Gemini Routing Risk After Google's $920M SpaceX GPU Lease

Google — widely estimated as the world's largest single owner of AI compute — is buying external GPU capacity. On June 5, 2026, SpaceX disclosed in an SEC filing that Google will pay $920 million per month from October 2026 through June 2029 for access to roughly 110,000 NVIDIA GPUs, CPUs, and memory at SpaceX-controlled data centers.

This is the second SpaceX compute lease in two weeks. Anthropic announced a $1.25 billion/month deal in late May for the full Colossus 1 facility in Memphis — 220,000+ NVIDIA processors and 300 MW of power. Google's contract covers approximately half that footprint.

For engineering teams routing AI traffic, this is not background news. It is a structural signal about where GPU supply is tightening, who depends on whom, and what your fallback policy should assume about Gemini capacity through 2029.

What the filing says

The filing includes several operational details that matter:

Pricing and term. $920M/month from October 2026 through June 2029, with a ramp-up period through September 2026 at a reduced fee. That works out to roughly $8.35/GPU-hour at full rate — competitive with on-demand cloud GPU pricing but structured as committed capacity.

Cancellation clauses. Either party can terminate with 90 days' notice after December 31, 2026. This is a hedge, not a marriage. If demand softens or SpaceX needs to reallocate capacity, both sides have an exit.

Delivery SLA. If SpaceX fails to deliver committed GPU capacity by September 30, 2026, Google can either terminate immediately (after a one-month grace period) or accept fewer GPUs at reduced fees.

No data center identified. Unlike the Anthropic deal, which names Colossus 1 specifically, the filing does not disclose which SpaceX facility Google will use. Elon Musk has said Colossus 2 is reserved for xAI.

Why Google needs external GPUs

Google's public statement frames this as "bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise." But the capex context tells a sharper story:

Alphabet has committed to more than $180 billion in capital expenditures this year and has signaled that figure will "significantly increase" in 2027. It recently announced an $80 billion equity sale to fund infrastructure buildout. Google is spending aggressively — but the demand curve for Gemini Enterprise is outpacing even that trajectory.

This is a meaningful signal for anyone routing to Gemini APIs. When the provider itself is capacity-constrained enough to lease external GPUs at nearly a billion dollars a month, the probability of rate-limit tightening, latency spikes, or quota adjustments during peak periods goes up.

SpaceX as a compute marketplace

SpaceX is positioning itself as something new in the AI infrastructure stack: a neutral GPU landlord. Both Anthropic and Google — direct competitors in the model and agent platform wars — are buying compute from the same seller, with similar terms, at roughly the same time.

This creates a new dependency layer that did not exist six months ago. If SpaceX encounters operational issues at its Memphis or other facilities, two of the top three AI API providers are affected simultaneously. For routing teams, this is a correlated-failure scenario worth modeling.

The structural comparison:

Both contracts include 90-day exit windows starting January 2027. If the AI demand cycle softens — or if a new GPU generation makes these clusters less cost-effective — both providers could unwind these leases within a quarter.

What to watch in your routing policy

1. Gemini capacity signals. Google's own infrastructure is maxed. Watch for Gemini API rate-limit changes, quota reductions, or latency increases during Q4 2026 as external capacity comes online and internal demand continues growing.

2. Provider correlation risk. Anthropic and Google now share upstream compute infrastructure via SpaceX. If you run both as primary and fallback, a SpaceX facility-level incident could degrade both simultaneously. Consider an independent third provider (DeepSeek, Azure-hosted OpenAI, self-hosted open-weight models) as your true disaster-recovery path.

3. Pricing trajectory. $8–9/GPU-hour for committed H100-class capacity is the new floor. If your cost model assumed declining GPU prices through 2027, the SpaceX deals suggest demand is holding prices flat or pushing them higher.

4. Cancellation window in Q1 2027. Both contracts unlock 90-day termination rights after December 31, 2026. If either Anthropic or Google exercises the clause, it could release significant GPU capacity back to the market — or signal that the demand peak has passed.

The bottom line

Google buying external GPU capacity is the clearest signal yet that the AI compute market is a seller's market, even for hyperscalers. For teams running multi-provider routing, the operational takeaway is straightforward: your Gemini fallback assumptions need to account for a provider that is itself capacity-constrained, and your disaster-recovery plan should not assume that Anthropic and Gemini fail independently.

SpaceX, a company that was not in the AI infrastructure conversation twelve months ago, is now a single point of capacity dependency for two of the three largest AI API providers. That is a risk worth routing around.