Source Foundry's $400M — your token price starts at lithography

The most interesting AI story this week is not a model. It's a hedge fund that just lost most of its assets writing a $400 million check into a stealth chip startup — because the founder concluded the constraint on AI is not algorithms. It's the machines that print the chips. Source Foundry is a bet on lithography, and lithography is the floor under every token price you will ever pay.

What actually happened

The Wall Street Journal reported on August 8 that Leopold Aschenbrenner's fund Situational Awareness invested $400 million into Source Foundry this week, taking its total position to $500 million — confirmed by TechCrunch and Bloomberg. Source Foundry was founded roughly a year ago by two Stanford materials-science PhDs, Abdulmalik Obaid and Joe Berg, and says almost nothing in public about how its technology works. The goal: semiconductor manufacturing equipment that is simpler, cheaper, and faster than what exists — starting at the lithography step. Reporting puts the round at roughly a $5 billion valuation, with Sequoia an early backer.

The timing is the story. Situational Awareness sold the majority of its public portfolio to Citadel in late July after steep losses on AI infrastructure stocks. Assets under management went from about $20 billion to $10 billion. Weeks later it concentrated half a billion dollars into one pre-revenue hardware company.

Why lithography? Because ASML has an effective monopoly on the EUV machines required for leading-edge chips, and those machines run north of $400 million each and ship at a rate the industry does not control. Every GPU, every inference dollar, every price cut a model vendor announces sits downstream of that queue.

Why it matters for your business

You will never buy a lithography machine. You are still exposed to it, through four layers of markup: equipment maker → fab → chip designer → cloud → the API you actually pay. When the bottom layer is a single-supplier chokepoint, price relief at your layer is a marketing decision, not a physics one.

That reframes what the current token price war actually is. GPT-5.6 Luna's 80% cut, DeepSeek's fourteen-cent coding tier, Meta pricing Muse Code below cost for training data — none of those reflect cheaper silicon. They reflect vendors buying market share with capital that has to be repaid. Capital that, as this week demonstrates, can evaporate 67% in a month.

So don't architect around today's price. Architect around swap cost. Keep the model call behind an interface. Log prompts and outputs in your own store, not the vendor's. Know what a migration to a different provider actually costs you in hours, and keep that number small. If it's under a day, a price hike is an inconvenience. If it's a quarter-long project, your vendor sets your margin and you'll find out when they need to.

Sources: TechCrunch, Bloomberg.