Reflection SpaceX compute deal: startup secures GB300 chips for $150M-per-month
Reflection’s $150M/month SpaceX compute deal gives the open-weight AI startup immediate access to Nvidia GB300 chips at Colossus 2, starting July 1, 2026.
Reflection has agreed to pay SpaceX $150 million a month beginning July 1, 2026 for access to Nvidia’s latest GB300 AI accelerators and supporting hardware housed at the Colossus 2 data center near Memphis, Tennessee. The contract runs through 2029 and is valued at up to $6.3 billion, with either party able to terminate after the initial three months with 90 days’ notice. This Reflection SpaceX compute deal positions the company among a growing list of major labs renting capacity from SpaceX’s chip-heavy data centers.
Deal terms and scale
Reflection’s commitment starts on July 1, 2026 and carries a monthly price that is substantially lower than the headline rents reported for other tenants of the Colossus facility. Anthropic and Google reportedly arranged larger monthly arrangements with SpaceX, but Reflection’s agreement is notable for being among the largest announced by an open-source AI lab to date. The three-year structure provides predictability while preserving flexibility through an early exit clause after the initial quarter.
The deal grants Reflection immediate access to GB300 accelerators and the system integration needed to run large-scale training and inference workloads. By securing continuous hardware throughput, Reflection aims to accelerate development of publicly released, open-weight models while managing costs and timetable risks tied to on-premise procurement or cloud market volatility.
Open-weight strategy and market positioning
Reflection, founded in 2024 by two former DeepMind researchers, has pitched itself as an open-weight alternative to closed frontier labs. The company argues that publishing trained model parameters reduces vendor lock-in and broadens scrutiny and innovation across academia and industry. The SpaceX compute pact is being framed internally as essential runway to train competitive, widely accessible models at scale.
This positioning comes amid increased attention to model openness after recent regulatory actions and debates over national security and model access. Reflection’s leadership says expanded compute capacity will allow larger, transparent models to compete on both performance and governance grounds, though it will still need to demonstrate parity with the best closed systems to win broad adoption.
Colossus 2 and SpaceX’s compute strategy
Colossus 2, built originally by xAI and now operating within SpaceX’s infrastructure portfolio, was designed around dense deployments of Nvidia accelerators. As xAI’s own ambitions shifted, SpaceX repurposed the facility to lease capacity to external labs, effectively turning a stranded asset into a revenue stream. The availability of GB300 chips — Nvidia’s latest high-end AI silicon — makes Colossus attractive to organizations that require raw training throughput.
SpaceX’s approach of leasing entire racks and dedicated capacity to frontier labs has reshaped the compute market by offering long-term block reservations rather than on-demand cloud instances. That model gives tenants predictable performance and potentially lower unit costs, while giving SpaceX steady, high-value income streams from its hardware investments.
Comparisons with Anthropic and Google agreements
While Reflection’s monthly payment is smaller than the reported $1.25 billion-per-month arrangement with Anthropic and the $920 million-per-month agreement with Google, it still represents a major infrastructure commitment for an open-source lab. Those larger contracts attracted headlines for their scale and for the broader implications about who controls the fastest training pipelines. Reflection’s contract underscores that access to frontier-level chips is no longer limited to a handful of hyperscalers.
Market watchers note that the headline numbers do not fully capture relative access and prioritization; cancellation options and implementation details can affect effective duration and utilization. The mutual-termination clause in Reflection’s deal provides a pragmatic balance, allowing both parties to reassess commitments as models, chips, and market conditions evolve.
Implications for the open-source AI ecosystem
Immediate access to GB300-class compute could accelerate the pace at which open-weight models close the performance gap with closed frontier systems. If Reflection can scale training runs and publish robust, reproducible weights, downstream developers and researchers could gain alternatives to proprietary offerings. That in turn could shift competitive dynamics around model licensing, deployment, and safety research.
However, the deal also highlights the deep capital intensity of modern model training and the strategic dependency on scarce hardware resources. Smaller labs and academic teams without multimillion-dollar contracts remain reliant on shared cloud providers or public compute grants, which could widen the gap between well-funded projects and the broader research community unless complementary funding mechanisms emerge.
The commercialisation of Colossus-style capacity may also influence geopolitical and regulatory debates, as nations and enterprises examine the concentration of high-end chips and the entities that control them. How firms like Reflection balance openness with commercial sustainability will be closely watched by industry observers and policymakers alike.
Reflection’s immediate priority will be to deploy the newly allocated capacity toward training cycles that showcase the practical advantages of the open-weight approach. Observers will watch model releases, benchmark performance, and whether the company’s outputs catalyze broader adoption among developers and customers.
The Reflection SpaceX compute deal is a significant milestone for an open-source-forward lab, but its ultimate impact will depend on execution, model quality, and how the company navigates the same cost and competition pressures that shaped the strategies of larger, closed labs.
