Blog

Nvidia will guarantee $105 billion of the debt behind OpenAI’s next mega data center, and that number keeps shrinking. If you are recruiting into banking, private credit, or infrastructure investing, this is the live case study in circular financing that you should expect to be asked about: the chipmaker is underwriting the borrowing capacity of the customer who buys its chips. On Monday, Nvidia signed a partnership backstopping SB Energy’s Ports Pike campus in Pike County, Ohio, disclosing in an SEC filing an “aggregate payment obligation” capped at $105 billion, plus a $1.5 billion investment in SB Energy itself. OpenAI leases the site for as long as 20 years and uses Nvidia chips exclusively in the first phase; the campus is expected to reach as much as 8 gigawatts, with the first 800 megawatts online in 2028. Watch the trend line on the guarantee: reporting in July put it near $250 billion, the Wall Street Journal reported on August 14 that it had been cut to “less than $120 billion,” and it landed at $105 billion. When the $250 billion figure first surfaced, Nvidia shares fell about 4.5% intraday on circular-financing concerns.

Goldman put real numbers on the AI hiring squeeze, and the honest version is narrower than the headlines. This is the week’s most useful labor-market read for anyone heading into recruiting, and its value is in the calibration: the effect is real, it is concentrated at the entry level, and it is smaller and narrower than most coverage implies. Goldman Sachs economists analyzed employment growth across more than 800 occupations in developed economies. Industries with greater exposure to AI automation have seen slower job-openings growth since the second half of 2022; US call center employment now runs roughly 39% below trend. The seniority split is the finding to remember: a 10% increase in occupational AI exposure is associated with about a 0.1 percentage point drag on annual headcount growth overall in the US, France, and Canada, but with more than 0.2 points for US entry-level workers and more than 0.6 points in Australia. Software publishing, management consulting, and advertising services are among the sectors diverging most from historical hiring norms. Goldman’s own conclusion is measured: the pressure is clearly visible in the data but “limited to a relatively narrow set of industries and workers.” Note also that adoption across major developed economies still sits at only about 15% to 20%, and that these are associations, not proven causation.

Two deals this week showed how the big AI players now buy capability without buying companies. The structure is the story here, not the size, and both templates are worth being able to explain in a corporate strategy or M&A interview. On August 19, Marvell granted Google a warrant to buy up to 58,970,907 of its shares at $206.58 apiece, roughly $12.2 billion, with most of the warrant vesting only against purchasing targets running through fiscal 2033; if Google hits them all, the arrangement could generate roughly $120 billion in cumulative Marvell revenue. Marvell closed up nearly 10% at $237.27, while Broadcom, which co-designed Google’s TPUs for about a decade, fell about 5%. Then on August 20, Nvidia agreed to pay Poolside $6 billion for a non-exclusive license to the “Model Factory” software the startup used to build its models, to hire 109 of its staff, and to invest $1 billion at a $12 billion pre-money valuation. Poolside’s letter to investors insists the deal “is not an acquisition and it is not an acquihire.” It is the third time Nvidia has used this structure, after Groq (about $20 billion) and Enfabrica (about $900 million): roughly $27 billion committed, three companies still standing, and no acquisition for a regulator to review.

Anthropic still leads with business buyers, OpenAI is closing the gap, and enterprise AI spending looks a lot less sticky than either would like. Until both labs file public financials, third-party spending data is the best proxy we have for who is actually winning the enterprise, and the most interesting signal in it is churn. New data from Ramp, covering more than 70,000 American businesses, shows Anthropic at nearly 44% of companies as of July against OpenAI’s nearly 40%; Anthropic first took the lead in May at 41% to 39% and OpenAI has not regained it, but Ramp economist Ara Kharazian says OpenAI is growing faster quarter to date. Two caveats matter: Ramp’s customer base skews toward tech, and it excludes large enterprises that use other spend-management providers. Kharazian credits OpenAI’s GPT-5.6 Sol and points to price and regulator-imposed data-retention requirements as drags on Anthropic’s Fable tier; OpenAI leaned directly into that opening on August 19 by announcing zero data retention for its frontier models. The read for anyone modeling these businesses: buyers will move between labs with each model release, and “sticky enterprise AI spend” is an assumption, not yet a fact.

Anthropic moved its agent tooling out of beta, which is the unglamorous step that makes agents actually deployable at work. “Generally available” sounds like a footnote, but it is the moment a regulated company can put a tool into a real workflow, and this release targets exactly the back-office processes that consulting and financial-services firms bill for. On August 20, Anthropic made computer use, a new browser use tool, the Skills API, and the Files API generally available on the Claude Platform; computer use now takes multiple actions per turn and is eligible for HIPAA-regulated workloads under a business associate agreement. The customer numbers Anthropic published are the part worth interrogating: Asteroid says its longest insurance claims workflow fell from 32 minutes to 13, with cost per task down about 30% and completion at 100%, with no prompt changes. Box describes using the Skills API to encode a bank’s credit methodology and produce source-grounded credit memos for analyst review. Treat vendor-supplied metrics with the usual skepticism, but the direction is clear: the automation target has moved from drafting text to executing multi-step processes inside software that has no API.

OpenAI split ChatGPT by age and by geography on the same day. Two announcements on August 18 point the same direction: one product, increasingly different terms depending on who you are and where you live, which is what a consumer platform looks like once regulators and advertisers both have a seat at the table. OpenAI launched ChatGPT for Teens for users aged 13 to 17, placing users there automatically if its age-prediction system estimates they are under 18 or if they state their age; protections around self-harm, violence, eating disorders, and explicit content are on by default, alongside Study Mode, homework reminders that redirect shortcut attempts, scheduled Study Hours, and expanded parental notifications. The same day, OpenAI expanded its ChatGPT advertising product across Europe. If you are thinking about product strategy or platform regulation, the segmentation itself is the thing to notice, not either announcement alone.