Why Companies Want AI They Can Own

The case for open-weight AI is shifting from its role in competition with China and as shared infrastructure toward a practical question: can companies use models to build AI systems they control? Interest is being driven by lower costs and data sovereignty, but a forthcoming US model has not yet demonstrated its capabilities. Axios reported that several US labs are preparing launches and that Reflection AI plans to release its first model later this month. Sources described it as competitive with Chinese rivals, but there are no released-model results to assess.

Reflection AI, founded in early 2024, has reportedly reached a $25 billion valuation before releasing a product. It won a Pentagon contract in May and signed a $6.3 billion compute deal with SpaceX, as well as another $1 billion in contracts with Nebius. The company’s stated aim is to build an “AI factory” that helps enterprises create their own systems on Reflection models. It is piloting this approach with South Korea’s Shinsegae Group, making enterprise deployment—not just model release—a central part of its proposition.

The ownership argument concerns more than access to model weights. Alex Karp said technical customers want control of compute, models, data, and their “alpha.” Satya Nadella warned that companies risk handing proprietary knowledge to AI providers, and argued they need a trust boundary around data, evaluations, adapted weights, and memory. He also said companies need rights to use outputs to train their own models. The underlying trade-off is whether organizations can gain AI capabilities without surrendering control over the information and customized systems that may give them an advantage.

Microsoft does not sell open models, but offers its models as a base for customer-specific training with a trusted partner. Mustafa Suleiman described Microsoft Frontier Tuning as a way for customers to create company-specific agents they control. He said MAI models tuned for McKinsey tasks had the highest win rate, beat GPT-55 on quality, and cost 10 times less. Those are claims about a particular tuning effort, rather than evidence that customized models will outperform general models in every setting, but they show how providers are pitching controlled, task-specific systems.

Usage data and industry investment add to the case that open models matter commercially. Guillermo Roche said open models accounted for 78.4% of tokens on Vercel AI Gateway on a possible record day, compared with 21.6% for closed models. NVIDIA has backed Reflection AI and other open-model labs. Jensen Huang argued that researchers, developers, startups, and companies need open-source models, citing NVIDIA’s 23 models on leaderboards across language, physical AI, and biology. Signal on X said substantial new development is being built on Chinese open-weight models, sometimes running on US infrastructure without disclosure, and argued that Luna cannot be fine-tuned in response to a question about whether open models were cheaper.

US policy has yet to settle how to balance limiting China’s access to advanced-model inputs with building an American stack that includes hardware and both open and closed models. Howard Lutnick has considered incentives for US labs to produce open-weight models, while some have warned that placing them in a US testing regime could suffocate the sector. The Wall Street Journal reported that Jay Clayton would be empowered as the administration’s AI czar; Clayton has framed AI primarily as a national-security issue and argued that failing to lead increases risks from adversaries. The broader debate also concerns who sets limits: Sam Altman said society should accept some bad outcomes in exchange for AI’s benefits and people’s agency, while describing his disagreement with Anthropic as a dispute over concentrating control in one lab. Both companies support third-party safety auditors, and a White House AI summit accord reaffirmed third-party auditing and board-level oversight for frontier labs.