Three separate stories this week about one question: who has the right to deny you access to a model, and on what grounds. A company did it to a customer, a government tried it on a company and lost in court, and a chipmaker walked away from a contract clause that would have let it do it wholesale. Let’s start with the loudest one.
SkyNet News Tribune is now The Drone Report. Check out the new, faster site with updates on the 1/2 hour with AI News.
OpenAI pulls its models from Cursor
Cursor — the AI coding editor a very large number of working developers now live inside — was acquired by SpaceX. On Friday, OpenAI announced it will terminate Cursor’s access to its models effective November 12.
The stated reason is unusually blunt for a corporate notice. OpenAI says it “cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts,” and it names them: Twitter breaching OpenAI’s terms after Musk acquired it, and xAI using OpenAI data without authorization to train models. This is not a pricing dispute dressed up in policy language. It’s a company saying, on the record, that it does not trust a specific person to honor an agreement.
Two details are worth pulling out. First, the roughly two-and-a-half-month runway is the maximum notice the contract allows — OpenAI took the longest window available rather than the shortest, and said so. That’s a deliberate signal that the target is SpaceX and not Cursor’s users, and it costs OpenAI something to send it. Second, OpenAI went out of its way to praise Cursor’s team and product and offer transition support. Read charitably, that’s a company trying to punish an owner without punishing a userbase. Read less charitably, it’s damage control on a decision that strands a lot of developers regardless of tone.
The part that should concern people who write software for a living has nothing to do with Musk or Altman. Cursor’s users made a tooling decision; months later, their tool’s corporate parent changed, and the model underneath is being removed by a third party they have no relationship with. Nobody in that chain did anything wrong from the developer’s point of view, and the developer bears the whole cost. Model access is a dependency you don’t control, can’t insure, and can lose because of a transaction two levels above you. Anyone building on a single provider should be reading this as a design constraint rather than as gossip.
Whether OpenAI is entitled to do this isn’t really in question — it’s a commercial contract and the termination clause is being used as written. Whether it’s good that model providers can and will do it is a different question, and it’s the one the next two stories are also about.
A judge tells the Pentagon it can’t blacklist a critic
The most consequential ruling in AI this year landed Thursday with a fraction of the attention the Cursor story got. U.S. District Judge Rita Lin held that the Pentagon’s designation of Anthropic as a supply-chain risk was unlawful — a violation of both the First Amendment and Fifth Amendment due process, and “arbitrary and capricious” besides.
The background, briefly. Earlier this year Defense Secretary Pete Hegseth and President Trump designated Anthropic a supply-chain risk, an administrative label that directed every federal agency to stop working with the company. The underlying dispute was Anthropic’s refusal to remove guardrails that would have permitted fully autonomous weapons and mass surveillance of American citizens. Anthropic sued in March, in California and in D.C. This is the California case; D.C. is still live.
Lin’s language is worth quoting rather than summarizing. She found the government had acted to “make a public example out of Anthropic for its ‘arrogance’ in criticizing the government,” and wrote: “The empty invocation of national security is not a blank check to punish and retaliate against government critics.”
What makes this bigger than one company’s procurement status is the mechanism it shuts down. A supply-chain risk designation is fast, requires no trial, and is nearly impossible to appeal in any practical timeframe — it is the cheapest available tool for making a company’s life impossible. Establishing that it can’t be pointed at a firm for declining to build something, or for saying so publicly, matters to every AI company that will at some point be asked for a capability it doesn’t want to provide. The precedent runs to whoever is in office next, too. The Pentagon has not commented, and an appeal is likely.
Nvidia backs away from approving its partners’ customers
Friday’s post covered Nvidia admitting on its earnings call that it provides credit enhancement for other labs’ compute purchases. Here’s the sequel. The Wall Street Journal reports that Nvidia has paused its AI Compute Partnership, launched only in July, after internal employees raised antitrust concerns.
The structure: Nvidia extended credit support to cloud providers; the provider set a base hourly rate covering costs, and Nvidia took 50% of revenue above it. Nvidia disclosed $36 billion of commitments under six-year agreements in its latest quarterly filing. The clause that caused the trouble wasn’t the money — it was an approved-customer provision limiting which clients a provider could rent the chips to, paired with Nvidia’s stated preference for spreading capacity across several smaller AI firms rather than one large one. Providers objected to being told who they could sell to.
Nvidia disputes the framing. A spokesperson says the underlying business model “remains in place and continues to evolve amid high demand,” which is a denial of the word “paused” more than of the substance. Either way, the interesting fact is already on the record: Nvidia wrote a contract giving itself approval rights over its customers’ customers, and then its own employees flagged it as an antitrust problem. That is a company discovering the edge of its position by walking into it.
Sony and Warner Chappell sue Anthropic
Filed late Friday in the Northern District of California: Sony Music Publishing, Warner Chappell, and a long list of other publishers accusing Anthropic of a “brazen campaign of illegally torrenting, scraping, and downloading copyrighted works” to train Claude. The complaint names Dario Amodei and Benjamin Mann personally alongside the company, which is an aggressive choice and a deliberate one.
The legal theory is narrower and stronger than “AI training is infringement,” and that’s what makes it dangerous. In Bartz v. Anthropic, the court held that training on copyrighted works was lawful but acquiring them through piracy was not — a distinction that cost Anthropic $1.5 billion. The publishers are running straight at that holding: not arguing that Claude shouldn’t have learned from lyrics, but that the books containing them were obtained by torrent. Anthropic says it disagrees with the claims and will defend itself robustly.
Expect more of these, at every rightsholder category with an organized licensing body. The music publishers have watched a $1.5 billion settlement establish the theory and are now applying it. Anthropic’s exposure here is a function of how it built its corpus years ago, which no amount of current-day policy can retroactively fix.
The Fed puts AI in the model
At Jackson Hole on Friday, in his first keynote as chair, Kevin Warsh called AI “a new variable — potentially a new factor of production,” and the current moment “a hinge point in history.” He noted that progress “has been faster even than its evangelists predicted,” and cited roughly $100 billion in annualized token sales across the leading labs, up more than 500% year over year.
“Factor of production” is not a rhetorical flourish; it’s a term of art that sits alongside labor and capital. A Fed chair using it signals that the institution now believes AI may change the productivity denominator rather than merely being a large sector. Warsh was careful to add that a task force is still studying whether AI complements or substitutes for workers, and how much of the value concentrates among chipmakers and infrastructure providers — and that none of it will drive near-term rate decisions.
That last caveat is the honest part. The Fed is saying this is probably the most important economic variable of the decade and that it does not yet have a framework for it. Both halves of that sentence are worth taking seriously.
Tencent open-sources 770 billion parameters
Hy4-preview, released Friday: 770 billion total parameters with 49 billion active, a context window over a million tokens, open weights, and API pricing at $0.83 per million input tokens and $2.50 per million output. Free on Tencent’s WorkBuddy and CodeBuddy for two weeks, with API access via Tencent Cloud and OpenRouter.
Tencent’s own blind evaluation — 163 experts across 203 engineering tasks — puts it at 2.99 out of 4.00 against GLM-5.3 at 2.92 and Kimi K3 at 2.94. That is a self-reported result on a self-designed benchmark and should be treated accordingly; the gap is inside the noise. The claim I’d want independently checked is the other one: Tencent says the model participated in the automated optimization of its own training pipeline and autonomously improved inference throughput by 31.8%.
Second major Chinese open-weight release in a week, after Zhipu’s GLM-5.3-Flash. The pattern is now hard to miss: the frontier-adjacent tier of open models is increasingly Chinese, and the American labs releasing weights at comparable scale can be counted on one hand.
And one good thing
Slightly older than the rest — it published in Nature on August 6 and I don’t think it got the attention it deserved. Google DeepMind’s WeatherNext Cyclones forecasts a tropical cyclone’s track, intensity, and wind structure together, and its three-day forecasts are as accurate as what the previous state of the art managed at two days. That single extra day represents roughly a decade of normal progress in the field.
It isn’t a demo. The National Hurricane Center used it operationally through the 2025 season, where it helped forecast Hurricane Melissa’s rapid intensification and Jamaica landfall — a day of extra warning for people who needed to move. The work was done with Google Research, the NHC, CIRA, and the UK Met Office, and the code and weights are on GitHub under Apache 2.0, including a mini version that runs in free Colab.
A week largely about who can cut off whom is a reasonable week to note that some of this is just straightforwardly good.
Sources
- Our decision on Cursor following its acquisition by SpaceX — OpenAI
- OpenAI to end model access to Cursor after acquisition by Elon Musk’s SpaceX — CNBC
- OpenAI will pull its models from Cursor due to SpaceX acquisition — Engadget
- Anthropic gets its first court win over the Pentagon’s supply-chain risk label — TechCrunch
- Judge says the Pentagon can’t designate Anthropic a ‘supply chain risk’ — NPR
- Judge rules the Pentagon’s supply chain risk label for Anthropic unlawful — CNN
- Nvidia pauses AI cloud revenue-sharing deals over antitrust concerns — Quartz, on WSJ reporting
- Sony Music, Warner sue Anthropic, alleging a “brazen campaign” of IP theft — TechCrunch
- Sony, Warner sue Anthropic, alleging “blatant theft” of intellectual property — Axios
- Keynote remarks by Chairman Warsh at the 2026 Jackson Hole Economic Policy Symposium — Federal Reserve
- Tencent releases and open-sources Tencent Hy4 preview — Tencent
- AI model achieves breakthrough in forecasting cyclones — Google DeepMind
- The underlying paper — Nature
Leave a Reply