The AI Brief

OpenAI CFO Sarah Friar told employees at a Wednesday all-hands that the company "will be a public company in 2027," but could make its public market debut sooner if "our business continues to inflect." Friar cast the listing as another milestone rather than a finish line, "another fundraise", pointing to the $122 billion OpenAI closed in March as the reason it can pick its moment.

The all-hands carried unusually specific financial disclosures. Friar said OpenAI's revenue run rate is up 35% so far this quarter, its enterprise revenue run rate is up 50%, and its AI coding and work product has reached 20 million weekly active users. Friar showed employees slides confirming OpenAI generated $6.7 billion in revenue for Q2, up 18% from the prior quarter. Separately, Friar told investors that enterprise and consumer revenue "have now crossed," after starting the year at a 60-40 split favoring consumer.

Friar addressed the Anthropic IPO race directly. She told staff: "As you know we are confidentially under file, and Anthropic is also under file. There is a chance they pull the cover off that confidential file in the coming weeks and become public in September. That's OK, we are running our own race." The 2027 date settles an internal argument: Sam Altman had been pushing for a Q4 2026 listing, while Friar had pressed internally to wait due to public-disclosure readiness concerns. The public S-1 has not yet appeared on SEC EDGAR as of today; SEC rules require the registration statement to be made public at least 15 days before an IPO roadshow begins.

Nvidia is in discussions to invest in Mercor, an AI data supplier and labeling provider, in a funding round that would value the three-year-old startup at roughly $20 billion, according to a report from The Information citing people familiar with the matter. The potential investment would double Mercor's valuation from its $10 billion Series C completed in October. The funding round is being led by General Catalyst, a current investor in Mercor.

Mercor reported $614 million in gross revenue during the first half of the year alone, and by June its annualized gross revenue run rate had surpassed $2 billion. Nvidia paid Mercor tens of millions last quarter, and a handful of Mercor staff now work almost entirely on the account, but Nvidia also buys from Turing and Scale and relies heavily on synthetic data for Nemotron. According to company disclosures, Mercor's data was integrated into Nvidia's two most recent Nemotron releases.

The potential investment underscores Nvidia's evolution from a pure AI chip supplier into a full-stack AI ecosystem player; the company is increasingly deploying capital across the AI value chain, and in the fiscal quarter ending in April, it invested $18.6 billion into private enterprise investments and infrastructure funds, exceeding its total investment output for the entire previous year. Major clients served by Mercor include OpenAI, Google DeepMind, and Anthropic , meaning a Nvidia stake would make the chip giant a co-investor in a supplier shared across the frontier lab landscape.

OpenAI's Zero Data Retention gives eligible API customers a clear promise: OpenAI does not retain their prompts or model responses after a request is processed. Customer content is not available to OpenAI personnel for review. The structural problem: as models take on longer, more complex tasks, some serious risks may only become visible across multiple interactions, and existing ZDR-compatible safety systems evaluate each interaction individually.

OpenAI is previewing Private Safety Processing, designed to identify patterns across related interactions without giving OpenAI personnel access to the underlying content. For ZDR deployments, customer content remains on infrastructure the customer controls. The program is set to start in September; OpenAI has not yet said what constitutes eligibility for the zero data retention program and will share details in a forthcoming technical white paper.

The competitive framing is explicit. Much of OpenAI's revenue flows not directly from enterprise but through partners like Microsoft and AWS. Zero data retention has technically been available to certain API customers since 2023, but this reaffirmation signals OpenAI is making it a cornerstone of its enterprise strategy rather than a niche offering. Procurement teams building direct-API contracts will need clarity on whether PSP is compatible with their own data-residency obligations, the technical white paper is the document to watch.

SpaceX approached AI coding startup Cognition AI about a potential acquisition, according to people familiar with the matter, in what would have been its second large takeover in recent months to gain ground in the AI race. Cognition didn't engage with the takeover approach; there are ongoing discussions about working together, including potentially arranging for Cognition to use SpaceX's computing capacity.

Cognition CEO Scott Wu disputed the report soon after it published, writing on X that the story was inaccurate and that Cognition "is not for sale," adding that the two companies haven't been in talks. The report comes a few days after SpaceX's $60 billion acquisition of Cursor, another AI coding startup, whose deal closed last week. Cognition, founded in 2023, builds Devin, an AI agent designed to automate programming tasks for software engineers; the startup was valued at $26 billion in a May funding round.

Cognition has since opened early talks for fresh financing at a valuation of at least $40 billion, a trajectory that gives it less reason to sell. Devin is used by Mercedes-Benz and GE Aerospace, and Bloomberg reports that Cognition's commercial progress interested SpaceX as much as its technology. Musk told staff that AI revenue would out-earn rockets by September, a target that requires enterprise customers rather than research prestige.

Gemini 3.5 Pro was first promised by Google CEO Sundar Pichai at Google I/O on May 19, 2026, with a June general availability target. Google's highly anticipated Gemini 3.5 Pro is experiencing significant, repeated delays. Dubbed the "longest-awaited model of 2026," reports cite persistent coding and reliability problems, senior researcher departures, and a possible complete retraining from its foundational pre-training phase due to a "structural problem."

The model missed the June target, the July 17 target, which prediction markets had assigned high probability, and the Made by Google hardware event in New York on August 12 that the brief first identified as the live window (Vol. I, No. 85). As of August 8, Gemini 3.5 Pro remained in limited preview on Vertex AI and had not launched publicly. No public API entry, no model card, and no confirmed launch date have appeared as of publication today. The prolonged absence of 3.5 Pro is impacting market confidence and raising doubts about its eventual release.

Google has shipped interim releases, Gemini 3.6 Flash (July 21) and Gemini 3.7 Flash (August 13), to keep the Gemini API competitive on price-performance while 3.5 Pro remains in development. The net effect: the field around Google already has newer flagships in production, and a Gemini 3.5 Pro slip carries greater competitive consequence than the June slip did , every additional week compresses Google's window to set developer stack preferences before rivals consolidate them.