AI startup Reflection has signed a deal worth more than $1 billion to secure computing capacity from Nebius, including access to Nvidia’s latest-generation chips. The agreement gives Reflection additional infrastructure for training and operating its artificial intelligence models as competition for advanced computing resources intensifies across the industry.
The Nebius deal follows a separate computing agreement Reflection reached with SpaceX in June. That arrangement was reported to involve payments of approximately $150 million per month through 2029, illustrating the scale of Reflection’s infrastructure commitments as it seeks to expand its model-development capabilities.
Together, the agreements position Reflection among a growing group of AI companies making large, long-term commitments to secure access to high-performance computing capacity. Training advanced AI models requires enormous clusters of specialized chips, substantial energy supplies and sophisticated data-center infrastructure. Once those models are released, companies also need additional computing power to support inference, the process through which models generate responses and perform tasks for users.
Demand for this infrastructure has continued to rise as businesses integrate AI into software development, customer support, research, financial services and other operations. However, the construction of new data centers and the deployment of advanced chips have not kept pace with demand in many markets. That imbalance has pushed AI developers to reserve capacity years in advance rather than risk being unable to access the computing resources needed to train new models or serve customers reliably.
For startups, the challenge can be particularly significant. Large technology companies often operate their own data centers or have longstanding relationships with major cloud providers, giving them greater control over computing supply. Smaller AI developers must frequently negotiate with cloud infrastructure companies, chip providers and specialized data-center operators to secure enough capacity to remain competitive.
Reflection was launched by two former Google DeepMind researchers and is developing open-source AI models positioned as alternatives to systems offered by OpenAI and Anthropic. Rather than keeping their model weights entirely private, open-source developers typically make more of the underlying technology available to customers and researchers.
This approach can give companies greater flexibility to customize models for their own data, workflows and security requirements. Open-source models may also be less expensive to operate than proprietary alternatives, particularly for organizations that have the technical resources to deploy models on their own infrastructure or through a cloud provider of their choice.
Interest in open-source AI has grown as businesses look for ways to control rising technology costs and reduce their dependence on a small number of model providers. Companies using closed AI platforms may face recurring usage charges, limited visibility into how models operate and restrictions on how the technology can be modified or deployed.
Recent U.S. restrictions affecting access to Anthropic’s advanced models have also highlighted the risks associated with relying heavily on a single provider. Policy changes, commercial disputes or access restrictions can potentially disrupt companies that have built critical products around externally controlled models.
Reflection is betting that open-source models can provide businesses with greater control, portability and resilience. Customers may be able to run the models across different infrastructure providers, adapt them for specialized use cases and avoid becoming locked into a single AI ecosystem.
Securing substantial computing capacity from Nebius and SpaceX could help Reflection train larger models, improve their performance and support broader commercial adoption. It may also allow the company to offer customers more predictable access to its technology as demand grows.
The agreements demonstrate that access to computing infrastructure has become one of the most important competitive factors in artificial intelligence. AI companies are no longer competing solely on model quality, research talent or software capabilities. They are also racing to secure the chips, energy and data-center capacity required to develop and operate those systems at scale.
