Gavin Baker says Anthropic added $11bn of annual recurring revenue in March, which he argues equals the combined businesses of Palantir, Snowflake and Databricks in one month.
Baker calls Anthropic's $900bn valuation on $50bn of ARR reasonable, citing a much lower cost per token and roughly 80% less cash burn than OpenAI.
Baker predicts the watts shortage will ease in 2027 and 2028, while he flags TSMC's capacity decisions as the key wafer variable to watch for bubble risk.
Power and chips, not demand, will set AI's pace
AI demand is outrunning supply, so the pace of the next phase depends on power and chip capacity. Anthropic's jump in revenue shows the demand, while the power shortage and TSMC's capacity choices set the limits. Baker's worry is that supply eventually overshoots, as it has in past technology bubbles.
Takeaways
Anthropic's growth justifies its $900bn valuation. Anthropic added $11bn of ARR in March, matching Palantir and two other top SaaS companies combined in one month. Baker calls $900bn on $50bn of ARR reasonable, given lower cost per token and perhaps 80% less cash burned than OpenAI.
Watts shortage eases by 2028, orbital compute follows. Baker expects the power shortage to ease in 2027 and 2028, with orbital compute solving it later. He sees racks in space linked by lasers, not giant data centers, and notes SpaceX operates 98 or 99% of satellites in orbit.
TSMC capacity decisions are the wafer variable. Baker says TSMC's expansion choices matter most, and Intel and Samsung will eventually break discipline. Terafab, a planned SpaceX venture with an Intel partnership, gains 50 years of institutional knowledge, though Intel is 3 to 5 quarters behind the leading edge.
Frontier models capture the returns, leadership shifts. Google dominated the Pareto frontier nine months ago, while Anthropic and OpenAI lead it now, and Google lost cost leadership with TPU v8. Baker says a violation of Richard Sutton's bitter lesson is the biggest risk to the AI trade.
Buildout differs from 2000, but bubble risk lingers. Today's buildout is cash-funded with full GPU utilisation, unlike 2000. Baker still worries about a breakdown in market diversity, the point where everyone turns bullish, since bubbles let supply outrun demand and end in a crash.
Inference splits into prefill and decode chips. Prefill is memory-capacity-bound while decode is memory-bandwidth constrained, and the two are being disaggregated. Nvidia acquired Groq's LPUs for prefill, Cerebras needed three generations to get wafer-scale right, and Baker sees Trainium as the strongest custom alternative.
SaaS loans hurt private credit, GPUs may help. Private credit is in pain from SaaS loans that probably need larger markdowns. Baker argues GPU useful lives may stretch to 10 to 15 years, and financing GPUs at 5% or 6% would change the cost of the buildout.
Other calls Baker makes
Amazon will see real P&L efficiencies from robotics in its retail business over the next 18 months.
Astera Labs' biggest product will be a switch, so it cannot be a copper loser.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Anthropic could raise money at a large premium and would earn far more revenue with unconstrained compute.
The Strait of Hormuz closure favored America, since US natural gas fell about 20% while Asian and European gas doubled or tripled.