Two American labs made big moves this week around the same conviction: that the open-weight model, the kind you can download, inspect and fine-tune, is now a strategic asset worth fighting for, and that the fight is currently being lost to China. Reflection announced a deal to buy $1 billion of compute from the neocloud Nebius, on top of an existing arrangement to pay Elon Musk's Colossus data center $150 million a month through 2029. Mira Murati's Thinking Machines, meanwhile, released Inkling, its first model, with the full weights posted to Hugging Face.
What makes the pairing striking is how little either company has shipped. Reflection, a Brooklyn startup founded in early 2024, has not released a model publicly at all; it has grown from 30 employees last September to 230 and plans its first open-source release later this year. "We are flying the plane as we're building it," cofounder Ioannis Antonoglou said. His stated goal is blunt: "to ensure that there is a competitive frontier open lab in the Western world." Thinking Machines raised a record $2 billion seed at a $12 billion valuation before it had any product, and Inkling is explicitly not pitched as the smartest model available. It is pitched as the most customizable, something enterprises can own and tune rather than rent from a frontier lab.
Both bets are a direct response to the same map. The most capable open models today come from Chinese firms, and the money is following: DeepSeek is reportedly in talks to raise $1.5 billion at a $71 billion valuation ahead of an IPO next year, and Moonshot AI is said to be raising $2 billion at a $20 billion valuation. Because those models are cheap and freely adaptable, US startups have leaned on them heavily, a defection on price that the frontier labs have struggled to answer.
Here is the awkward part. The Western counteroffensive is being built partly on Chinese foundations. Thinking Machines used data generated by existing open models, including Moonshot's Kimi K2.5, in Inkling's final training phase. The open ecosystem these labs want to reclaim is one whose current center of gravity they are still borrowing from.
There is a strategic urgency underneath the commercial one. Chinese authorities are reportedly weighing whether to curb overseas access to the country's leading models. If the cheap, capable Chinese weights that so much of the American startup stack now runs on were restricted, the value of a home-grown open champion would jump overnight. That is the scenario Reflection and Thinking Machines are, in effect, insuring against.
The open question is whether the economics work. Open weights can be inspected, self-hosted and fine-tuned, which enterprises increasingly want. But giving the model away makes the billions in compute harder to recoup, and Murati has been careful to leave herself room: Inkling is open, she has signaled, but future Thinking Machines models may not be. It is the same case-by-case logic she watched play out at OpenAI in 2019, when the lab founded on openness held back the full GPT-2. The ambition to build a Western open frontier is real. Whether it can be a business, rather than a subsidy, is the part nobody has answered yet.