Investment & Market Intelligence
The Investor
Private marks on OpenAI still assume a $70B run-rate that is closer to $50B.
Investors and reporters took one month from second-hand investor letters and multiplied it by twelve. That implies about $5.8B a month against roughly $4.2B. Nvidia, CoreWeave and Nebius fell on the news. Listed shares reprice when the story changes. Private marks and secondary bids built on the same arithmetic reprice when someone has to trade, and the company has not revised the figure, so if you hold that paper it is still marked to a multiplication.
In Play
OpenAI's Run-Rate Correction
The Financial Times reported Thursday that OpenAI's widely cited $70B annualized revenue is closer to $50B. The Information concedes it also ran the wrong figure. Nvidia, CoreWeave and Nebius fell on the news. Any private AI mark, secondary bid or infrastructure deal you priced off month×12 run-rates carries the same error.
Ask ClaritySmall-Model Pricing: Cost Per Task, Not Per Token
Anthropic priced Claude Haiku 5.5 at $0.10/$0.50 per million tokens, the same as OpenAI's GPT-6 Luna. That is a 90% list cut, which Anthropic itself puts nearer 75% on average. Artificial Analysis token counts compiled by AINews show Haiku using about 3.2x Luna's output tokens per task at max effort. Your app-layer margin models need cost per task, not cost per token.
Ask ClarityAI Component Exit Band Gets Tested
The Information reports that Eliyan, a chiplet-interconnect startup valued at $1B in July, has a takeover bid and will likely ask about $3B. Arm, an existing investor, held talks and did not bid. For your component holdings, $3B is an asking price, not a completed sale. The comparable deals that actually closed are Empower at $1.5B and Celestial AI at up to $5.5B.
Ask ClarityAgent Harness Becomes Table Stakes
Cloudflare data cited by Daniel Miessler shows agents made up 48% of Wrangler CLI usage last week. In the same cycle, Anthropic's SDK took over the computer-use action loop (AINews), and OpenAI shipped sessions, compaction and MCP Events natively (Artificial Ignorance). Many orchestration startups in your pipeline are now selling features the labs ship by default. Agent identity, granular permissions and verification remain unsolved.
Ask Clarity
Deep Dives
- ●
OpenAI's Run-Rate Miss Exposes How AI Revenue Gets Measured
The error was arithmetic, not demand. Every mark built the same way carries it, and the two benchmark labs don't even define revenue alike.
Where the $20B went A run-rate is one month of revenue multiplied by twelve. On the corrected figure, OpenAI's implied monthly revenue is about $4.2B , not the ~$5.8B the $70B headline implied. That is a gap of roughly $1.7B…
3 action items
- ●
Haiku 5.5's 90% Cut Shrinks Fast Once You Price the Task
Tokenizer inflation, a 5x long-context price step and heavy reasoning output mean the agent workloads investors prize get the smallest share of the savings.
What the 90% leaves out Three pricing mechanics sit below the sticker price. Tokenizer. Haiku 5.5's new tokenizer counts the same text as roughly 1.25x more tokens. That is why Anthropic's own blended estimate is ~75% cheaper, not 90%. Long-context…
3 action items
- ●
Eliyan's $3B Ask Tests an Exit Band Insider Arm Won't Pay Up For
Strategic buyers now pay billions for pre-revenue interconnect and power IP, but each deal removes a buyer along with a target, so entry price decides returns.
Two tiers of AI hardware exit Eliyan is the third AI component company in play in eight months. Together they define a strategic exit band, priced on how scarce the IP is rather than on revenue. Company Layer Buyer /…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn