the thing.

Madrona's letter this week called out the same thing from a different angle. Their take: "the next AI bottleneck is security", meaning enterprises are ready to deploy agents broadly, but the friction now is around governance, not capability.

Even Lenny's interview with Benedict Evans hit the same note. Evans called this "a 1997 internet moment", meaning we're at the point where everyone agrees the technology is real, the only argument is about where the value accrues.

Bessemer just disclosed an average Series A check size of $18.3M, with deep concentration in AI-native enterprise companies. The enterprise AI agent market grew from $5.25B in 2024 to ~$7.84B in 2025, projected to hit $52.62B by 2030.

What this means for founders:

If you're pitching anything SaaS-shaped right now, your deck needs to answer one question: what do you charge for that an agent can't replace?

It's not enough to say "we have AI features” because every SaaS has AI features. The real question is whether the agent version of your product replaces you or runs on top of you.

The founders raising successfully this quarter are building the infrastructure agents call out to. Look at Cognition (reported by Neo this week): $1B raise at $26B valuation, $492M run-rate. They're the agent layer for engineering work. That's the model VCs are now writing checks against.

2. The secondary market is now bigger than the IPO market

US venture secondary transaction value hit $112.2 billion annualized this quarter. That's larger than VC-backed IPOs over the same period.

The PitchBook Q1 2026 outlook puts the question: what happens to a $112 billion secondary market when its most valuable constituents (SpaceX, OpenAI, Anthropic) go public? The top 20 names currently capture 81% of secondary trading value. Once they exit, returns get held in lockup, get directed toward addressing venture's four-year distribution deficit, and don't recycle back into secondaries for a long time.

So we're in an unusual moment. Secondaries are bigger than IPOs, but they're also concentrated in the same names that are about to leave the secondary market. The middle of the unicorn pool, thousands of companies, still has no real liquidity path.

20VC's newsletter this week touched on the institutional adoption angle. Goldman Sachs bought Industry Ventures last year. Morgan Stanley acquired EquityZen. Schwab moved for Forge Global. Secondary SPVs increased 682% from 2023. This is becoming infrastructure, not workaround.

What changed: over 40% of active unicorns raised their first round more than a decade ago. Cumulative cash flows from US venture funds to LPs have been negative by $197 billion since 2022. Holding period for VC investments is now 10+ years on average. Equity became a long-term retention mechanism that needs liquidity built in.

If your investors are pushing you toward profitability and away from another priced round, the reason might be that they need to set up a secondary in 12-18 months and need a clean cap table to do it.

If you have early employees who've been with you 4+ years, tender offers are now a realistic retention tool. The Founder Collective letter this week made the point that "the founder factories" of the next decade will be companies that built liquidity events for their teams before the IPO. That's a competitive advantage in hiring.

If you're at Series B or later, ask your lead what their secondary strategy is. If they don't have one, that's a yellow flag - they may not understand the new exit reality and you'll be the one pushing them to figure it out.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3. Three jurisdictions on crypto regulation

Japan, June 1. The FSA formally included foreign trust-type stablecoins in its electronic settlement regulatory framework. As of this week, foreign stablecoin services have a legal basis in Japan, with equivalency standards for providers. This is the clearest institutional path for stablecoin adoption in a major economy, and the template other G7 jurisdictions are watching.

United Kingdom, June 3. The UK FCA’s consultation on a comprehensive crypto framework closes today. The iea - covering stablecoins, trading, custody, and staking -is set to take effect October 25, 2027. The UK is going slower than Japan but building broader. The deliberate pace suggests they want one well-designed framework rather than incremental rules.

United States, June 5. The House Oversight Committee’s subpoena deadline for Polymarket and Kalshi hits this Friday. The committee is demanding full documentation on identity verification, geographic restrictions, and anomalous transaction monitoring, framed around national security concerns about suspicious activity preceding military and geopolitical events. This is the most aggressive regulatory action on prediction markets to date.

Three completely different regulatory philosophies.

Japan opens the door for stablecoin issuers. The UK is building a fence carefully and slowly. The US is sending out subpoenas with national security framing attached.

What’s notable is the speed and the simultaneity. Three major jurisdictions making consequential moves in the same week, while Telegram became the largest validator on TON and the US Treasury continues digesting last year’s GENIUS Act. The regulatory environment isn’t drifting toward consensus. It’s hardening into incompatible tracks.

For stablecoin issuers and infrastructure: Japan just became the most legible market. Mastercard’s $1.8B BVNK acquisition this week is a tell - incumbents are positioning for the regulated stablecoin window opening across G7 markets.

For prediction market operators and anyone doing event-based trading: the Polymarket/Kalshi subpoena response on June 5 will set the operating model for the entire category in the US for the next 18 months. If they push back hard, expect formal enforcement. If they comply, expect every prediction market to adopt similar identity verification and geographic restriction infrastructure as table stakes.

For DeFi protocol founders generally: jurisdiction arbitrage is real this quarter, but it has a cost. Operating from a clear-rules jurisdiction (Japan, UK by 2027) costs more in compliance but unlocks institutional capital. Operating from a permissive jurisdiction is faster and cheaper but caps your eventual fundraising ceiling at the level where institutional LPs get uncomfortable.

One thing worth flagging: G7 regulatory frameworks tend to move from “consultation” to “enforcement” in 18-month cycles. Japan’s framework is live now. The UK framework is targeted for October 2027 - that’s the window where founders building today need to be compliant or grandfathered.

4. Fundraising itself is being rebundled

This is the practical consequence of everything above, and it's the most underreported shift in the venture stack.

The old playbook: pre-seed (~$500K-1M) → seed ($2-4M) → Series A ($8-15M) → Series B ($20-40M) → growth. Each round on a ~18-24 month cadence.

The new playbook that's actually working in 2026: pre-seed → bridge → Series A directly, with the seed round either skipped or absorbed into either an extended pre-seed or a pre-A bridge.

First: Polsia closed $30M at a $250M valuation with 0 employees, running the fundraise itself via AI on a public Twitter dashboard. That's not a normal seed. That's a pre-seed startup pricing like a Series A, skipping every intermediate step. Reported in The GTM Newsletter this week.

Second: Cognition's $1B raise at $26B valuation. Founded in 2024, now at $492M run-rate. They went from Series A to Series C in 18 months. The traditional stage progression — A, then B, then C — is collapsing into "raise as much as you can while you have momentum, before the next round becomes harder."

Third: Edith Yeung's roundup of last week's Silicon Valley deals shows 14 startups raised $1.36B combined. The check sizes cluster around two patterns - very small (under $5M, pre-seed) and very large (over $50M, late stage). The $10-20M Series A round, once the median, is becoming the exception.

The GTM Newsletter piece "The Distribution Era" this week captured what's driving this. The thesis: distribution is now the moat, not product. The companies winning in 2026 are the ones building distribution before product. That changes the fundraise - you're not raising to build, you're raising to acquire users at scale before someone else does.

This rewires the fundraising stack. If distribution is the moat, you need a lot of capital upfront, fast. You can't take three years to get from seed to Series A while a competitor sprints. You raise huge or you don't raise.

What this means for founders:

Stop planning a 4-stage raise. Plan a 2-stage raise.

Stage 1: Pre-seed plus bridge plus extended runway, ideally from specialist micro-funds who can write follow-on checks. Aim for 18-24 months of runway from $1-3M total.

Stage 2: A meaningful Series A only when you have either real ARR ($1M+) or genuine distribution scale (~100K users with engagement). Skip the seed-then-A two-step.

If you can't get to either of those milestones on Stage 1 capital, the seed-then-A path is no longer reliably available to you. The fund that would have written your seed in 2021 doesn't exist anymore, or is too underwater to write checks.

The implication: be much more cautious about scaling team headcount on pre-seed money. The bridge cycle is now longer than founders are budgeting for. Founders who keep their team small and their burn low — even at the cost of slower growth — are the ones surviving the gap between pre-seed and the new Series A bar.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

FUNDING TRENDS

Series A Down $45M median (was $52M) -13% month over month

Bridge Rounds Standard 29% of deals now seed+A bridge Up from 22% Q1 2026

Seeds Reaching A <40% of 2025 seeds Selectivity increased

1. Founder Selection is Wrong

Arian Ghashghai invests in cold inbound founders without networks or credentials. They raise tier-1 rounds within 12 months. The pattern: these founders have unconventional lives. They climbed Everest, lived abroad, coded at 10, speak five languages, quit safe jobs. One VC analyzed 15,000 founders across 70 variables. The single best predictor of success wasn't IQ, Stanford, Google, or previous exits; it was Outlier Quotient - how much the founder deviated from the mean. OQ had zero correlation with elite schools. The highest OQ scorers took risks everyone else called insane.

This means VCs are filtering out their best candidates: Stanford degrees, Google credentials, perfect pitch decks - all conventional signals. The founders they pass on are the ones who lived in Chile instead of the Bay Area, who built something weird instead of something obvious. Soma Pirityi says founder underwriting is now a psychology question, not a spreadsheet question. When the input is human intelligence and the clock is 12 months, you can't predict outcomes with credentials.

If you're a founder: don't optimize your resume. Have a real life. If you're a VC: stop asking about degrees. Ask what they sacrificed, what they built alone, what obsession made them leave the safe path.

2. One Mode or Nothing

Polsia hit $10M ARR in three months: one human, zero employees. Ben Cera didn't raise to build - he built first, then raised. Commure raised $7B valuation. Tanay Tandon said: "We didn't need the capital. We raised for pricing purposes." They had traction, had proven the model.

One founder posted the framework: "There are only two modes. Fully in or not fundraising at all. No in between." When fundraising, you're in back-to-back meetings, calendar stacked, moving so fast VCs keep up. When building, you're available but not pitching, working on hard problems. Founders who blur these modes come across desperate without urgency. They lose to founders who pick one and execute completely.

Hyperliquid built the clearing engine first, proved it worked, then unlocked layers. HIP-1 was asset standards. HIP-2 was bootstrapped liquidity. HIP-3 broke listing monopoly. HIP-4 came last. Every step was prerequisite, not feature. The boring layer before the flashy product. On-chain insurance started with balance sheet graphs, not products.

If you're fundraising: go fully in. If you're building: stay quiet until you have traction. Build the prerequisite before the feature.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3. Capital Flows Direct Now

Sam Altman offered $2M in tokens to every YC startup. Compute flows direct from model providers to founders. Nick Tomaino did the math on launchers: they take 20-60% of everything you raise forever. A founder raising $10M gives away $2-6M permanent fees. Compute-for-equity replaces traditional VC capital. Tokens replace cash. Founder pools replace VC networks.

Soma Pirityi said: "If I ran a VC fund today, I would fire my LPs. This is compute for equity. That's the whole fund." The 2-3% management fee made sense when VCs were gatekeepers. Model providers now have capital, so do protocols. Tokens are capital. What's not commoditized: judgment and real access. The VCs who survive have built companies and know founders.

If you're a founder: get compute from model providers, get capital from multiple sources, get network from founder peers.

Three moves the winners are making: (1) Stop looking like a founder. Have an unconventional life and genuine expertise. (2) Pick one mode completely: fundraise fully or build fully. Build the prerequisite first. (3) Take capital direct.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

Here's what's actually moving, why it matters, and what to do with it.

1. Per-seat pricing is dying — and the funds know it

Sierra just raised $950 million at a $15.8 billion valuation. Tiger Global and GV led, Sequoia and Benchmark followed. The headline got the rounds of applause but the part nobody's talking loudly enough about is how Sierra makes its money. The startup employs an outcomes-based pricing model, charging customers for completed work rather than charging flat subscription fees.

They get paid per resolved conversation.

That's the model that just got priced at $15B. And it's not just Sierra. HubSpot has moved to outcome-based pricing with its "Breeze AI Agents." The Customer Agent now costs $0.50 per resolution (halved from the $1.00 launch price). The Prospecting Agent costs $1.00 per recommended lead. Established SaaS players are scrambling to match.

Here's why this matters for you, even if you might be nowhere near a Series E. The standard VC diligence question for any AI startup now includes some version of: "what happens to your ARR when agents replace seats at your customers?" If your answer is "we charge per seat, but we promise our customers don't need fewer seats" - you've already lost the meeting. Rigid per-seat pricing: AI agents are replacing human SaaS operators. Therefore, the per-seat models face structural revenue risk that cautious buyers now account for explicitly.

The lesson. Your pricing model is now part of your defensibility story, whether you realized it or not. If you can't tie your revenue to something an agent can't replace — outcomes, infrastructure usage, data volume, transactions completed - your unit economics get marked down by every VC who's been reading the same headlines. Reread your pricing page tonight. Ask yourself what gets cheaper for your customer when AI gets cheaper. If the answer is "everything we charge for" - you have a problem.

2. Enterprise budgets are concentrating, and the pilot era is ending

For three years, the enterprise AI playbook for founders was simple: get into 10 pilots, prove value, convert one or two, ride that into your Series A. That playbook just expired.

TechCrunch surveyed 24 enterprise-focused VCs and an overwhelming majority predicted enterprises will increase their budgets for AI in 2026 but not for everything. Most investors said this budget increase will be concentrated and that many enterprises will spend more funds on fewer contracts.

Read that twice. Bigger budgets. Fewer contracts.

Andrew Ferguson at Databricks Ventures called it directly: "2026 will be the year that enterprises start consolidating their investments and picking winners. Today, enterprises are testing multiple tools for a single-use case, and there's an explosion of startups focused on certain buying centers like [go-to-market], where it's extremely hard to discern differentiation even during [proof of concepts]."

If you're a founder sitting on 8 active pilots with 0 conversions, you don't have 8 opportunities — you have a problem. The buyer is about to pick one vendor per category, and "in pilot since Q3" is not a winning position. Founders who built decks around "Fortune 500 logos in pilot" need to redo their decks around one logo in production.

The flip side: if you do win the consolidation, you win big. The companies operating hard-to-replicate products such as vertical solutions or those built on proprietary data will likely still be able to grow. Vertical depth beats horizontal breadth at every turn this year. Vertical AI is winning because general-purpose tools can't compete on depth.

Stop optimizing for "more pilots." Start optimizing for "fewer pilots, deeper integration, one signed contract." The diligence question is changing from "how many enterprises are testing you?" to "which one renewed and expanded?" If you can't name one logo that's gone from pilot to production in the last 6 months, your fundraise just got harder. Pick your two best pilots and put 80% of your GTM energy there.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3. The Series A timeline doubled while you weren't looking

Here's a number that should change how you plan your runway: the average time between seed and Series A has stretched to around 616 days.

That's about 20 months. If you raised your seed with 18 months of runway because that's what your accelerator told you to do, you're going to run out of cash 2-4 months before the average raise even happens. And it gets worse - less than 40% of seed-funded startups successfully raise Series A. Most don't make it.

The market knows. According to Carta, about 29% of all deals at both the seed stage and Series A were bridge fundings in 2025. Bridges used to be a yellow flag. Now they're standard. Many bridges today are being raised opportunistically by founders who are not quite ready for a Series A or B process and need some extra runway to reach the next milestone.

What this means in practice: the founder who plans for "raise A in 12 months" is planning to fail. The founder who plans for "raise a bridge at month 14, then Series A at month 22-24" is planning realistically. The math has shifted but most fundraising advice hasn't caught up.

The other thing that's changed - the bar is higher. Pre-money valuations at this stage generally fall between $25M–$50M, with the median sitting at approximately $45M for primary rounds. But to get those numbers, you need real revenue (most A's now need >$1M ARR), real retention (NDR >100%), and real momentum. The "we have product-market fit and need to scale" pitch doesn't work alone anymore.

The lesson. Three things to do right now if you raised a seed in the last 12 months. One: extend your runway calculations to 24 months minimum, even if it means tighter cost discipline today. Two: stop being embarrassed about bridges. If your seed plan was 18 months and you're at month 14 hitting your milestones but not Series A-ready, raising a smaller bridge round is the standard move now, not a confession of failure. Three: stop asking "when can I raise A" and start asking "what does my metric look like at month 22 if I don't raise A." The answer is your real cash plan.

4. The "pure SaaS" gospel is being rewritten

For a decade, "we're not a services company" was something founders said with pride. Services = lower margins, harder to scale, less venture-fundable. You wanted to be pure SaaS, even if you actually shipped consultants in your duffel bag.

A subset of enterprise AI companies will shift from product businesses to AI consulting. These companies may start with a specific product, such as AI customer support or AI coding agents. But once they have enough customer workflows running off their platform, they can replicate the forward-deployed engineer model with their own team to build additional use cases for customers. In other words, many specialized AI product companies will become generalist AI implementers.

Forward-deployed engineering, which used to be a dirty phrase, is now a standard play. Palantir wrote the playbook. OpenAI's enterprise team is running it. Sierra, back to Sierra, operates on it: Sierra positions itself as a strategic partner rather than a vendor of plug-and-go solutions. The company collaborates closely with enterprises to create customized AI agents that embody each client's unique brand voice, policies, and workflows.

This isn't a contradiction of the consolidation point. It's the answer to it. The reason a customer consolidates onto you instead of your competitor is because your forward-deployed engineers built your product into their workflow. The implementation depth is the moat. When asked how they know that an AI startup has a moat, multiple VCs said companies with proprietary data and products that can't easily be replicated by a tech giant or large language model company are the most defensible.

The lesson. If your unit economics depend on customers self-serving onto your product, in 2026 you're competing on price against open-source alternatives and AWS. If your unit economics include some forward-deployed engineering work to embed deeply into customer workflows, and you price for it, you're now playing a structurally better game. The thing to stop doing is hiding it. "We do implementation services to deepen integration" is now a strength in the pitch, not a weakness.

Connecting the dots

The SaaS playbook that built the last decade - per-seat pricing, broad pilots, fast Series A, no services - is breaking down in four places at once. And the funds picking the next winners are pricing in the new playbook before founders have noticed it's changed.

The founder who'll win the 2026 fundraise is the one whose deck answers four questions:

1. What do we charge for that an agent can't replace? (pricing)

2. Which two enterprise logos do we have in production, and what makes us un-rip-out-able? (GTM)

3. What does our company look like at month 22 if Series A doesn't happen? (runway)

4. What forward-deployed work do we do that competitors can't match, and how do we price it? (services-as-moat)

None of those questions were standard 18 months ago. All four are now table stakes.

The fund's IC has already started writing their memo by question 2.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

Our X (for timely updates and founder/investor opps)

Our Linkedin (for connections and professional insights)

Our app (for VC index cards, the feed of newsletters, and upcoming AI summaries)

That’s the data. Now go build something.

— Byblos.digital

Subscribe now

289 emails. 47+ unique sources. Here is the honest breakdown:

The inbox was roughly 8% actual VC fund content with real thesis thinking. 14% crypto media. 6% research. 12% content creators. 60% noise. We read all of it to find the 25% that matters.

a16z was the most active publisher in our inbox - 8 pieces in 7 days across stablecoins, retention frameworks, prediction markets, AI x crypto, and infrastructure, followed by the $2.2B fund announcement. This was visible in our tracking before the announcement. When a fund doubles its publishing cadence across multiple topics, capital movement usually follows.

Sentiment across VC sources shifted bullish - 47-50% bullish, highest in 3 weeks. The shift was driven by the CLARITY Act breakthrough, BTC touching $80K, and the DTCC tokenization announcement. Tradfi-adjacent funds were overwhelmingly bullish (52 bullish, 0 bearish). Crypto-native funds were nearly split (30 bullish, 29 mixed). The mood depends on who you are meeting with.

Five normally-active funds went quiet this week simultaneously: Canonical Crypto, CoinShares, Collaborative Fund, Elad Gil, and Maven 11 Capital. When active publishers go silent at the same time, it usually means end-of-quarter portfolio review, an unannounced raise, or a strategy shift. We track the silence because the absence of signal is itself a signal.

Vocabulary tracked: “outcome-priced services” appeared for the first time in 4 independent sources on May 1. “AI infrastructure” turned negative across multiple funds the same morning. “buy and use” entered stablecoin vocabulary after the Tillis-Alsobrooks compromise. a16z published that “stablecoin” itself is becoming like “horsepower” - the technology outgrew the label. And YC introduced “tokenmaxxing” as the framing for AI-first resource allocation.

1/ The Settlement Layer is Being Rebuilt

The most important thing that happened this week was three separate companies announcing deals that bypass DTCC - the company that has cleared virtually every stock trade in America for 50 years.

Bullish acquired Equiniti for $4.2 billion. That is the largest crypto M&A deal in history. Equiniti is a transfer agent with 3,000 issuer clients, 20 million shareholders, and $500 billion in yearly payments. Bullish now has the infrastructure to tokenize real securities for real companies at real scale. Tom Farley, Bullish's CEO, called it "a once-in-a-generation shift in how capital markets operate."

The same week, Computershare, the transfer agent that manages roughly 60% of the S&P 500, including Apple, Tesla, Microsoft, Nvidia, and Coinbase, partnered with Securitize to bring stocks on-chain natively. And NYSE tapped Securitize for a 24/7 tokenized securities platform. The stock exchange itself.

All three bypass DTCC. Then DTCC announced its own tokenization platform - pilot in July, full launch in October, with 50+ firms in the working group including BlackRock, Goldman Sachs, JPMorgan, Circle, Nasdaq, and Morgan Stanley.

To understand why this matters: DTCC processed $4.7 quadrillion in securities transactions in 2025. After every trade, every participant in the chain maintains separate records of the same transaction. The reconciliation layer alone employs thousands of people whose entire job is making sure two copies of the same number match.

The irony is that crypto has the opposite problem. In crypto, execution is the inefficiency - fragmented liquidity, MEV, slippage. But settlement and clearing are near-instant, verifiable, and do not require reconciliation. Tokenized securities take the part crypto does well and apply it to the part tradfi does worst.

Three external challengers plus the incumbent itself, all in one week. Kraken also acquired Bitnomial, the first crypto-native platform with all three US derivatives licenses, to build a full-stack derivatives platform. The settlement monopoly is not being disrupted from one direction. It is being surrounded.

For founders: if you are building anything in the settlement, registry, custody, or transfer agent layer, this is the hottest infrastructure category in crypto right now. $4.2B in M&A in a single deal. Computershare connecting 60% of the S&P 500 to on-chain rails. NYSE going tokenized. The capital and the institutional partners moved this week.

2/ $2.2B Deployed - 14% Laid Off.

a16z closed Crypto Fund V at $2.2 billion on May 5. The same morning, Coinbase cut 14% of its workforce citing rough market conditions and AI-driven efficiency improvements. The largest crypto VC fund launch and the largest crypto exchange layoff happened on the same day.

This is the same market seen from two different positions. The capital layer is expanding. The operating layer is compressing. More money, fewer people.

a16z's fund is intentionally half the size of Fund 4 ($4.5B in 2022). They named exactly where the money will go: stablecoins, onchain lending, perpetual futures, prediction markets, tokenized assets, creator platforms, AI-agent infrastructure, and decentralized infrastructure. They published 8 thesis pieces in 7 days before the announcement - the publications were the roadmap.

The same pattern showed up across every major AI deal this week. SpaceX signed a multi-year deal renting its entire Colossus 1 supercomputer to Anthropic - 300 megawatts, 220,000 GPUs, worth $3-4 billion in annual revenue. The backstory: xAI's 550,000-GPU fleet was running at 11% utilization. Industry standard is 40%. Elon built too much, used too little, and is now making more money renting GPUs to a competitor than running them himself. The most profitable AI strategy of 2026 might be building excess capacity and renting it out.

DeepSeek is raising its first outside money at a $45 billion valuation. The founder owned 89.5% of the company before this round - funded entirely from his hedge fund's balance sheet for three years. The valuation moved from $10 billion to $45 billion in weeks after DeepSeek V4 launched. China's state semiconductor fund is leading. Tencent and Alibaba co-investing.

Meanwhile, the Pentagon approved 8 AI vendors for classified defense networks: Google, Microsoft, AWS, Nvidia, OpenAI, SpaceX, Oracle, and Reflection AI (a two-year-old startup with no public model). Anthropic - which has $30 billion in annualized revenue and the most advanced safety research in the industry - was excluded after declining to give the government unrestricted use of Claude. Trusted enough to evaluate. Not trusted enough to procure.

The concentration is unprecedented. Anthropic at $30B ARR with 1,000+ enterprises paying $1M+ each. OpenAI at 900 million weekly active users. Cohere at $20B valuation (from $1.85B in May 2023). Sierra at $15.8B (from $1B in under two years). The top 5 AI companies have a combined valuation north of $200 billion. Capital is flowing at record levels and concentrating in fewer hands with smaller teams.

For founders: a16z just told you exactly what $2.2B is going to fund. If you are building in any of their 8 named sectors, read their 8 publications from last week - those are the thesis documents. If a fund publishes about your space, reach out that week. They are in research mode and will take meetings they normally skip. Also: the spacex/anthropic deal confirms that excess infrastructure is a monetizable asset. If you have compute, power, or data center capacity you are not fully using, that idle capacity has a buyer now.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ The 1-Person Company Arrived

YC published today: "how the top builders use AI to do the work of 400 engineers." They are calling it tokenmaxxing - the idea that spending on AI tokens (compute) can replace spending on engineering salaries. A single person working with AI agents building products that previously required entire teams.

This is not a theoretical framing anymore. It showed up in actual companies this week.

Memelord raised $3 million and reached $100K ARR with zero engineers. Built entirely on Bubble and AI tools. A newsletter turned into a multi-million dollar API business by one person using the tools that exist right now.

Coinbase cut 14% of its workforce and restructured around what they called "AI-native pods" - smaller teams augmented by agents. The reason they can cut 700 people is not because the business is shrinking. It is because they believe smaller teams with AI tooling can do the same work.

Microsoft published their Q1 2026 AI Diffusion Report: developer git pushes increased 78% year-over-year as Claude Code, OpenAI Codex, and GitHub Copilot moved into routine workflows. 17.8% of the world's working-age population now uses generative AI. The tools are the baseline.

Greg Isenberg published this week that maybe 1,000 companies on earth are truly AI-native at $5M+ ARR. Not "we added copilots" but companies where the workflows were designed from scratch for agents to execute and humans to supervise. His distinction: AI-assisted companies add AI at the edges. AI-native companies redesign the center. The field is almost empty despite all the noise, which means the opportunity is still wide open.

And from YC's current batch: Klaimee is building insurance for AI agents. That is the missing piece. We tracked the agent financial stack assembling over 10 days but nobody was insuring the agents. YC just funded the company that does.

Garry Tan is framing this as personal AI infrastructure. His thesis: "the future belongs to individuals who build compounding AI systems, not to individuals who use corporate-owned centralized AI tools." He is building GBrain in open source and giving it away. The YC president is saying out loud that individual builders using personal AI systems will outperform teams using corporate tools.

For founders: the team size assumption just changed. If you are raising pre-seed and your deck says you need to hire 8 engineers, the VC is going to ask why. Memelord has zero engineers at $100K ARR. The question shifted from "how big is your team" to "how much are you spending on tokens vs salaries." If you can show that a 2-person team plus AI tooling achieves the output of a 10-person team, that is a capital efficiency argument that makes your raise easier, not harder.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

We processed 300+ emails between April 28 and May 3. Here’s the honest breakdown:

• VC fund letters with real thesis content: ~25 (chamath, a16z, QED, madrona, nascent, fifth wall, gradient, lemniscap, castle island, redalpine, 91 capital, entree, tundra angels, NFX, battery, canonical, south park commons, and others)

• Crypto media: ~14 (bankless published 5 pieces alone, blockworks 3, cryptoslate 2, lark davis 2, coindesk, galaxy research)

• Research / data: ~4 (goldman sachs, S&P global, NYDIG, blockworks research)

• Content creators / general tech: ~12 (lenny, not boring, newcomer, GTM newsletter, 20VC, hard fork)

• Portfolio updates / events: ~8

• Noise / marketing / irrelevant: ~40+

We’ve been tracking convergence since mid-April. When a16z publishes something, the ecosystem usually echoes it within 48 hours - we saw that pattern three times last month. But this was four independent sources arriving at the same conclusion at the same time. Our cross-newsletter consensus metric would flag this immediately.

The rhetoric matters for founders. Here’s the specific language shift we noticed: none of these four pieces used the word “infrastructure” as a positive pitch framing. All four explicitly said infrastructure is being commoditized, value is moving up. If you walked into a pitch meeting last month saying “we’re building AI infrastructure,” you’d get attention. This week that phrase entered the same territory as “we’re building a platform” - too vague, too crowded, too many decks in the queue. The new vocabulary: “outcome-priced services,” “vertical AI,” “service-as-software.” Three terms that appeared in all four publications independently.

And then the market confirmed the thesis the next day: MARA, one of the biggest bitcoin miners, spent $1.5B on an energy asset to pivot into AI data centers. The thesis went from newsletters to M&A in 24 hours.

1/ Six Companies Gave AI Agents Wallets

Stripe launched Link - wallets that let AI agents hold funds, make payments, and stream stablecoin micropayments in real time via Tempo's Machine Payments Protocol. Circle shipped multichain USDC nanopayments for agents. Coinbase added Polygon and Solana to its agentic wallet. MoonPay launched MoonAgents Card - a virtual Mastercard for AI agents. OKX introduced Agent Payments Protocol for agent-to-agent transactions. Nansen deployed an onchain analytics agent at $0.10 per query.

They all independently decided that agents need their own financial infrastructure at the same time. When six competitors ship the same feature in the same week, the thesis is over. The land grab is on.

We've been tracking agent infrastructure hitting consensus across VC newsletters for two weeks - 7 independent funds publishing on the same topic without coordinating. Stripe didn't read the Lemniscap newsletter about "the agentic control plane" and decide to build it. It was already happening.

The numbers: 40 cents of every crypto VC dollar in 2025 went to AI-adjacent firms, up from 18 cents the prior year. The agentic economy is projected to reach $30 trillion by 2030. Crypto platforms moved faster than traditional finance because they have always-on markets and programmable infrastructure: no market hours, no intermediary layers for agents to navigate.

But here's the part that should make you nervous. The same week agents got wallets, a startup called Nava raised $8.3M (Polychain and Archetype co-led) to build the exact opposite - escrow that holds agent funds until a verification framework confirms the transaction matches the user's intent. If the agent tries something unauthorized, the money stays locked. Every decision gets posted onchain for other AIs to reference. Polychain's CEO compared it to building "SWIFT for the agent economy." QED Investors called this the "napster era" of agents two weeks ago. Nava is trying to build iTunes before the lawsuits hit.

Meanwhile: April was the worst month for crypto hacks in history, $625M across 30 exploits. The industry's response was to give AI agents their own credit cards while simultaneously building the brakes to stop them from going rogue. The optimism in this industry is genuinely inspiring and also slightly concerning.

2/ MiCA's Deadline Decides Survivers in Eastern European Crypto

The MiCA transitional period ends July 1, 2026. After that date, any company providing crypto services to EU clients without a MiCA license is operating illegally. The EU's first comprehensive crypto regulatory framework becomes fully enforceable in 57 days.

The cost of compliance: licensing runs €250,000 to €500,000. Ongoing compliance eats up to 15% of revenue for smaller firms, compared to under 2% for large exchanges. Almost 70 companies hold a MiCA license so far. For context, that licensing cost is more than most Eastern European seed rounds.

Eastern Europe built some of the strongest crypto developer communities in the world. MultiversX is Romanian. RedStone Oracles is Polish. Aave was started by a Finn. Solana has a Superteam Balkan community. Bucharest hosts Eastern Europe's flagship crypto conference. Warsaw's Next Block Expo attracts 2,000+ builders. Estonia was one of the first countries to offer crypto licensing. Poland, Romania, and Ukraine have deep engineering talent pools that powered much of DeFi's growth.

But the builder culture that thrived in low-regulation environments is about to collide with EU-grade compliance. A Romanian startup that built a protocol with 3 engineers now needs to budget €250K-€500K for a license before it can legally serve the person sitting in Paris using its product. The math doesn't work for most seed-stage companies. Remember when we tracked that april crypto funding dropped 85% from march? Imagine raising in that environment AND budgeting half a million for a license.

The result will be consolidation. Larger firms absorb smaller ones. Some teams relocate outside the EU - probably to Dubai, Singapore, or the UK which is running its own regulatory sandbox. Some shut down. The irony: MiCA was designed to protect consumers and create market stability. It will also price out the small teams that built the products consumers actually use. The same regulatory speed that makes Europe look responsible compared to the US (still debating the CLARITY Act at 47% odds on Polymarket) might kill the very ecosystem it's trying to legitimize.

We've been tracking the global regulatory divergence for weeks: the US debates, Europe builds, India deploys, Canada bans crypto ATMs. Add to the list: Japan's SBI Holdings just made $561M in crypto profits and is buying every exchange in sight while preparing for securities classification. Every country is choosing a different path. The founders who survive are the ones who structure for regulatory optionality.

For founders in Eastern Europe: 57 days to file, find a licensed partner, restructure to serve non-EU markets, or wind down EU operations. ESMA will publish a non-compliant entities list. Being on it means banks cut you off.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ AI Ate 80% of All Venture Capital Last Quarter

Q1 2026 was the biggest quarter in venture capital history. Crunchbase data: $300 billion invested into 6,000 startups globally. Up 150% quarter over quarter. More than any full year before 2018. And 80% of it, $242 billion, went to AI.

The concentration is staggering. The four largest venture rounds ever recorded all closed in Q1: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). Those four companies alone absorbed 65% of all global venture funding for the quarter. We wrote about Anthropic's revenue going from $1B to $30B in 16 months, now you can see WHY. When one company absorbs 10% of global venture capital in a single quarter, the revenue trajectory makes more sense.

Seventeen US AI startups raised $100M+ rounds in the first two months alone. Remember when we said $145B in AI infrastructure deals happened in one month (SpaceX $60B, Google $40B, Nvidia $20B, OpenAI $20B, Amazon $5B)? That was just the tip. The full quarter was twice that.

Crypto's share of venture capital has been shrinking for quarters. April crypto funding was $466M, down 85% from March's $3B. But instead of competing for a smaller slice, crypto started merging with AI.

MARA Holdings, one of the biggest publicly traded bitcoin miners, announced a $1.5B acquisition of Long Ridge Energy, a 505 MW gas plant in Ohio, to pivot into AI data centers. Their CEO Fred Thiel said "power is the scarce input in AI." Stock jumped 13%. They already have inbound interest from investment-grade AI tenants. The deal lifts MARA's power capacity 65% to 2.2 GW, with AI data center operations targeted for mid-2028. They also bought a majority stake in French computing firm Exaion earlier this year.

The logic is beautifully simple: bitcoin miners survived the last crypto cycle by building massive power infrastructure: contracts, cooling systems, land, grid connections. That's exactly what AI hyperscalers need and can't build fast enough. The miners aren't pivoting away from crypto. They're discovering that the same infrastructure serves two markets. It's like finding out the factory you built for making cars also happens to be the perfect factory for making tanks.

The same day MARA announced, Chamath, Madrona, Nascent, and a16z all published variations of "AI value is migrating up the stack" within a 90-minute window. No coordination. Four funds, same thesis, same morning. We've been tracking a16z publishing 7+ pieces in 14 days and the ecosystem echoing them within 48 hours, but this was four different funds converging independently. That's not echo. That's consensus forming in real time.

For founders: if you're pitching "AI infrastructure" you're in a queue of hundreds of decks. The bid for tools is dead. But if you're sitting on power, compute, or grid infrastructure, especially in regions with cheap energy (parts of Eastern Europe, Latin America, US midwest), the AI data center opportunity is real and doesn't require you to build AI. It requires you to build the thing AI runs on. MARA's CEO said it best: power is the scarce input.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

1/ Stablecoins became infrastructure

Look, we’ve all heard the “stablecoins are the killer app” pitch for years. This week it stopped being a pitch.

Stripe went on stage at the RWA Summit in Cannes and said in public that they’re building the “AWS for money.”

They process $2 trillion a year. They acquired Bridge for $1.1B. They built Tempo, a payments-focused blockchain.

And they launched Open Issuance - any business can now launch its own stablecoin with a few lines of code. Metamask is next. Hyperliquid’s USDH is on the platform.

The company that powers checkout for half the internet is now a stablecoin issuance platform.

Meanwhile in Europe, as we reported in our daily digest. twelve banks formed a consortium called Qivalis to launch a MiCA-compliant euro stablecoin in H2 2026. The CEO they hired? Former head of Coinbase Germany. Europe’s approach to competing with American stablecoin dominance is beautifully pragmatic: just poach from it.

Circle launched a Payments Network for banks. The pitch: stablecoin settlement without ever touching “digital assets.” That framing is doing a lot of heavy lifting — and it’s exactly right. You sell crypto to a bank by never calling it crypto.

Toss, the Korean super-app with 30 million users (60% of South Korea), filed 24 KRW stablecoin trademarks and started building an L1 mainnet. Y Combinator is paying startups in USDC. Rain raised $250M for stablecoin payment rails. Mastercard acquired BVNK for $1.8B — the largest stablecoin infrastructure acquisition ever.

Stablecoins aren’t a crypto use case anymore but a payments infrastructure category. The people building them aren’t crypto founders. They’re Stripe, Mastercard, and a consortium of banks that probably still misspell “blockchain” in internal memos.

For founders: if you’re building anything that touches cross-border payments, settlement, or B2B transactions and you’re NOT thinking about stablecoin rails — you’re building on the wrong stack.

SWIFT took 50 years to build its network. Stablecoins are replacing it in 5.

2/ Everyone's fighting over the checkout for AI agents with wallets

Six weeks. That’s how long it took for the entire agent payment stack to go from “interesting research” to “five trillion-dollar companies shipping competing protocols.”

Here’s are the milestones:

— Stripe and Paradigm launch Tempo’s mainnet, and it ships with the Machine Payments Protocol (MPP). In its first week, MPP processed 34,000 transactions. Fees as low as $0.003. Stablecoins are a default payment method. Mastercard, UBS, Klarna, and Visa are already on the testnet.

— Ant Group (yes, Alibaba’s fintech arm) launches Anvita. Agents can hold assets, trade, and make payments autonomously. They call it an “agent-to-agent economy.” Unveiled at a summit in Cannes, because apparently even AI agents need a French Riviera launch party.

— Visa drops Intelligent Commerce Connect. First major payment network to formally open its infrastructure to AI agents. Supports four different protocols simultaneously (TAP, MPP, ACP, UCP). This is Visa saying: “We don’t care which protocol wins. We’ll be the layer that connects all of them.”

— Nava raises $8.3M from Polychain to build agent guardrails. Their product: an escrow service that holds funds until an agent’s proposed transaction matches the user’s intent. If it doesn’t pass the check, the money stays in escrow. Think of it as a seatbelt for your AI’s wallet.

— Stripe and OpenAI have already shipped the Agentic Commerce Protocol (ACP). You can now buy things through a conversation with an AI.

And sitting underneath all of this: Coinbase’s x402 protocol, which has quietly processed $600M in cumulative volume across 500,000 active AI wallets. Google’s AP2 standard has 60+ organizations signed up.

Gartner says 58% of crypto trading volume is already bot-driven. McKinsey projects $3-5 trillion in agent-mediated commerce by 2030. 80% of organizations have already observed “risky AI agent behaviors” including unauthorized transactions.

For founders: the agent payment stack is assembling in real-time and it needs middleware. Guardrails (Nava), interoperability (Alchemy), identity verification, insurance, fraud detection — all of it is greenfield. If you’re building at the intersection of AI and payments, you’re not early.

The last time this many companies raced to own a checkout experience, it was the browser wars of the 2000s. This time the buyer isn’t a human with a credit card. It’s a piece of software with a stablecoin balance. And it doesn’t sleep.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ The attack vector shifted from code to people

Crypto spent the last five years getting really good at auditing smart contracts. Turns out the attackers just switched targets.

Hacken published their Q1 2026 security report this week. The number: $482 million lost across 44 incidents.

63% of all losses, $306 million, came from phishing and social engineering. The thing your grandmother gets warned about.

A single hardware wallet scam in January accounted for $282 million. One person clicked the wrong thing and lost more money than most Series B companies have ever raised.

4/ AI is eating crypto's VC lunch

Let’s talk about the elephant in the room that nobody in crypto wants to acknowledge.

AI companies raised $242 billion in early 2026, 80% of all global venture capital. Gartner projects $2.52 trillion in total AI spending this year. Trillion. With a T.

Crypto raised about $5 billion in Q1. That’s roughly 2% of what AI pulled in. The funding gap is 48x. If crypto VC were a basketball player, it wouldn’t be on the bench — it wouldn’t be in the building.

Five VC newsletters landed in our inbox within 24 hours, all independently arriving at the same conclusion: AI compute is getting scarce, the cloud rental model is breaking, and the builders who own their infrastructure will have the advantage. Kerman Kohli published a deep-dive on building a $15K inference machine , 96GB VRAM, assembled by hand, zero prior hardware experience. His thesis: “worst case I resell it for more in 6 months given the way compute prices are heading.”

That’s a DePIN pitch disguised as a weekend project.

Here’s where it gets interesting for founders. The smartest crypto companies aren’t trying to compete with AI for VC attention. They’re positioning themselves as the layer AI needs to function.

AI agents need programmable payments? x402 has 500K active wallets and $600M in volume. Alchemy launched AgentPay as the interoperability layer across every competing agent payment protocol.

AI agents need identity? Onchain registries, verifiable credentials, “Know Your Agent” standards.

AI agents need settlement? Stablecoins. Already doing $33 trillion annually. Already faster and cheaper than SWIFT.

AI needs decentralized compute? DePIN networks, inference marketplaces, the entire Bittensor subnet ecosystem — all crypto infrastructure.

You can’t outbid a $242 billion industry for LP attention. But you can build the plumbing that industry depends on. That’s the entire payments stack for the machine economy.

Dragonfly’s Rob Hadick said it best: the distraction of DAT companies has faded, and VC can refocus on operating businesses. Galaxy’s Mike Giampapa added: “despite lacklustre price performance, we’ve seen unprecedented improvements in underlying fundamentals.”

Translation: the people writing checks don’t care about the chart. They care about whether your company sits at the intersection of the two biggest capital flows in technology.

AI has the funding. Crypto has the rails. The founders who figure out how to connect those two sentences will build the next decade of financial infrastructure.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

1/ The AI Pricing Floor Just Moved

Anthropic ended 2025 at $9B ARR. By February it was $14B. By March, $19B. This week: $30B. SaaStr reported that Salesforce took 20 years to reach $30B in annual revenue, which Anthropic did it in under three years.

The AI Corner

Anthropic Just Passed OpenAI in Revenue. Spending 4x Less

Let me give you the number first…

Read more

3 months ago · 78 likes · 7 comments · Ruben Dominguez

The growth is almost entirely enterprise. Sacra reported 80% of revenue comes from business customers. Over 1,000 companies now spend more than $1M annually on Claude, that number doubled in under two months.

But the same week the $30B number landed, Madrona’s newsletter flagged something buried in the headline. Anthropic cut off flat-rate access for third-party agent frameworks. Heavy users are being pushed to pay-per-token API billing. Anthropic also launched Claude Managed Agents — a signal, as Madrona put it, that they want to own the agentic layer, not just the model underneath it.

The flat-rate era is ending.

On April 2nd, Google DeepMind released Gemma 4: four sizes all under Apache 2.0 — unrestricted commercial use, no revenue thresholds, no custom clauses. According to Google’s official blog, the 31B model ranks third among all open models globally. It outperforms models 20 times its size.

Dapta reported the benchmark jumps vs the previous generation: math from 20.8% to 89.2%, coding from 29.1% to 80%. The smaller versions run on a Mac Mini, while the 2B model runs on a smartphone, with no API bills and cloud dependency.

This is a new alternative.

A Databricks 2026 report found that 75% of enterprises now use two or more LLM families in production. This is the multi-model era.

Foundation Capital’s 2026 AI predictions were direct: founders who can prove measurable value per outcome will compound. The ones building on assumptions that pricing stays flat are already exposed.

So basically check your AI cost assumptions: if your product runs heavy workloads on Claude, figure out which parts actually need it and which ones don’t. Gemma 4 handles most standard tasks and costs nothing to run. Splitting your model usage by task type is the easiest margin improvement available to you right now.

When you go to raise your next round, an investor is going to ask you what your gross margins look like at scale. You should know exactly how much of your workload actually needs Claude versus what a free open source model handles just as well.

2/ The Companies Doing "Okay" Are The Most At Risk

There’s a pattern showing up in the newsletters this week that’s worth paying attention to. The companies getting disrupted aren’t the ones in trouble. They’re the ones with decent numbers, a stable team, and no real reason to panic yet.

Zapier’s CEO Wade Foster told GTMnow this week that when GPT-4 launched, he called a company-wide “Code Red.” It’s because he could see what was coming and knew that good-enough numbers would kill the urgency to move. He ran a company-wide hackathon function and went from under 10% daily AI usage to over 50% in a single week. They’re now at 97% company-wide as AI handles 50% of their customer support tickets. His sharpest observation: the leaders who aren’t using AI themselves are the single biggest bottleneck in their companies. You can’t define what “good” looks like if you’ve never done it yourself.

Ramp is the other side of the same story. Ed Sim's What's Hot in Enterprise IT newsletter reported 6,300% growth in AI usage at Ramp and 1,500 internal apps built in six weeks by non-engineers.

What's Hot 🔥 in Enterprise IT/VC

What’s 🔥 in Enterprise IT/VC #493

The discussions on how to go wall-to-wall with agent building and deployment across every organization continues every week. I’ve already written about this a couple of times in the last month from What’s 🔥 #488 (When Product Velocity Breaks the Company) and…

Read more

3 months ago · 22 likes

Keith Rabois said in Lenny's Newsletter this week that AI is collapsing the PM role. If one engineer can spec, build, and ship a feature in a day, the person in the middle translating business requirements into tickets stops being necessary. That's already happening at the companies that moved early.

Look at your last sprint. Count how many tickets needed a PM to translate a business problem into an engineering task. That number is your exposure.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ The VC Market Split In Two

So Q1 2026 was technically the biggest venture quarter in history. $300 billion raised globally according to Crunchbase. Record-breaking, unprecedented, all the superlatives. But if you actually look at where that money went, the number is almost misleading. Four deals, OpenAI at $122 billion, Anthropic at $30 billion, and a couple of others, absorbed the vast majority of it. Strip those out and you're looking at a pretty normal, actually fairly tight market for everyone else.

Last Money In's newsletter this week published probably the most honest breakdown of what's actually happening. Their argument is that xAI's exit, where SpaceX acquired xAI at $250 billion, got misread by the entire industry as a template. Investors saw a $6 billion seed round turn into a $250 billion outcome and thought "okay so that's what billion-dollar AI seeds can return." The problem, as Last Money In laid out clearly, is that the xAI outcome required Elon Musk to control both sides of the transaction. He founded the company and he owned the acquirer. This is while Machines Lab raised $2 billion at seed, Safe Superintelligence - $1 billion, World Labs - $1 billion. None of them have meaningful revenue. Most haven't shipped a product.

A $1 billion seed investment into a company valued at $5 billion needs a $25 billion exit to return 5x gross, and that's before dilution, fees, and carry. To reach a $25 billion valuation as a software company in public markets today, you need somewhere north of $1.5 billion in ARR at a best-in-class revenue multiple. For context, Samsara, Veeva, Zoom, and HubSpot — genuinely exceptional public software businesses — can't clear that bar at their current valuations. So the exit math on most of these mega-seed rounds doesn't work unless you assume outcomes that the entire history of the software industry says are extremely rare.

Meanwhile seed is actually fine. Moxxie Ventures founder Katie Stanton told Crunchbase that it's now a barbell market: "you're either an AI elite team growing really fast and you'll raise a ton of capital at Series A from one of the big firms, or you're everybody else." The number of active crypto investors collapsed from 5,345 unique names in deals in 2022 to just 377 in the last 90 days according to Last Money In, and they noted the real picture is probably even starker because that data lags. The Series A to B conversion rate is declining. Series B and C rounds are getting harder to close if you don't have a clean AI narrative. QED Investors said it plainly in their April newsletter: "AI companies are getting funded very easily with high valuations, and regular fintechs, even when performing well, are struggling to get funded."

Wellington's VC outlook report put it this way: only companies with the strongest competitive positions are attracting substantial funding. That sounds obvious until you realize it means a company growing 40% year-over-year with solid retention might still struggle to raise a Series B if it doesn't have a clear answer to "what happens when OpenAI ships this feature."

If you’re at seed, the market is reasonably open but the bar has moved. TechCrunch’s investor roundup found that VCs are now looking specifically for what they’re calling “high-context founders” — people who have spent years inside a specific industry and can combine that domain knowledge with AI. The pitch that works is “I spent seven years running operations at an X company, I know exactly where the pain is and exactly who is going to buy this on day one.”

If you’re at Series A or B, the question you need to answer before any investor asks it is what your defensibility looks like when a larger player enters your market. A specific answer about the data you have, the customer relationships you’ve locked in, the workflow integrations that would be painful to replace. If you can’t answer that clearly in two sentences, that’s what’s going to kill the round.

And if you’re thinking about a mega-seed round because the market looks open at the top — read the Last Money In piece first. The math on those rounds requires outcomes that almost no software company in history has delivered.

4/ Nobody Built The Agentic Governance Layer

HumanX just wrapped in San Francisco, 6,714 AI leaders from 79 countries, 500+ sessions. The thing that came through across almost every session is that we’ve crossed a threshold. Agents are running in production inside real companies handling real workflows and real money.

Jensen Huang mapped it out at the keynote. He called it three waves: generative, reasoning, agentic. His prediction, via the GTMnow newsletter recap, is that within one to two years a non-technical CEO will describe a business problem and an agent will integrate the systems, read the documentation, iterate, and report back.

Mayfield surveyed 266 CIOs and CTOs. 42% are already running agents in production. 91% are increasing their agentic AI budgets. Salesforce’s Agentforce is at $800M ARR with 25,000 customers. Slack crossed a billion messages a day with a 1,000% increase in AI apps being built on the platform.

But here’s the part that stands out. While 91% of companies are increasing their agentic AI budgets, 60% have early-stage or no formal AI governance framework. That’s most companies running autonomous systems in production workflows without a real structure for what the agent can and can’t do, what happens when it gets something wrong, or who’s responsible when it drifts.

The agent wars are also moving fast, four distinct camps shipping to real users. OpenClaw runs locally with 100+ built-in skills and persistent memory. Anthropic’s Claude Cowork is the managed, safe-by-design alternative. Base44 Superagents lets non-technical users describe tasks and spins up purpose-built agents running 24/7. Perplexity’s Comet turns the browser into an action engine.

Rob Seaman from Slack made the most practically useful point of the whole conference: the gap between an agent people use and an agent people ignore is almost entirely placement, not capability. When Slack tested agents as a separate tool users had to navigate to, adoption was low. When they embedded the same agents invisibly into the flow of work, so asking a question in Slack just surfaces the right agent contextually, adoption jumped 25% immediately.

Salesforce talked openly at HumanX about the fact that building their Agentforce agent took one and a half months. Actually operating it — measuring it, catching drift when business conditions changed — took two more months and is ongoing. The vendors pitching "build an agent in hours" are selling you the first mile. The other 99 is Day 2 operations.

Foundation Capital said the pricing implication clearly in their 2026 predictions: the market is moving from per-seat to per-outcome. You pay per task completed, per workflow resolved. The opportunity is aligning revenue with value. The trap is that you need instrumentation to prove it, and 70% of enterprise buyers want to test in their own environment before committing anything.

Your agent's capability is probably fine. Your actual problems are placement and Day 2. Where does your agent live relative to where the work already happens? And what happens when it starts doing the wrong thing at scale, do you have observability built in to show a customer exactly what their agent did and where it drifted?

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

1/ Tariffs rewrote the fundraising map

April 2, 2025. Trump announces sweeping import tariffs from the White House Rose Garden, Liberation Day. Markets crater. Policy changes 50+ times over the following eight months. February 2026: the Supreme Court rules the tariffs unconstitutional. Refunds are issued and the legal chapter closes. But the structural effect on how VCs deploy capital, which stages they favour, and how they diligence founders has not reversed. A year later the map looks different and most founders haven't updated their mental model of it.

The most concrete effect of the tariff year on venture isn’t in due diligence or hardware costs — it’s in the exit pipeline. Klarna, StubHub, and Chime all pulled their IPO plans. WHOOP CEO Will Ahmed said publicly the target is 12–24 months. Rebellions, the Korean AI chip company, did a $400M pre-IPO round instead of listing. CoreWeave completed its IPO in Q1 and is essentially the only exception.

This matters beyond late-stage founders. The chain runs upstream:

no IPOs → VC funds don’t generate liquidity → LPs get cautious → VCs raise smaller follow-on funds → fewer dollars available at every stage

Torch Capital put it directly: “when you have an IPO and a listing, it has multiple beneficial effects on the venture ecosystem: liquidity for funds, optionality for employees, cycle refresh. When that gets clogged, all of those things back up.”

Even with the IEEPA tariffs struck down by the Supreme Court, the diligence habits they created stuck. VCs added tariff exposure as a standard due diligence question in early 2026. Even as specific tariff rates were refunded and rolled back, the underlying concern (supply chain fragility, cross-border revenue concentration, infrastructure cost exposure) remained valid and in use.

Hardware founders now need to show alternatives outside China, a cost model that survives a 25% tariff scenario, and lead time contingencies if chip supply disrupts. SaaS founders need to show enterprise customer retention data under budget pressure and their AWS cost exposure. These questions existed before, but they’re now first-meeting questions, not late-stage diligence ones.

EU VC deal volume dropped to its second-lowest level in a decade in 2025 according to KPMG. US VCs pulled back from EU-exposed sectors as cross-border risk became harder to model. 50% of German companies with US business planned to invest less in the US or postpone, per the German Chamber of Commerce.

But something else happened at the same time. Gleamer, a Paris-based AI radiology company, exited at €230M on €50M raised, 4.6× return. Elaia doubled its fund to €254M. Tech talent flow reversed US → Europe for the first time, per The Economist. European founders who read that gap early had a window that US-facing founders missed.

Seed stage is the safest place to be in this environment. Early investors don’t need IPOs to return capital to LPs — they’re on 7–10 year fund cycles. Torch Capital said explicitly: “seed is getting a free pass.” The data matches: check sizes up 31% YoY even as deal count fell 30%. Fewer companies funded, larger rounds when they are.

If you’re series B or later with hardware exposure, adjust your timeline. The large round window is not open. VCs with cross-border exposure are being cautious. If you can run lean for another 12 months you probably should.

The pitch framing that’s working this month: domestic-first, global optionality. Six months ago it was the opposite. VCs are repricing “global-first” startups downward — a company built to scale everywhere now reads as a company with concentration risk everywhere. Know your revenue geography before you walk in.

2/ Defense tech stopped being a niche

Defense tech has been "the next big VC category" for three years running. This week it actually became one. Not because of the dollar amounts, the amounts have been building since 2023, but because of what changed underneath them. The Pentagon moved from pilot programs to enterprise contracts. Kleiner Perkins, a16z, and Bessemer all wrote checks into the same sector in the same week. And a consortium of VC-backed companies, Anduril, Palantir, Saronic, SpaceX, OpenAI, Scale AI, is now jointly bidding for government contracts against the traditional defense primes. That's a different world from the one that existed 18 months ago.

The signal is the Pentagon’s contracting shift. In March 2026, the US Army signed a 10-year enterprise contract with Anduril, ceiling of $20 billion, consolidating more than 120 separate procurement actions into one vehicle. The same model was used for Palantir last year: a $10B, 10-year enterprise agreement wrapping 75 existing software contracts. They’re the kind of fixed-price, multi-year deals that traditional defense primes like Lockheed Martin and Raytheon used to be the only companies capable of winning.

PitchBook senior analyst Ali Javaheri called the Anduril deal “a meaningful signal that the government increasingly sees Anduril’s stack as repeatable and scalable, rather than bespoke R&D.” That distinction matters enormously for investors. Bespoke R&D = unpredictable revenue. Repeatable, scalable enterprise contract = predictable revenue. The VC math changes completely.

Anduril and Palantir are forming a joint consortium with SpaceX, OpenAI, Saronic, and Scale AI to bid for government contracts together. The stated goal: deliver AI infrastructure to the military that no single company can build alone. The unstated reality: six VC-backed companies locking arms to compete directly against Lockheed Martin, Boeing, and Raytheon for the contracts that define defense budgets.

US defense spending is $900B in 2026. The Trump administration has plans to increase it to $1.5T in 2027. But despite that headline, VC-backed defense tech companies currently account for less than 1% of total defense contract dollars — and 88% of that tiny share goes to Anduril, Palantir, and SpaceX. The TAM is enormous. The access to it is currently very narrow.

The Saronic round is the clearest signal of who moved into this category. Kleiner Perkins led it. a16z, Bessemer Venture Partners, DFJ Growth, Caffeinated Capital, and Advent International all participated. These are generalist and tech-first firms, not defense specialists. The fact that Kleiner Perkins led a $1.75B round into an autonomous shipbuilding company means their investment committee approved “we build warships” as a VC thesis. That’s genuinely new.

Kleiner Perkins partner Ilya Fushman explained the thesis directly: “Maritime dominance isn’t just about technology — it requires the production capacity to field it at scale. Those two things rarely come together.” That framing — technology AND manufacturing scale, simultaneously — is exactly what the traditional defense industry couldn’t do. It’s what these companies are betting they can.

The most relevant thing about this week isn’t the valuations. It’s the supply chain that forms around these companies. Saronic needs AI systems, simulation software, sensor manufacturers, logistics tools, and cybersecurity layers. Shield AI just acquired Aechelon for its simulation technology. Anduril has made multiple acquisitions of adjacent tech companies. Every one of these “neoprime” platforms is a customer for founders building the technical layers underneath autonomous systems.

The pitch for those founders has changed. Six months ago, “our customer is the US military” was a hard conversation with most VCs. Now the pitch is “our customer is Anduril / Saronic / Shield AI, who has a $20B enterprise contract with the US Army.” That’s a different conversation. Know which VCs moved into this category and which ones haven’t. The Kleiner Perkins, a16z, Bessemer move this week is the permission slip for a lot of downstream pitches.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ The personal health OS thesis

WHOOP raised $575M at $10.1B and announced a Stanford partnership to build the first FDA-grade longevity score. Oura is interviewing IPO bankers and pursuing FDA clearance for blood pressure detection. Anthropic paid $400M for an 8-month-old biotech company to build AI models for biological research.

Newcomer

Anthropic Buys Stealth Dimension-Backed Coefficient Bio in $400M+ Stock Deal

Read more

3 months ago · 90 likes · Eric Newcomer

Three organisations, three different starting points, all moving toward the same place: continuous wearable data powering clinical-grade health predictions. The companies building in the gap between those three — the clinician platforms, the employer health tools, the drug trial acceleration layer — are looking at the first large healthcare TAM that doesn't require building hardware from scratch.

These three companies are building toward the same point from opposite directions. WHOOP and Oura are moving up the clinical credibility stack: from consumer wellness tracker toward FDA-grade clinical screening tool. Anthropic is moving down the abstraction stack: from frontier AI toward biological research and eventually clinical applications. They are going to meet in the middle, and the meeting point is predictive health: continuous physiological data plus AI model plus clinical validation.

The WHOOP Stanford move is the clearest signal. The THRIVE coalition, backed by $34.5M from ARPA-H (the US government’s health research agency), is building the first FDA-grade “Intrinsic Capacity” score, a predictive model designed to forecast major health outcomes up to 20 years in advance. WHOOP’s wearable provides the continuous physiological data feed. Stanford builds the model. ARPA-H funds the validation. If it works, it means a consumer wearable can power a clinical-grade health prediction. That changes the regulatory category, the billing category, and the insurance reimbursement category simultaneously.

On January 6, 2026, the FDA revised its General Wellness Policy for low-risk devices, broadening the pathway for wearables to make health-related claims without going through full medical device clearance. Oura’s chief medical officer described it as “a pivotal milestone.” WHOOP is operating on both sides simultaneously: it received an FDA warning letter for Blood Pressure Insights (marketing without clearance) while also formally partnering with Stanford on a federally funded clinical validation program.

The regulatory boundary between consumer wellness and medical device is the most commercially important line in health tech right now. Apple cleared it for hypertension notification. Oura is in active IRB study to cross it for blood pressure. WHOOP is trying to cross it for a 20-year predictive longevity score. Whoever clears that line first sets the benchmark for what every insurance company, employer health programme, and clinical trial operator will pay for the data.

WHOOP and Oura are on almost identical trajectories and they both know it. WHOOP: $1.1B run rate, +103% YoY, IPO in 12–24 months. Oura: $1B revenue in 2025, projecting $2B in 2026, interviewing bankers as of March 24. Whoever goes out first sets the financial benchmarks that define how public markets value the entire personal health OS category, growth rate, retention, data asset multiple, clinical partnership valuation. The first mover sets the comp.

Oura’s largest customer is the US Department of War, which uses the rings for military personnel monitoring through a Palantir software contract. That’s institutional health monitoring at scale. WHOOP’s investors include Mayo Clinic, Abbott, Qatar Investment Authority, and Mubadala, not typical consumer fitness investors. The signal in the cap table: both companies are being underwritten as healthcare platforms, not wearables companies.

The hardware and the AI are already built. WHOOP and Oura have the data platforms. Anthropic and OpenAI have the AI models. The regulatory pathway just got clearer with the January 2026 FDA guidance. The gap is in the application layer: the tools that sit between the wearable data and the clinical outcome.

Three categories where that gap is widest right now: employer health programmes (Oura’s B2B contracts represent 23% of new ring sales), clinical trial acceleration (if ARPA-H’s FDA-grade IC score works, you can run 5-year trials in 6 months), and insurance reimbursement infrastructure (Truemed’s Garmin deal in 2025 was the first wearable reimbursed through HSA — that playbook needs to be built for WHOOP and Oura at scale). None of these require building a ring or a frontier model. They require knowing how to connect the two in a way that a payer or a clinician will pay for.

4/ General Catalyst's AI roll-up thesis

Every VC is investing in AI companies that sell software to service businesses. General Catalyst is doing something different. They are buying the service businesses: call centres, law firms, IT providers, property managers — deploying AI to automate 30–70% of the work, and compounding through acquisition until the margin profile looks like software. They call it the Creation Strategy. $1.5B committed from their $8B fund. Eight companies built or actively rolling up. The early results are real enough that Lightspeed, Thrive Capital, and 8VC have launched their own versions. It is a structurally different model for what AI can do to a $16 trillion industry that most VC firms haven't touched.

The traditional PE roll-up model: buy a business at 8× EBITDA, cut costs, add some adjacent acquisitions, exit at 12× EBITDA. The margin improvement is incremental and the value is mostly in the multiple arbitrage. What GC is doing is structurally different because the margin improvement isn’t incremental — it’s transformational. A call centre running at 15% gross margin that automates 80% of its routine work becomes a 60–65% gross margin business.

Marc Bhargava, the GC partner leading this, is explicit about the model: “Services is a $16 trillion-a-year industry globally. When we AI-enable a workforce, you can now take on twice as much work with the same team because they have more tools and they’re automating away the easier tasks.” The key point: GC is not laying people off to cut costs. They’re automating tasks so the same team can handle more volume. Crescendo doubled resolved calls in the first week at a regional telecom after deployment — not because they cut headcount but because customers had stopped calling due to busy signals. The demand existed; the capacity to serve it didn’t.

The financial arbitrage underneath this strategy is the gap between how service businesses are valued today and how they should be valued after AI transformation. A call centre trading at 8× EBITDA on 15% margins becomes a 60% margin business after AI deployment. Same revenue base. The EBITDA grows dramatically. If you’re buying it at pre-AI pricing and your transformation works, the return is the kind of multiple that makes VC funds work.

GC is also explicitly building data moats. Every acquired company contributes proprietary operational data to the platform. A call centre for a telecom produces different training data than one for a healthcare company. Over time, GC’s portfolio companies know things about their specific service verticals that a generic AI cannot replicate from outside. That’s the defensibility argument that separates this from “AI SaaS with a services business attached.”

GC ran this largely alone through 2024. By mid-2025 it had “officially gone mainstream” per Newcomer.

Newcomer

Inside the VC Roll-up Craze That Has Taken Silicon Valley by Storm

Read more

7 months ago · 43 likes · 1 comment · Madeline Renbarger

Thrive Capital launched Thrive Holdings: a $1B+ evergreen vehicle dedicated to AI-enabled service roll-ups. Joshua Kushner’s bet is specifically on accounting, IT services, and insurance. Lightspeed moved into engineering services and healthcare. 8VC has backed Sequence Holdings, which is attempting the model across multiple unrelated industries simultaneously. Total identified capital across all players: over $3B committed specifically to this strategy.

The competition is already reshaping acquisition dynamics. Multiple buyers pursuing the same service businesses means target valuations are rising. Founders of profitable service companies now have more exit optionality than at any point in the last decade, and the buyers aren’t legacy PE firms, they’re VC-backed platforms that will pay a technology premium for businesses that can be AI-transformed quickly.

If you run a profitable service business — accounting, IT support, legal operations, property management — you now have institutional buyers offering technology premiums on top of traditional PE multiples. The conversation used to be “a PE firm will buy you at 8× EBITDA if your numbers are clean.” Now it’s “a VC-backed platform will pay a technology premium if your business can be AI-transformed quickly.” Know which players are active in your sector before you take the first call.

If you’re building AI tools for service industries — workflow automation, document processing, client management — your potential customers just got a lot wealthier and more motivated. GC’s portfolio companies have capital and mandate to deploy technology. Titan MSP needs tools to automate the other 62% of tasks it hasn’t automated yet. Eudia needs tooling to scale fixed-fee legal services across more practice areas. These aren’t companies watching their budget; they’re companies actively assembling their stack.

If you’re thinking about starting something in this space: the model is being replicated by Thrive, Lightspeed, and 8VC. Second-time founders who understand both AI and operations are the profile being sought. GC’s Marc Bhargava said explicitly: “The hardest part is getting proprietary data. The hardest part is go to market. The best founders coming to us have done M&A themselves.” That’s the talent specification.

Below, is the data summary of the most important numbers and deals this week, accompanied by the crypto VC funding trends spotted so far.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

A single leaked draft text from Washington did more damage to Circle in one afternoon than any competitor has managed in three years. On March 23, Senate staffers shared the latest CLARITY Act language with crypto industry reps on the Hill. The text bans passive yield on stablecoin balances and anything "economically equivalent to interest." By Tuesday March 24, Circle stock was down 20%, its worst single day since IPO. Coinbase dropped 10% in sympathy. Then the same day, Tether announced it had hired KPMG for its first-ever full financial audit. Circle's entire competitive advantage, being the "clean" stablecoin, just got flanked from both sides at once.

The reaction split the industry in two. Coinbase quietly told Senate staffers it cannot support the current draft.

On X, Crypto Banter's Ran Neuner backed Armstrong's position, saying the banks lobbied hard for this and got exactly what they wanted.

But Frax Finance founder Sam Kazemian went on The Rollup's Stabled Up podcast and said the opposite: accept the yield compromise, pass the broader bill, come back for yield in the next cycle. His argument: a passed law is almost impossible to undo, a regulatory guidance can be reversed the next day.

Keyrock's Amir Hajian was blunter: "It pulls the rug on the pass-through model that has been driving stablecoin adoption." Mizuho analyst Dan Dolev told CoinDesk the ban "weakens a key part of the bull case" for USDC.

Not everyone panicked. Bitwise CIO Matt Hougan called the market reaction overblown, noting interest income was never the primary driver of stablecoin growth. The real read came from Kazemian's most underappreciated point: Tether has never paid passive yield to holders. This ban changes nothing for them. It changes everything for their competitors.

Bloomberg ran a piece on March 26 with a headline that framed the whole week: "Crypto VCs Are Abandoning Web3 for the Dependability of Stablecoins." Dragonfly's Haseeb Qureshi, closing a $650M fund, said it plainly: "non-financial crypto has failed."

a16z's Arianna Simpson called stablecoins the "belle of the ball." The capital is going into plumbing.

a16z crypto

Privacy trends for 2026

Read more

6 months ago · 41 likes · 4 comments · a16z crypto

The yield ban doesn't kill stablecoins but one business model and opens another. Activity-based rewards stay legal. That's your product brief. Senate markup is late April. Polymarket gives it 49% odds of passing this year. If it misses the pre-midterm window, this goes dark until 2027.

2/ Wall Street decided tokenization is inevitable

Five separate institutions made major on-chain moves this week. The NYSE signed an MOU with Securitize to build a blockchain-based trading platform for tokenized US equities and ETFs, 24/7, T+0 settlement, no DTCC. NYSE Group President Lynn Martin said it's about preserving "the trust and transparency investors expect" while modernising the rails. Securitize manages $4B+ in tokenized assets and is planning a SPAC listing at a $1.25B valuation.

Franklin Templeton, a $1.7 trillion manager, partnered with Ondo Finance to offer five tokenized ETFs tradeable 24/7 from crypto wallets, covering growth equities, large-cap, high-yield bonds, equity income, and gold.

Sandy Kaul, Franklin Templeton's head of innovation, called it "a good test case to see what appetite looks like from this new audience." Ondo now controls roughly 60-70% of the tokenized equities market with $700M+ TVL. BlackRock added Chronicle Protocol as a proof-of-asset verification layer for its $2.1B BUIDL fund. Larry Fink’s annual shareholder letter compared tokenization to “the internet in 1996.”

On March 25, Congress held what FinTech Weekly called its most important tokenization hearing yet. Both parties agreed the actual debate about framework was coming. Witness Amir Banaei flagged something most founders aren’t watching: under a 1982 tax law called TEFRA, peer-to-peer token transfers are functionally indistinguishable from bearer bonds, triggering a 30% withholding tax. That’s a structural barrier sitting inside the plumbing right now. Bernstein projects $80B by end of 2026. McKinsey’s base case is $2 trillion by 2030.

The institutional buyers are real now but they're choosing partners today. NYSE picked Securitize. Franklin Templeton picked Ondo. The gaps that remain are compliance tooling, privacy layers, cross-chain settlement, and wallet UX for allocators who have never touched crypto. That's where early-stage capital is still available.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

---

3/ The age of vertical models is here

On March 26, Intercom CEO Eoghan McCabe posted something that should be required reading for every founder with an AI product in their roadmap. Intercom quietly shipped Apex, a brand new model trained entirely by their own AI team on customer service data. It beats GPT-5.4 and Opus 4.5 on their domain. A company that started as a chat widget just out-modelled the frontier labs at their own use case.

McCabe’s argument is simple and uncomfortable for a lot of builders: the general model race is not your race. The founders who win are the ones sitting on proprietary domain data and actually using it. Intercom has processed billions of customer service interactions. That data trained Apex. No amount of clever prompting on a general model closes that gap.

This is not an isolated case. Harvey Team built a legal reasoning model that outperforms GPT on contract review.

Sierra built a customer agent that runs on its own fine-tuned stack.

The pattern is consistent: vertical models trained on real domain data are pulling ahead of horizontal foundation models for specific tasks. The frontier labs keep winning on benchmarks. The verticals keep winning on production.

The numbers behind Intercom’s move tell the full story. McCabe announced a $250M debt raise on March 9, hiring 650 people globally. Intercom crossed $400M ARR with Fin, their AI agent, on track to hit $100M ARR as a standalone product. The company has a 60-person AI group doing nothing but model training.

The take that cuts deepest came from McCabe’s own post: “Answering service questions is a neat trick. But it’s not nearly taking advantage enough of the tech we’ve got in 2026.” His point is that most companies are still treating AI as a support cost-saver. Intercom is building agents that sell, onboard, retain and advise, and using their own model to do it.

If you are building a web3 product with any AI layer, your moat is not your API call. Every competitor has the same API. Your moat is the data only you have: transaction history, user behavior, protocol interactions, on-chain patterns. The founders who start collecting and structuring that data now will be the ones training vertical models in 18 months. The ones who don’t will be paying OpenAI forever and wondering why they can’t differentiate.

4/ Binance is losing its grip and founders have to pay the price

Two Bloomberg pieces dropped this week that would have been unthinkable a year ago. The first: "Binance's Grip on Crypto Loosens as Market Crash Rattles Faith." The second: "Binance Tightens Market Making Rules in Wake of Crash Criticism." Read together, they tell you the world's largest crypto exchange is on the defensive, and the fallout is landing directly on founders trying to launch tokens.

The backstory matters. On October 10 last year, crypto markets experienced a $28 billion liquidation cascade. Cathie Wood said publicly on Fox Business that it was “the flash crash associated with a software glitch on Binance.” The OKX CEO piled on. Binance denied everything, calling it “overall market conditions.” That dispute never got resolved. It just went quiet until this week, when three back-to-back token manipulation scandals made it impossible to ignore.

The SIREN token is the one everyone is talking about. It launched on BNB Chain as an “AI agent analyst” in early 2025, sat dormant, then mysteriously surged from $0.08 to $3.61 between February and March 22, a 4,500% move. Bubblemaps issued a warning on March 22: a cluster of 200+ wallets held roughly 50% of circulating supply.

Hours later the crash began. Chain analyst EmberCN dug deeper and found that 52 of the top 54 holders of circulating tokens were connected. Within 72 hours SIREN fell 71% from its peak. On X, people called it “the biggest scam of 2026.” This came after similar episodes with GPS and MOVE tokens. Three in a row, all on Binance-listed contracts.

Binance’s response on March 25 was a blog post titled “A Red Flag Guide for Crypto Market Makers”, banning revenue-sharing models between projects and market makers, prohibiting price manipulation arrangements, and threatening to blacklist bad actors. The crypto community’s reaction was split. Some read it as a genuine cleanup. Others were less charitable. Earlier on analyst Ignas from DeFi posted on X: “First Cathie Wood, now the OKX CEO coming after Binance.”

Mike Dudas of 6th Man Ventures said Binance collects a “10% tax” on every project that lists. One analyst cited that nearly three-quarters of Binance-listed projects have dropped to near zero. Deep Tide TechFlow summed it up bluntly: “GPS was the beginning, MOVE was the upgrade, and SIREN was a complete public humiliation, and it happened on Binance’s own listed futures contracts.”

The structural problem is real and worth understanding. In a bear market with low organic volume, a new project listing on Binance often can’t support its daily trading volume with real buyers. So “active market makers” fill the gap but unlike traditional market makers who earn from the spread, these ones take zero-cost tokens upfront and sell under the guise of providing liquidity. It’s legal on paper but extraction in practice. Binance’s new rules target this model directly. If you are planning a TGE, the Binance listing playbook is being rewritten in public right now.

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

1/ Agentic Payments: The Race to Build Money Rails for AI

On March 18, Stripe and Paradigm’s blockchain startup Tempo went live on mainnet. The headline number is $500M raised at a $5B valuation, but the more interesting thing is what they actually shipped alongside it.

Tempo launched the Machine Payments Protocol, an open standard for letting AI agents pay for things autonomously. The way it works is pretty elegant. Instead of requiring human approval on every transaction, an agent authorizes a spending limit once upfront, and then streams micropayments continuously as it uses services: data APIs, compute, model calls, whatever it needs. One authorization, thousands of payments. No human in the loop for each one.

The design partner list gives you a sense of how seriously this is being taken. Anthropic, OpenAI, Visa, Mastercard, Revolut, Shopify, Standard Chartered, DoorDash, Ramp, and Nubank all came in as partners. Over 100 services were integrated on day one. Klarna launched its first-ever stablecoin on the network during testnet. There’s no native token as gas fees are payable in any major stablecoin, which is a deliberate choice to avoid the speculation circus.

The reason crypto makes sense here, structurally, comes down to economics. Card rails charge a minimum of around $0.30 per transaction. If an AI agent is calling a specialized API ten thousand times to complete a research task, the total revenue for that API provider might be $10 or $15. That math simply doesn’t work on traditional payment infrastructure. Onchain, those same payments cost fractions of a cent each and settle instantly.

Tempo isn’t the only one building this. Coinbase has been at it for a while with x402, a protocol that revives an old HTTP status code, “402 Payment Required”, and turns it into a programmable payment rail for AI agents. It’s already logged over 50 million total transactions and is used as the crypto settlement layer inside Google’s Agent Payment Protocol. On March 17, Sam Altman’s World project launched AgentKit in beta, combining x402 with World ID so that AI agents can carry zero-knowledge proof that a real human is behind them. Visa has its own Trusted Agent Protocol running on card rails. Mastercard ran the first live AI agent bank payment inside Santander’s infrastructure last week. AWS published an x402 enterprise reference architecture. The whole space moved at once this week.

The thing worth paying attention to, though, is the gap between the narrative and the actual numbers. X402 currently processes around $28,000 in real daily volume, according to Artemis data.

A16z’s Noah Levine shared analysis showing that roughly 81% of December’s volume was artificial, wash trading, self-dealing, gamed stats.

The merchants that x402 was built to serve, the small pay-per-use API businesses that only make sense with sub-cent payments, are still pretty rare. Circle’s co-founder Sean Neville said it plainly in an interview this week: a lot of AI developers have a low opinion of crypto because of memecoins and Ponzi schemes, and that perception doesn’t disappear just because the underlying infrastructure is now genuinely useful.

None of this means the thesis is wrong. Every infrastructure cycle looks like this in the early stages, the rails get built before there’s traffic to run on them. Email servers existed before email was a thing people used daily. What it does mean is that founders building in agentic payments right now are operating in a 2026 funding environment for what is likely a 2028 adoption story. That’s not necessarily a bad position to be in, but it’s worth being clear-eyed about. The projects that will raise in this market are the ones that can show real agent transactions on real infrastructure, not just a compelling slide about where this is all heading.

2/ The VC Split: AI, Blockchain, Infrastructure

The biggest thing that happened in AI x crypto infrastructure this week was a philosophical split between the two largest funds in the space, playing out in public.

Fortune reported that a16z crypto is raising its fifth fund, which we mentioned on time on our X, targeting around $2 billion, with plans to close by end of first half 2026 - less than half the $4.5 billion Fund IV raised in 2022. The fund will be 100% blockchain focused, according to sources. Meanwhile, Paradigm is seeking to raise as much as $1.5 billion for a new fund that would expand its scope into frontier technologies, including artificial intelligence and robotics. Two firms that built their names doing the same thing are now betting on meaningfully different futures.

The reaction on X was immediate. A lot of people read Paradigm’s move as a quiet admission that crypto-only investing is running out of runway. Others pointed out that Paradigm and OpenAI literally built a tool together in February, EVMbench, a benchmark for testing how well AI models detect security flaws in smart contracts, so the expansion isn’t exactly a pivot away from crypto, it’s more like following where the work was already going.

Chris Dixon got ahead of the a16z narrative with a long post on X. He wrote: “We are clearly in the financial era of blockchains. But the core idea was never that every crypto application would emerge all at once, or that finance wouldn’t come first. Finance isn’t separate from the broader thesis; it’s part of it. It’s the foundation and proving ground for everything else.” He also added, pointedly, that blockchains are increasingly useful for coordinating AI agents too, threading his original thesis into the current moment.

The Crypto Valley Journal covered it drily, noting that a16z is not the only crypto VC firm seeking fresh capital, but Paradigm is also expanding its focus for the first time systematically into AI and robotics.

The Bankless Empire newsletter captured the community tension well, pointing out that most crypto Twitter hot takes on this are wrong in both directions: it’s not that crypto VCs are abandoning Web3, and it’s not that AI and crypto are fully merged yet either. Messari put it more bluntly in their latest podcast coverage: the next wave looks like application-specific infrastructure, stablecoins at scale, and energy/DePIN feeding AI, not generic infra or meme “vibe-coded” apps.

That DePIN point is worth sitting with. Networks such as Akash and io.net are attracting As Silicon Valley Bank said in their newsletter, AI compute workloads as miners shift from token incentives to actual revenue. Enterprise cloud buyers are tapping these networks for compute overflow capacity, edge computing, and distributed storage. SVB’s 2026 crypto outlook frames this as a genuine second act for DePIN, a sector that nearly died when token prices fell and subsidy-driven supply dried up, now finding real paying customers in AI companies that need GPU access they can’t get from AWS at a price that makes sense.

The numbers are more real here than in most corners of crypto. The DePIN sector has stabilized into a $10 billion market, generating $72 million in verifiable on-chain revenue over the past year, according to Messari’s State of DePIN 2025 report.

That’s not huge, but it’s actual revenue from actual compute work, not token emissions dressed up as traction. Bittensor sits at a $3.4 billion market cap, Render surged roughly 40% last week on strong volume, and the AI crypto token sector as a whole sits above $28 billion in total market cap as of this week.

The community reaction to Render’s move was split along predictable lines. CryptoLogicHQ (39.8K followers) posted: “After an 81-session floor at $1.18 and a textbook bear trap, Render has pierced its one-year resistance. Institutional conviction we haven’t seen in months.”

The more interesting long-term question, one that a16z spent a good chunk of their recent report on, is identity. Non-human agents now outnumber human employees in financial services, and a16z argues that cryptographically signed credentials, what they’re calling “Know Your Agent”, will be required for agents to transact. The idea is simple: KYC for bots. Just as you need a financial history to get a mortgage, AI agents will need verifiable on-chain credentials to access banking rails, execute contracts, or trade assets. The projects building that identity layer are the ones that a16z says are positioning for the biggest opportunity in the space. Blockworks covered the KYA thesis more skeptically, noting that it assumes a regulatory framework that doesn’t exist yet and a level of coordination between AI developers and crypto protocols that has so far been more aspiration than execution.

Coincub’s DePIN report made the most interesting point of the week, and it didn’t get nearly enough attention: DePIN becomes infrastructure once crypto disappears from the UX.

Real scale arrives when developers pay with credit cards, SLAs look familiar, and blockchain mechanics stay invisible in the background. That’s essentially the thesis for the entire AI x crypto convergence in one sentence. The winning products won’t announce themselves as crypto products at all.

For founders, the honest read from this week is that two things are both true at once. The VC money is real: a16z, Paradigm, Polychain, and Pantera are all deploying into this space with fresh capital and genuine conviction. And the infrastructure is also genuinely earlier than the funding narrative suggests. According to TechFlow, the era when startups could raise funds solely on narratives and PowerPoint decks is over. The projects getting checks right now have live networks, paying customers, and developer activity they can point to.

All of this was pulled from the Byblos live feed - 600+ VC firms, 27500+ newsletters, one place.

See what else came in today at byblos.digital →

3/ Europe's Biggest Money Just Moved On-Chain

The Chainlink newsletter dropped a headline on Saturday that most of crypto Twitter barely noticed: Amundi, Europe’s largest asset manager with €2.3 trillion under management, launched a tokenized mutual fund this week.

The Spiko Amundi Overnight Swap Fund uses Chainlink for automated NAV reporting and cross-chain interoperability. That’s not a pilot. That’s a €2.3 trillion institution putting real infrastructure on-chain.

The same week, CoinShares, a European firm that’s been building regulated crypto products for years, published a newsletter titled “Our on-chain era begins” and meant it literally. They launched their first on-chain asset management strategy, allocating across DeFi lending protocols, institutional secured lending, tokenized bond funds, and delta-neutral basis strategies inside a single vehicle. CEO Jean-Marie Mognetti put it plainly: “On-chain is not another crypto product. It’s a new distribution rail.” CoinShares also listed BNB and Toncoin staking ETPs on Deutsche Börse Xetra this week, quietly expanding the menu of regulated crypto products available to European institutional investors on one of the world’s oldest exchanges.

The institutional movement isn’t just European. In the US, Nasdaq received formal SEC approval on March 19 to trade tokenized versions of Russell 1000 stocks and select index ETFs. EPOCH Digital Credit launched a tokenized fund across six global markets simultaneously using Chainlink infrastructure.

S&P 500 perpetual futures went live on Hyperliquid this week under a licensing agreement with S&P Dow Jones Indices, 24/7 USDC-settled exposure to the world’s most-watched index, on a decentralized venue, for the first time.

The most useful thing to read alongside all of this is Multicoin Capital’s essay published Thursday. Titled “RWAs Are Just Built Different,” it’s the clearest framework the industry has produced for understanding how different asset classes will actually come on-chain and why they’ll each do it differently. Multicoin’s argument is that the industry has made a mistake by treating RWAs as one category.

Treasuries are wrapped funds and will stay that way, the underlying settlement infrastructure is too entrenched. Private credit can leapfrog to native on-chain origination faster because there’s no centralized clearing house to fight. Synthetics dominate equities and commodities because traders want exposure, not ownership. Real estate remains the hardest because tokenization doesn’t fix the underlying illiquidity. The framework matters because it tells you where to build, not just that the category is big.

The distribution problem is the real conversation though. Our Byblos signal on Friday flagged a notable piece from Electric Capital: they mapped 501 sources of real-world yield and found only 2 of 35 major RWAs above $50M have crossed 2,000 holders. Treasuries lead at $11B in tokenized supply. Private credit sits at $2.8B. But the product is ahead of the distribution.

The same week, BlackRock’s HLEND private credit fund gated redemptions, honoring only $620M of $1.2B in redemption requests and enforcing its 5% quarterly withdrawal cap for the first time in four years. The liquidity mismatch that hit HLEND investors this week is precisely what on-chain settlement is designed to solve, and yet most tokenized RWA products have barely a fraction of the holder base needed to prove the model at scale.

The macro backdrop makes the timing interesting. The CryptoSlate newsletter this Saturday led with British bond yields climbing toward levels not seen since 2008, describing it as an unintentional argument for why programmable, 24/7 settlement rails matter. Whether or not you buy the thesis that gilts panic proves Satoshi right, the point is that traditional fixed income infrastructure, the backbone of most pension funds and insurance companies, is creaking under macro stress at exactly the moment on-chain alternatives are hitting product maturity.

The angle that connects all of this for founders is MiCA. We also covered the CBC Summit Europe in London, which gathered 100+ executives and regulators the same day the SEC was making headlines in Washington. FCA’s Head of Digital Assets keynoted. Coinbase UK’s Senior Country Director was on stage. The whole conversation circled back to one date: July 1, 2026. Every EU crypto firm operating under transitional national arrangements has to be fully MiCA-licensed or shut down. CoinShares built its new on-chain strategy with dual AIFMD/MiFID and MiCA authorization from day one. The founders who treated it as a distribution advantage are now inheriting a market.

Receive up-to-date unique statistics of what VCs think about founders in Internet Capital Markets from our X - Byblos.Radar

4/ The Great Stablecoin Land Grab

Three of the world’s largest payments companies moved on stablecoins this week in ways that, taken separately, look like product decisions. Taken together, they look like panic.

Mastercard agreed to acquire BVNK, a London-based stablecoin infrastructure company processing $30B annually across 130+ countries, for up to $1.8B. That’s 2.4x the $750M valuation BVNK got six months ago. Coinbase reportedly entered exclusivity at around $2B and walked. The 51 Insights newsletter pointed out what makes this strategic rather than tactical: in 18 months, Stripe paid $1.1B for Bridge, MoonPay paid $100M+ for Iron, and now Mastercard paid $1.8B for BVNK. Every deal prices stablecoin infrastructure higher than the last.

The same week, PayPal expanded PYUSD to 70 countries in a single release, from 2 markets to 70 in one go, targeting remittance-heavy corridors where cross-border fees still average 6.49%. PYUSD’s market cap has grown from under $500M in early 2025 to $4.1B. PayPal also quietly launched PYUSDx in February, a framework that lets any app issue its own branded stablecoin on top of PYUSD reserves. They’re not just distributing a stablecoin. They’re becoming a stablecoin infrastructure provider for other companies.

And HSBC and Standard Chartered are expected to receive Hong Kong’s first stablecoin issuer licenses as early as March 24. 51 Insights made the sharpest observation on this: Hong Kong gave its first stablecoin licenses to the same banks that print Hong Kong’s physical banknotes. When the institution trusted to print paper money is also issuing digital money, stablecoins stop being a fintech category and start being monetary infrastructure.

The reason all of this is happening at once is in a Ripple survey published this week of 1,000+ global finance leaders: 74% now use stablecoins as cash flow management tools. 72% said offering digital asset solutions is required to remain competitive.

The Byblos daily on Friday flagged this as the clearest signal yet that stablecoins have decoupled from the crypto narrative entirely and crossed into corporate treasury. Meanwhile, stablecoin transfers on public blockchains hit $27.6 trillion in 2024, more than Visa and Mastercard’s combined annual card volume. The card networks aren’t fighting stablecoins but buying their way onto the new rails.

There’s a real founder stress buried in this narrative that doesn’t get discussed enough. If you’re building stablecoin payments infrastructure right now, the most acquisitive buyers in the world are circling your category. It means your potential distribution partners are becoming your most well-funded competitors. Mastercard isn’t just a partner anymore, it owns BVNK. Stripe owns Bridge and Tempo. PayPal is offering white-label stablecoin issuance to third parties.

The window where a startup could build something the card networks can’t is still open but it requires being further up the stack than payments infrastructure, or operating in corridors the incumbents won’t touch, like EM currencies and the long tail of cross-border flows that are still too fragmented and too small to be worth acquiring.

Today Signal’s Question: The payments giants just spent $3B acquiring stablecoin rails. The AI funds are splitting between pure crypto and frontier tech. Europe’s biggest money moved on-chain. And the infrastructure for AI agents to pay each other autonomously is live.

If you had to bet on which of these narratives produces the most funded web3 startups by end of 2026 - which one is it?

Everything in today’s digest came from the Byblos live feed. The full newsletter archive, the VC directory, and tomorrow’s signal are all waiting for you at Byblos.digital - the investor interned, curated.

That’s the week. Now go build something.

— Byblos.digital

Subscribe now

1/ BlackRock's ETHB: The ETF That Pays You to Wait

On March 11, BlackRock's iShares Bitcoin Trust (IBIT) pulled in $115.51 million in a single day, nearly every dollar that flowed into US spot Bitcoin ETFs that day. The next morning, as we reported in Byblos Digest, the firm did something it hadn't done with any previous crypto product. It launched one designed to pay you while you hold it.

The iShares Staked Ethereum Trust, trading under the ticker ETHB on Nasdaq, began trading on March 12. It's BlackRock's third crypto ETF and its first to incorporate staking, meaning the fund doesn't just track Ethereum's price. It locks up 70–95% of its ETH through Coinbase Prime, then passes roughly 82% of the resulting staking rewards to investors as monthly income.

At current yields of around 3.1% annually, BlackRock is effectively turning Ethereum into something that looks a lot like a digital bond. Jay Jacobs, BlackRock's US head of equity ETFs, told CoinDesk this was "really about investor choice." Bloomberg ETF analyst James Seyffart called the debut "very very solid for a day 1 ETF launch", $15.5 million in first-day volume on $100 million in initial assets.

Two things made ETHB possible that simply didn't exist when BlackRock launched its spot Ethereum ETF (ETHA) in July 2024. First, the GENIUS Act cleared the regulatory runway for yield-generating crypto products. Second, former SEC Chair Gary Gensler, who had explicitly instructed firms to strip staking from ETF filings, was gone. Under Chair Paul Atkins, the SEC approved the structure.

In terms of structural implication, Solana and Cardano staking ETF filings are already in front of the SEC.

BlackRock hasn't filed for either, and, as FinTech Weekly noted, it doesn't need to. The "spot ETF era" of crypto appears to be ending. What replaces it is productive capital: assets that earn while you hold them, packaged into wrappers that fit cleanly into brokerage accounts. The more uncomfortable question, raised across CT, is what happens to Ethereum's decentralization when a single Wall Street product potentially controls 5–10% of all staked ETH routed through a handful of validators. Whether that's a triumph for adoption or a quiet erosion of crypto's original ethos is a debate that isn't going away.

2/ Arthur Hayes' Big Macro Short: AI Is Going to Break the Banks

Bitcoin fell 52% from its October all-time high of $126,000 while the Nasdaq barely moved. For most analysts, that divergence was a crypto-specific story: cycle dynamics, ETF outflows, profit-taking. For Arthur Hayes, BitMEX co-founder and Maelstrom CIO, it was something else entirely.

In his Substack essay "This Is Fine," published in mid-February and still dominating CT discussion this week, Hayes argued that Bitcoin is functioning as a the most sensitive major asset to tightening credit conditions, pricing in a crisis that equity markets haven't acknowledged yet. The culprit, in his framework, is AI.

Hayes estimates there are roughly 72.1 million knowledge workers in the US, many carrying significant consumer debt and mortgages. If AI displaces just 20% of them fast enough to impair their ability to service debt, he models approximately $557 billion in consumer credit and mortgage defaults, roughly half the severity of the 2008 financial crisis, enough to put a meaningful portion of regional US banks underwater. Hayes wrote bluntly: "AI took der jobs... Sell and withdraw first; investigate later. This is how investors and depositors will approach this situation as soon as everyone knows that everyone knows."

The Fed, in his view, won't act until a visible banking crisis forces their hand as political dysfunction will cause them to "dither until shit gets really bad." But once they do print, Hayes expects Bitcoin to rip to new all-time highs. His playbook for Maelstrom: stay liquid, avoid leverage, and wait for the Fed to blink. When it does, he said he'll be deploying into Zcash and Hyperliquid's HYPE token. The most concrete update came on March 11, when Hayes told CoinStories that if he only had $1 to invest right now, he would not buy Bitcoin.

The article received very mixed impressions. Along with some negative comments on Substack from users, Ryan McMillin, CIO of Merkle Tree Capital, told Decrypt the thesis was "intellectually coherent but overstates the speed of near-term disruption?. Labor markets don't produce synchronized defaults in weeks, they grind down over quarters through attrition and hiring freezes.

That said, McMillin acknowledged "the directional concern isn't wrong: rising credit card delinquencies are already real, SaaS valuations are under pressure, and a rolling deterioration in consumer credit quality is plausible." The timeline, he argued, is probably more stretched than Hayes suggests. DL News noted pointedly that Hayes had predicted Bitcoin would end 2025 at $200,000, then revised to $150,000 but it ended the year at $87,500.

Whether this plays out as the macro call of the cycle or another Hayes overshoot, CT clearly can't stop talking about it.

3/ The Great VC Squeeze: Fewer Deals, Bigger Checks, and No Fresh Capital

The headline number looks healthy: crypto fundraising rose nearly 50% year-over-year in the 12 months through March 2026, hitting over $25.5 billion. But digging deeeper, the number of individual deals dropped 46%. Average deal size climbed 272% to $34 million. Active investors fell 34.5% to just 3,225. Messari described it plainly: "capital concentration is heavily skewed by late-stage and strategic mega-rounds."

February told the story in miniature - three deals alone accounted for 44% of the $795 million raised that month, per Messari. Then February's total itself was a 65.3% drop from January. The week of March 3–7 saw startups raise just $135 million, one of the slowest weekly totals of 2026. This week's Messari Venture Weekly logged 18 projects raising a combined $220 million, led by KAST ($80M), Cryptio ($45M), and Zcash Open Development Lab ($25M), which are solid deals, but almost entirely in infrastructure and compliance tooling.

Messari CEO Eric Turner flagged the structural issue directly: outside of Dragonfly Capital, no major crypto VCs have closed new fund raises recently, and "the industry needs some fresh capital."

The other pressure is where money is going instead. In February, global venture investment surged to $189 billion, a 780% year-over-year jump, with 83% going to just three companies: OpenAI, Anthropic, and Waymo, per CrunchBase Report. AI has captured roughly 50% of all global VC funding.

For crypto founders raising at seed or Series A, they're competing for the remaining scraps from a depleted, more disciplined investor pool that, as BeInCrypto noted, now demands real revenue and institutional clients before writing checks: "No L1s. No DEXs. No 'community-driven' anything."

The counternarrative getting traction this week came from Multicoin Capital's Shayon Sengupta, who published a thesis on March 10. He argued that "Internet Labor Markets", networks where users earn crypto through work like data labeling or bandwidth sharing rather than buying tokens, represent the next major onboarding wave.

"There are only two ways to enter crypto: buy or earn," Sengupta told CoinDesk. "The first path has been followed for the last ten years. The next decade, most will take the second." It's a compelling reframe of where crypto's next billion users come from. But for founders trying to raise right now, the philosophical long game feels distant. The money is in late-stage infrastructure, and early-stage crypto is largely on its own.

4/ Aave's 12 Days of Chaos and What It Means for DAO Governance

Aave is DeFi's biggest lending protocol. $27 billion in deposits. 65% market share. And in the span of 12 days this week, as we were following through Byblos Digests, four separate crises hit it one after another.

March 3 - The Governance Firm Walks Out

Marc Zeller, founder of the Aave Chan Initiative (ACI), posted on the governance forum that his team was leaving. ACI wasn’t a minor player because it drove 61% of all Aave DAO governance actions over three years and helped grow Aave’s stablecoin GHO from $35 million to $527 million in supply.

The trigger: a $42.5 million budget proposal from Aave Labs that passed its preliminary vote, allegedly backed by voting power from addresses linked to Aave Labs itself. “When we applied those same standards to the entity requesting the largest budget in DAO history, the system stopped working,” Zeller wrote. AAVE dropped 11% the same day.

Per DL News, this followed BGD Labs, the team that built Aave V3, announcing its own exit by April 1, citing Aave Labs accumulating control over the brand, social accounts, and voting power while pushing a V4 roadmap they viewed as hostile to the existing V3 system.

March 10 - The Oracle Gets It Wrong

A misconfiguration inside Aave’s CAPO oracle, a system designed to prevent price manipulation, caused it to undervalue wstETH by 2.85% against its real market rate. Per a post-mortem from Chaos Labs, Aave’s risk management partner, that small gap was enough to push 34 highly leveraged positions below their liquidation thresholds.

Around 10,938 wstETH was automatically sold off from accounts that were perfectly healthy under correct pricing. Liquidation bots captured 499 ETH in profits. Kulechov confirmed no bad debt was incurred and all 34 affected users would be fully reimbursed.

Chaos Labs CEO Omer Goldberg noted on X thread that the system had processed over 1,200 payloads without incident before this.

March 12 - $50 Million Becomes $36,000

A wallet attempted to swap $50.4 million USDT for AAVE through the Aave interface. The trade was routed through CoW Protocol into a SushiSwap pool with just $73,000 in total liquidity. The wallet received 327 AAVE tokens, worth roughly $36,000. MEV bots led by Titan Builder extracted approximately $34 million in ETH from the price dislocation within the same block.

It wasn’t a hack but the proper protocol work. Aave engineer Martin Grabina clarified on X: "It was just the accepted quote with 99% price impact." The interface had shown a warning and required a checkbox confirmation on mobile. Kulechov said Aave would refund roughly $600,000 in fees it collected.

Post Mortem was published by Aave later on as well.

CZ posted seven words: "Sad to see this. Liquidity is the best user protection."

The debate on X was quite active as Aave's senior VP of engineering Emilio Frangella defended the architecture. But Marc Zeller, now departing, argued earlier versions of the protocol had stronger guardrails that had weakened over time. DeFiLlama founder 0xngmi pointed out that LlamaSwap would have blocked the trade entirely: "If you try to make this swap on llamaswap the UI won't let you at all, buttons get locked."

DeFi Rate framed it best: permissionless and unprotected are not the same thing.

Four events, 12 days. Together, as The Coin Republic noted, they raise a harder question: when a protocol loses its governance team and its core development team in the same month, does the institutional knowledge needed to prevent the next failure leave with them?

5/ Project Crypto": The SEC and CFTC End Their Turf War

For years, the biggest cost of building a crypto company in America wasn't compliance but not knowing which regulator was coming for you. The SEC treated most tokens as unregistered securities. The CFTC said Bitcoin and Ether were commodities. Companies got sued by both, registered with both, and often just moved offshore.

On March 11, that era officially ended. SEC Chairman Paul Atkins and CFTC Chairman Michael Selig signed a Memorandum of Understanding establishing a shared framework: BTC and ETH explicitly classified as digital commodities under CFTC jurisdiction, securities-like tokens stay with the SEC, and both agencies now share data, coordinate enforcement, and conduct joint examinations.

Atkins said in the press release that "regulatory turf wars, duplicative agency registrations, and different sets of regulations have stifled innovation and pushed market participants to other jurisdictions."

Selig posted on X: "The days of turf battles between the CFTC and SEC are over."

The MOU also came with a to-do list from the CFTC. Per CoinDesk, Selig announced four items in active rulemaking: whether DeFi software developers must register with the agency, formal rules for prediction markets, classification of perpetual derivatives, and a framework for AI-driven trading systems. On prediction markets specifically, the CFTC is asserting exclusive federal jurisdiction, putting it on a direct collision course with states like Nevada and Massachusetts that have been trying to classify event contracts as gambling.

Notably, the agencies aren't waiting for Congress. The CLARITY Act, which would codify all of this into law, remains stalled in the Senate. The MOU is the bridge. And the same week it was signed, Custodia Bank lost its years-long court battle to get a Fed master account, the same access Kraken just received. Two paths for crypto banking access diverged in the same 48 hours. The message was the same either way: the rules are being written now, and they favor firms willing to comply.

The Numbers:

• BTC: ~$71,490 | ETH: ~$2,097 | BTC dominance: 59.3%

• Fear & Greed Index: 15 - Extreme Fear

• Total market cap: $2.41T

• Crypto VC funding: $895M in February, down 40% from January and down 77% since October 2025

• Stablecoin supply: $310B

Deals this week:

• KAST — $80M

• Cryptio — $45M

• Zcash Open Development Lab — $25M

• Metaplanet Ventures — $25M new fund (Bitcoin-focused, Japan)

• Strategy — $1.28B BTC purchase (17,994 BTC)

• Bitmine — $120M ETH purchas5

That’s the week. Now go build something.

— Byblos

Subscribe now

Here are the four narratives that actually mattered this week.

1/ Kraken Gets A First for Crypto Fed Master Account

On March 4, Kraken Financial received approval for a Federal Reserve master account through its Wyoming-chartered banking arm, making it the first crypto-native institution to secure direct access to the Fed's core payment rails. That means Kraken can now move US dollars directly through systems like Fedwire, without routing through an intermediary bank.

The application was filed in October 2020. It took over five years. The account grants direct entry into Fedwire, a real-time gross settlement network that processes trillions of dollars in transfers each day. Until now, every dollar Kraken moved went through a partner bank first.

Kraken co-CEO Arjun Sethi said: “This milestone marks the convergence of crypto infrastructure and sovereign financial rails. With a Federal Reserve master account, we can operate not as a peripheral participant in the U.S. banking system, but as a directly connected financial institution.”

The account comes with constraints. Blockworks Research pointed out it's a "skinny" account - no interest on reserves, no access to the discount window - and it's approved on a one-year term. But as Blockworks Research analyst Luke Leasure noted, those limitations don't undercut the significance: it eliminates reliance on correspondent intermediary banks, providing a more durable and regulated foundation for custody, payments, and institutional cash management tied to its exchange.

The approval also marks something broader. The decision signals a notable shift in the Fed's posture toward regulated crypto institutions after years of resistance and legal battles. The crypto industry has spent the better part of a decade trying to get inside the same plumbing that powers traditional finance. Kraken just walked through the door.

Not everyone's treating it as routine infrastructure news. Jeff Park at Bitwise called the integration a moment that is "redrawing the architecture of modern finance," framing stablecoins as a direct structural threat to fractional reserve banking. That's a stronger read than most but it points at the real anxiety underneath the banking lobby's objections. Several banking groups have already raised concerns, arguing the decision poses systemic risk. The Lark Davis newsletter noted the reaction plainly: the banking industry doesn't welcome the competition.

2/ NYSE's Parent Bets on OKX at $25B, Tokenized Stocks Are Next

On March 5, Intercontinental Exchange, the company that owns the New York Stock Exchange, took a minority stake in OKX at a $25 billion valuation and secured a board seat. Bloomberg reported the investment was roughly $200 million. The terms weren't officially disclosed, but the product roadmap was: OKX will let its 120 million users trade tokenized stocks and derivatives listed on the New York Stock Exchange, in a feature expected to launch in the second half of 2026.

Star Xu (OKX founder/CEO) on X said: issuers may in future be able to bring securities directly to global investors through modern digital infrastructure, while still benefiting from the governance and regulatory frameworks that have long defined traditional exchanges and that OKX's focus is not simply on new technology, but on building durable infrastructure for the global financial system.

The deal started, according to Fortune, as a 30-minute meeting between OKX's global managing partner Haider Rafique and ICE chairman Jeffrey Sprecher that turned into a four-hour conversation. Rafique's framing afterward: "There was great chemistry in how we looked at the world and the future of tokenized securities, how derivatives should make it to the global stage, how TradFi and digital assets should merge together."

This isn't ICE's first move in this direction. BeInCrypto noted that back in January, NYSE had already announced a platform for on-chain settlement of tokenized securities with 24/7 operations, instant settlement, and stablecoin-based funding, pending regulatory approval. The OKX deal adds the distribution layer to a vision ICE was already building in-house. ICE is also working with BNY and Citi to support tokenized deposits across its clearinghouses, enabling collateral transfer outside traditional banking hours. OKX's 120 million users are the global reach that NYSE's own infrastructure can't replicate alone.

The Decrypt Morning Minute newsletter put the week in context well: "Two major TradFi moves for crypto over the past two days. First, Kraken gets a Federal Reserve master account. Now, NYSE's parent company takes a board seat inside OKX. The pieces of a crypto-native financial system are assembling faster than most people realize and it's being overshadowed by the bear market."

The nuance worth holding onto: everything here is contingent on U.S. regulatory approval, and OKX is still rebuilding credibility after its $500 million DOJ settlement in 2025. As Cryptopolitan noted, OKB's 24-hour trading volume hit $470 million after the announcement, about 1,657% above its usual daily volume, reflecting a market conviction that ICE's backing will make OKX a legitimate TradFi-to-crypto bridge. Whether that conviction holds through H2 2026 depends entirely on whether the product actually ships.

3/ Curve vs PancakeSwap - DeFi's Code Theft Problem Isn't New

On March 6, Curve Finance posted a public callout directed at PancakeSwap. Curve shared a screenshot of a file named CLStableSwapHook, attributed to PancakeSwap as author, containing the comment @license # TODO: Which license should we use? - an unresolved placeholder indicating no license had been assigned at the time of publication. Curve's message was direct: you copied our code without asking, it's a license violation, and historically that has not ended well for projects that went that route.

DeFi code theft has a straight timeline. Let’s quickly walk over it together.

2020 — SushiSwap vampires Uniswap

An anonymous developer called Chef Nomi forked Uniswap V1 wholesale, added a governance token and yield incentives, then ran a “vampire attack”, incentivizing Uniswap liquidity providers to migrate over. In 2020, SushiSwap reached a TVL of over $8 billion at its peak as it challenged the dominance of the protocol whose code it had copied. Uniswap had no legal recourse because it had published under a fully open license. The response was to build Uniswap V3 with a Business Source License, a time-restricted commercial protection preventing forks for two years.

2021 — Saddle Finance copies Curve line by line

Curve publicly accused Saddle Finance of porting its code line by line into a different programming language, a direct translation with no meaningful independent contribution. Saddle’s own security audit confirmed it: the implementation was ported from Curve’s codebase and was “also vulnerable to the same attack” vectors as the original. Saddle launched anyway. It got hacked in 2022.

2021 — HolaSwap forks Uniswap V3 days after launch

Uniswap had literally just introduced the BSL specifically to prevent another SushiSwap situation. Anonymous developers launched a fork of Uniswap V3 on Binance Smart Chain, defying the business source license that restricts any use of the source code for two years. No consequences.

2021 — Uranium Finance copies SushiSwap, changes one number wrong, loses $57M

The most operationally spectacular example. Uranium developers copied the code of SushiSwap, replacing the number 1,000 with 10,000 everywhere, except for one instance. As a result, the attacker could swap a tiny amount of input tokens for $57 million in output cryptocurrency. The developer who spotted it called it “one of the most expensive typos in blockchain history.” Users lost the money, not the devs who copied the code.

2021 — Andre Cronje, on the builder’s perspective

Cronje wrote: “I can build the superior product even, but a competitor can just fork my code, add a token that infinitely mints, and they’ll have twice the users in a week. Motivating to keep building right?” (now post is deleted). He wasn’t describing an edge case as he was describing standard DeFi competitive dynamics.

2024 — Andre Cronje again, Aerodrome Finance steals his GPL code

Cronje claimed someone used his work to create an unauthorized copycat that amassed $1 billion in TVL and FDV and they “slapped on a restrictive license” over his open-source GPL3 code (post isn’t available anymore). The project was reportedly Aerodrome Finance, one of the largest protocols on Base. Taking GPL3-licensed code and re-licensing it under BUSL violates the GPL3 explicitly derivative works must be licensed under the same terms. From a legal standpoint, this constitutes copyright infringement. Aerodrome is still running.

2025 — Balancer gets exploited for $128M via Curve-derived code

Not a copy dispute but a consequence. Curve cited the $128 million hack of DeFi protocol Balancer in 2025 as an example of swap-based code exploits, specifically when warning PancakeSwap that copying StableSwap without deep expertise creates attack surface, not just legal exposure.

2026 — PancakeSwap, license field reads “TODO”

DefiLlama's forks dashboard shows Uniswap V2 is the single most-forked protocol in all of DeFi, accounting for nearly 35% of all tracked forks. Aave V3 and Compound V2 round out the top three. The forks collectively hold more TVL than many of the originals.

The industry has essentially normalised code appropriation as a business strategy. Uniswap tried to fight it with BSL licensing; when the two-year license expired in April 2023, developers forked Uniswap V3 within days, attracting $123M in TVL almost immediately.

Andre Cronje’s framing from 2021 is still the most honest read on why this keeps happening: the incentive structure rewards shipping over building. Fork fast, add a token, capture liquidity. The original developer gets credit on crypto Twitter and nothing else. The pattern from 2020 → 2021 → 2024 → 2025 → 2026 isn’t a series of isolated incidents but it’s a cultural norm that the industry has never seriously confronted.

Curve’s message to PancakeSwap ended with “better to be friends and build together.” Contrary, some users expressed their negativity in regards to such a mild public resolution, calling out Curve Finance for not escalating this issue to legal matters.

The question is whether the industry ever develops enforcement mechanisms that make friendship optional rather than the only available path.

4/ AI Agents Need Passports and Ethereum Is Building Them

AI agents are no longer running in sandboxes. They’re signing transactions, managing treasury positions, submitting governance votes, and interacting with lending protocols without direct human initiation. The infrastructure question nobody fully answered until recently: how do you know which agent you’re dealing with, whether it’s the same one across sessions, and who’s accountable if it does something wrong?

On March 5, ENS published ENSIP-25, a new standard that addresses exactly this. If an AI agent is registered in a system like ERC-8004 and claims an ENS name, the name owner can explicitly confirm that association by setting a standardized text record. Wallets, explorers, and frontends can then independently verify the link using a single resolver lookup. The design is deliberately minimal as ENS doesn't become the agent registry, just the identity bridge. Same thing it's always done for humans and DAOs, now extended to autonomous software.

The underlying standard ENSIP-25 went live on Ethereum mainnet in January, built by a team from MetaMask, the Ethereum Foundation, Google, and Coinbase. It creates three on-chain registries: Identity (a global directory assigning each agent an on-chain ID via ERC-721), Reputation (a persistent, queryable history of ratings after interactions), and Validation (a standard for recording independent verification of agent actions). The Allium blog put it plainly: lending protocols can check an agent's reputation before extending credit; DAOs can hire agents for treasury proposals once their history of producing correct financial reports is validated; agents can offer different pricing tiers based on their track record.

The problem this solves is structural, not cosmetic. According to a Cloud Security Alliance survey cited by Strata, only 21% of organizations maintain a real-time inventory of active agents, and nearly 80% of those deploying autonomous AI cannot tell you in real time what those systems are doing or who's responsible for them. The current workaround: 44% authenticate agents with static API keys, and 43% use username and password combinations as methods designed before the concept of an autonomous on-chain actor existed.

The deeper issue is legal as one Medium analysis by Jung-Hua Liu noted, an on-chain identity for an agent does not confer legal identity. It is a technical handle. Current law doesn't recognize AI agents as persons, so any action an agent takes gets attributed back to a human or corporate principal. That works fine for simple automation. It gets complicated fast when an agent is autonomously rebalancing a $25M vault or casting a governance vote on a protocol treasury.

A Substack from PhD Research in DeFi framed the coordination gap well earlier this year: existing AI agents optimize narrow objectives, react locally, and operate without shared context which works in simple environments but breaks down completely once you introduce cross-protocol complexity. Identity continuity is the prerequisite for coordination. You can't build agent-to-agent trust, portable reputation, or accountability without a stable way to refer to the same entity across interactions.

ENSIP-25 is a small standard. ERC-8004 is still a draft. But the stack being assembled: identity, reputation, validation, payments via x402 - is the plumbing for an agent economy that actually functions at scale. The Ethereum Foundation's decentralized AI team has made it a core part of their 2026 roadmap. The question isn't whether AI agents will participate in on-chain finance. That's already happening. The question is whether the infrastructure catches up before something consequential goes wrong.

The Numbers:

• BTC: ~$67K | ETH: ~$1,932 | BTC dominance: 56.7%

• 38% of altcoins trading near all-time lows

• Total market cap: $2.37T

• Crypto VC funding: $895M in February, down 40% from January and down 77% since October 2025

• Stablecoin supply: $310B

Deals this week:

• a16z Crypto — $2B Fund V targeting H1 2026 close

• Paradigm — up to $1.5B new fund

• Crossover Markets — $31M Series B led by Tradeweb, with Ripple, Virtu, Wintermute, XTX

• Unknown stablecoin project — $70M, Sequoia + Founders Fund + Brevan Howard + YC

• Kazakhstan central bank — $350M crypto infrastructure allocation

  ---

That’s the week. Now go build something.

— Byblos

Subscribe now