.jpg?w=1024)
Reflection AI has signed a $1 billion compute capacity deal with neocloud provider Nebius, its second major agreement despite having yet to launch a public AI model.
It is another sign of the growing gap between supply and demand for AI resources, with several AI model makers announcing billion-dollar deals with neoclouds and hyperscalers to lock down capacity. The Nebius agreement is smaller in scope than the $6.3 billion deal Reflection signed with SpaceX last month, for which the startup is paying $150 million a month through 2029.
Open-source AI models, built in America
Reflection AI was launched by two former Google DeepMind researchers to develop open-source models as alternatives to those from OpenAI and Anthropic. With the market shifting from tokenmaxxing toward more careful budgeting AI use, a US-based open-source alternative could prove valuable, especially as many American companies are currently turning to cheaper Chinese AI models.
“The need for open models is clear, and this additional compute capacity will allow Reflection to continue to build and train frontier AI models at scale,” Reflection chief technology officer and co-founder Ioannis Antonoglou told Reuters.
However, Reflection AI has yet to launch a public AI model despite operating for two years. It has released a code-comprehension agent called Asimov, but this is chump change compared with the products offered by its larger competitors.
It has also been listed as one of the AI companies granted classified privileges by the Pentagon, indicating that its models or applications are highly valued by partners with access to them.
Neoclouds are king in times of high demand
The major neoclouds are signing large agreements with AI model makers and hyperscalers, both of which are hungry for additional compute capacity. While some hyperscalers, especially Google, are struggling to supply customers while also running their own AI training workloads, neoclouds are not burdened to the same extent.
At the same time, neoclouds are more nimble than hyperscalers and can retrofit data centers to meet the custom demands of individual customers. Nebius, which was spun out of the Russian search engine Yandex, has signed a deal worth up to $27 billion with Meta and another worth up to $19.4 billion with Microsoft.
Anthropic, the leader in the enterprise AI market, recently signed a $19 billion deal with TeraWulf, a neocloud only beginning its transition from crypto mining to AI infrastructure. This was perhaps the clearest indication yet that AI model makers will take capacity deals wherever they can find them.
The market is becoming so valuable that operators with additional capacity, such as SpaceX, have signed agreements with Anthropic, Google and Reflection AI. Meta, which has spent billions building out its own AI infrastructure, is reportedly also exploring the creation of a cloud computing business serving either partners or rivals.
With the amount being invested in securing capacity, it doesn't seem likely that the price war currently being fought by the top AI model makers will continue, especially if they go public. For operators like Reflection AI, finding ways to generate revenue while growing their market share will be even more difficult, despite the strong demand for AI tools.
Related reading: As AI companies commit billions to growth, Anthropic is also strengthening oversight through its governance trust.