Starting in October 2026, Google begins paying SpaceX $920 million every month for AI compute capacity. The full rate sits on a contract that runs through June 2029 and carries a total value of roughly $30 billion, and it hands Google access to about 110,000 Nvidia GPUs plus CPUs, memory and related components inside SpaceX data centers (CNBC). The headline number is startling, but the strategic signal matters more. A hyperscaler that spends more than $180 billion a year on its own infrastructure concluded it could not build enough capacity fast enough, so it started renting from a rocket company. Compute, not capital and not model quality, is the binding constraint on the 2026 AI build-out, and the firms that control energized, GPU-filled buildings now hold the leverage.
What Google is actually buying
The agreement was disclosed in a SpaceX regulatory filing on June 5, 2026, days before the company's initial public offering. Google will lease approximately 110,000 GPUs along with central processors, memory and other components housed in SpaceX data centers. The contract runs from October 2026 through June 2029 at the $920 million monthly rate, with capacity "ramping up through September at a reduced fee," according to the filing (CNBC; Data Center Dynamics).
The public reason is demand, not scarcity strategy. A Google Cloud spokesperson told CNBC the deal exists "to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected." That phrase, bridge capacity, is the honest description of what nearly every large AI buyer is now doing: buying time from any supplier with spare GPUs while its own data centers catch up.
Why the October trigger matters
Two dates in the filing define the risk on both sides. The first is September 30, 2026. SpaceX stated that if it fails to "deliver access to the committed amount of GPUs" by that date, Google can immediately end the agreement, or accept the number of GPUs actually provided with a proportional reduction in monthly fees after a one-month grace period. That deadline has now passed, and the clock has moved to the full $920 million monthly billing that begins in October.
The second is December 31, 2026. After that point, either party can terminate the contract on 90 days' notice. Together the clauses show how these deals are structured in 2026: enormous headline commitments, short leash, generous exit. Google is not buying a data center. It is buying an option on capacity with a defined escape hatch, because no buyer trusts that demand, power and delivery schedules will line up exactly as planned.
The scale is easier to grasp annually. At $920 million a month, Google commits more than $11 billion a year to a single supplier through 2029. The contract sits alongside Alphabet's broader ramp: the company revised its 2026 capital expenditure forecast to a range of $180 billion to $190 billion, up from a prior estimate of $175 billion to $185 billion, and said it plans to sell $85 billion in stock, including a $10 billion investment by Berkshire Hathaway, to meet "unprecedented customer demand" (CNBC). Even for Alphabet, renting is now part of the strategy.
SpaceX's pivot from rocket maker to AI landlord
The supplier side of this deal is barely recognizable as the company that launched Starlink. SpaceX merged with xAI, Elon Musk's AI company, in February 2026 in a transaction that valued the combined entity at $1.25 trillion. That merger folded Grok's data centers, including the Colossus campus in Memphis, Tennessee, into a business whose primary revenue was rockets and satellite internet. SpaceX's annual revenue was only about $18 billion per year before the compute deals, which is what makes the hosting business so material (Axios).
The economics cut both ways. In the first quarter of 2026, SpaceX's capital expenditures totaled $10.1 billion, more than double the prior year, with $7.7 billion of that committed to AI, while the AI segment recorded an operating loss of $2.5 billion on just $818 million in revenue (CNBC). Filling that gap with paying tenants is the point. "We believe our compute infrastructure and related strategy provides us with substantial flexibility in how we allocate and monetize capacity," SpaceX wrote in its IPO filing.
There is a physical nuance worth noting. The Colossus 2 facility came online in January 2026. Musk claimed it offered 1 gigawatt of capacity at launch, but satellite imagery taken that month reportedly showed cooling equipment capable of managing about 350 megawatts, a reminder that announced AI capacity and energized AI capacity are different numbers (Data Center Dynamics).
Anthropic's larger bill and the copycat pattern
Google is the second major tenant, not the first. In May 2026, Anthropic agreed to pay SpaceX $1.25 billion per month through May 2029, a contract that could generate more than $40 billion for the supplier. The deal gives Anthropic full access to the Colossus 1 data center in Memphis, with more than 300 megawatts of capacity and over 220,000 Nvidia GPUs across H100, H200 and GB200 accelerators (TechCrunch; Axios). Anthropic co-founder and chief compute officer Tom Brown said the company was expanding into Colossus 2 on Nvidia GB200 capacity throughout June.
The subtext is hard to miss. Reporting on the filing noted that Grok usage had dropped in recent months, freeing servers that xAI then sold to one of the closest competitors in frontier AI, and the filing framed the arrangement as a way to "monetize unused compute capacity in our infrastructure" (TechCrunch). This is the defining trade of the compute crunch: an AI lab that overbuilt capacity becomes a landlord, and a rival that underestimated demand becomes a tenant. The pattern has continued, with reporting pointing to additional hosting agreements and other model makers signing multi-year GPU leases of their own.
The neocloud economy: renting GPUs at scale
SpaceX is a newcomer to a market that already has specialists. The leasing firms are collectively known as neoclouds, and the two public benchmarks are CoreWeave and Nebius. CoreWeave (Nasdaq: CRWV) posted about $2.08 billion in first-quarter 2026 revenue and an order book near $99.4 billion, while still recording a net loss of roughly $740 million for the quarter. Nebius (Nasdaq: NBIS) reported about $400 million in first-quarter revenue, up roughly 684 percent year over year, and second-quarter AI cloud revenue of $575 million, up 514 percent (Penchan; Data Center Dynamics).
Growth and losses arrive together because these companies are in permanent build-out. CoreWeave's full-year 2025 net loss was about $1.17 billion on roughly $5.1 billion of revenue, and its second-quarter 2026 net interest expense reached $640 million, up from $267 million a year earlier, against a debt load reported at $35 billion or more (MarketWise). Nebius is funding expansion differently. Roughly 70 percent of its signed deals include customer prepayments, with $9 billion in prepayments reported for 2026 alone, which reduces its reliance on debt (MarketWise). When Google signed with SpaceX, it was choosing among these options, and SpaceX was named in its own filing as a competitor to CoreWeave and Nebius.
Why compute sold out
The deals only make sense against a market where on-demand GPU capacity is effectively gone. Analysis cited by Apollo found that on-demand capacity is sold out and even older-generation chips are seeing rental rates rise to levels not seen since 2024. H100 rental rates, which fell from about $3.00 per hour in early 2023 to roughly $1.85 in late 2025, climbed back to about $2.40 by March 2026 (Apollo).
Supply cannot move fast enough to clear the gap. Data center GPU lead times now run from 36 to 52 weeks, constrained by high-bandwidth memory and advanced packaging as much as by wafer capacity (VEXXHOST). The demand side is equally rigid: OpenAI's Stargate program alone has been tracked at roughly 7 gigawatts of planned active capacity and more than $400 billion in commitments, including a $300 billion Oracle compute purchase (Presenc AI). When every large buyer signs multi-year commitments years ahead of deployment, the buyers who did not lock in capacity early are the ones now paying $920 million a month for bridge supply.
The risks and what to watch
Three risks define the next year of this market. The first is concentration. Neocloud revenue leans heavily on a handful of customers, so a single contract decision at one buyer moves an entire supplier's financials, a dynamic already visible in CoreWeave's early dependence on a single large client. The second is financing. Debt-funded build-outs met rising interest costs in 2026, while prepayment-funded models shift risk onto customers who may themselves be unprofitable. The third is the overbuild question. If AI demand growth slows, the market could flip from scarcity back to surplus, and long-term take-or-pay style leases become liabilities rather than assets.
For anyone planning enterprise AI infrastructure, the practical read is to stop treating GPU capacity as a utility that will always be available at a falling price. Capacity is contracted, priced and rationed, and the firms signing multi-year leases today are locking in both cost and availability. Watch three signals over the next two quarters: whether SpaceX delivers the committed GPUs to Google at the promised scale, whether hyperscaler capital expenditure guidance keeps rising or finally plateaus, and whether GPU rental rates on the spot market keep climbing toward their 2023 highs. Those three numbers will tell you whether 2026 was the year compute became the scarcest resource in technology, or the year the industry financed one build-out too many.