Open models. Competing ambitions.

Open-weight AI models give startups and established businesses a foundation for developing capabilities of their own. The work that follows can improve an operation, open a route to customers or provide expertise and training material for another company’s AI models. The resulting value gives suppliers, partners and buyers reasons to cooperate and reasons to compete.

Customers pay Cursor to help them write software. In August 2026, SpaceX acquired the company in a deal announced at US$60 billion. By then, Cursor was contributing to something much larger than a coding tool: its work with developers was helping train SpaceX’s Grok models for tasks extending into research, finance and other professional work.¹²

The developer buying help with a difficult piece of code and the company buying Cursor could value the same business for different reasons. A useful product had gathered customers, skilled people and material for developing further products. Some of that value arose through people using what they had bought.

Someone else’s model. Cursor’s product.

In 2023, Cursor built its editor on the existing Visual Studio Code ecosystem. It used external AI models while developing parts of its own. Its early research agenda identified problems a better prompt would not readily solve: speed, cost and finding the right information in a large codebase.³

One of those problems was the wait between a model proposing a change and the change appearing in the developer’s files. In 2024, Cursor trained a separate model for applying edits, using Meta’s Llama 3 as its foundation. It could change how this model ran in ways unavailable through a model supplier’s standard service. The external model planned the edit; Cursor’s adapted model applied it quickly.⁴

The distinction mattered commercially. Cursor could buy an important part of the capability while investing in the part that made its own product more useful. A model released by another company became a foundation for development it could direct itself.

Meanwhile, the economics of the service were becoming harder to package. In July 2025, Cursor apologised for a poorly communicated pricing change. A simple question and a long task in which an agent repeatedly used tools consumed very different amounts of computing. Customers who thought they understood their allowance encountered unexpected charges.⁵

By October 2025, it had launched a coding model called Composer, designed to work across files while responding quickly enough to keep the developer involved.⁶ In February 2026, Cursor reported that the additional training of one Composer release had consumed more computing than the original training of its base model.⁷ Starting with an existing model had given it a route into development; pursuing better performance was taking it towards much larger demands on computing.

The following month, Cursor released Composer 2 using Kimi K2.5, developed by China’s Moonshot AI, as its foundation. Kimi was an open-weight model, available for others to obtain and adapt under its licence. Cursor assessed potential foundations for coding knowledge, keeping track of changes to files and efficiency on its own infrastructure. It deliberately excluded coding-agent benchmark results from that selection because further training would substantially change those abilities.⁸

The best model to use immediately was not necessarily the best model to build on.

Moonshot’s licence allowed commercial modification without requiring Cursor to publish its improvements.⁹ A Chinese developer’s investment could therefore support an American company’s proprietary product. The companies could benefit from the same development without owning the same part of it.

In April, Cursor identified computing capacity as a constraint and announced a training partnership with SpaceX. A month later, while releasing another Kimi-based model, it disclosed joint work on a much larger model being trained from scratch.¹⁰¹¹

Cursor was pursuing several routes at once. It continued adapting an available model, began work on a new foundation and kept using outside suppliers. Its subsequent routing system selected among external models and Grok according to the task, observed performance and cost.¹² Its own development expanded what it could offer without ending its reliance on other companies.

What the product helped create

The SpaceX partnership brought together different assets. SpaceX supplied computing capacity. Cursor contributed personnel, datasets, code, workflows and know-how to joint model development. The agreement preceded the acquisition.¹³

In July, the companies released Grok 4.5. Cursor reported training it with trillions of tokens of data capturing interactions with codebases and software tools, alongside scientific, research and other material. Its team also constructed difficult tasks in which models had to investigate a problem, use tools, recover from mistakes and check their work. Those abilities had uses beyond coding, across other work performed on computers.²

Developers were paying for assistance. Permitted records of their interactions could also help train a wider capability. The work of selecting material and constructing useful training tasks was part of what Cursor brought to the partnership. Its value lay in a combination of participation, engineering and experience.

That participation depended on the customer’s agreement and settings: Cursor’s published terms distinguish training permissions, and Privacy Mode excludes customer data from training.¹⁴ A service can deliver what someone bought while helping its supplier develop something else.

Acquisition gave SpaceX the operating business as well as the collaboration. SpaceX subsequently reported US$608 million of Cursor revenue between completion on 14 August and the end of September. Its acquisition accounting also described awards intended to retain people after the transaction.¹⁵ There were customers to keep and a team to retain, alongside capabilities that could contribute to Grok.

Ownership changed another relationship. On 28 August, OpenAI announced that it intended to wind down its Cursor contract, proposing 12 November as the end date. It invoked change-of-control provisions and cited concerns about compliance under the new owner.¹⁶

OpenAI was supplying a customer that had become part of a rival model business. Developing its own models had widened Cursor’s options. Becoming part of SpaceX also put its access to an existing supplier at risk.

The incumbent’s opportunity

A Walmart customer planning a football watch party can ask for help with the occasion rather than search separately for snacks, decorations and tableware. Walmart describes combining a fine-tuned open-source model with OpenAI models to interpret those requests and suggest products.¹⁷

Across its catalogue work, models extract and check product information, while human quality assurance specialists define what counts as correct and supply validated examples. Walmart reports fewer inaccuracies and lower model-running costs. Better product information helps customers find suitable goods; clearer packaging images help employees pick orders. The benefits reach beyond maintaining the catalogue into selling and delivering what it contains.¹⁸¹⁹

Walmart also wants to carry that retail capability into other companies’ AI services. Its AI and product chief, Daniel Danker, described plans for Sparky, Walmart’s shopping assistant, to manage the shopping experience inside ChatGPT and Gemini, recognising customers and combining baskets. A Walmart-run experience launched inside ChatGPT in March 2026.²⁰²¹

The ambition includes winning purchases beyond customers’ usual Walmart shopping and retaining a role in recommendations, which Danker linked to advertising income. Its partners offer access to customers; Walmart brings product knowledge, customer relationships and the ability to fulfil the purchase. Where discovery ends and the retailer takes over matters to who earns from the transaction.²⁰

Walmart: when shopping starts somewhere else follows the investment, the commercial adjustments and what customers’ and employees’ experience reveals about the work.

The business of helping others build

Developing these capabilities also creates demand for tools, computing and services. Hugging Face has built a business around helping developers share, find and deploy models, datasets and applications. NVIDIA agreed to acquire it in September for approximately US$11.9 billion payable to shareholders, with up to US$1 billion in employee retention. Completion was expected in the first half of 2027.²²

NVIDIA already benefited when developers needed computing. The two companies also had a direct commercial relationship: in June 2025, they introduced a service through which customers could rent computing capacity for model training, with the companies sourcing and provisioning it together.²³

What, then, would ownership add?

Hugging Face develops and maintains tools such as Transformers, which helps software use many different models. It has brought in engineering teams working on the storage of large model files and on running models across different computing systems. It sells enterprise services, storage and computing.²⁴²⁵ NVIDIA would acquire the company responsible for that work, its commercial relationships and a continuing say in what it develops next.

The community’s contributions are a different matter. Hugging Face’s terms preserve contributors’ rights, and individual models and datasets retain their licences.²⁶ The company’s value depends partly on people continuing to find its tools and services useful enough to bring their work there.

In a joint interview, Hugging Face chief executive Clément Delangue said he had approached NVIDIA about gaining greater resources. Jensen Huang said he had initially wanted Hugging Face to remain independent and acknowledged other bidders.²⁷ The proposed acquisition would secure NVIDIA’s position in the company and prevent another bidder from taking that ownership.

NVIDIA’s announcement commits to preserving choice of models, cloud services and competing hardware.²⁸ If steering developers towards NVIDIA made the platform less useful, it could damage the participation on which its value depends. Supporting competitors can be part of the commercial logic of ownership.

The geopolitical tension follows from those same relationships. NVIDIA’s regulatory filing links demand for open models with use of its products. It warns that restrictions affecting models derived from China could materially harm both Hugging Face and NVIDIA.²²

A US Department of Homeland Security order published on 20 October 2026 required federal civilian agencies to remove DeepSeek products and third-party products and services developed using DeepSeek implementations.²⁹ The restriction reached beyond the original developer into the market for products others had built.

A Chinese model can support an American product, activity on an American platform and demand for American computing. Security concerns remain, but the commercial interests do not divide neatly by nationality. Some of the companies a government seeks to strengthen may be earning from the participation it seeks to restrict.

When the interests pull apart

A developer can want its technology widely used and still object to the businesses others build around it. That tension predates AI.

In 2021, Elastic changed the terms for future releases of its search software, Elasticsearch, to restrict competing hosted services. It argued that AWS was benefiting without contributing fairly. AWS disputed that account and committed to maintaining an alternative using an earlier open-source version.³⁰³¹

Elastic could change the terms for future releases. AWS could build a competing service around rights already granted and its own ability to maintain the software. The dispute concerned the commercial bargain that had developed around shared technology: what each party contributed, what each could sell and who should benefit.

Model developers face related choices. Encouraging others to build can widen a model’s use and create opportunities its developer could not pursue alone. Those businesses may buy services, bring new users or become competitors. The invitation to participate can serve several interests while those interests still fit together.

The Resource Some assembly required explains adaptation, operating choices and the permissions attached to particular model releases.

What else are we building?

An AI investment may be justified by a better service or a lower operating cost. As the work develops, it can also create training material, expertise, customer relationships and the ability to pursue a different opportunity. The original business case may capture little of that value.

Some of it may also accrue elsewhere. A supplier can learn from serving the business. A platform can become more useful through contributions from companies that compete with one another. Those exchanges can be worth making: they may give the company capabilities and access to customers it could not readily obtain alone. But what is the company becoming better able to do, and what is its participation helping others build?

The answer can change as the models improve. A better general model may remove the reason to maintain one specialist component while making another product possible. As more businesses gain access to capable models, the company’s understanding of its customers, the reliability of its service and its ability to deliver may become more important than which model it uses.

An acquisition price, a benchmark score and a saving on computing each reveal something. None settles what will remain valuable.

As the models improve, what will customers keep choosing us for?

Notes

Cursor, Cursor is now a part of SpaceX, 14 August 2026; SpaceX, third-quarter results, 20 October 2026, Cursor merger discussion. The US$60 billion figure is the transaction’s implied equity value; purchase accounting separately treats awards tied to subsequent employee service. ↩
Cursor, Introducing Grok 4.5, 8 July 2026. Describes joint training with SpaceXAI, developer interactions and a broader mix of training material and tasks. The release preceded the acquisition. ↩ ↩²
Cursor, version 0.2.0, 6 April 2023, and Our problems, 12 October 2023. The editor used VSCodium, derived from Visual Studio Code; the research agenda discusses custom models, retrieval, speed and cost. ↩
Cursor, Editing Files at 1000 Tokens per Second, 14 May 2024. Compared adapted DeepSeek Coder and Llama 3 models and selected Llama 3-70B for applying edits. Custom inference engineering, undertaken with Fireworks, was part of the speed improvement. ↩
Cursor, Clarifying our pricing, 4 July 2025. Explains the move from request allowances towards usage-based charging, acknowledges poor communication and offers refunds for unexpected charges in the specified period. ↩
Cursor, Composer: Building a fast frontier model with RL, 29 October 2025. This is the coding-model release; the Composer name had previously been used for an editor feature. ↩
Cursor, Introducing Composer 1.5, 9 February 2026. Compares post-training computing with base-model pretraining computing. It does not disclose an equivalent comparison of total expenditure. ↩
Cursor, Composer 2 Technical Report, version 2, 26 March 2026, especially §§3–6 and Appendix B. Sets out the base-model selection criteria and subsequent training. ↩
Moonshot AI, Kimi K2.5 licence, checked 23 October 2026. Modified MIT terms permit commercial modification and distribution, subject to notices and naming conditions. They do not require improvements to be published. Separate agreements between the companies are not established by this public licence. ↩
Cursor, Cursor partners with SpaceX on model training, 21 April 2026. Identifies computing capacity as a constraint and announces access to Colossus infrastructure. ↩
Cursor, Introducing Composer 2.5, 18 May 2026. Retains Kimi K2.5 as its base while announcing separate work on a larger model trained from scratch. ↩
Cursor, How Cursor Router chooses the right model for the task, 6 August 2026. Describes combining Grok, Sol, Opus and Fable. Its performance signal is inferred from subsequent user behaviour, including corrections, rather than an independent test of every answer’s correctness. ↩
SpaceX, EU prospectus, 5 June 2026, §4.3.4.4, printed pp.63–64. The collaboration agreement was dated 19 April. It describes the companies’ contributions, retention of pre-existing intellectual property and shared rights in jointly developed models. ↩
Cursor, Data Use and Terms of Service, especially §§1.1–1.5, checked 23 October 2026. Privacy Mode excludes customer data from model training; the terms require express agreement for training uses and distinguish customers governed by a separate commercial agreement. Current terms do not establish every contributor’s historical permissions or awareness. ↩
SpaceX, third-quarter 2026 results, 20 October 2026, Cursor acquisition discussion. The revenue covers 14 August–30 September. Acquisition accounting distinguishes purchase consideration from awards attributable to subsequent employee service. ↩
OpenAI, Our decision on Cursor following its acquisition by SpaceX, 28 August 2026. Gives OpenAI’s stated contractual and compliance reasons and proposed termination date. At this paper’s date, that date remains in the future. Its current guidance, checked 24 October 2026, describes separate ways for customers to use OpenAI models in supported local Cursor features. The proposed withdrawal concerns OpenAI’s supply contract with Cursor. ↩
Walmart Global Tech, Walmart’s Generative AI search puts more time back in customers’ hands, 30 January 2024. Describes occasion-based shopping and a fine-tuned open-source model used alongside OpenAI. The search account and the later catalogue article describe related areas of work, without establishing that they use the same models. ↩
Binwei Yang, Walmart Global Tech, How Walmart uses LLMs to manage its massive product catalogs, 20 May 2025. Describes extraction and quality-control models, human-validated examples, separate evaluation data and larger models generating targeted training examples for smaller models. Walmart does not name the underlying models or quantify the reported savings and accuracy improvements. ↩
Walmart, Second-quarter FY2025 earnings call, 15 August 2024, corrected transcript, pp.4 and 10, remarks by Doug McMillon and John Furner. Connects catalogue improvements to product discovery, the expanding assortment and packaging images that help employees identify items. These are results described for the wider programme, not a measurement of the particular model configuration outlined in the May 2025 article. ↩
Walmart, Morgan Stanley Technology, Media & Telecom Conference transcript, 4 March 2026, pp.5–9, Daniel Danker. Contrasts existing native checkout with plans for Sparky to take over the shopping journey inside ChatGPT and Gemini. Discusses shared baskets, membership benefits, broader shopping occasions, product selection and advertising. These are management’s plans and expectations, not measured returns from the proposed integrations. ↩ ↩²
OpenAI, Powering Product Discovery in ChatGPT, 24 March 2026. Announces the Walmart-run experience on the web, with mobile access forthcoming, including account linking, loyalty benefits and Walmart payments. It establishes the launch and changed approach, not a measured improvement in conversion or profitability. This source is also recorded in the Walmart case study. ↩
NVIDIA, Form 8-K, event dated 2 September, signed 3 September 2026, Item 8.01 and acquisition risk factors. Separates shareholder consideration from retention. Expected completion remains subject to approvals and other conditions; the filing explicitly identifies restrictions on Chinese-derived models as a business risk. ↩ ↩²
Hugging Face, Training Cluster as a Service, 11 June 2025. Describes the joint route for sourcing, pricing, provisioning and configuring GPU clusters using NVIDIA DGX Cloud Lepton. ↩
Hugging Face, Transformers v5, 1 December 2025; XetHub acquisition, 8 August 2024; ggml and llama.cpp team announcement, 20 February 2026. These document maintenance and engineering capabilities within the company; publicly licensed tools remain distinct from exclusive ownership of community contributions. ↩
Hugging Face, pricing and services, checked 23 October 2026. Includes enterprise accounts, storage and hosted computing. ↩
Hugging Face, Terms of Service, content and intellectual-property provisions, checked 23 October 2026. ↩
CNBC, interview with Jensen Huang and Clément Delangue, 3 September 2026, published unofficial transcript. Huang acknowledges other bidders and rejects the suggestion that the deal is simply about keeping Hugging Face out of other hands. ↩
NVIDIA, NVIDIA to Acquire Hugging Face, 3 September 2026. Promises continued model, cloud and hardware choice. These are commitments for the proposed acquisition. ↩
US Department of Homeland Security, Exclusion and Removal Order 2026-01, issued 24 July 2026; published in the Federal Register on 20 October, pp.66346–66350. Covers federal civilian executive-branch systems, excluding national-security and compartmented systems, and provides an exception process. Its scope expressly includes third-party implementations; it does not resolve every possible form of model adaptation. ↩
Shay Banon, Elastic, Doubling down on open, Part II, 14 January 2021; Amazon: NOT OK - why we had to change Elastic licensing, 19 January 2021. The new licensing options restricted competing hosted services; the SSPL option required disclosure of modifications and service-management code. Elastic’s account of AWS’s contribution was disputed. ↩
AWS, Stepping up for a truly open source Elasticsearch, 21 January 2021. Disputes Elastic’s account and commits to a fork under the earlier Apache 2.0 licence. Software source and model weights are different forms of technology; the comparison concerns commercial rights and incentives around reuse. ↩