Tech • AI • Robotics • Game
Gavin Baker on Anthropic's S-1, GPU Prices, and SpaceX
9.3/10TL;DR
Atrades Management’s Gavin Baker said recent disclosures from Anthropic point to stronger-than-expected AI economics, while demand for compute remains so tight that GPU capacity into 2027 is still being priced at a premium.
KEY POINTS
Anthropic’s figures surprised to the upside
Baker said Anthropic’s latest reported metrics showed healthier cash burn, improving profitability and stronger unit economics than many investors expected. He highlighted a striking adoption gap: roughly half of large US companies have already adopted the company’s tools, yet AI still accounts for only about 0.1% of those customers’ total expenses, suggesting substantial room for expansion.
Enterprise AI adoption is still early
Despite rapid revenue growth, Baker argued corporate deployment remains in the early stages. He pointed to annualized revenue running near $6.7 billion while penetration inside the Fortune 500 is still limited, a sign that many large companies are only beginning to work out how to deploy AI broadly across their organizations.
Revenue concentration may be less severe than feared
Baker pushed back on the view that AI revenues are driven almost entirely by a handful of giant accounts. He said Anthropic had only about 250 customers spending more than $10 million annually, while roughly 11,000 to 12,000 customers were spending more than $100,000, indicating a more diversified customer base than some recent reports suggested.
Open-weight and infrastructure startups are scaling quickly
Baker said a growing group of AI infrastructure companies built around open-weight models is expanding rapidly, naming Fireworks, Base10, Together, Modal and Cognition. He said many of these businesses were already at scale and in some cases growing faster than the best-known foundation model providers, making valuation difficult even during fundraising because revenue was rising so quickly.
GPU markets are tightening, not easing
Contrary to hopes that new supply would loosen the market, Baker said the price to rent GPUs has been rising steadily and that both spot and forward compute markets remain tight. He cited data indicating the market has moved into contango, meaning future compute is priced higher than current supply, and said companies seeking major clusters in 2027 are often shocked by how much they must pay.
Long-term compute contracts have become highly valuable
Baker said companies that locked in compute before the market tightened are now sitting on valuable contracts. He cited comments from OpenAI CFO Sarah Friar that some commitments could effectively be resold at 3x to 5x current value, underscoring how sharply market pricing has moved.
Execution is becoming a differentiator for cloud operators
Baker compared operating large GPU fleets to running a Formula 1 team, arguing that success depends on power, deployment speed and operational discipline, not just access to land and financing. He named CoreWeave, Crusoe, Nebius, SpaceX and hyperscalers as examples of operators with proven ability to energize and run clusters at scale.
New hardware packaging could reduce costly delays
He also pointed to Nvidia’s DGX Pod approach as a potentially important improvement because pre-assembled systems could shorten the time between delivery and full operation. That matters because some operators currently take three to six months to bring GPUs online, tying up large amounts of capital before assets begin generating revenue.
AI is already producing real-world benefits in biology
On healthcare, Baker argued AI’s impact is not purely theoretical or long-dated. He said researchers are already using AI to identify older drugs that can be repurposed to treat rare and devastating diseases, in some cases producing life-saving outcomes that have not been widely communicated to the public.
Future financing may rely more on cash flow and public markets
Baker said the next stage of AI financing will involve mutual funds, sovereign wealth funds and large institutional investors, especially as major private companies eventually list publicly. He added that repricing underpriced cloud contracts and rising cash flow from AI infrastructure could allow a larger share of the 2027 buildout to be funded internally rather than through constant external capital raises.
CONCLUSION
Baker’s outlook remains strongly bullish: AI demand is broadening, compute remains scarce and expensive, and the industry’s next phase may be shaped as much by operational execution and financing structure as by model performance alone.
Ask a question