Lex Sokolin | Generative Ventures

Here are all my secrets We cover crypto, AI, and the machine economy. The weather changes but the stars point in the same direction. Really enjoyed this conversation with @francescoswiss

@MarkYusko You’re free!

So funny that day I pick you to test how deep I was in Twitter jail is the ay they change the algorithm and we can all see each other again… 😂🧡🙏🏼🚀

@sreeramkannan What a great bet, would love to see it

Vorflux. Autopilot for Software. Prasanna comes with an exceptional background in coding and entrepreneurship. Now he has built a product that helps small teams ship big products. Coding agents on the cloud that scale infinitely. Coding agents that are neutral and track SOTA across model providers. Coding agents that find every single human bottleneck and apply intelligence till it’s gone. Software autopilot that handles everything from code to deployment to iterative improvement. Excited to back Prasanna on his new adventure, Vorflux!

@cyntro_py Yes

Let's build cybernetic economy together

@LorenzoARK @santiagoroel When you use a tradfi broker the price a user base for a single transaction vs a large institution pays for a prime relationship are very different Anchoring L2 data to the L1 should be a more expensive endeavor, ie L1 charging the L2, than somebody doing a regular trade

@LexSokolin @santiagoroel what do you mean by institutioanl pricing from retail?

A few thoughts on @santiagoroel country and taxes analogy. People love to say blockspace is commoditized and switching costs are basically zero. The numbers say otherwise. We're 10 years into this experiment and only 3-4 L1s actually matter. Same story as AI models: everyone calls them commoditized, yet everybody uses either Chatgpt or Claude. Coming back to countries and taxes, California is the best counterexample here to Santi's argument, and probably the model ETH is going for. A 30% tax is too much, but the current take rate is almost certainly too low. Why don't people leave California despite outrageous taxes? Weather, quality of life, the AI job market. No single factor, the bundle. ETH's real moat is the assets on platform plus ETH the asset. If Ethereum had $3-5T of AOP instead of $250B, this would be a very different conversation, and Hood probably never leaves. You need to coordinate 2-3 outstanding qualities at once. Good weather alone (Portugal), not enough. Low taxes alone (Dubai), not enough. Great quality of life alone (Japan), still not enough. @HyperliquidX aggregated so much demand precisely because it coordinated three things: great UX, deep liquidity, and strong execution tech. Any one of the three alone doesn't cut it. One more thing. There's the @ethereumJoseph theory: subsidize blockspace to attract applications, then raise prices once network effects are real. The problem is that Ethereum the blockchain needs a strong ETH asset in the meantime. Hard to do that with ETH at $1,500. And for everyone saying we need thousands of Robinhood L2s: there just aren't that many Hood-like companies to go around. Robinhood has 30M accounts and ~$300B in deposits. At that scale you're not closing a Robinhood L2 every week. The math matters here.

@sgoldfed You are right on this Institutional pricing should be 1000x higher than retail gas fees but also be that much more valuable Otherwise it’s a parking lot with tragedy of the commons (ie memes, inscriptions)

I’ve proposed what I believe is a win-win fix for Ethereum’s rollup revenue a few times: Ethereum needs to provide more for its most successful L2s, and charge them a premium for it. Ethereum should adopt its largest rollups in the sense that a critical bug in Arbitrum, Base, or Robinhood Chain should be treated as an Ethereum vulnerability and trigger an L1 fork just like an L1 bug would. This fixes the ecosystem's biggest pain points in three ways: 1. It's an incredibly valuable feature, and Ethereum could name its price. 2. The main reason L2s keep security councils around today is to handle emergency vulnerabilities. If L1 acts as the security backstop, L2s can safely get rid of them. 3. Several L1 folks have been pushing for based/native rollups, motivated by the belief that L1 should have more control of and value capture from its L2s. But the technical designs are problematic and none of them have any commercial traction. Rather than trying to build momentum from scratch, Ethereum should just monetize the thriving L2s that already exist.

@JasonRaznick What are you up to?

And, one day, I will share a lot more stories about the behind-the-scenes of the financial media companies for the last 20 years. I even remember when there was a company Wall Street cheat sheet that was killing it back in the day.

I have to say one of the more gratifying things is the team I hired & trained along w others @Benzinga now lead the ad/biz dev teams at most of the successful financial media companies these days. Hmm, maybe I need to be trainer?? :-)

@ccatalini Preach

There is no “mandate of heaven.” A frontier market position is not legitimacy. Status is just deference, aggregated from the bottom up. Right now, that deference comes from our collective awe at the magic. But awe is not a mandate over everyone’s future.

The next neobank success story may look like a pharmacy with a mascot. Farmacias Similares and Stori just launched a Dr. Simi credit card in Mexico. The partnership combines more than 11,000 pharmacy locations with a fintech serving 5 million customers, many with limited or no formal credit history. Fintech has spent a decade moving financial distribution onto smartphones. And this move puts it inside an existing retail habit: buying affordable medicine from one of Mexico’s most recognizable brands. The credit product itself is easy to copy. The difficult assets are trust, foot traffic, and a customer relationship that already exists. There is a delicate edge here. Better access to credit can help people manage healthcare expenses. Poor underwriting can turn medical necessity into expensive debt. Approval rates and repayment behavior will tell us which version this becomes.

Robinhood chain launch has been incredibly successful. This is very bullish for Ethereum.

@mikebelshe They money is in the money

When I was working on HTTP, I used to say, "there's no money in protocols". Then crypto seemed to prove that wasn't entirely true. But as the real applications emerge on chain, we realize it is in fact true. Similarly, tether & usdc accrue the value as applications. The money has always been in the application, not the protocol.

@mdudas @bchesky The AIs have discovered crypto

@bchesky Great thread Claude Chesky

Economics of block space trending to commodity because retail and institutional are treated the same

Even though I'm more bullish than David, his estimate $1.5T estimate may actually be low. His concern about the scale of AI capex holds, but the comparison still does not show whether that capital earns its cost which I think is the more correct investor focus. Before we jump into the nuances, the NVIDIA route starts with data-center revenue that already includes networking, then applies a TCO multiplier that includes energy and other operating costs. That risks counting some interconnect twice and mixes some capex with opex. The alternative starts with ~$750B of total hyperscaler capex, including replacement and non-AI investment, so treating all $750B as one AI growth cohort is a demanding upper case. The 50% margin has also changed meaning across the series. The 2023 post excluded cloud-vendor margin. The 2026 post uses the same 2x to cover the hyperscaler and AI product company together. My own bottom-up work puts a frontier facility GW at ~$50.5B, with a $42.5-62B range and roughly 65% in shorter-lived IT versus 35% in physical infrastructure. Custom-silicon campuses can land around $30-35B/GW. $80-100B/GW is a forward stress case (will link my thread again). Assume a 5.5-year life for IT, a 20-year life for physical infrastructure, a 10% after-tax project return, a 50% full-stack pre-D&A operating margin after recurring opex, and a 21% tax rate. The 4.08 (5.5 years) and 8.51 (20 years) are capital-recovery factors @ 10% WACC: Post-tax return(n) = [1 − (1.10)^-n] / 0.10 For one $50.5B facility GW: >Required after-tax cash flow: $32.8B / 4.08 + $17.7B / 8.51 = ~$10.1B/year >Annual depreciation: $32.8B / 5.5 + $17.7B / 20 = ~$6.85B >Required revenue (21% tax rate): (0.50R − $6.85B) × 0.79 + $6.85B = $10.1B R = ~$22B/GW/year TLDR: The $42.5-62B frontier range needs roughly $19-27B of annual revenue per facility GW. Custom silicon at $30-35B needs roughly $13-15B using the same asset mix. The $80-100B stress case needs roughly $35-44B. On the full $750B denominator, the model requires approximately ~$330B (~15GW × ~$22B/GW) of annual revenue during the first 5.5 years and $2.8T over the asset lives ([$326B × 5.5 years] +[$71B × 14.5 years]). His $1.5T is simple payback at a 50% margin, with no required return or time value. His framework therefore understates the revenue required to earn an acceptable return. All that said, the operating evidence is improving. @ExponentialView estimates $110B of trailing GenAI revenue and a $175B annualized run rate, growing roughly 35% quarter over quarter. Quarterly GenAI revenue first exceeded quarterly depreciation in Q4 2025 (will link their report). In Q1 2026, depreciation still absorbed 68% of total GenAI revenue and 81% of hosting revenue before power, labor, training or any return on capital. Cumulative revenue remains around half of cumulative depreciation. Current contracts provide a direct per-GW test. Anthropic is paying SpaceX $15B annually for more than 300MW of compute capacity. That is below ~$50B per IT GW, or less than ~$44B per facility GW at the 1.15 PUE used in my work, approaching twice the ~$22B facility-GW hurdle. The contract runs through May 2029 with a 90-day termination right that Musk says SpaceX requested in case compute stays tight. It demonstrates current scarcity pricing; normalized full-cycle pricing remains unknown. Another example is TeraWulf’s $19B Anthropic headline uses a different layer. The 20-year lease covers 401MW of critical IT load, equivalent to approximately $2.1B of annual landlord revenue per facility GW. TeraWulf supplies shell and power while Anthropic supplies the compute. GPU rental pricing is firmer than the simpler collapse narrative argues. The July 9 Silicon Data screen put non-hyperscaler on-demand B200 pricing at $5.81/GPU-hour, the broader Blackwell index at $5.08 and H100 at $2.67. B200 pricing rose from roughly $4.30 in February while fast availability fell to zero. H100 one-year term pricing moved from $1.80 in November 2025 to $2.40 on July 8, with the curve holding around $2.34-2.40 through 36 months. Composite availability has risen because older A100, H100 and H200 supply loosened, while frontier B200 capacity remains tight. Older GPUs also continue earning rental yields beyond six years. The current data do not support immediate obsolescence or collapsing rental values. Token pricing has split by tier. @3F_Research's OpenRouter estimate shows the weighted price for OpenAI, Anthropic and Gemini rising from roughly $1.07 per million tokens in January to $1.62 in late June. All other models moved from roughly $0.22 to $0.18, while the overall weighted price held at $0.71. Frontier models carried a roughly 9x premium to the rest of the market. @WarrenPies is all over this (will link) The series is mix-weighted and moves with model selection, reasoning usage and the input/output mix, so it does not isolate like-for-like list-price changes. It does show frontier capability retaining pricing power while open models commoditize. Anthropic officially crossed a $47B annualized revenue run rate in early May. A July 8 SemiAnalysis estimate cited by Cahn puts it ~$60B (maybe $70BN now?). Pre Reuters reporting, Anthropic told investors it expected $10.9B of Q2 revenue and $559M of adjusted operating profit, a roughly 5% margin that includes the full cost of training new models but excludes stock compensation. Anthropic does not disclose its training/inference split. Paid inference growing faster than lumpy training spend provides a path to margin expansion even if training dollars keep rising. These revenue figures cannot be added together because one company’s revenue is another company’s compute bill. They are evidence about demand, pricing and margin at different layers of the stack. To recap, five full $750B incremental cohorts would require roughly $1.6T of annual revenue once all five are operating, about 74% annual growth from today’s $175B run rate through 2030. That is an upper case because some later capex replaces compute inside existing facilities rather than building a new shell and power path each year. Of note, Cahn’s tone has softened with the evidence: GPU capacity was 'overbuilt' in 2023, the bubble was at a 'tipping point' in 2024 alongside a B100 demand surge he called final that wasn’t, 'nothing short of AGI' could justify the spend in October 2025, and the latest post credits coding and Anthropic with creating a clearer path to monetization. The bull case requires revenue per GW, utilization and margins to persist as supply grows. Better models and agentic workloads need to expand demand faster than unit prices fall, inference contribution needs to outgrow training spend, and physical infrastructure needs to support several compute generations. The bear case requires scarcity pricing to normalize, GPU availability to rise, frontier-token premiums to narrow, chip lives to shorten or new capex cohorts to stack faster than end revenue. At $13-15B of annual revenue per GW, lower-cost custom silicon can work while a $50B frontier build does not. At $30-44B, the frontier build earns an acceptable return if margins hold. Cahn likely understates the hurdle, while current contracts, rentals, token pricing and lab margins increasingly show it can be met at the unit level. The next two to three years will show whether those economics survive normalized supply and stacked cohorts.

You gotta hand it to them

One of my proudest tweeting moments by far

"Open USD" is less interesting when viewed as “another stablecoin” and more interesting as a fight over the float. Stripe, Visa, Mastercard, BlackRock, Coinbase and a long list of partners want a dollar that partners can mint and redeem without paying the old issuer tax, with reserve yield pushed back into the network. That's the real product. Not the logo on the token. The economics of money at rest. Circle felt it immediately. Good. The market is finally pricing distribution power into the stablecoin stack. My test is simple: who actually mints at scale, who holds the reserves, who settles the boring parts, and who keeps the basis points when this stops being a press release. Stablecoins win by becoming plumbing. Plumbing gets fought over by the people who already move money.

I was blown away by @zapper_fi when it launched Programmable financial flows completely permissionless It became synonymous with DeFi and open source investment selection That it is shutting down is a strong signal about the viability of what we once all believed

The free office agent may be a short-lived thing. This week Anthropic put Cowork on your phone and Microsoft teased an always-on Copilot tier called AutoPilot. Both companies looked at the busywork nobody wants to own (status updates, meeting recaps, the orphan spreadsheet) and decided it's the most valuable thing to automate. So they're charging for it. Here's what that means for your fall budget 👇