The $60B Deal That Ended AI Neutrality

Welcome to Memorandum Deep Dives. In this series, we go beyond the headlines to examine the decisions shaping our digital future. 🗞️

This week, we're unpacking the four days in June that reshaped the AI industry. It started with the largest IPO in history and ended with the largest acquisition of a venture-backed startup ever, a $60B all-stock deal that stunned even seasoned deal watchers. The buyer was a rocket company. The target made software for writing code. On the surface, the two had almost nothing in common.

The company being acquired was, by most measures, thriving. It was growing faster than any business software company on record, its product sat on developer machines across the Fortune 500, and investors were lining up to fund it at ever-higher valuations. Companies like that don't usually sell. They go public, they raise another round, they keep compounding. This one signed away its independence in the same week its acquirer rang the opening bell.

Understanding why requires looking past the headline numbers and into the machinery underneath: who supplies whom, who competes with whom, and who can turn off whose access with 90 days' notice. Because the real story of this deal isn't about one company's exit. It's about a structural problem that every AI application company is quietly living with, and what happens when one of them stops pretending otherwise.

The most valuable surface in AI

For much of the past three years, the AI industry has measured competition by the quality of its frontier models, with each new release judged by benchmark scores and reasoning ability. Yet as those models have become increasingly capable, the industry's center of gravity has begun shifting away from the models themselves and toward the deployment layer, the applications through which AI reaches the people who use it. The reason is simple: a frontier model without distribution remains little more than a research achievement, while the product that sits between the model and the customer ultimately determines which AI millions of professionals rely on every day. Nowhere has that shift become more apparent than in coding, the first market where generative AI evolved into a meaningful software business, and where Anthropic, OpenAI, and Google have spent the past 18 months discovering that owning the intelligence matters less than owning the surface where that intelligence meets a paying customer. One company came to embody that shift more clearly than any other, despite never training a frontier model of its own.

For most of that period, the most valuable surface was owned by a company that made no models. Cursor, built by the startup Anysphere, is an AI coding tool built on a modified version of Microsoft's VS Code (a free, widely used program for writing software), letting developers generate, edit, and review code via plain-language instructions. Its pitch was neutrality: developers could route their work to whichever model performed best, Anthropic's Claude, OpenAI's GPT, or eventually Cursor's own, and the tool competed on experience rather than allegiance. That pitch carried the company from $500M in annualized revenue in June 2025 to $1B by November and past $2B by early 2026, a trajectory enterprise software had never produced before. Then, in the space of four days this June, the neutral layer stopped being neutral. The catalyst was not a competitive product launch or a pricing war, but one of the largest corporate transactions the technology industry has ever seen.

A record IPO, then a record acquisition

On June 12, 2026, SpaceX completed the largest initial public offering in history, raising $75B at a valuation of $1.77T and closing its first day of trading with a market capitalization above $2T. On June 16, it announced it would spend some of that new currency to acquire Anysphere in an all-stock deal valuing the company at $60B, the largest acquisition of a venture-backed startup ever, and roughly 3.4% dilution at the IPO price. The market approved, sending SpaceX shares up about 16% on the day.

Although the acquisition looked impulsive, SpaceX had spent months arranging it. The company had merged Elon Musk's AI venture xAI into itself in February, pitched IPO investors on an addressable market of roughly $28T, with its AI division as the centerpiece of the offering, and disclosed in April that it held an option to buy Anysphere outright for $ 60B. If it walked away, it would owe roughly $10B, a $1.5B termination fee, plus $8.5B in computing resources. The acquisition was engineered before the offering was priced, meaning the biggest IPO in history also doubled as the financing round for the biggest startup takeover. The scale of the transaction explained how the acquisition happened, but not why a company growing faster than almost any software business before it agreed to sell in the first place.

What the deal reveals about the squeeze

The more interesting question is why a company growing that fast was available at all, and the answer is that Cursor's independence had been failing for a year. Its business model required renting intelligence from the same labs it competed with, paying retail prices for Claude and GPT access, while Anthropic ran its rival, Claude Code, on wholesale internal economics.

TechCrunch's reporting found that Cursor achieved only slight gross-margin profitability after launching its own Composer model in late 2025, and that it still lost money on individual developer accounts. The market data told the same story from the outside. Corporate spending figures from Ramp showed Cursor's share of the AI coding category falling from 41% in June 2025 to about 26% by May 2026, while Anthropic came to control roughly half of it.

The suppliers were also willing to pull the plug. In January 2026, Anthropic blocked xAI staff from accessing its Claude models via Cursor, citing terms of service that prohibit building competing systems. The precedent was already established: Anthropic had cut off the coding startup Windsurf in 2025 when OpenAI moved to acquire it, and later that year, restricted OpenAI's own access. A company whose product depended on rivals' models, and whose rivals had demonstrated they would sever access when it suited them, had a shrinking set of futures: it could be squeezed slowly, or it could find an owner with its own compute and its own models, and it chose the owner. Yet solving one economic problem created another, one that struck at the very reason customers had embraced Cursor in the first place.

The contradiction at the center

Cursor's entire appeal for enterprises was that it was model-agnostic. It let them pick the best model for each task, keep sensitive code pointed at providers they had vetted, and avoid betting their engineering organization on any single lab. SpaceX has bought that trust and placed it within a conglomerate that competes directly with Anthropic, OpenAI, and Google. SpaceX intends to route Cursor's training onto its Colossus supercomputer, and its AI division has a documented record of safety failures that the company itself listed as a business risk in its IPO filings. Neither company has publicly committed to preserving multi-model access after the deal closes, and Anthropic and OpenAI now face a customer owned by a competitor, exactly the circumstance under which both have cut access before. That tension is not unique to Cursor but reflects how the AI industry has become intertwined across applications, models, and infrastructure.

The contradiction extends beyond one product, because the industry's structure makes neutrality unstable everywhere. Anthropic currently pays SpaceX $1.25B per month through 2029 to rent the entire Colossus 1 data center, and Google pays an additional $920M per month, meaning SpaceX's fiercest rivals are also its largest infrastructure tenants. Musk has said SpaceX retains the right to reclaim that capacity, and both leases carry termination clauses. No compute cutoff has happened, and none may ever happen, but the arrangement means every major lab now holds a contractual choke point over at least one competitor, whether it's model access or the hardware underneath it. Those overlapping dependencies raise a broader question that extends beyond any single acquisition.

Why does the system keep producing this outcome?

The forces holding this contradiction in place are structural. Training frontier AI models costs billions of dollars, so every lab is under pressure to own the products that customers use directly, rather than simply supplying the technology behind them. That puts the labs in direct competition with the very companies that rely on their models. American antitrust law, as Brookings has documented, offers little remedy when a platform refuses to do business with a competitor, even one with substantial market power.

Venture capital cannot solve the problem either. Cursor was reportedly in talks to raise $2B at a $50B valuation when SpaceX's option superseded the round, and one source suggested that even that sum would not have been enough to make the business sustainably profitable. When your suppliers are also your competitors, and there is little legal or financial protection against that, staying independent becomes much harder.

The playbook question

The acquisition now sits with regulators under the standard pre-merger review, and the break provisions suggest both sides took that risk seriously. What regulators decide matters well beyond this deal, because approval without conditions would validate a new playbook: merge the model maker into a capital-rich parent, use a public listing to create acquisition currency, buy the application layer, and retain contractual leverage over the models and compute that rivals depend on. Anthropic has filed confidentially for its own listing, and OpenAI is expected to follow, meaning the two companies with the strongest incentive to replicate this strategy will soon have public stock of their own. The remaining independent AI application companies, whether in design, writing, autonomous agents, or whatever comes after coding, are watching closely.

Over the coming months, Cursor's customers will find out whether the tool they chose for its neutrality continues to offer the same freedom of choice. The rest of the industry will discover something more consequential. At the beginning of the AI boom, the race was about building the smartest models. Increasingly, it is a race to own the products people use. Cursor's sale suggests that winning the application layer may be too important for the frontier labs to leave in independent hands. If that proves true, then the fastest-growing software company ever was not an exception. It was the first proof that, in AI, distribution has become as valuable as intelligence itself.