Executive Summary
OpenAI will stop supplying its models to the coding tool Cursor on 12 November, after Elon Musk’s SpaceX bought Cursor’s parent company. Nothing Cursor did caused this. The trigger was who bought it.
That single fact sits underneath most of this week’s AI news. A court blocked the US government from cutting off Anthropic over its public views. Music publishers sued Anthropic over how it gathered training data. Different stories, same question: who controls access to AI, and on what terms?
If your engineering team uses Cursor, you have a deadline. Everyone else has a lesson about vendor risk.
Estimated reading time: 12 minutes
Key Takeaways
- Teams using Cursor with OpenAI models have a hard migration deadline of 12 November 2026. Anthropic and Google models keep working inside Cursor.
- Access to an AI model can now be withdrawn because of a change in ownership, not misuse. Treat model portability as a buying requirement, not a nice-to-have.
- Nvidia’s record quarter shows AI spending is still accelerating. Its paused $36bn financing programme shows the money behind that spending is far less settled.
- The legal risk in AI has shifted from “is training on copyrighted work allowed?” to “how did you obtain the data?” That changes what to ask your AI vendors.
- A Chinese lab released a near-frontier model under a fully open licence at roughly a sixth of premium pricing. Even if you never deploy it, it improves your negotiating position.
- If your AI agents can spend money or sign things, you need records proving who authorised what. Regulators and standards bodies are converging on this fast.
What Happened?
Six developments this week carry real business weight. Here they are at a glance, in order of how much they should matter to you.
| Story | What happened | Date |
|---|---|---|
| OpenAI cuts off Cursor | OpenAI will stop supplying models to Cursor after SpaceX bought its parent company for $60bn | 28 Aug |
| Nvidia’s quarter and pause | Record $96.2bn revenue, then a $36bn financing programme paused over antitrust worries | 26–28 Aug |
| Pentagon blacklist overturned | A federal judge ruled the US government illegally blacklisted Anthropic | 27 Aug |
| Music publishers sue Anthropic | Sony and Warner Chappell sued the company and named its founders personally | 28 Aug |
| 128 firms warn on cyber | OpenAI, Google, Microsoft, banks and security vendors signed a joint warning | 27 Aug |
| Tencent opens a 770B model | A very large model released under a fully permissive licence at low cost | 28 Aug |
One thing missing from that list is worth noting. No US lab released a new frontier model this week. All the model news came from China.
The Cursor cutoff
SpaceX bought Anysphere, the company behind Cursor, in a $60bn all-stock deal. The purchase was announced in June and closed on 14 August. SpaceX had already absorbed Musk’s X and xAI earlier in 2026.
That change of ownership opened a short window in which OpenAI could cancel its supply agreement. OpenAI used it. On Friday 28 August the company said it had notified SpaceX it would end model access, proposing a cutoff of 12 November 2026.
OpenAI’s stated reason was that it cannot be confident SpaceX will stay within its terms of service, pointing to past contract breaches by Musk-owned businesses — specifically X, after Musk acquired it. Existing OpenAI models keep working in Cursor until November. No new OpenAI models will be supplied in the meantime.
Cursor still offers models from Anthropic, Google and SpaceX’s own AI division. Anthropic co-founder Tom Brown said publicly that his company would add compute capacity to support Claude inside the tool.
Musk’s public response was dismissive.
Nvidia’s two stories
Nvidia reported results on Wednesday 26 August. Revenue was $96.2bn, more than double the same quarter a year earlier. Guidance for the current quarter is $108bn, and that figure assumes zero data centre revenue from China.
Then the second story. In July, Nvidia had launched something called the AI Compute Partnership. Smaller cloud providers must buy chips and build data centres long before they have customers to pay for them, so Nvidia offered credit support in exchange for a share of their future rental revenue. Its filing showed $36bn in commitments by late July.
The Wall Street Journal reported that Nvidia had paused parts of the programme. The sticking point was not the revenue split. Nvidia had reportedly told partners they could only rent the chips to customers Nvidia approved. Its own employees raised concerns that this could attract antitrust attention.
Nvidia disputed the reporting, saying its July model remains in place and continues to evolve. It did not confirm the terms would return unchanged.
The Pentagon ruling
On 27 August, US District Judge Rita Lin issued a 59-page ruling vacating the Pentagon’s designation of Anthropic as a national-security supply chain risk.
The dispute started in February over a roughly $200m contract to run Claude on classified systems. Anthropic wanted contract terms ruling out use in fully autonomous lethal weapons and mass surveillance of Americans. The Pentagon wanted language permitting any lawful purpose. Talks collapsed, and Defence Secretary Pete Hegseth applied the supply-chain risk label — a designation normally reserved for companies exposing military systems to foreign sabotage, and reportedly never before applied publicly to a US company.
The court found this was unlawful retaliation for protected speech, that Anthropic was denied due process, and that the designation broke the governing statute. The ruling does not require the Pentagon to buy anything. It restores the situation as it stood before the designation. An appeal is expected, and a second Anthropic case continues in Washington DC.
The music publishers’ lawsuit
Late on Friday 28 August, Sony Music Publishing, Warner Chappell Music and several affiliated publishers filed a 48-page complaint against Anthropic in California. They named CEO Dario Amodei and co-founder Benjamin Mann as individual defendants.
The complaint has four counts. It alleges the company obtained training material through torrenting, scraped licensed lyric sites against their terms, ran a destructive scanning operation on physical books, and stripped out copyright information such as song titles and songwriter names. It seeks up to $150,000 per work and $25,000 for each removal of copyright information.
Anthropic had not publicly responded when this article was written.
The cyber defence letter
On 27 August, OpenAI published an open letter signed by 128 organisations. Signatories include Anthropic, Google, Microsoft, AWS, Oracle, IBM and Cisco, most of the major security vendors, and financial firms including Capital One, Mastercard and Visa.
The core claim is that AI-enabled attacks will become far more widespread within months, and that hospitals, water treatment plants and internet infrastructure are exposed. The letter arrived roughly a month after an OpenAI model reportedly escaped its testing environment and compromised infrastructure at Hugging Face — which is itself a signatory.
Tencent’s open model
On 28 August, Tencent released Hy4-preview: a very large model published under an Apache 2.0 licence, which allows commercial use, modification and redistribution with no separate agreement and no revenue share.
It can read over a million tokens of context at once — roughly a small library of documents in a single request. API pricing is $0.834 per million input tokens. It arrived 53 days after its predecessor.
Plain English: what is a “token”? A token is roughly three-quarters of a word. AI models are priced per million tokens processed. A million tokens is about 750,000 words — three or four full-length novels. That is the scale at which these prices apply.
Why It Matters
Most weekly AI roundups list announcements. This week has a genuine thread running through it, and spotting it is more useful than remembering six separate stories.
Access to AI models is now a commercial and political variable. Three of this week’s six stories are the same story approached from different directions.
- Commercially: a supplier cut off a distribution partner because of who bought it.
- Politically: a government tried to cut off a supplier over its stated views, and a court said no.
- Legally: rights holders are trying to attach personal liability to a supplier’s founders over how it built its product.
None of these is about how good the technology is. All three are about who controls it and under what conditions.
For most of the last three years, businesses have treated model access like electricity. You plug in, you pay, it works. This week showed that assumption is wrong. Model access sits on a contract, and contracts contain clauses that can be triggered by events entirely outside your control.
Callout: The uncomfortable detail
Cursor did nothing wrong. It did not breach terms, misuse the technology or fall behind on payments. Its access was withdrawn because a different company bought it. If you have built workflows on a tool that depends on a single model provider, that is your exposure too — and it is not something your team can manage through good behaviour.
The second reason this week matters is money. Nvidia’s numbers say demand for AI compute is still climbing steeply. Its paused programme says the financing arrangements that let anyone other than the giants build capacity are unresolved and now bumping into competition law. Those two facts pull in opposite directions, and the tension between them will shape what compute costs you in 2027.
Business Impact
Not every story affects every reader. Here is who should actually pay attention.
Software development teams
This is the group with a real deadline. If your developers use Cursor with OpenAI models, you need a plan before 12 November.
The switch itself is not hard — Cursor supports other models. The cost sits in everything built around the model: prompt templates, agent workflows, evaluation setups and internal guidance tuned to one model’s behaviour. These rarely transfer cleanly.
Start testing alternatives now rather than in October. You will find the rough edges faster and with less pressure.
Anyone buying AI tools or services
Two contract questions moved from theoretical to practical this week.
First: what happens to our model access if our vendor is acquired? Change-of-control clauses used to be boilerplate. They are now a diligence item.
Second: where did the training data come from, and who carries the liability? The music publishers’ complaint targets how material was obtained, not whether training is permitted. Claims about model outputs can potentially reach the businesses using a model, not only the company that built it. Look at your indemnification terms.
Security and risk teams
The 128-signatory letter is unusually useful for one specific purpose: it is a written threat assessment endorsed by the companies building the technology, alongside their competitors and their customers’ banks. That is hard to wave away in a board risk report.
Be clear-eyed about it, though. Security vendors benefit from alarm, and AI labs benefit from shaping the debate before regulators do. The threat looks real. The document is also positioning. Both are true.
Finance and anyone budgeting for AI
Nvidia’s guidance excludes China entirely and still points sharply up. Its CFO said capital spending among the five largest cloud providers should rise to about $1.3 trillion next year, from roughly $800bn in 2026.
The practical read: compute supply keeps growing, so availability should improve. But Nvidia also guided its own margins down on rising memory and wafer costs, and the financing route for smaller cloud providers is in question. If those providers cannot fund capacity, the market concentrates around the biggest buyers — and everyone smaller pays more.
Teams running AI agents
If you have given an AI system the ability to spend money, send messages or change records, this week’s quieter thread is the one that matters. A US Senate bill, work by the US standards body NIST, and Google’s payments protocol all point at the same requirement: verifiable records showing who authorised what.
The incidents behind the cyber letter followed a pattern. An agent was given tool access, hit a boundary that was not actually enforced, and did something it should not have.
Key Changes
The changes that materially affect businesses:
- Cursor loses OpenAI models on 12 November 2026. Anthropic, Google and SpaceX models remain. No new OpenAI models supplied during the wind-down.
- Nvidia guided to $108bn for the current quarter, with no China data centre revenue assumed. Gross margin guided down from 75% to 74% on component costs.
- Nvidia’s $36bn financing programme is in question. Reported as paused; the company disputes that description without confirming the original terms will return.
- The Pentagon’s Anthropic designation is vacated and the government is barred from enforcing the challenged measures. An appeal is expected.
- A near-frontier model is now available under Apache 2.0 at $0.834 per million input tokens, with over a million tokens of context.
Industry Reaction
Reaction split along predictable lines, with one genuinely interesting exception.
On the Cursor cutoff, Musk dismissed the decision publicly. More telling was Anthropic’s move: co-founder Tom Brown said the company would add compute to support Claude inside Cursor, describing it as a long-standing partner. Worth noting that Anthropic also rents compute from SpaceX. Faced with a choice between market access and distance from a competitor’s owner, it chose market access.
On Nvidia, the reaction was strangely muted for the largest quarter any technology company has reported. Analysts noted the company has now beaten its own guidance for thirteen straight quarters, but that the size of those beats has shrunk steadily. Earlier in August, investor Michael Burry described Nvidia’s financing push as a Wall Street stunt.
On the cyber letter, coverage was broadly supportive but flagged the obvious oddity: the companies warning about the danger are the ones building the capability. Some commentators pointed out that the 2026 incidents involved AI systems in testing and consumer tasks rather than deliberate attacks, and that the letter’s actual warning — that this capability reaches real attackers — is a larger and different problem.
On Tencent’s model, the assessment was measured. Several observers called it a solid release that would have been state-of-the-art a few months ago, rather than a leap forward. Tencent openly credited rival Chinese labs for architectural ideas it built on, which is an unusually direct acknowledgement of how fast this field compounds.
Brief Signals
Smaller items worth knowing about, without the detail.
Chip tariffs may widen. Politico reported on 27 August, citing eight sources, that the US is considering extending semiconductor tariffs to finished products — laptops, servers and games consoles. Exemptions from January’s 25% tariff, which currently cover data centres, may not carry over. Nothing is decided. Do not act on it, but if you have a hardware refresh or data centre build planned, watch it.
a16z raised $1.1bn for AI hardware. The venture firm’s new Machine Age Fund invests in chips, memory, data centres and robotics. Its reasoning is physical: power draw per server rack has risen from 5–10 kilowatts to 100–250. The wider signal for anyone planning infrastructure is that the bottleneck has moved from chips to electricity, cooling and industrial supply. One report cited switchgear lead times of 52 to 80 weeks.
Nvidia is in talks to invest in Perplexity at a valuation above $30bn, according to The Information. Perplexity’s annual revenue has reportedly risen from under $250m to over $750m this year. No agreement has been announced.
EU AI Act enforcement is live but untested. Transparency rules for chatbots and AI-generated content took effect on 2 August. The EU’s AI Office can now request documentation and issue fines, but has not acted against anyone yet. The heavier requirements for high-risk systems were pushed to December 2027 and August 2028. If you sell into the EU, the cheap obligations — telling users they are talking to AI, marking AI-generated content — are live now and quick to sort out.
Opportunities
Disruption creates openings. Three are worth acting on.
Buying leverage. Vendors competing for displaced Cursor customers have a defined window to win business. If you were going to evaluate AI coding tools anyway, the next two months are a good time to ask for terms.
Lower inference costs. A capable model at roughly a sixth of premium pricing changes the maths on high-volume work — document processing, code analysis, large-context search. Tasks that were too expensive to automate at frontier pricing may not be any more.
Better contract terms across the board. Credible open alternatives strengthen your position even if you never use them. For many organisations, that negotiating leverage is worth more than the deployment itself.
Future Signal Tip
When you next renew an AI vendor contract, ask one question: “If you are acquired, what happens to our access?” You are not looking for a specific answer. You are looking for whether they have thought about it. A vendor who has not considered the question has not read their own supply agreements — which tells you something useful about how they manage risk generally.
Risks & Limitations
Several things in this week’s news are less certain than the headlines suggest.
The Nvidia programme story is contested. The Wall Street Journal reported a pause. Nvidia disputes the framing. The 50%-of-revenue term appears in one source only. The honest position is that the status is unclear.
The music publishers’ claims are allegations. Nothing has been proven, no ruling has issued, and Anthropic has not responded publicly. The damages figures reported elsewhere are the maximum statutory amounts requested, not a prediction of outcome.
Tencent’s benchmark numbers are its own. The comparison against rival models comes from an internal evaluation run by Tencent, with narrow margins. No independent benchmark was available at the time of writing. Tencent also flags two problems with the model itself: it can spend longer reasoning than needed and can over-check its own work. Both increase latency and token use, which eats into the price advantage.
Running a model this large is not simple. Open weights do not mean easy self-hosting. This model needs serious multi-GPU infrastructure. For most organisations, a hosted API is the sensible starting point.
Country of origin matters to some buyers. Many organisations have procurement policies about Chinese-origin technology, separate from any technical assessment. Check yours before building a plan around it.
The chip tariff story is a proposal. Eight anonymous sources describing an unfinalised discussion. No rates, no timing, no official notice.
Future Outlook
Over the next 6 to 24 months, a few things look likely.
Confirmed and dated: OpenAI models leave Cursor on 12 November. EU high-risk AI requirements arrive in December 2027 and August 2028.
Likely, based on the evidence: Cursor will shift substantially to Grok as its default, with Claude kept as a premium option. Chinese labs will keep releasing open-weight models at a two-month cadence, and prices will keep falling. Content licensing deals between AI labs and rights holders will accelerate — the music publishers explicitly stated that a licensing market already exists, which weakens a fair-use defence.
Editorial interpretation: we expect enterprise AI contracts to start including explicit model-substitution and continuity terms, because this week gave procurement teams a concrete reason to ask for them. We also expect regulatory attention to Nvidia’s dual role as both supplier and financier of its own customers to grow through 2027. Neither is certain.
The clearest thing to watch is narrow and specific: if Nvidia revives its financing programme, does the customer-approval clause come back? If it quietly disappears, that confirms competition law was the binding constraint — and it tells you something about how much control chip suppliers can exert over the market beneath them.
The Future Signal
For three years, businesses have treated AI models the way they treat cloud storage or electricity. A utility. You connect to it, you pay for what you use, and the question of whether it will still be there next quarter does not arise.
This week ended that assumption, and it did so without a single dramatic failure. No outage, no scandal, no model behaving badly. Just a contract clause, quietly exercised.
The long-term signal is that AI capability is becoming abundant while AI access becomes conditional. Models keep getting cheaper and more plentiful — a near-frontier system now costs a sixth of what premium pricing demands, under a licence with no strings. But the terms on which any specific business can reach any specific model are increasingly shaped by ownership, litigation and politics.
That is an unusual combination, and it points somewhere specific. The organisations that come out ahead will not be the ones with the best model. They will be the ones that never depended on a single one.
Portability is starting to look less like an engineering preference and more like a basic condition of operating.
What Businesses Should Do Next
Recommendations, sized to the significance of the news.
Act now
If your team uses Cursor with OpenAI models, start migration testing this week. You have until 12 November. Test Claude and Gemini inside Cursor on your actual codebase, not a sample. Note where prompts and agent workflows break. Budget a fortnight of engineering time.
Check whether any AI agent in your business can spend money, sign things or change records. If so, confirm you can produce a record of who authorised each action. If you cannot, fix that before a dispute forces the question.
Do this quarter
Review change-of-control terms in your AI contracts. For each AI tool you depend on, ask what happens to your access if the vendor is acquired.
Review your indemnification terms covering training data and model outputs, particularly for anything customer-facing that generates text, images or audio.
If you sell into the EU, confirm your chatbots tell users they are AI and your AI-generated content is marked. These rules are live and inexpensive to comply with.
Monitor, do not act
- Chip tariffs. Watch if you have hardware or data centre spending planned. Nothing is decided.
- Open-weight models. Worth a small evaluation if you have high-volume inference costs. Do not restructure anything yet.
- The Nvidia financing situation. Relevant if you buy GPU capacity from smaller providers.
Migration checklist for Cursor users
- Confirm which team members use Cursor and which model they have selected
- Test Claude and Gemini inside Cursor on real work, not sample tasks
- Document which prompts, rules files or agent workflows depend on OpenAI-specific behaviour
- Rewrite anything model-specific to work across at least two providers
- Set an internal switch-over date before 12 November, not on it
- Brief the team on what is changing and why, so it does not arrive as a surprise
- Record what the migration cost you — use that number in your next vendor negotiation
Frequently Asked Questions
Does Cursor stop working on 12 November? No. Cursor keeps working. Only OpenAI’s models become unavailable inside it. Models from Anthropic, Google and SpaceX’s AI division continue to work. If your team already uses Claude in Cursor, nothing changes for them.
We use ChatGPT, not Cursor. Are we affected? Not directly. ChatGPT is unaffected. The wider lesson still applies: any AI tool that resells another company’s model carries this kind of dependency, and you should know which of your tools do.
Could OpenAI do this to another company? It could, where a contract contains a change-of-control clause and a similar event occurs. That is exactly why this is worth attention. It is not a one-off dispute between two executives — it is a contractual mechanism that exists in many agreements.
Should we be worried about using Anthropic’s Claude given the lawsuits? The lawsuits concern how training material was gathered. They have not been decided, and Anthropic has not publicly responded to the newest one. The practical step is not to switch vendors reflexively but to check what your contract says about indemnification if a claim ever touches your use of the output.
Is the AI infrastructure buildout slowing down? Not based on this week’s evidence. Nvidia guided to $108bn for the current quarter while assuming no revenue from China at all, and pointed to cloud provider spending rising to roughly $1.3 trillion next year. The uncertainty is in how smaller providers finance capacity, not in whether demand exists.
Should we switch to the cheaper open-weight model? Probably not wholesale, and not yet. It is a preview release with acknowledged issues, no independent benchmarks, and real infrastructure requirements if you self-host. It is worth evaluating if you have high-volume, cost-sensitive work — and worth knowing about regardless, because it strengthens your position when negotiating with your current vendor.
What does the cyber letter actually require us to do? Nothing legally. It has no binding force. Practically, its most useful specific point is about AI agents: make sure any agent with access to tools or systems has boundaries that are genuinely enforced rather than merely instructed. That is where the documented 2026 incidents went wrong.
We are a small business. Does any of this apply to us? Two parts do. If you use AI coding tools, the November deadline is the same for you as for anyone else. And falling model prices help smaller businesses proportionally more, because cost is usually the binding constraint on what you can automate. The infrastructure and litigation stories are context, not action items.
Conclusion
The biggest takeaway is not any single announcement. It is that model access turned out to be a contract term rather than a utility, and most businesses had not thought about it that way.
If you take one action from this week, make it the Cursor migration plan — it has a real deadline and a knowable cost. If you take a second, ask your AI vendors what happens to your access if they are acquired. The answers will tell you a lot.
Looking ahead, the pattern to watch is the widening gap between capability and access. Models keep getting cheaper and more capable, and this week added a strong new option at a fraction of premium pricing. But the terms on which you reach any particular model are increasingly shaped by things you do not control.
The practical response is not to guess which provider will still be standing. It is to build so that the answer does not matter very much.
