Cursor Loses OpenAI in 75 Days: Who Owns AI Wins

If this landed in Promotions or Other, move it to your main inbox. One move keeps it there.

So, What Actually Happened?

Sunday, and the thing I cannot put down is a date: 12 November. That is the day Cursor loses access to OpenAI's models, not because the product broke but because SpaceX bought the company. We scanned 190,000 articles this week so you don't have to. In the same 48 hours, Sony and Warner sued Anthropic over training data while quietly backing a different AI company with equity instead. And a16z closed $1.1 billion for the physical buildout, steel and power rather than software. I kept scrolling for the story where a model got better and somebody won on quality. I did not find it. I found three stories about who is allowed to use what, and one engineer somewhere being told to rebuild a system that was already working.

The Bottom Line: None of this week's money moved on how good the technology was. It moved on who owns the thing underneath it and what the contract says.

The Tracks That Matter

1. Your AI Vendor Can Switch You Off In Seventy-Five Days

OpenAI told Cursor its model access ends on 12 November, the maximum notice the contract allowed, after SpaceX closed the acquisition it had agreed back in June. Nothing about the product changed. The editor works this morning and it will work in October. What changed is the cap table. Cursor had built a business reselling third-party models alongside its own, and future OpenAI releases are now off the table too, which leaves it leaning on Colossus and whatever xAI ships next. Musk answered by calling OpenAI's founders untrustworthy, which is the part everyone quoted and the least useful part of the story. The original deal carried a $60 billion purchase option and a $10 billion break-up fee. That is how much of the value sat in access rather than in code.

Here's what works: Open your three largest AI contracts and find the termination clause. If a change of control at your vendor ends your access, name a fallback model before renewal.

2. The Labels Sued One AI Company And Funded Another

Sony and Warner sued Anthropic over mass copyright theft this week. The same two labels, joined by Universal, had spent the days before that taking a position in Stability AI rather than taking it to court. Read those as one move instead of two and it is a licensing strategy with a stick attached: back the companies willing to pay for the catalogue, litigate the ones that helped themselves. The music business has run this play before, against Napster and then with Spotify, and it took roughly a decade to work out that owning a slice of the distributor beats winning a judgment against it. This is the compressed version. Meanwhile almost 40% of music released globally in July had AI somewhere in it, which is the number making everyone negotiate faster.

Here's what works: If you train or fine-tune on licensed material, write down the provenance list this quarter. Settlements are being priced on what you can prove you paid for.

3. a16z Puts $1.1 Billion Into Steel, Not Software

a16z closed a $1.1 billion Machine Age fund pointed at the physical side of AI: hardware, robotics, manufacturing, the parts that show up on a truck. The firm frames it as a supply crunch rather than a software opportunity, which is a striking thing to say out loud for a house that made its name on software margins. The chip calendar backs them up. TSMC's A14 process is not in volume production until 2028, with pre-production a year ahead of that, so the capacity that constrains you in 2027 has already been allocated to somebody. Venture money has worked out that the queue is the product.

Here's what works: Take your 2027 AI capacity plan and mark every line that depends on hardware arriving. Those are your real dates. Software lines move, trucks do not.

Quick hits:

• ServiceNow's AI business crossed a billion dollars in contracted value. Its AI annual contract value passed the $1 billion mark with agentic deployments up ninefold in nine months, the first clean sign that agent spending is landing in signed renewals rather than pilots.
• Schools are buying AI faster than they can measure it. District contracts are going out with no way to show the tools help, a procurement gap that will read badly the first time a state auditor pulls one.
• Altman says cyber defence is at a critical moment. He warned there is not much time to act on AI-driven attacks and invited competitors to join him, notable mostly because it cuts against his own commercial interest.

Signal vs. Noise

🟢 Signal: Agent spending showing up in signed contracts. ServiceNow's AI contract value crossed a billion dollars with agentic deployments up ninefold in nine months, and its renewal rate ticked up to 98%. That is the number that matters: buyers renewed. Most coverage is still counting product launches and demos instead of watching what survives a renewal conversation.

🔴 Noise: ”Agentic AI” as a category label. It pulled one of the heaviest volumes of the day while steadily losing its hold on the stories actually moving underneath it. Box's CTO says the architecture beneath the label now has a half-life measured in months. When the word stays fixed and the thing keeps changing shape, the word stops carrying information.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

OpenAI cut off a customer over who bought it, Sony and Warner sued one AI company while funding another, and a16z raised $1.1 billion for hardware instead of software, all inside two days.

Each one goes to a different desk. The developer-tools press writes the Cursor cutoff as a Musk feud. The entertainment lawyers write the label suits as a copyright story. The venture press writes the a16z fund as a portfolio shift. Put them on the same morning and the shared subject is permission. Model access, catalogue rights, physical capacity: three different assets, all of them things you rent from somebody who can stop renting to you. Nobody in this set of stories lost because their technology was worse. They lost, or paid, or hedged, because of a relationship they did not control.

What changes Monday is small and unglamorous. Take your top three AI dependencies and write one line each on what would actually happen if the counterparty walked away in ninety days. Not whether they would. Whether you would still be operating on day ninety-one.

By The Numbers

• SpaceX's Cursor deal carried a $60 billion purchase option and a $10 billion break-up fee — the price of a code editor once you count the model access attached to it.
• ServiceNow's AI annual contract value passed $1 billion with agentic deployments up ninefold — subscription revenue hit $3.877 billion in the quarter and renewals rose to 98%, so this is contracted, not projected.
• a16z raised $1.1 billion for the physical buildout of AI — a software-native firm putting a fund-sized bet on things made of metal.
• Almost 40% of music released globally in July 2026 involved AI — which explains why the labels stopped debating the principle and started pricing the deal.
• TSMC's A14 will not reach volume production until 2028 — 10-15% more speed at the same power, and a two-year wait before anyone can buy it at scale.
• See what's rising across AI and data this week →

Deep Dive: The Track You Licensed, Not The Track You Own

I have around four thousand records in the house. Every one of them, I can play tonight, at 3am, in a field, on a boat, without asking anyone. Then streaming arrived and I built playlists instead, and a few years in I went looking for a track I had used in a set for years. Gone. Pulled from the platform in a rights dispute I was never party to. I had not lost the music. I had lost the permission, and it turned out I had never held anything else.

The set list is not yours

Cursor did not get worse this week. It got orphaned. A supplier decided that the identity of its new owner was reason enough to end model access, with the maximum contractual notice offered as courtesy. The product is unchanged. The permission is gone on a date somebody else picked.

The labels learned this twenty years ago

Sony and Warner are not confused about what they own. They sued one AI company and took a stake in another in the same week, because they worked out during Napster that the fight is never about the file. It is about who gets to license it, and at what point you would rather own part of the distributor than beat it in court.

Which is why the money moved to concrete

a16z reading the moment as a supply crunch is the same insight wearing work boots. Software can be reissued. A fab slot in 2028 cannot. When permission is the scarce thing, capital walks toward the assets nobody can revoke by email.

What Actually Works

1. Map your AI stack by who can revoke it: model access, data rights, compute capacity. Not by vendor logo, by who holds the off switch.
2. Get a change-of-control clause into every AI contract: if your supplier is acquired, you want notice measured in quarters, not the seventy-five days Cursor got.
3. Write your data provenance down before someone asks: the labels are pricing settlements on documentation, and yours is either written or it is a guess.
4. Put dates on the physical layer: hardware, power, fab capacity. Those constraints do not negotiate, and they are already committed two years out.

A record you are not allowed to play is a frisbee with nice artwork.

What's Coming

Change-Of-Control Clauses Become Standard

The Cursor cutoff is the reference case every procurement team needed. Expect AI contracts signed this autumn to carry acquisition-triggered notice terms as a default line item, and expect the first vendors to offer them to win deals on it.

The Label Playbook Spreads Beyond Music

Sony and Warner running equity and litigation in parallel is a template, not a one-off. Publishers, stock image houses and broadcast archives are all sitting on the same asset with the same problem. Watch for the first non-music rights holder to take a stake instead of filing.

Physical Capacity Becomes The Funding Story

With a16z pointing $1.1 billion at hardware, other generalist funds will follow rather than explain why they did not. The interesting consequence is on the other side: application startups will find late-stage capital harder to raise while the money is busy buying the floor beneath them.

For Your Team

Monday's meeting prompt: ”Name our three biggest AI dependencies. For each one, if the supplier were acquired tomorrow by someone who did not want us as a customer, how many days of notice does our contract give us, and what do we switch to? If we cannot answer that in the room, who owns finding out by Friday?”

Share-worthy stat: A code editor came with a $60 billion purchase option and a $10 billion break-up fee, and its model access still got cancelled with seventy-five days' notice because a different company bought it. The technology worked the whole time.

Go deeper: Track where AI money and access are actually moving →

The Track of the Day

”Wait, what? It's working.”
An engineer at Box, on being told to rebuild a system that had just shipped

Ben Kus told that story to make a point about how fast agent infrastructure turns over. It reads differently this week. Half the teams who woke up to the Cursor news had exactly that thought, about something they did not build and cannot fix.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 30, 2026 | Curated by Yves Mulkers @ Ins7ghts