As reported by Bloomberg.
Bloomberg reports SpaceX approached Cognition about a potential acquisition that never closed — the second attempted move into the agentic-coding layer in one quarter, after the completed ~$60B Cursor deal, and a clearer signal of strategic intent than any single transaction.
What Happened
According to Bloomberg, SpaceX approached Cognition — the AI startup behind Devin, an agent built to automate software-engineering work — about a potential acquisition. That approach matters, but the qualifier matters just as much: the talks are not currently active. No deal closed, no deal is imminent, and this should not be read as SpaceX acquiring Cognition. What the two companies are still discussing is a working relationship, potentially including Cognition using SpaceX’s computing capacity.
The acquisition attempt would have been SpaceX’s second large AI takeover in a matter of months. The first — the roughly $60 billion purchase of Cursor (Anysphere), the developer environment where millions of engineers write code — closed in June, roughly eight weeks after a compute partnership between the two companies began in April. Cognition was valued at approximately $26 billion in a closed May funding round, and is in early talks to raise new financing at a target valuation of at least $40 billion — a figure that represents a negotiating benchmark in ongoing discussions, not a set price, and one that could move in either direction.
Devin is an autonomous coding agent: software designed to take on engineering tasks end-to-end, operating in an automated loop rather than responding to prompts in a chat window. It is an ambitious product with real limitations, and it is best understood as a tool that automates specific software-engineering tasks — not as a replacement for engineering teams. What it represents structurally is a distinct tier above the developer harness: the layer where AI executes work rather than assisting with it.
The key insight: The story is not that a deal failed to close. It is that SpaceX tried, for the second time in one quarter, to acquire the layer where AI does software work — and that this time, the target had the leverage to say not yet. Two moves in one quarter, one completed and one rebuffed, make the strategic intent legible in a way that a single transaction never could.
The Structural Read
The approach makes structural sense the moment you stop reading it as a transaction and start reading it as a layer-by-layer integration. SpaceX, with xAI’s Grok folded in, already owns a frontier model. Through Cursor, it owns the harness — the environment where developers spend their working hours. Adding Cognition’s Devin would have added the autonomous-agent tier on top: the software that executes the task rather than assisting with it. And, critically, the stream of coding traces that agent generates.
That last point is not incidental. Grok’s most recent model was trained in part on Cursor workflow data — a dynamic that made the Cursor acquisition as much about data as about product. Every coding surface SpaceX owns is also a training-data pipeline for its model. Devin’s traces — the full decision sequences of an autonomous agent working through real engineering problems — would have been among the most valuable fine-tuning signals available anywhere. The roll-up is about data architecture as much as product ownership. (See the Grok 4.6 post-training analysis for the mechanics.)
Map of AI — The Coding-Agent Stack
Model → Harness → Agent → Data: SpaceX is assembling a vertical
Grok is the model layer. Cursor is the harness layer — the environment developers live in. Devin (attempted) is the agent layer — where software executes work autonomously. Each surface also feeds training data back to the model. The integration logic is vertical: own the stack, own the signal. The Harness War essay on Business Engineer maps this in full — the harness-war thesis is not a metaphor, it is the deal thesis.
Compute as the lever. The instrument SpaceX keeps reaching for is compute access. Even without an acquisition, it is offering Cognition the use of its computing capacity — which is exactly how the Cursor relationship started. A compute partnership in April became a $60 billion acquisition by June. That precedent is a real pattern, not a prediction: Cognition’s path may look nothing like Cursor’s. But the mechanism is worth naming clearly. Offering scarce compute to a fast-growing AI startup is how an integrator pulls a company into its orbit, deal or no deal. Compute is the scarce resource, and controlling it is a gravitational force.
Independence has value — and Cognition has it. The stalled talks point at the other half of the lesson. Cursor, before the acquisition, was a harness without a model: it owned the environment but depended on third-party frontier models to make it useful. That dependency created the conditions for absorption. Cognition, raising at a target valuation of at least $40 billion and growing fast, is not in that position. It has the capital, the leverage, and the standing to take the compute and keep the company. Not every harness gets absorbed. A target with options is a structurally different negotiation than one without them — and as long as capital continues flowing to the best agentic-coding startups, some of them can maintain that position.
Three Implications
IMPLICATION 1 — THE AGENTIC-CODING LAYER IS THE CONTESTED GROUND
SpaceX’s two moves in one quarter define where the competitive pressure is concentrating: not at the model layer (Grok is already there) and not at the IDE-harness layer (Cursor is acquired), but at the autonomous-agent tier — the software that executes engineering work end-to-end. Every major model owner will face the same logic. Expect more approaches, more compute offers, and more consolidation attempts in this layer over the next 12 months.
IMPLICATION 2 — COMPUTE PARTNERSHIPS ARE NOW DUE-DILIGENCE SIGNALS
The Cursor sequence — compute deal to acquisition in eight weeks — means that any compute partnership between a frontier-model owner and an agentic-coding startup now carries a different reading than it did a year ago. It is not just infrastructure sharing; it is a structural on-ramp. Investors, founders, and boards in this layer should price that dynamic into any compute arrangement they accept. The terms of the partnership shape the terms of any eventual negotiation.
IMPLICATION 3 — VALUATION IS A DEFENSE, NOT JUST A METRIC
Cognition’s ability to decline — or at least defer — an approach from SpaceX is a function of its capital position and growth trajectory, not its product maturity alone. The $40B+ target valuation in early talks is, among other things, a negotiating posture. For the agentic-coding startups that follow: the fastest path to maintaining independence may be the most aggressive fundraise, not the most cautious one. Capital is leverage when the buyer is also the infrastructure provider.
The Bottom Line
SpaceX did not acquire Cognition. What it did was make its strategic logic visible for the second time in one quarter: own the model, own the harness, own the autonomous-agent layer on top, and use compute access as the instrument that pulls each piece into orbit. The completed Cursor deal shows what absorption looks like when a target has a structural dependency and no leverage. The stalled Cognition approach shows what it looks like when a target has neither. The roll-up intent is real; whether the agentic-coding layer actually consolidates under a single owner is the question this non-deal leaves exactly where it found it — open, and worth watching closely.
Sources:
Bloomberg — SpaceX Attempted to Acquire AI Coding Startup Cognition ·
FourWeekMBA — SpaceX / Cursor Merger Close & Grok 4.6 Data Absorption ·
FourWeekMBA — xAI Grok 4.6 Post-Training Turn ·
Business Engineer — Grok Bot and the Fourth Moment of AI ·
Business Engineer — Beyond NVIDIA’s Moat ·
Business
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.More Resources
