Cognition raised $2 billion this week at a $48 billion valuation — about 53 times its revenue. Six days later Bending Spoons agreed to buy Miro outright for $1.355 billion, about 2.3 times its ARR. Every account of the week, including the first draft of this one, treated that as two prices for two kinds of company. It is not. Cognition sold a minority slice of preferred stock carrying a liquidation preference, at a valuation nobody paid for the whole company. Miro sold the whole company, for cash, with control. Those are not two prices. They are two different kinds of number, and dividing either by revenue produces a figure that does not mean what the other one means.
Harvey settles it. Its revenue is the most legible in the week — $400 million of ARR, 80% of the Am Law 100, named institutional buyers — and it priced at 38.8 times, up with Cognition rather than down with Miro. If verifiable revenue were what set the level, Harvey would sit near Miro. It does not, because Harvey also sold a minority preferred round. Sort the week by how legible the revenue is and the four transactions scatter. Sort it by what was actually sold and they separate cleanly: preferred slices at 38.8 and 53.3, whole companies at 2.3 and 2.7. The instrument decides the number. The revenue decides whether anyone signs.
Cognition closed a $2 billion Series E on September 8 at a $48 billion valuation, led by Andreessen Horowitz and Accel with Founders Fund, General Catalyst and Avenir returning. Its annualized run-rate revenue grew from $492 million at May's Series D to roughly $900 million — figures the company reports itself, which are not audited. Both rounds land in the low 50s as a multiple of that number, though the difference is smaller than the precision of the input: on a figure described as “roughly $900 million,” a few percent either way moves the multiple by more than the gap between the two rounds. The $48 billion is a post-money price for preferred stock carrying a liquidation preference, not a price anyone paid for the whole company. Devin runs inside Nvidia, GE Aerospace, Citi, Mercedes-Benz and Modal — and Nvidia is also an investor in the round.
Harvey raised $550 million on September 9 at a $15.5 billion valuation, co-led by Lightspeed and Diffusion, with Sequoia, Kleiner Perkins, Andreessen Horowitz, Coatue and Goldman Sachs Alternatives joining. ARR crossed $400 million, having run $100 million in August 2025 and $190 million in January 2026. Eighty percent of the Am Law 100 use it, alongside 20% of the Fortune 500. At 38.8 times ARR it sits with Cognition, an order of magnitude above Miro — despite having far more verifiable revenue than Cognition does. What it shares with Cognition is not the revenue. It is that both sold a minority preferred round rather than the company.
Bending Spoons entered a definitive agreement on September 10 to acquire Miro at a $1.355 billion enterprise value, roughly $1.79 billion in equity value including net cash. Miro runs about $600 million in ARR, nearly 90% of it from business and enterprise customers, across 250,000 organizations and almost 4 million paying users. That is about 2.3 times ARR. Five weeks earlier the same buyer took Airtable at 2.7 times, and Airtable's own sequence is the clearest illustration of the week's point: marked at $11.7 billion in a 2021 preferred round, trading near $4 billion on secondary markets earlier this year, sold for $2.25 billion of equity in cash. Five years of business reality sit inside that fall as well as the change of instrument — but only the last number was ever paid. Certain Miro shareholders are rolling $295 million of proceeds into new Bending Spoons equity; the deal is expected to close in Q4.
A cash buyer paying a low multiple has to earn its return from costs, and Bending Spoons has been explicit about the model: acquire, transform, reinvest, repeat. After buying WeTransfer it cut roughly 75% of staff. After acquiring Vimeo for $1.38 billion in late 2025 it laid off almost everyone in January 2026, including the entire video engineering team, keeping a skeleton crew into April and closing the Israeli office. It has deployed about $4.5 billion across Evernote, Vimeo, AOL, Eventbrite and now Airtable and Miro, and states internal return hurdles of 65% levered. That is what the low band costs, and who pays it.
Mistral announced a €3 billion Series D on September 8 at a post-money valuation above €21 billion, led by Samsung with EQT's Scaleup Europe Fund and PSG Equity as co-leads. Advent, BlackRock-managed funds and the Grand Duchy of Luxembourg joined existing backers a16z, ASML, Nvidia and Bpifrance. It is the largest equity round ever raised by a European technology company; Mistral was valued at €11.7 billion a year ago. CEO Arthur Mensch said the capital goes toward owned data centres and rented compute capacity.
Ramp data covering roughly 70,000 companies shows AI spend per employee at the top 1% of AI-using firms fell nearly 10% in August, to $7,205. The share of customers paying for any AI product reached 56%, up just 0.4% on the month. Average token costs have fallen to $0.68 per million, against a 2026 peak of $1.15 in March. Note the two windows differ — the spend figure is month-on-month, the price figure runs from March — and that spend fell far less than price, which means volume rose even as the bill shrank.
Nvidia is in talks to invest as much as $10 billion in Anthropic's initial public offering, Reuters reported on September 11. Anthropic is weighing a raise of up to $100 billion at a valuation near $2 trillion, which would make it the largest IPO on record. Talks are preliminary. Anthropic's annualized run rate passed $65 billion by the end of July, up from about $9 billion at the end of 2025, and it has committed $30 billion to Microsoft Azure, which runs on Nvidia silicon.
Sam Altman told Fortune in an interview published September 12 that OpenAI will not hold an IPO this year, saying that “given everything happening with safety, right now would be an ill-advised moment to go public,” and that the company feels no pressure on timing. He pointed to 2027 as a more plausible window. The week before, OpenAI said it was lobbying for mandatory national AI safety requirements in the United States.
OpenAI announced ChatGPT for Financial Services on September 10, pairing GPT-6 Astra with financial data, modelling tools and client-material generation. Initial data partners are Daloopa, PitchBook, LSEG News and Crunchbase, and OpenAI indexes and hosts those datasets on its own infrastructure — removing, in its words, the need for firms to negotiate separate contracts or configure connectors. Morgan Stanley and Evercore were design partners. Initial workflows target company research, valuation, LBO modelling and pitchbook preparation.
TAR announced a $120 million Series A on September 10 led by Spark Capital at a $1 billion post-money valuation, with Buckley Ventures and Align Fund joining. The company builds self-contained modular renewable generation and battery systems in West Texas, aimed at data centres that cannot wait on interconnection. Capital goes to its Austin headquarters, a San Francisco engineering office and West Texas manufacturing and logistics.
Baseten acquired Blaxel on September 10 for undisclosed terms. Blaxel builds microVM-based sandboxes that start and resume up to five times faster than alternatives and support agent code execution, tool use and MCP servers. Baseten intends to pair that execution layer with its own inference platform so models and the environments they act in run on one runtime. It is Baseten's second acquisition since its Series F at a $13 billion valuation, after the reinforcement-learning startup Parsed.
HelmGuard raised $7.3 million in seed funding announced September 9, co-led by Infinity Ventures and Frontline with FinTech Collective, Stage 2 Capital and Entrepreneurs First participating. Founded in 2024 by former Palantir executive John Daley and Jack Miller, the UK company runs an agentic GRC and security platform that continuously collects and assesses risk signals from source systems. The capital funds US expansion and engineering and go-to-market hiring.
Lightfield raised a $47 million Series A led by Andreessen Horowitz on September 9, with Maverick Capital, Coatue, Audacious, Alumni Ventures, Greylock and Lightspeed participating. The AI-native CRM reports 5,000 business customers since launching in November 2025 and 400% net dollar retention. It is the same team that built the AI presentation product Tome, which reached 25 million users on $80 million raised; the pivot took headcount from 70 to 7 before rebuilding to about 40. Net dollar retention of that order is the cleanest evidence a company can offer that its revenue compounds without being re-sold.
Latitude raised a $35 million Series A on September 9 led by Oak HC/FT, with NEA, Coinbase, Lightspeed Faction and OpenFX participating. Founded by alumni of Stripe, Uber, Coinbase and Meta, the company settles stablecoins into local bank accounts and mobile wallets in emerging markets, so recipients receive local currency in minutes without holding crypto. It follows an $8 million seed earlier in 2026, bringing the total to $43 million, and the capital funds compliance, legal and licensing across 45 US markets and abroad.
Automattic's board put co-founder and CEO Matt Mullenweg on paid leave on September 9, naming CFO Mark Davies interim CEO. Mullenweg said in Slack that he had 50 minutes' notice of the vote and that his request for legal review was refused. By September 12 Automattic confirmed he had returned as chairman and CEO with the board's support. The reversal took three days. Where the rest of the week showed how firmly the market now prices revenue, this showed how loosely a board can hold control when the founder holds the relationships.
Every headline this week reported two prices. There was only one. A price is what somebody paid to own a thing. The other number is what a minority investor accepted for an option, on terms that pay them before they pay you.
I should have seen it in issue #023 →, when Bending Spoons bought Airtable. I read that markdown as a verdict on the business. It was largely a change of instrument — the first buyer who wanted the whole thing, in cash, was the first whose number meant what the word price means.
Harvey is the case, and I nearly filed it as an oddity. The most legible revenue of the week — $400 million ARR, 80% of the Am Law 100 — priced at 38.8 times, nowhere near Miro. Growth and sector sit in that gap too. But Harvey sold preferred stock and Miro sold itself, and no amount of growth makes those the same kind of measurement.
Which makes the useful question a different one entirely.
The number to distrust is your own. Multiply your ARR by 2.3 to 2.7 — the band Bending Spoons actually wired, twice, in cash — and set it beside your preference stack. If the stack is larger, your common is worth nothing today, and so is every option you granted at a strike derived from the other number. That gap is measured and mailed to you annually, in the 409A most founders file without opening. And the preference is a paragraph, not weather: its multiple and seniority were drafted by someone, so they can be drafted differently. Belief becomes capital. But only at the moment somebody wires it.