Google’s SpaceX GPU Lease Has a Sept. 30 Deadline

SpaceX must deliver Google about 110,000 Nvidia GPUs by Sept. 30, 2026, under a deal that pays $920 million a month at full rate. The June 5 cloud contract also covers CPUs, memory, and related parts, with a reduced-fee ramp through September. Google Cloud described the arrangement as bridge capacity for Gemini Enterprise, and either side can walk after Dec. 31 with 90 days’ notice.

That combination, a giant monthly check sitting next to an easy exit, is the SpaceX Google deal investors actually bought when the rocket company went public in June.

Google Pays $920 Million a Month for 110,000 GPUs

SpaceX’s free writing prospectus, filed under Rule 433, says the company entered a Cloud Service Agreement with Google LLC on June 5, 2026. The compute on offer includes approximately 110,000 NVIDIA GPUs, plus CPUs, memory, and other related components. Google agreed to pay $920 million a month from October 2026 through June 2029, which is 33 months at the full rate, or $30.36 billion if every one of those months is collected.

The filing does not name the campus, the chip generation, or a service-level target. It does say Google keeps ownership of, and intellectual property rights in, its content, AI models, and related data. This is a hardware lease. It is not a model partnership, and it is not a joint lab.

THE GOOGLE CONTRACT TRIPWIRES

• The ramp: Capacity rises through September 2026 at a reduced fee before the $920 million monthly rate begins in October.
• The deadline: SpaceX has to deliver access to the committed GPU count by Sept. 30, 2026.
• The grace period: If it misses that date, Google waits one more month, then may end the deal at once or take fewer GPUs with a matching cut in the monthly fee.
• The walkaway: After Dec. 31, 2026, either party may terminate on 90 days’ notice.

SpaceX’s second-quarter 10-Q treats these cloud jobs as stand-ready service contracts with a single promise: continuous access to reserved compute, paid as fixed monthly fees. Revenue follows the ramp, then runs straight-line once the cluster is steady. The accounting matches a rental, not a construction project.

Bridge Capacity on a Multi-Year Invoice

Google Cloud’s own words cut against the IPO-week reading of a locked, multi-year backlog. A spokesperson said the companies are long-time partners and tied the spend to a product that had already outrun internal forecasts.

Google Cloud and SpaceX are long-time partners. This is a short-term, timely agreement 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.

Google Cloud spokesperson, statement on the SpaceX compute deal

The Gemini Enterprise agent platform is Google Cloud’s workplace agent stack, the product line that needs extra inference nodes when corporate customers show up faster than new halls can be powered. Google already designs its own TPU chips and still went outside for a block of Nvidia gear. That is a supply gap, not a strategy essay.

The calendar in the filing still runs into mid-2029. The language from Mountain View treats the same paper as a stopgap. Both can be true because of the December exit. After Dec. 31, 2026, Google can give 90 days’ notice and leave. SpaceX can do the same and pull the GPUs back for Grok, Cursor, or the next tenant who will pay more.

What Happens if SpaceX Misses Sept. 30?

If SpaceX cannot put the committed GPUs in Google’s hands by Sept. 30, 2026, the filing gives Google a one-month grace period, then a choice: kill the contract or keep whatever chips showed up and pay less. Full-rate billing is still scheduled to start in October, so a shortfall would hit the first invoices, not a distant option.

THE 2026 COMPUTE CALENDAR

1. February 2, 2026: SpaceX closes its acquisition of xAI, folding Grok, X, and the Colossus campuses into the rocket company.
2. May 2026: SpaceX signs cloud deals with Anthropic covering about 325,000 Nvidia GPUs across Colossus and Colossus II at $1.25 billion a month, with a reduced-fee ramp in May and June.
3. June 5, 2026: SpaceX and Google LLC sign the $920 million-a-month Cloud Service Agreement.
4. June 2026: SpaceX completes its IPO, selling 638.9 million Class A shares at $135.00 and taking in $85.675 billion net of fees.
5. June 30, 2026: Nameplate compute stands at 1.4 gigawatts, and the June quarter books $7.81 billion of group revenue.
6. September 10, 2026: CFO Bret Johnsen tells a Goldman Sachs conference that another hosting deal will pay about $1.11 billion a month from Dec. 1.
7. September 30, 2026: Google delivery test. Miss it, and the one-month grace starts.
8. October 2026: Google’s full $920 million monthly rate is scheduled to begin.
9. December 1, 2026: The unnamed hosting deal is scheduled to start.
10. December 31, 2026: The 90-day termination window opens for Google and SpaceX.

None of those later dates are in force yet. The date that is live is the September delivery test, because the reduced-fee ramp is the month now on the calendar.

Four Tenants and a 90-Day Out

Google is the named hyperscaler on the June 5 paper. It is not the only lab renting the same fleet. Anthropic is already on Colossus. Reflection AI has been paying $150 million a month since July. On Sept. 10, Johnsen added a fourth ticket without naming the buyer.

THE FOUR COMPUTE LEASES

Those monthly fees do not all hit in the same month, and they are not guaranteed through 2029. Stacked at full rate they would be $3.43 billion a month, but Johnsen has been explicit that SpaceX does not want long leases that trap chips inside someone else’s training run.

Almost all of them, if not all of them, have been, in essence, 90-day with a 90-day out, so roughly six-month commits.

Bret Johnsen, chief financial officer, Goldman Sachs Communacopia and Technology Conference, Sept. 10, 2026

That structure is the product. SpaceX can rent spare racks when Grok does not need them, then take the same silicon back when a new model run starts. Tenants get GPUs without waiting on their own substations. Nobody on either side is locked into a 33-month marriage unless they choose to stay.

Why the Grok Clusters Went on the Market

Colossus was built so xAI could train Grok, not so a rocket company could become a landlord. After SpaceX took in xAI on Feb. 2, 2026, the combined group leased Colossus 1 in full to Anthropic, the lab behind Claude, once its own teams had trouble using the Memphis campus as one coherent training cluster, including latency when tying it to halls more than 10 miles away. The machines still ran. The in-house model work did not get first claim on every rack.

The June quarter already shows what that choice did to the income statement. Group revenue was $7.81 billion, up from $4.07 billion a year earlier. The AI segment contributed $2.56 billion, of which $2.19 billion of AI infrastructure sales sat in the Solutions and Infrastructure line, versus $311 million a year earlier. One unnamed AI customer, labeled only as Customer B, accounted for 19.5 percent of group revenue, about $1.52 billion. Advertising on X added $367 million. The racks out-earned the feed.

Renting those racks to Google and Anthropic is a cash decision with a competitive twist. Grok still competes with Gemini and Claude. The same company is now selling them the electricity, cooling, and Nvidia boards those models need, while keeping a 90-day path to yank the boards back. The bet is that a dollar of rent arrives faster, and with less model risk, than a dollar of token revenue. It also means SpaceX’s chatbot is training on whatever capacity the tenants are not using.

The Memphis Campuses Were Built for Speed

The Google clock is tight because the campuses were stood up fast. SpaceX has been moving launch-side operations talent into the Tennessee and Mississippi AI halls after reliability problems, including sites that ran for months without backup cooling and power. Temporary chillers and rushed electrical work are fine when the only customer is your own research group. They are a different problem when Google and Anthropic are on the contract and an outage can idle someone else’s training job.

The company told investors it ended June with 1.4 gigawatts of nameplate compute, up from 0.4 gigawatts a year earlier, and that it wants a little more than 2 gigawatts by the end of 2026, then 5 to 10 gigawatts in 2027. Johnsen also said the new $1.11 billion-a-month hosting deal, starting Dec. 1, gave him “even more conviction” in a $100 billion annualized revenue run rate by year end. That $100 billion figure is an exit run rate, not 2026 recognized sales. The June quarter, at $7.81 billion, annualizes near $31 billion before the Google full-rate month and the December tenant even begin.

Backlog in the 10-Q was $47.46 billion at June 30, and that number mixes rockets, Starlink, and AI. It is not a substitute for the Google fee, which still has to clear Sept. 30, then survive a winter in which either party can give 90 days’ notice. If Google takes fewer GPUs, or walks after the grace period, the cleanest named contract from the IPO week shrinks on the first invoice. If the cluster is on time, SpaceX starts collecting $920 million a month in October for boards it can still reclaim in 2027. That is the lease. The model lab is the tenant that did not get a long-term claim on its own machines.

Disclaimer: This article is news reporting and analysis of SpaceX’s disclosed cloud contracts, public filings, and related company remarks, and it is for information only. It is not investment advice, a recommendation to buy or sell SPCX, GOOGL, NVDA, or any other security, and it is not a forecast of delivery, termination, or 2026 or 2027 revenue. Readers should consult a licensed financial adviser or securities professional before making any investment decision based on contract terms, run-rate targets, or capacity figures. Monthly fees, GPU counts, dates, and status details reflect the SEC filing, the June 30, 2026 Form 10-Q, and company comments cited above and may change if SpaceX misses delivery, if a party gives notice, or if later filings revise the numbers.