SpaceX Reflection AI deal vs Anthropic and Google — what's the difference?

Size and strategic intent. The Reflection AI deal (announced June 22, 2026) is $150 million per month from July 2026 through 2029, worth up to $6.3 billion total, for GB300 access at Colossus 2 near Memphis. By contrast, Anthropic pays SpaceX $1.25 billion per month (~$52.5B total) for 220,000 Nvidia GPUs, and Google pays $920 million per month (~$38.6B total) for 110,000 GPU equivalents. Reflection is the smallest of the three but strategically important: it's the first major SpaceX compute contract with an explicitly open-source frontier lab, not a closed-frontier lab. All three contracts share the same 90-day exit clause after the first three months.

Why did SpaceX stock fall 10% on news of a new $6.3B deal?

Three concerns hit at once. First, the 90-day exit clause means SpaceX's $80B+ in 'committed' Colossus revenue is actually $1.5B in firm commitment followed by rolling quarterly decisions — a structural disclosure problem. Second, compute-as-a-service margins are lower than AI product margins, and the Reflection deal underlines that SpaceX is becoming an infrastructure provider rather than an AI-product company. Third, the Cursor acquisition at $60B is not yet showing revenue synergy. SPCX closed near $170 after peaking near $225 in the first week post-IPO; still above the $135 IPO price.

Why would SpaceX rent out compute if Grok needs it?

This is now the question Wall Street is openly asking. Three plausible reads. (1) SpaceX is hedging: if frontier closed models become commoditized by Chinese open-weights releases (DeepSeek V4 Pro, OpenPangu-2, Kimi K2-7, GLM 5.2), the higher-margin path is infrastructure rental rather than competing with Anthropic and OpenAI on capability. (2) Grok's roadmap is intentionally narrower than rivals' — focused on X integration and developer tools (via Cursor) — and doesn't need every GPU. (3) The Reflection deal is loss-leader pricing to keep a strategic open-source lab from going to AWS or Azure. SpaceX has not commented on the tension.

Is Reflection AI actually a frontier lab or a paper company?

Real lab, no public frontier model yet. Reflection was founded by two former Google DeepMind researchers, last valued at $25 billion (March 2026 WSJ report on a $2.5B round), backed by Nvidia, with JPMorgan considering investment via its Security and Resiliency portfolio. The deal stack tracks like a real frontier lab. But: no public frontier model has shipped from Reflection yet. The thesis is American open-source frontier capability for governments and enterprises that won't use closed US labs or Chinese open-weights — a real gap, and Reflection has cultivated US Department of Energy Genesis Mission ties to support it. Verdict: real lab with strategic positioning, still has to ship to validate the bet.

Quick Answer

SpaceX Reflection AI ($6.3B) vs Anthropic ($52.5B) vs Google ($38.6B) Colossus Deals

Published: June 23, 2026

SpaceX Reflection AI vs Anthropic vs Google: The Colossus Customer Stack

On June 22, 2026, SpaceX signed a $6.3 billion, three-year compute deal with Reflection AI — the third major Colossus 2 contract after Anthropic and Google. Combined committed compute revenue is now $97.4B+ through 2029, transforming Colossus from an internal training cluster into a $2.3B+/month recurring revenue infrastructure business. Here’s how the three deals compare and what the structure says about where AI infrastructure is going.

Last verified: June 23, 2026.

The customer stack

Customer	Monthly	Total (to 2029)	Compute	Public model status
Anthropic	$1.25B	~$52.5B	220,000 Nvidia GPUs (H100/H200/GB200)	Claude Fable 5, Mythos 5, Opus 4.8
Google	$920M	~$38.6B	110,000 Nvidia GPU equivalents	Gemini 3.5 Pro (GA), 3.5 Flash
Reflection AI	$150M	~$6.3B	GB300 chips, Colossus 2 (Memphis)	None public yet
Total committed	$2.32B+/mo	$97.4B+	—	—

All three contracts run through July 2029. All three contain a 90-day exit clause after the first three months.

What each deal tells us

Anthropic ($1.25B/month)

The anchor tenant. Anthropic’s contract is roughly 5x Google’s monthly spend and gives Anthropic dedicated capacity for training (Mythos / Fable / Opus / Sonnet families) plus inference at scale. The size signals that Anthropic’s compute thesis is “build through 2029” — consistent with the $965B valuation reporting and the late-2026 IPO race coverage. If frontier capability and inference cost both keep improving, this is a defensible commitment. If model gains plateau, it’s a heavy fixed cost.

Google ($920M/month)

Notable because Google has its own TPU stack. The Colossus contract is a hedge — additional Nvidia capacity for workloads where TPU portability is hard or where geographic placement matters. The 110,000 GPU equivalents number is small relative to Google’s total compute footprint but symbolically large: Google is willing to be a Colossus customer, validating SpaceX as a tier-1 AI infrastructure provider.

Reflection AI ($150M/month)

The strategic interesting one. Reflection is the first open-source-frontier customer of Colossus 2. The smaller commit ($150M/mo) is appropriate for a pre-public-model lab — most of the spend is training rather than inference. The GB300 chip access is meaningful: GB300 supply is constrained, and Reflection getting GB300 first signals Nvidia-as-investor influence on allocation.

The 90-day exit clause everyone is ignoring

Here’s the disclosure issue analysts are raising. All three Colossus contracts include a 90-day exit clause after the first three months. That means:

Reported “committed” revenue: $97.4B+ through 2029
Actually contractually firm: ~$7B (first three months × all three customers)
Beyond that: rolling quarterly decisions

This is closer to a SaaS book of business than a long-term infrastructure contract. For analysts modeling SPCX, the question is what fraction of that $97.4B should be present-valued. The market said “less than the prior implied multiple” on June 22 — hence the 10% drop.

Why open-source matters here

The Reflection deal is the first signal that SpaceX is willing to fund open-source frontier compute at scale. Three implications:

Closed frontier labs lose a degree of pricing power. If American open-source frontier capability exists with GB300 backing, the floor on enterprise API pricing moves.
Government access changes. Reflection has cultivated DOE Genesis Mission ties. An American open-source frontier model is the natural counter to both the EU Mistral push and Chinese open-weights (DeepSeek V4 Pro, OpenPangu-2, Kimi K2-7, GLM 5.2).
SpaceX hedges its own model. If Grok plateaus or commoditizes, SpaceX’s higher-margin path is infrastructure rental — and being the home for the American open-source frontier lab is the strongest such positioning.

What this means for the rest of the market

For Amazon (Trainium) and Microsoft (Azure Maia): SpaceX has joined the top tier of AI infrastructure providers without prior cloud business. Trainium and Maia have to deliver more than chip parity — they have to deliver enterprise integration depth that SpaceX can’t replicate from a standing start.

For Nvidia: Three more years of guaranteed GB200/GB300 demand from SpaceX customers. Reinforces the durability of the data-center revenue base.

For Anthropic specifically: The Reflection deal is a long-run threat. If Reflection ships a credible frontier open-source model with Colossus-class compute behind it, Anthropic’s enterprise pricing power on coding and general-purpose intelligence comes under pressure. Anthropic’s response will likely be Project Glasswing-style defensibility plays — capabilities that benefit from restricted access, not commoditization.

Sources

CNBC, “SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion,” June 22, 2026
MLQ News, “SpaceX Signs $6.3B Compute Deal With Reflection AI for Colossus Data Center,” June 22, 2026
Yahoo Finance and Startup Fortune coverage on Reflection AI’s $25B valuation
AIToolsRecap, “AI News June 23 2026,” for SPCX trading data

Verified June 23, 2026.