We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Frontier labs: record revenue, trillion-dollar prices
OpenAI's annualized revenue is reportedly approaching $70B. That would be up more than 70% since the start of Q3, with enterprise sales more than doubling since July . OpenAI is also reportedly in talks to raise at least $30B at about $1.4T pre-money, as a bridge to an eventual IPO. Anthropic is reportedly targeting a listing as early as mid-November, with a roadshow the week of November 9 and a valuation discussed at $1.8T–$2T .
When asked about the $70B figure, Sam Altman would not confirm or deny it and said he was "very pleased with the growth" . He said OpenAI wants to be public eventually, but not while it adjusts to new capability levels and safety requirements. He described investors as "very happy with us and very patient" . In the same interview he referenced holding back Astra 6.1, and said OpenAI will slow training or releases to focus on safety when needed .
On the 20VC/SaaStr panel, Rory O'Driscoll framed the competitive picture this way:
- OpenAI grew only 18% quarter over quarter from Q1 to Q2.
- By his numbers, Anthropic's GAAP revenue overtook OpenAI's in Q2, at roughly $11B against $6.8B.
- Anthropic's Q3 number will be in its prospectus before the stock trades. That matters because teams are reportedly moving work to Codex .
His advice to a hypothetical buyer choosing between the two was to lean toward liquidity .
The State of AI Report 2026 came out the same day. Its headline figures:
- Lab revenue: OpenAI's and Anthropic's combined run rates reached about $105B by late summer, up from about $30B at the start of the year.
- Application companies: Legora and Sierra each doubled to $200M in about six months; Harvey reached $400M, Lovable reports $600M, and Cursor has been reported above $4B. The report notes that metrics differ across companies .
- AI-native vs. AI-enabled: at the 75th percentile in the $1–20M revenue band, AI-native companies grew revenue 256% year over year, versus 90% for AI-enabled firms .
Nathan Benaich's other takeaways from the report:
- Anthropic's internal index says Claude led 26% of measured model R&D in August, with humans supervising .
- He thinks robotics is approaching its "GPT-2 moment" .
- AI-discovered drugs are reaching late-stage trials: Insilico's rentosertib is recruiting for Phase 3. Human trials still have to show these drugs are safe and effective .
Agents are becoming a buying channel
Vercel says agents went from under 3% of its deployments to more than 50%. On the podcast, Harry Stebbings cited $600M ARR with agents driving half of new business . Jason Lemkin asked an agent live on the show what to use for hosting. It ranked Render first and Vercel fifth . His advice to founders is to find 10 people with agents in production and ask them every two weeks what their agents recommend in your category . Dev Ittycheria warned that the data agents learn from favors incumbents, so startups need documentation and APIs that agents can read easily .
Infrastructure providers are building for this traffic:
- AWS: plans $220B in 2026 capex, has two million Nvidia GPUs on order, and says its AI chips are sold out through next year. It is adding a 30-second account signup with no credit card so agents can start using it . AWS also says it deliberately holds back accelerator capacity for startups rather than selling it all to frontier labs, and says yes in some form to about 60% of startup requests .
- Token usage: a16z says agents use nearly 5x the tokens people do, and that this is up 14x in six months .
- Cloudflare: non-human traffic passed human traffic on its network in May 2026, more than a year earlier than its own forecast . It is building agent payments (Monetization Gateway and Wallets), but this layer has no revenue yet .
Agent reliability is still a weak point. Microsoft authors introduced ThinkingBox, a benchmark of 507 business workflows that runs each task 20 times:
- Kimi-K3 solved 93.89% of tasks at least once but only 13.41% on all 20 attempts. Claude Opus 5 solved fewer (79.09%) but repeated far more reliably (47.53%). Ranking models by either measure gives nearly opposite leaderboards .
- 67.24% of failures ended cleanly, so a simple "did it finish" check would have scored them as successes .
- The tasks are synthetic reconstructions of enterprise workflows, not production traffic .
For diligence, this means one-off agent demos say little about whether an agent works reliably.
Venture: bigger funds, big seeds, mixed exits
USV raised $900M in new funds. Its last core fund was $275M . USV's explanation is that AI has made building easier, which has raised demand for the best companies and led to "bigger rounds and higher prices" . It will make about the same number of investments, but lead more rounds and back capital-intensive robotics, manufacturing and energy companies . Its energy thesis, previously run through a dedicated fund, moves into the core fund .
Hone launched Engines: the customer points the product at a business metric and it builds its own agents to move it. Hone announced a $60M seed led by Benchmark and Index .
Exits and secondaries:
- Listen Labs: Salesforce agreed to buy the three-year-old AI market-research company for a reported ~$2B. Listen Labs had walked away from a signed Series C term sheet at $1.5B, and its revenue was estimated at about $30M annualized . Dev Ittycheria expects more deals like this, because many AI apps are features built on someone else's platform .
- ElevenLabs: closed a $300M employee tender at $22B, double its February Series D valuation .
- Oura: postponed a $2.2B IPO that was about four times oversubscribed, despite $1.21B in revenue for nine months, up 74% . Jason Lemkin called it a hole in the "liquidity is back" story . Ittycheria said price was the likely reason, but stressed that he was speculating .
- Groq/Nvidia lawsuit: two former Groq engineers sued over Nvidia's $20B license-and-hire deal ($17B license, $3B in RSUs for 150–200 engineers) . Panelists said founders should expect employees to ask what happens to their equity in deals like this .
- Secondaries demand: in SetterVC's Q3 ranking, Anthropic was #1 for a third straight quarter. Etched debuted at #13 and Hadrian rose nine spots to #11. Kalshi fell 16 spots and Polymarket 24 .
An unverified anecdote: Kenan Saleh says an unnamed company selling "frontier data" to AI labs went from $0 to $60M in three months and will reach a $100M run rate this month .
Model pricing: a caveat on Haiku 5.5
The prior brief covered Haiku 5.5's launch. The new detail is Artificial Analysis's evaluation. It scored the model 43 at max effort, ahead of GPT-6 Luna at 38 . But at max effort Haiku used about 162k output tokens per task, roughly 3x Luna . Cost-per-task numbers are still pending because the evaluation does not yet account for the 5x price step above 100K tokens . When comparing per-token prices, check cost per completed task.