Business of GTM

OpenAI Became a B2B Company the Same Week Cursor Stopped Being One

OpenAI CFO Sarah Friar told investors enterprise revenue passed consumer on a $40 billion annualized run rate, two quarters early. Cursor, with 50,000 teams by its own count, closed its sale to SpaceX.

Black-and-white halftone of OpenAI CFO Sarah Friar speaking with her hands raised, in front of a wide red arch on a beige background.
Sarah Friar, CFO of OpenAI, at the World Economic Forum Annual Meeting in Davos in January 2026. Photo illustration by Business of GTM. Original photo: World Economic Forum / Thibaut Bouvier (CC BY-NC-SA 4.0).

On Thursday, OpenAI CFO Sarah Friar told a room of investors that the company's enterprise business now makes more money than ChatGPT's consumer side. The numbers she put behind it, reported by CNBC: a $40 billion annualized run rate as of July, up 20 percent in a single month, with business customers up 32 percent to more than 2 million. OpenAI entered the year with revenue split 60-40 toward consumer and had forecast the crossover for the end of 2026. The lines crossed in August, two quarters early. "The majority of our revenue is now enterprise," Friar said.

About twenty-four hours earlier, Cursor CEO Michael Truell published the shortest big announcement of the year: "Cursor is now part of SpaceX." The acquisition closed Thursday, completing a process that began in April as a model-training partnership with SpaceXAI. Cursor's own explanation for selling fits in a clause: SpaceX owns the largest GPU fleet in the world, and Cursor's models need it.

Read separately, that is a strong earnings note and a surprising acquisition. Read together, they are one event. The two most-watched product companies in AI concluded, in the same week, that this market is won with enterprise contracts and compute. One of them became the buyer's side of that trade. The other became the asset.

The distribution machine is the story

Start with OpenAI, because the obvious reading of Friar's number, that enterprise AI adoption is booming, is true and misses the point. Adoption doesn't book itself. Three years ago OpenAI had no enterprise sales organization worth the name. It now runs a motion serving 2 million business customers, credits the surge to the GPT-5.6 model series, an enterprise agent called ChatGPT Work, and Codex, and is rebuilding that motion's leadership while it scales. Denise Dresser is out as chief revenue officer after eight months, replaced this week by Dali Rajic, the former president and COO of Wiz. A consumer company can tolerate churn in its sales leadership. A company that swaps CROs mid-surge, and reaches for the operator who scaled the fastest-growing security company in history, is telling you where it believes its future revenue lives.

The surge is also partly manufactured, in the way good land grabs are. A TD Cowen analysis reported by Business Insider found OpenAI's model price cuts drove roughly 14x usage growth and revenue gains of 34 to 45 percent across products. Cutting prices 80 percent to win enterprise workloads only works if your compute costs keep falling, which only happens at the scale of the largest GPU fleets on earth. The pricing strategy and the infrastructure strategy are the same strategy. Bloomberg reports the company on track to double revenue ahead of an IPO, which makes the enterprise line the one investors will price.

No enterprise software company has built a motion this big this fast. Salesforce needed roughly a quarter century to reach the run rate OpenAI hit inside four years of ChatGPT's launch. Whatever you think of OpenAI's models against Anthropic's or Google's in any given month, the distribution machine is now the story, and it compounds independently of the benchmark table.

The rest of the market is arranging itself around the same conclusion. IBM struck a partnership with OpenAI this week to deploy its models across financial services, government, telecom and retail, joint go-to-market included, mirroring the alliance IBM built with Anthropic last year. When the largest enterprise services firm on earth becomes a distribution channel for both leading labs at once, the labs' revenue question stops being model quality and becomes channel coverage. Enterprise distribution is the scarce asset now. Everyone holding a frontier model is buying it, renting it, or building it at OpenAI's pace.

Cursor did everything right and still sold

Cursor is the same lesson read from the other side, and it stings more, because Cursor was doing everything right. It was the fastest-growing developer tool anyone had measured. It had built a real enterprise motion of its own: 50,000 teams and nearly 70 percent of the Fortune 1000, by the company's own count, with a President of Global Revenue hired from Rubrik in February to scale it. Product working, enterprise GTM working, brand beloved; it still sold, and the reason is in its own trajectory. Cursor stopped being an editor and started being a model company: first Composer, its own coding model, then a partnership to train Grok 4.5 with SpaceXAI in July, then Grok 4.6, released two days before the close at half the price of comparable frontier models. Training frontier models is a capital competition. The currency is GPUs, and GPUs pool on a handful of balance sheets. The moment Cursor crossed from renting models to training them, its independence had a clock on it.

An Nvidia filing that landed the same day completes the picture. Nvidia disclosed a $21 billion stake in SpaceX, per the FT, after Musk announced SpaceX would outfit its data centers exclusively with Nvidia. Follow the loop: the chip maker invests in the compute owner, the compute owner acquires the application company, and the application exists to sell the compute. Money is circulating among four balance sheets and calling itself a market.

Lovable stays independent by renting models

The counter-example arrived the same week. Lovable raised $400 million led by Menlo Ventures at a $13.3 billion valuation on Tuesday, doubling its value and staying independent. So the app layer isn't doomed as a category. Look at what Lovable doesn't do, though. It trains no frontier models. It rents from every lab, sells to people who never wrote code, and competes on product and distribution. Independence in AI applications appears to be available on exactly those terms: stay out of the training game, and own a customer the labs don't already consider theirs. Cursor's customers were developers, the labs' first and favorite market. Lovable's are everyone else.

Power pools where the enterprise contracts and GPUs sit

An engineering team that standardized on Cursor bought an independent editor. It now holds a contract with SpaceX, which owns the GPU fleet, sells the Grok models inside the tool, and counts Nvidia as a $21 billion shareholder. Cursor announced the change in one line.

OpenAI's version of the same pull runs through its sales organization. The majority of its $40 billion run rate is now enterprise, Bloomberg reports it on track to double revenue ahead of an IPO, and the CRO who will defend that number came from Wiz five days ago.

By Business of GTM