On Tuesday June 16, 2026, at the inaugural Compile conference held at Fort Mason in San Francisco, Cursor announced four things in one keynote. A 1.5-trillion-parameter foundation model, trained from scratch on more than 100,000 GPUs on xAI's Colossus supercluster. A GitHub-rival Git platform called Origin, demonstrated at 22.6 commits per second with an AI-powered conflict-resolution engine. A Cursor Mobile iOS beta for steering remote agents from a phone. And — confirmed by SpaceX's parallel announcement — a $60 billion all-stock acquisition of Cursor's parent company, Anysphere, days after SpaceX's own IPO. Closing is expected in Q3 2026, pending regulatory approval.

Eight months ago, Cursor was a $9.9 billion AI-powered code editor running on other people's models. In June 2026 it is a $60 billion vertically integrated software stack with its own foundation model, its own Git layer, its own mobile control plane, and a rocket-and-satellites company as its parent. The arc the Loop has been writing for the last fortnight is harness-not-the-model. Cursor's answer is to own the harness and every layer underneath it.

Four products, one week, one buyer

The day's announcements pull in opposite directions until you stack them.

The model. Co-founder and CEO Michael Truell described a model "matched in scale" with Anthropic's Claude Opus and OpenAI's GPT-5.5 — 1.5 trillion parameters, no open-source base, ten to twenty times more compute than the Composer 2.5 model Cursor shipped in May. Composer was an iteration on Kimi; this one is a clean training run. Training is underway. Ships within weeks. Built explicitly to do more than code.

Origin. A clean-sheet Git forge built on top of the Graphite acquisition Cursor closed in December 2025. The pitch: GitHub was designed for one human, one branch, one pull request, one merge at human cadence. Origin is designed for dozens of agents on the same repository, doing the cloning, branching, conflict resolution, and CI fixes at machine cadence. The demo hit 22.6 commits/second on a single repo. Sub-400 ms global synchronisation latency. 296,000 clones per hour. Origin goes GA in fall 2026; the waitlist is open at cursor.com.

Cursor Mobile. iOS beta. Lets a developer review, approve, and steer remote agents from a phone, including a "remote control" mode for locally running agents. The deliverable is not yet another mobile coding IDE — it is a glass surface on top of the agents Cursor's customers are already running.

The acquirer. SpaceX, in an all-stock deal valuing Anysphere at $60 billion. The structure carries forward an April 2026 partnership option — buy outright at $60 billion, or pay $10 billion for the collaboration. SpaceX chose to buy. The merger ties Cursor to xAI's Colossus supercluster in Memphis — which is where the new model is being trained — to SpaceX's IPO-fresh balance sheet, and to a parent company whose other businesses are rockets, satellites, and a frontier AI lab.

The cleanest summary is the one explainx.ai's writeup puts in the lead: in one keynote, Cursor reframed itself from an editor on top of other people's stacks into a stack of its own — running on its acquirer's compute.

The harness, inverted

The Loop's last fortnight has been about the model layer ceasing to be where the customer chooses — Samsung running ChatGPT and Claude and Gemini side by side on a Samsung SDS-mediated harness; OpenAI's $150M Partner Network trying to lock the implementation harness with channel money; Cognition's Devin at $26 billion for orchestrating models it does not own. The thesis is that the harness is now the more valuable layer because the model is a commodity behind a multi-vendor API.

Cursor's June 16 announcement is the inverse argument. Customers are not buying just the harness — they are also paying for what is in the harness, and the harness is in turn pinned to a Git layer, a CI loop, and a chat surface. If a single vendor owns all of those layers, the customer's switching cost is total. Origin owns the source-of-truth. The new foundation model owns the inference. Cursor Mobile owns the control plane. The IDE is now a thin presentation on top of three layers Cursor controls.

This is the Microsoft-1995 playbook with two adjustments. The first is the speed: a stack that took Microsoft a decade to assemble around Windows, Office, and Visual Studio, Cursor is shipping in four quarters. The second is the parent. Microsoft built its stack with its own retained earnings. Cursor is building it on SpaceX's all-stock acquisition value and on xAI's compute.

What Origin actually changes

The argument inside Origin's marketing is that GitHub's twenty-year design assumes the human is the bottleneck. Pull requests sit unreviewed for hours because the reviewer is asleep. CI runs ten minutes because the test suite was sized for the rate at which a person produces commits. Merge conflicts wait for one developer to rebase by hand.

If the developer producing commits is fifty agents working on different parts of the same codebase at the same time, every one of those assumptions becomes a load-bearing bottleneck. Cursor's own community forum thread on the keynote is full of engineers parsing the 22.6-commits/second number as a deliberate provocation: at that rate, a single Origin repository can absorb commits faster than the GitHub Enterprise instance backing a Fortune-500 product.

The harder question — and the one aiinsiders.net flagged — is what happens to the code once it is in Origin. When one company owns the editor that writes the code, the forge that stores it, and the model that will be trained on the next generation of it, the distinction between hosting and training becomes a terms-of-service question, not a technical one. Origin's pricing, data-retention, and export policies are not yet published. They are the policies the buyer needs to read more carefully than the performance metrics.

The acquirer math

SpaceX is paying $60 billion of its own stock for a company that closed its Series D at $29.3 billion in November 2025 — a doubling in seven months. The Series D was led by Accel and Coatue. The October 2023 seed was led by the OpenAI Startup Fund. The arc of who has financed Anysphere's growth is itself the story: from OpenAI's investment vehicle, through Thrive Capital's growth round, to an all-stock merger into Sam Altman's largest model-layer competitor.

Cursor's revenue trajectory makes the multiple at least defensible: $100 million in annualised revenue in early 2025, $2.6 billion at the time of the deal — a 26x in eighteen months, per Techzine's reporting. The Q1 2026 enterprise customer book reportedly covers more than half the Fortune 500. The valuation is one a normal acquirer would haggle. The acquirer here is a $2-trillion newly public rocket conglomerate paying in its own stock, not cash. That is a sentence that costs $60 billion.

The $60-billion number is also a number chosen for vibes — it lines up so precisely with the April 2026 option price that nobody on either side appears to have re-negotiated when Cursor's annualised revenue tripled in the intervening two months. Either April was too high or June is too low. Whichever it is, both sides got the result they signed for at the option strike.

What this is

The clean reading is that the AI dev-tools layer has consolidated into three vertical stacks: Microsoft (GitHub + VS Code + Copilot + Azure), Anthropic (Claude + Claude Code + the model-as-platform play of Claude Tag), and now Cursor (Origin + foundation model + IDE + Mobile) — with Cursor sitting under SpaceX/xAI as the most aggressively vertical of the three. Cognition's Devin is the orchestration outlier; Google's Gemini Code Assist is the search-company outlier. The middle of the table is now stacked.

The dirtier reading is that Origin and the new model are bargaining chips, not products. Both ship in the future tense — waitlist open, ships within weeks, fall 2026 GA. The acquisition closes Q3 2026. By the time Origin ships, the deal is done. By the time the new model ships, the customer base has been re-papered onto SpaceX terms.

Both readings are simultaneously true. They will resolve in the order the regulators allow.

What to watch

1. The model's actual benchmark. Truell named Opus and GPT-5.5 in the same breath. If the new Cursor model ships at parity with either on coding evals — SWE-Bench Verified, LiveCodeBench, the internal Cursor agent harness — the company has the technical claim it needs. If it ships materially below either, the model story collapses back to a procurement asset for SpaceX, and the harness-is-the-product thesis returns intact.

2. Origin's data policy. Pricing matters less than what the terms of service say about whether code stored in Origin can be used to train future Cursor or xAI models. If the answer is opt-out-by-default, the enterprise contracts will follow. If it is opt-in-by-default, GitHub keeps the Fortune 500.

3. The regulatory pathway. A $60-billion all-stock deal between a $2-trillion newly public rocket-and-AI conglomerate and the AI dev-tools market leader is not a transaction the FTC will wave through on a thirty-day clock. The stated closing target is Q3 2026. Whether the deal slips into 2027, and what concessions get attached at closing, is the read on how serious the antitrust review gets.

4. Anthropic and OpenAI's response. Both have been the named default model behind Cursor in different periods. Both are competitors of xAI. Watch for an Anthropic-with-Cursor-rival or an OpenAI-with-Cursor-rival announcement inside ninety days — Codex Studio, Claude Code IDE features, a Replit deal, a Vercel acquisition. If neither comes, the market has accepted that Cursor's stack is unassailable and the competition has moved to the seam between Microsoft and Cursor.

The week Cursor went vertical, every conversation about whose model wins the developer tier got a new answer: nobody's, because the developer tier is now a stack, and the stack is owned. The vendors that thought the model layer would always be where the customer chose are about to find out whether the customers agree.