Scott Wu's $26B coding-agent company stayed independent, while the two sides kept discussing a compute deal that echoes SpaceX's Cursor playbook.
By Ryan Merket · Published
Primary source: Bloomberg Technology
Why it matters #
SpaceX is using compute as a route into AI applications, developer distribution and coding data. Cognition's independence now tests whether a $26B agent company can buy scale without surrendering control.
Scott Wu (@ScottWu46) built Cognition around the premise that software engineers would direct autonomous agents instead of manually implementing every task. SpaceX tried to buy that premise along with the company.
SpaceX approached Cognition about an acquisition, Bloomberg reported on August 19th, citing people familiar with the matter. The acquisition discussions are inactive. The companies continue to discuss working together, including an arrangement that could give Cognition access to SpaceX's computing capacity.
The approach came into view five days after SpaceX completed its acquisition of Cursor, the AI coding product built by Anysphere. Cursor announced the closing on August 14th, following an April compute partnership and a deal that valued Cursor at $60 billion. RuntimeWire first reported SpaceX's agreement to acquire Anysphere in June.
SpaceX's interest in a second major AI coding company suggests Cursor was an opening move. Coding products give SpaceX a direct route from its GPU infrastructure and Grok models into the daily work of developers. They also produce the technical feedback that model builders need: prompts, tool calls, accepted code, rejected code, test results and the long chains of actions behind completed engineering tasks.
Cognition would have added a second route into that workflow. Its core product, Devin, is designed to plan, write, test and ship code inside a customer's repositories and existing tools. Cognition also owns Windsurf's editor, intellectual property and commercial operation after buying the remaining business in July 2025.
SpaceX has already documented the playbook
The proposed Cognition compute arrangement matters because SpaceX's relationship with Cursor began the same way.
SpaceX and Cursor signed a compute agreement on April 19th, according to SpaceX's June prospectus. SpaceX agreed to provide GPU-cluster capacity for training and improving AI models. Cursor agreed to contribute personnel, datasets, technical knowledge, workflows, prompts, specifications and software code.
The companies simultaneously signed an option agreement that gave SpaceX a path to acquire Cursor. The prospectus said Cursor was also subject to exclusivity obligations. If SpaceX abandoned the option under specified circumstances, Cursor could collect a $1.5 billion termination fee and an $8.5 billion deferred-services fee. The acquisition consideration was tied to a $60 billion implied equity value and paid primarily in SpaceX shares.
That sequence turned scarce compute into a transaction tool. SpaceX offered the resource AI application companies struggle to secure at scale. In return, it gained contractual access to the people, data and operating knowledge produced by a fast-growing developer product, plus an option to buy the entire company.
Bloomberg's report does not establish that the proposed Cognition relationship included comparable data-sharing, exclusivity or acquisition rights. It does establish that SpaceX again discussed compute after acquisition talks. The order has changed, while the strategic pieces remain familiar.
SpaceX has been explicit about the objective. In its prospectus, it described the Cursor relationship as part of a plan to integrate compute infrastructure, models and applications. The filing said Cursor would strengthen SpaceX's position in AI-assisted developer productivity and help improve existing models, including Grok.
The company already controls a large physical AI stack through the xAI business it absorbed in the first quarter of 2026. SpaceX said its Colossus and Colossus II data centers collectively provided about 1 gigawatt of compute power when the prospectus was filed, with a further expansion planned. Cursor now gives that infrastructure an application used by professional developers. Cognition would have brought an autonomous agent, another editor and a separate enterprise customer base.
Cognition has enough capital to make independence credible
Wu, Steven Hao and Walden Yan founded Cognition in 2023. The three had competed at the International Olympiad in Informatics, and Cognition says its wider founding team holds 10 IOI gold medals. Wu had previously co-founded professional-networking service Lunchclub. Hao was an early engineer at Scale AI, while Yan worked on early versions of Cursor.
The founders were betting on long-running software agents before the category became a default product line for every major AI lab. When Wu introduced Devin in March 2024, he described a system that could operate a shell, editor and browser, plan across thousands of decisions and correct its own mistakes. The initial benchmark and product claims came from Cognition, and Devin's early performance remained far short of a human engineer across the full range of software work. The underlying product direction proved durable: developers increasingly delegate complete tasks to cloud agents rather than using AI solely for inline completions.
Cognition has since accumulated the financing and revenue base required to reject acquisition interest. On May 27th, Cognition said it raised more than $1 billion at a $26 billion valuation. Lux Capital, General Catalyst and 8VC led the financing, with participation from Founders Fund, Elad Gil, Ribbit Capital, Atreides, Layer Global and other investors.
Cognition also reported $492 million in annualized run-rate revenue and enterprise usage growth of more than tenfold since the start of 2026. Those figures are company-reported snapshots rather than audited annual revenue. They still set a demanding baseline for any buyer: a credible proposal would need to compensate shareholders for a company that had repriced itself less than three months earlier and was telling investors it still had steep growth ahead.
Windsurf made Cognition harder to replicate. Cognition's July 2025 acquisition announcement said the assets brought $82 million in annual recurring revenue, more than 350 enterprise customers and hundreds of thousands of daily active users. The deal paired Cognition's cloud agent with an integrated development environment and a larger sales organization.
One year later, Wu and Windsurf executive Jeff Wang said Cognition had expanded from 44 to 350 employees and pushed its revenue run rate above $500 million since combining the businesses. Cognition attributed those results to its own internal reporting. The growth explains why SpaceX called. The combination of an autonomous coding agent, an editor, proprietary models and enterprise distribution is expensive to recreate, even for a buyer with abundant compute.
Independence is part of Cognition's product strategy
Wu has spent 2026 positioning Cognition as an independent application and agent company that can work across model providers. In a July interview, he said Cognition had become known as "the folks betting on independence". He declined to say how many acquisition approaches Cognition had received.
That stance carries commercial weight. Engineering buyers frequently use multiple models because the strongest option varies by repository, task, latency requirement and price. Cognition has described itself as the "Switzerland of AI" while building evaluation and routing systems across competing models. Its own SWE models add another option, but Cognition's products still benefit from access to outside frontier models.
A sale to SpaceX would have changed that positioning. SpaceX owns the Grok model family, the Colossus compute infrastructure and Cursor. Adding Cognition would place two major coding interfaces under a vertically integrated model and infrastructure owner.
That structure can produce lower inference costs, tighter model-product integration and faster training loops. Cursor said the SpaceX combination would give it access to a huge GPU fleet and allow it to deliver stronger models at lower cost. The same structure creates a channel conflict for customers and model suppliers. An application that presents itself as neutral has different incentives once its owner is funding and training a competing model.
Cognition's continued independence preserves its ability to route work among model providers and sell that neutrality to enterprises. SpaceX's approach shows how valuable the opposite strategy has become. The owner of compute wants control of the products where developers generate training signals and spend money.
Compute agreements are becoming acquisition funnels
The lesson for AI founders sits in the contract, rather than the headline price.
Compute partnerships can resemble infrastructure purchases while transferring strategically important rights. SpaceX's original Cursor agreement involved personnel, data, prompts, workflows, code and technical knowledge. It included an acquisition option and exclusivity terms. Cursor received compute and economic protection if SpaceX walked away. SpaceX received a structured path from supplier to collaborator to owner.
A founder considering this kind of agreement has to price the full dependency. Dedicated compute can accelerate training, lower unit costs and remove the capacity planning that consumes engineering time. It can also bind the company's model roadmap to one provider, expose proprietary operating data and narrow the field of future buyers or partners.
Cognition has its own reason to keep talking. Training coding models and running autonomous agents are compute-intensive businesses. Cognition's SWE-1.7 model, released in July, used additional reinforcement learning on a Kimi base model. Devin also performs long-running tasks in hosted environments, making the cost of inference and virtual machines part of the product margin.
SpaceX can offer capacity at a scale few counterparties can match. Cognition can offer high-value software-engineering workloads, enterprise distribution and feedback from agents operating inside real repositories. A commercial agreement could benefit both sides without an acquisition. The Cursor precedent means the governance around such a partnership will matter as much as the GPU allocation.
SpaceX is assembling an AI distribution stack
Buying Cursor gave SpaceX an editor and an installed developer channel. Pursuing Cognition indicates a broader appetite for the agent layer where software work is delegated and completed.
That is a different acquisition thesis from buying talent or a promising model laboratory. Coding products generate revenue, own customer relationships and sit at the point where models encounter measurable work. A code change either passes tests, meets review standards and ships, or it does not. That feedback is valuable for improving agents and proving their usefulness to paying customers.
SpaceX's prospectus already described this loop: compute flows to the coding company, while data and technical knowledge flow back. Ownership closes the loop and prevents a rival model provider from capturing the same distribution.
Cognition remains outside that structure. Wu has raised enough money to keep building, assembled a product suite through acquisitions and made neutrality part of the sales pitch. SpaceX has enough compute and stock-market currency to keep asking whether independence has a price.
The acquisition talks ended without a deal. The continuing compute discussions leave SpaceX close to Cognition's most immediate constraint and Cognition close to SpaceX's most valuable AI asset. Cursor showed where that arrangement can lead.