⚡️Anthropic: Know Before You IPO — Tanay Jaipuria, Wing VC
Anthropic's S-1 shows an unusually capital-efficient, profitable lab, but with risks from model-release dependence, API-driven revenue, a slowing growth rate and huge compute commitments.
Watch on YouTube ↗Charlie's summary
- ▶ 11:47
Anthropic burned under $10 billion to reach this scale, which Tanay thinks is the S-1's biggest surprise
Anthropic raised about $140 billion and still holds roughly $132 billion in cash. Tanay says almost no company of this scale that went public in the last five to seven years comes close on cash burn per dollar of run rate. Airbnb is the only one near it. He adds that Anthropic has burned fewer dollars than Uber had at its IPO. The host notes OpenAI is far less efficient, closer to Snowflake, Cerebras and CoreWeave.
- ▶ 17:25
Skipping a free consumer tier keeps Anthropic's inference margins high
Tanay says Anthropic's reported 85% gross margin covers paid inference only, which is about 15% of revenue. Free-tier inference is usually booked as sales and marketing. Because Anthropic focused on coding and enterprise, only about 5-6% of its compute goes to free users. Tanay expects OpenAI's figure to be much higher given its user base. He ties this to OpenAI dropping Sora, which used a lot of compute without monetizing.
- ▶ 21:02
The next token-hungry workload after coding is an open question, and security may be a candidate
Tanay points to Q2's huge run-rate gain followed by a slowdown from over 100% quarter-over-quarter growth to about 40-45%. Skeptics read Q2 as peak token-maxing before rivals caught up. Coding is the clear product-market fit, with customer support and legal next. Tanay thinks security could be the next big source of token spend, since attackers now have powerful models too. Anthropic itself talks most about science and life sciences.
- ▶ 24:41
About 70% of Anthropic's compute goes to training and research, and data spend is ramping fast
Tanay was surprised that only about 30% of compute serves inference, down from 50% at the start of last year. He estimates inference costs at only $10-12 billion a year. Data spend was around $900 million to $1 billion run rate until a jump to roughly $600 million in the June-September quarter, now perhaps $2-2.5 billion a year. He reads this as a sign that RL environments and post-training are working.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.