SpaceX signs computing deal with Reflection AI

The more durable story here is Colossus 2's evolution from SpaceX-internal infrastructure into a commercial multi-tenant compute platform - this Reflection deal joins previously reported arrangements with Anthropic, Google, and Cursor - which matters more for practitioners tracking compute access than the specific dollar figure attached to any single contract.

According to CNBC, SpaceX signed a multi-year agreement to give the open-source AI startup Reflection access to Nvidia GB300 GPUs and related hardware at the Colossus 2 data center in Memphis. CNBC reports Reflection agreed to pay $150 million per month starting July 1, 2026, through 2029 - about $6.3 billion over the full term - with a 90-day termination clause exercisable by either party after the first three months. A Reflection spokesperson told CNBC and Axios the deal was strategically important, stating, "Recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models."

Reflection was founded in 2024 by Misha Laskin, who led reward modeling for Gemini at Google DeepMind, and Ioannis Antonoglou, a co-creator of AlphaGo. The company raised a $2 billion round at an $8 billion valuation last October, led by Nvidia, which Quartz and Yahoo report included roughly $800 million directly from Nvidia - the same company supplying the GB300 hardware Reflection will run under this deal, a capital-compute loop frequently observed across the sector. Access to high-end accelerators such as the GB300 materially shortens wall-clock time for large-model experiments; companies that cannot source immediate GB300-class capacity often face multi-quarter delays, making leased colocated capacity at established sites a common workaround in 2026.

Quartz reports prior multi-year compute commitments tied to SpaceX's Colossus facilities, citing roughly $30 billion for Google and $45 billion for Anthropic, framing this Reflection deal as one more data point in SpaceX's broader commercialization of Colossus capacity alongside a reported deal with Cursor. Quartz notes Reflection has not yet released a public frontier open-source model at scale, making this a forward-looking infrastructure bet rather than a deal tied to a shipped product.

The transaction illustrates that immediate access to next-generation accelerators increasingly depends on commercial arrangements with data-center operators rather than only ownership of on-prem hardware or standard cloud quotas. Teams evaluating similar arrangements should track contract structure - month-to-month pricing, exit clauses, and power/networking terms - alongside raw chip access, since those terms determine practical flexibility if pricing or availability shifts.

Whether Reflection publishes model checkpoints, training logs, or benchmarks demonstrating utilization of the acquired GB300 capacity; how pricing and contract structure at large colocated sites evolve as more entrants compete for scarce GB300-class inventory; and whether SpaceX discloses additional Colossus tenants beyond the Anthropic, Google, Cursor, and Reflection deals reported so far.