Current Contracted Baseline Reflection AI’s current dedicated compute footprint relies on two binding agreements signed in mid-2026. The primary contract is with SpaceX at the Colossus 2 campus, where Reflection pays $150M/month from July 1, 2026, through 2029 for access to Nvidia GB300 chips and supporting infrastructure 2 sources. Crucially, this agreement allows either party to terminate with 90 days’ notice after the first three months, making it a rolling commitment rather than guaranteed long-term capacity 2 sources. The second is a cloud contract with Nebius, providing >$1B in GB300 capacity through 2029 2 sources.

Converting Lease Spend to Gigawatts Because these are compute leases rather than direct capital expenditures, capacity must be inferred from comparable tenant contracts and hardware power limits. SpaceX’s filing-based benchmarks show Google paying $920M/month for roughly 110,000 GPUs sec.gov and Anthropic paying $1.25B/month for roughly 325,000 GPUs sec.gov. Scaled to Reflection’s $150M/month, this implies an allocation of roughly 18,000 to 50,000 GPUs. Assuming standard Lenovo GB300 NVL72 rack specifications and applying Epoch AI’s facility overhead estimates yields roughly 3.0 to 3.4 kW all-in per GPU 2 sources. On this basis, the SpaceX contract secures approximately 0.06 to 0.16 GW, while the Nebius deal adds another 0.01 to 0.03 GW. The current live, dedicated baseline therefore sits around 0.08 to 0.18 GW.

Excluded Sovereign and Partnership Capacity Two other major initiatives do not contribute to Reflection’s dedicated 2028 tally under strict resolution criteria. The 250 MW Shinsegae "sovereign AI factory" in South Korea is currently a memorandum of understanding for a joint venture 2 sources. It is explicitly designed in phases to serve Korean enterprises and government agencies, meaning the capacity will not be for Reflection's exclusive internal use prnewswire.com. Similarly, Reflection’s foundational intelligence partnership for the DOE Genesis Mission utilizes the Department of Energy’s compute as deployed, rather than providing Reflection with a dedicated, exclusively controlled capacity block axios.com.

Trajectory and Future Outlook Reflection has raised approximately $4.6B to date but is already burning around $2.1B annually just to maintain its current compute footprint. Significantly expanding this capacity by 2028 would require massive new fundraising or heavy subsidies from Nvidia. Furthermore, positioned as a "Western DeepSeek," Reflection’s open-weight strategy may rely more on compute efficiency than the multi-gigawatt scaling pursued by OpenAI and Anthropic. A median expectation near 0.21 GW reflects a maintenance or modest expansion of their existing contracts. The lower tail (~0.05 GW) accounts for the risk of a capital squeeze triggering the 90-day SpaceX termination clause, while the right tail (~0.82 GW) accommodates the possibility of an unforeseen multi-gigawatt captive buildout heavily subsidized by Nvidia or sovereign wealth.