KEY POINTS
- Google has agreed to pay SpaceX $920 million per month to rent 110,000 NVIDIA GPUs, CPUs, and memory from October 2026 through June 2029 — a contract worth roughly $33 billion at full run rate.
- The deal signals that even Google, with its own TPU silicon program and massive internal infrastructure budget, cannot build or procure AI compute fast enough to meet internal demand.
- Traders should watch whether this lease structure — and SpaceX's parallel Anthropic GPU deal for 220,000+ chips — compresses available NVIDIA hardware supply for other buyers in the October 2026 ramp window.
Google has agreed to pay SpaceX $920 million per month to rent 110,000 NVIDIA GPUs, CPUs, and memory starting October 2026, a filing confirmed — a contract that runs through June 2029 and totals roughly $33 billion at full run rate, making it one of the largest compute infrastructure agreements ever disclosed.
What the SpaceX Filing Actually Reveals
The terms embedded in the SpaceX filing are more operationally specific than most infrastructure deals that reach public disclosure, and the specificity is itself the story. Capacity begins ramping in September 2026, with the full $920 million monthly obligation kicking in from October. SpaceX has a one-month grace period to cure delivery failures before Google can terminate — a clause that implies Google's legal team views non-delivery as a real operational risk, not a boilerplate hedge. Both parties also retain the right to walk away with 90 days' notice after December 31, 2026, which means the earliest either side can cleanly exit is around April 1, 2027. That termination window will matter to traders watching for any strategic pivot from Google toward its own TPU buildout or alternative GPU suppliers.
The hardware at the center of this deal sits at SpaceX's Colossus 1 data center, the same facility where SpaceX had previously agreed to rent its entire GPU capacity — more than 220,000 NVIDIA chips — to Anthropic. The geographic and logistical overlap between those two deals raises an immediate question that the filing does not answer: whether SpaceX is operating Colossus 1 as a multi-tenant facility with segregated workloads for Anthropic and Google, or whether the Anthropic arrangement applies to a different tranche of capacity. If it is the latter, SpaceX's total GPU estate would need to exceed 330,000 NVIDIA chips — a figure that would make it one of the largest private holders of NVIDIA accelerator hardware in the world, ahead of most sovereign AI programs and comparable to mid-tier hyperscaler clusters.
The pricing math is equally instructive. At $920 million per month for 110,000 GPUs, the implied monthly per-GPU cost is approximately $8,364. For context, NVIDIA's H100 SXM5 units were trading on the secondary market at roughly $25,000 to $30,000 per card earlier this year, implying a payback period for SpaceX of three to four months per GPU at this rental rate — an extraordinary return on hardware capital if the utilization assumptions hold. This is the GPU rental market becoming an asset class in its own right, and SpaceX is positioning itself as the dominant infrastructure landlord at a moment when NVIDIA's own production ramp cannot satisfy hyperscaler timelines.
Google's TPU Problem and What It Means for the Compute Race
The strategic embarrassment buried inside this deal is impossible to ignore. Google invented the tensor processing unit. The company has been designing custom AI silicon since 2016, has deployed TPU v5 clusters internally, and publicly touts its vertical integration as a competitive moat against Microsoft Azure and AWS. And yet, as of June 2026, Google is writing SpaceX a check for nearly $1 billion a month to rent another company's NVIDIA GPUs — hardware that Google's own chip program was theoretically designed to make unnecessary.
The explanation is scale and timing, not competence. NVIDIA's Q1 FY2027 data center revenue hit $75.2 billion, up 92% year-over-year, a number that reflects demand so far ahead of supply that even well-capitalized hyperscalers with internal alternatives cannot close the gap through organic buildout alone. Google's Gemini 3.5 Pro, currently in limited Vertex AI enterprise preview with a general availability window expected between June 23 and June 30, requires compute at a scale that makes the SpaceX arrangement a bridge strategy rather than a permanent solution — but bridges that cost $33 billion at full run rate have a way of becoming load-bearing structures. The 2-million-token context window that Gemini 3.5 Pro is confirmed to support at its $250/month Ultra tier demands inference compute at an order of magnitude beyond prior-generation models, and Google cannot afford to throttle capacity during a competitive window when Anthropic's Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, and OpenAI is simultaneously launching a self-serve Ads Manager inside ChatGPT.
Gartner's projection that 40% of domestic AI data centers will face severe power constraints by 2027, with up to 50% of facilities scheduled to open in 2026 already stalling due to grid connection delays, provides the structural context. More than $1.5 trillion in proposed physical infrastructure is currently trapped in permitting bottlenecks. When you cannot build fast enough and cannot connect to the grid, you rent from whoever already has the hardware plugged in — and SpaceX, apparently, has the hardware plugged in.
Antitrust Risk and the NVIDIA Concentration Problem
The SpaceX-Google deal does not exist in a regulatory vacuum. Democratic senators including Elizabeth Warren, Ron Wyden, and Richard Blumenthal have already written to the FTC and DOJ demanding scrutiny of NVIDIA, Meta, and Google for potential antitrust violations in AI infrastructure. The letter specifically flags Meta's $14.3 billion investment in Scale AI and Google's $2.4 billion nonexclusive licensing agreement with Windsurf as transactions that warrant examination. A $33 billion GPU rental agreement where a single private company — SpaceX — controls enough NVIDIA hardware to serve both Anthropic and Google simultaneously will not escape notice.
The DOJ's posture has already been telegraphed. A former DOJ prosecutor stated publicly that "algorithmic collusion and anticompetitive conduct involving AI will be a key area of enforcement for the DOJ over the next decade." The FTC has an active probe into Microsoft's cloud and AI businesses. The structural risk for investors is that GPU concentration creates antitrust surface area: if SpaceX controls access to compute for two of the three frontier AI labs competing directly with OpenAI, regulators will eventually ask whether that arrangement is competitively neutral. The Trump administration's AI Action Plan signals a pro-innovation posture and a preference for applying antitrust tools "with sensitivity to the unique characteristics of AI-driven markets," but that language offers no immunity — it offers deference, and deference has limits when market concentration becomes visible enough to generate Senate letters.
NVIDIA's $20 billion bond deal, launched to fund its own AI buildout with capex expected to reach $7.9 billion in 2026 versus $6 billion in 2025, suggests the company itself is accelerating supply as fast as its balance sheet allows. But the SpaceX-Google deal is evidence that the gap between demand and available supply is still wide enough for a rocket company to build a $33 billion annualized infrastructure leasing business on top of it. Qualcomm's reported $8 to $10 billion talks to acquire RISC-V chip designer Tenstorrent is the market's medium-term answer to NVIDIA dependency — but Tenstorrent hardware at scale is a 2028 story, not a 2026 one. For now, whoever controls the GPU real estate controls the AI race, and the October 2026 ramp date on the Google-SpaceX contract is the next concrete catalyst to mark on the calendar.
