The week the money moved and nobody shipped a model

Between 16 and 22 August, no major lab announced a new frontier model. The most recent one, Gemini 3.7 Flash, landed on 13 August, outside this window. What moved instead was capital and paperwork: a chip vendor guaranteeing a customer's rent, an inference startup doubling its valuation in a month, a Chinese platform publishing what its model business actually costs, and OpenAI cutting a price on a timer while posting eight announcements in four days, most of them about rules rather than capability.

The compute layer is being financed like real estate

On 17 August NVIDIA disclosed, in a Form 8-K, that it had entered into multiple residual value guaranties with SB Energy covering leases for approximately 4.25 gigawatts of IT load at the PORTS Technology Campus in Pike County, Ohio. An OpenAI affiliate is the tenant. NVIDIA's aggregate payment obligation is cumulatively capped at $105 billion for that initial commitment, and it can provide credit support for roughly 3.8 gigawatts more at its sole discretion. The filing is precise about the mechanics: if OpenAI becomes insolvent or fails to pay, NVIDIA covers the shortfall between a guaranteed minimum lease value and whatever a replacement tenant or sale recovers. Its obligations end if OpenAI achieves a satisfactory credit rating, and payment conditions are expected to begin as premises reach ready-for-service, from 2028.

The accompanying press release framed the same arrangement differently: NVIDIA securing land, power and shell capacity for its own AI factories, plus a $1.5 billion equity investment in SB Energy alongside SoftBank and OpenAI. Both descriptions are accurate. Together they say the party selling the accelerators is now underwriting its buyer's twenty-year property obligations.

The following day, Etched said it had raised $700 million led by Jane Street at a $21 billion valuation, with Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global and Blackstone among the participants. It says it shipped its first rack to Jane Street, which is deploying it in production, and holds more than $1 billion in customer contracts. Its own release notes that this doubles a valuation set less than two months earlier. These are vendor-supplied figures from a private round, not audited disclosure, and the performance claims for its Low Voltage Inference and Cluster Scale Memory designs come from Etched alone.

The token layer showed what it costs

Then on 20 August Alibaba reorganised its segment reporting and, for the first time, broke out AI Labs and Applications: the model labs, the Qwen consumer app and QwenWork. For the June quarter that segment reported RMB3,338 million of revenue (US$492 million), up 16%, against an adjusted EBITA loss of RMB13,861 million (US$2,043 million), a loss 330% wider than a year earlier. In the same filing, AI Cloud and Compute Services grew revenue 45% to RMB48,437 million with adjusted EBITA up 133%.

One company, two businesses, opposite signs. Selling compute is profitable and growing fast; giving away consumer inference is not. Few large platforms publish that split at segment granularity, and the number Alibaba published is the size of the subsidy.

OpenAI's move on 21 August rhymes. GPT-5.6 Sol dropped to $4 per million input tokens and $20 per million output tokens. The changelog describes it as promotional pricing available at least through 21 November 2026. A price with a stated expiry is a demand instrument, not evidence that serving the model got cheaper. If you are rebuilding a cost model this weekend, that distinction is the whole story: budget the promotional rate for the quarter, and keep the old rate in the spreadsheet as the fallback case.

What actually changed for people building things

The developer-facing changes were small and mostly plumbing, but worth wiring in:

• A prompt caching dashboard (20 August) that reports cache hit rate over time, cache reads per write, and the split between cache-read, cache-write and uncached tokens. If you have been estimating cache efficiency from invoice deltas, you can stop.
• Per-request regional processing (21 August), selectable through a prefixed domain with an API key from a project with Global geography, with existing eligibility, retention, endpoint and model support requirements unchanged.
• Transparent backgrounds in preview for gpt-image-2 (20 August), with png or webp output; jpeg is not supported.

Outside the API, Cursor launched Origin on 18 August, a code hosting platform aimed at what developers use GitHub for, with repos able to sit alongside GitHub rather than replace it and "agent native" features promised without detail. TechCrunch reported that GitHub suffered a lengthy worldwide degradation the same day. Cursor's own announcement page was not reachable from our newsroom, so these details rest on that reporting rather than the company's post. Google, separately, said on 20 August that Gemma has passed a billion cumulative downloads with over one hundred thousand community variants published, which measures distribution rather than use.

OpenAI spent the week writing policy. Anthropic said nothing.

Between 17 and 20 August, OpenAI's news feed carried, in order: New policy ideas for the Intelligence Age, The Defender's Window, Pacing model development in an era of cyber-critical capabilities, Strengthening democratic oversight in national security, ChatGPT for Teens, ChatGPT Ads expanding to European markets, Offering Zero Data Retention for frontier models, and AI Futures, a new blog on how transformative AI could reshape power and governance. We covered two in detail during the week: how OpenAI is pacing frontier training after July's evaluation-sandbox escape, and the Zero Data Retention post and its Private Safety Processing preview.

That is a company arguing about the rules while shipping a price cut and a teen product, not one arguing about capability. Anthropic, by contrast, published nothing to its newsroom in the window; its most recent post remains the 14 August explainer on Claude's text watermark. That is an observation about publishing, not about activity.

What we are not treating as settled

The window opened with a report, published on 16 August, that Stripe had finalised a deal to acquire the model-routing gateway OpenRouter at a price above $7 billion. OpenRouter raised a $113 million Series B in May at a reported $1.3 billion valuation and claims eight million users and access to more than 400 models. The account rests on anonymous sourcing, and a Stripe spokesperson said the company does not comment on rumours or speculation. If you route traffic through a gateway, ownership of that layer matters, but nobody has confirmed it. Two other widely circulated deals this week rested on sources we could not open at all, so they are absent here rather than reported thinly.

The throughline

For a week with no model launch, an unusual amount was decided. The layer that sells compute is being financed with instruments borrowed from commercial property: twenty-year leases, residual value guaranties, vendor credit support, capital committed against demand arriving in 2028. The layer that sells tokens absorbs the cost meanwhile, visibly in Alibaba's case and implicitly in OpenAI's dated discount. Capability was not the variable this week. Who carries the risk was.