OpenAI Cuts Off Cursor Over SpaceX Buyout

The company invokes Musk's history of broken contracts to justify a November shutoff, Anthropic quietly walks from a $7 billion chip deal, and a Beijing model-maker asks Microsoft, AWS, and Google for a 30% cut

Saturday, August 29, 2026 · 8 min read · Issue #251

Lead

OpenAI notified SpaceX on Friday that it will wind down the contract supplying OpenAI models to Cursor, the AI coding tool SpaceX acquired earlier this month, with a shutoff date of November 12, 2026. In a public statement, OpenAI said it "cannot be confident that SpaceX will use our technology within our terms of service," citing Musk's record: after acquiring Twitter, the company broke terms of its OpenAI contract, and Musk admitted under oath earlier this year that xAI — also now part of SpaceX — had distilled OpenAI's outputs to train its own models in violation of similar terms. OpenAI said it is giving the maximum contractual notice period specifically to preserve developer access as long as possible.

The company frames the decision as a governance problem, not a competitive one, tying it explicitly to the accountability bar it says it now has to clear for its next frontier model, Astra, which OpenAI paused earlier this month over unresolved cyber-capability concerns. A four-year commercial relationship with real developer dependency is being unwound over a change-of-control clause few outside the deal probably remembered existed — which is precisely the point: these clauses exist so a vendor doesn't wake up serving a company it doesn't trust.

The immediate effect lands on Cursor's user base, who now have a hard deadline to migrate workflows built around OpenAI models to Anthropic, Google, or open-weight alternatives before mid-November. The broader effect lands on every other AI company running a multi-model coding assistant: model access is not infrastructure, it's a relationship, and relationships get renegotiated when ownership changes hands. Anthropic, whose Claude models already power a share of Cursor's traffic, is the obvious short-term beneficiary. The long-term question is whether SpaceX tries to plug the gap with its own or xAI's models, turning Cursor from a neutral multi-model tool into another node in Musk's vertically integrated AI stack — the exact outcome OpenAI's statement suggests it's trying to pre-empt by leaving early rather than getting phased out later. (OpenAI, Reuters)

Briefs

Anthropic quietly abandons its $7 billion MatX acquisition, pivots to a partnership. Anthropic had been negotiating to buy AI chip startup MatX — founded by former Google TPU engineers — for close to $7 billion, but the talks stalled and the companies are now exploring a cooperation agreement instead, according to Reuters, corroborated by a follow-up Seeking Alpha report. The reversal fits a pattern: Anthropic wants to cut its dependence on Nvidia and Google TPUs for inference, but building or buying custom silicon is proving harder to close than to announce. Coming days after Anthropic's own $30 trillion addressable-market pitch to investors, the walked-back deal is a reminder that infrastructure ambitions and infrastructure execution are running on different clocks. (Yahoo/GuruFocus)

An Anthropic fellow shows automated researchers beating humans at alignment work — for $4 an hour. A new Anthropic paper, "Automated Researchers Can Reliably Mitigate Alignment Failures," describes a system that searches the literature, proposes a training method, runs it for 30 minutes, and keeps what works — improving performance on all 10 tested misalignment benchmarks without degrading the model elsewhere. Anthropic's own text is blunt about the comparison: "the best AAR method beats what experienced humans propose, on average within six hours," and the automated system runs at roughly $4 per hour in API inference against $150 per hour for a human researcher. The paper is careful to flag its own ceiling — this only works as well as the benchmarks it's optimizing against reflect real alignment goals — but it's a concrete data point in a debate that's mostly been theoretical: recursive self-improvement is not a someday scenario for at least one narrow research task. (TechCrunch, Anthropic)

Z.ai confirms its stealth "Ox Alpha" model is a new GLM release built to rival DeepSeek. After weeks of speculation, Beijing's Z.ai (formerly Zhipu) confirmed that the anonymously-tested "Ox Alpha" model circulating on leaderboards is its own, part of the GLM series, and will ship with open weights. The confirmation adds a fourth serious open-weight contender — alongside DeepSeek, Qwen, and Kimi — to a Chinese model landscape that is now shipping frontier-adjacent releases on a near-weekly cadence. (The Edge Singapore)

China's CXMT rides the AI memory shortage to a stock-market debut that makes it the country's most valuable listed company. Shares in memory-chip maker CXMT surged as much as 466% on its market debut, a rally driven by the global DRAM and HBM squeeze that AI training and inference demand has created. Apple is separately reported by the Wall Street Journal to be testing CXMT memory chips for iPhones and MacBooks as supply constraints bite Western chipmakers. It's a second-order effect of the AI boom that gets less attention than GPU shortages: the memory market underneath the compute layer is being reshaped just as fast. (MSN, Seeking Alpha)

China / East Asia

Moonshot AI is negotiating with Microsoft, Amazon Web Services, and Google Cloud to host and sell access to Kimi K3, its 2.8-trillion-parameter open-weight model, asking for revenue-sharing terms as high as 30%, according to a Techstrong.ai report. The pitch is straightforward: Kimi K3's weights are already downloadable for free, but almost no enterprise buyer wants to self-host a model that size, and the hyperscalers already have the billing, compliance, and inference infrastructure that turns a free download into a sellable service. If any of the three cloud providers signs, it would be the clearest evidence yet that Chinese open-weight models are moving past hobbyist and budget-constrained adoption into mainstream Western enterprise procurement — which is exactly the scenario Washington's export-control architecture was built to slow down, not accelerate. Separately, DeepSeek's revenue reportedly reached $70 million in July, a tenfold jump from 2025, according to The Information, as the company heads into a funding round that would value it near $74 billion — evidence that the "give away frontier-competitive weights, monetize the API" strategy that alarmed US labs a year ago is now producing real revenue, not just market share. (Techstrong.ai, MSN)

India

Porsche signed a five-year, €1.25 billion ($1.46 billion) AI deployment contract with Tata Consultancy Services, India's largest IT services firm — and as part of the deal, TCS will acquire Porsche's own IT consulting arm, MHP, for €320 million. The transaction is notable less for its size than for its direction: an Indian IT major is buying a European automaker's in-house digital consultancy outright, rather than simply staffing a contract, in order to "industrialize AI at scale" for a marquee Western industrial client. TCS says its annualized AI revenue hit $2.6 billion in the June quarter, up 13.6% from the prior quarter, even as the Nifty IT index — the benchmark for Indian IT services stocks — sits nearly 20% down for the year on investor fears that AI will hollow out the outsourcing model that built companies like TCS in the first place. The Porsche deal is the counter-argument in real time: AI is displacing routine IT labor, but it's also creating a market for firms that can actually implement it at industrial scale, and TCS is moving to be the counterparty of choice for European manufacturers making that transition. (CNBC)

Europe

Wallonia's regional government blocked a permit for a Google AI data center that would have drawn river water for cooling, citing drought concerns — a small decision with an outsized signal: European regulators are increasingly willing to say no to hyperscaler infrastructure on straightforward resource grounds, not just data-protection or antitrust theory. It lands in the same week Mistral announced a "hundreds of millions of euros" collaboration with Saudi Arabia's HUMAIN to build sovereign AI infrastructure and Arabic-language frontier models for the Middle East — Mistral's third major sovereign-AI infrastructure move this summer, following an expanded Microsoft compute partnership and the launch of "European Compute Units" for long-term enterprise capacity commitments. Read together, the two stories describe the same tension from opposite sides: Europe wants sovereign AI capacity built on its own terms and within its own resource limits, and it's more willing than the US to let local friction — drought, land, water rights — override a hyperscaler's site plan. Mistral, meanwhile, is exporting the "sovereign AI" playbook it developed for European regulators to a Gulf state that wants the same leverage over its own AI stack. (Wallonia/MSN, Mistral)

The View

OpenAI's Cursor decision is the clearest evidence yet that AI model access is becoming a instrument of corporate foreign policy, not a commodity service. The company didn't cite performance, pricing, or competitive strategy — it cited a change-of-control clause and Musk's personal record of violating similar contracts at Twitter and xAI. That's a legal and reputational calculation, not a business one, and it sets a precedent every AI lab now has to reckon with: if you're building a product on someone else's frontier model, your vendor can walk away not because of anything you did, but because of who now owns you. Cursor didn't breach anything. It got acquired by someone OpenAI doesn't trust, and that was enough.

The interesting second-order effect is who benefits. Anthropic doesn't need to compete for Cursor's business — it just needs to still be there when OpenAI leaves. The same dynamic played out with MatX this week in miniature: Anthropic wanted to buy custom chip capability outright, discovered the deal was harder to close than to want, and settled for a partnership instead. Access, ownership, and control are turning out to be three different problems with three different price tags, and every major AI player is now negotiating all three simultaneously — with hardware vendors, with cloud providers, and now, apparently, with each other's customers.

The Miss

The story that should have gotten more attention this week: Moonshot AI asking Microsoft, AWS, and Google Cloud for a 30% revenue share to host Kimi K3. Coverage largely treated it as a routine cloud-distribution deal — "Chinese open model gets enterprise channel" — but the number itself is the real story. A 30% cut is roughly what Apple charges App Store developers, and Moonshot is asking hyperscalers who already compete with each other on margin to hand over that share for a model whose weights are already sitting on Hugging Face for free. Either the hyperscalers see enough differentiated demand for Kimi K3 specifically to pay it, which would be a remarkable statement about Chinese open-weight quality, or Moonshot is testing a number it expects to negotiate down hard. Nobody covering the story asked which one it is.

Pull Quotes

"We 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."
— OpenAI, official statement

"The best AAR method beats what experienced humans propose, on average within six hours... An AAR costs roughly $4 per hour in API inference against the $150 per hour we pay our human researchers."
— Anthropic, "Automated Researchers Can Reliably Mitigate Alignment Failures"

"That's serious money even by AI standards."
— Techstrong.ai, on Moonshot's reported 30% revenue-share ask to hyperclouds, Techstrong.ai

Reads & Links

• Wired on Anthropic's emerging standard for how AI agents should operate in the physical world — robotics, permissions, and safety rails ahead of humanoid deployment: Wired
• OpenAI's Cursor decision in full, including the Astra accountability framing: openai.com
• Bloomberg's read on the same Cursor decision, framed as "OpenAI Dumps Cursor": Bloomberg
• arXiv: WikiSkill — compiling agent experience into persistent, reusable knowledge for skill evolution, relevant to the same self-improvement theme as Anthropic's alignment paper: arXiv:2608.27454
• arXiv: RedEvoAgent — automatic red-teaming agents that evolve their own attack skills through experience, a mirror image of Anthropic's automated alignment researchers: arXiv:2608.27439

Out

That's issue #251. OpenAI just proved model access is a foreign-policy lever, not a subscription; Anthropic couldn't close the chip deal it wanted so it's settling for the one it can get; and Moonshot wants a 30% cut for a model anyone can already download for free. Back tomorrow.