SpaceX’s IPO paperwork hides a contract that matters more to the AI hardware market than most of the rocket talk. On June 5, 2026, the company signed a Cloud Service Agreement with Google LLC, and the SpaceX filing on the SEC’s site says the deal covers approximately 110,000 NVIDIA GPUs, plus CPUs, memory and related components. The customer agreed to pay $920 million per month from October 2026 through June 2029.
Barron’s ran a story headlined “Hidden Gems in the SpaceX Filing: Soaring Compute Rates and Deep Tesla Ties,” which TradingView carried on October 28. That headline points to two threads. One is the size of the Google payment stream. The other is a set of Tesla transaction figures that jump from $4 million in 2024 to $147 million in 2025.
This analysis breaks down what the filing confirms, what the headline language adds on top, and what the numbers mean for GPU pricing, Tesla and the long-promised plan to put AI compute in orbit. If you want a baseline for what GPU time costs outside this deal, our guide to what renting an NVIDIA H100 costs per hour in 2026 gives the retail side of the same market.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The SpaceX Filing at a Glance: Key Terms of the Google Deal
Start with the contract itself, because every other number in this story hangs off it. The filing describes a compute-capacity arrangement, not a launch contract and not a Starlink service plan. SpaceX is renting out GPU capacity at scale, and the terms include a delivery deadline and an exit clause.
| Term | What the filing says |
|---|---|
| Signing date | June 5, 2026 |
| Counterparty | Google LLC |
| Capacity | Approximately 110,000 NVIDIA GPUs, plus CPUs, memory and related components |
| Monthly payment | $920 million per month, October 2026 through June 2029 |
| Ramp period | Capacity ramps through September at a reduced fee |
| Delivery deadline | September 30, 2026, followed by a one-month grace period |
| Termination | After December 31, 2026, either party may end the agreement on 90 days’ notice |
The filing does not say which NVIDIA generation the 110,000 GPUs belong to. Hopper, Blackwell and Rubin-class parts all carry very different price tags, so any per-GPU conclusion in this article stays a rough average and nothing more.
$920 Million a Month: The Math Behind the Headline
The monthly figure is easy to repeat and harder to put in context. Over the 33 months from October 2026 through June 2029, a flat $920 million works out to $30.36 billion. That is our arithmetic on the filed terms, not a number the filing states, and it excludes the reduced-fee ramp period before October.
| Metric | Calculation | Result |
|---|---|---|
| Monthly payment | Filed figure | $920 million |
| Annualized run rate | $920M x 12 | $11.04 billion |
| Total over the term | $920M x 33 months | $30.36 billion |
| Per GPU per month | $920M / 110,000 | About $8,364 |
| Per GPU per hour | $8,364 / 730 hours | About $11.46 |
Why the per-GPU figure is only a rough guide
The $8,364 monthly average divides one payment across GPUs, CPUs, memory and other components, so the true GPU-only rate sits lower. The unknown GPU generation also blurs the picture. Still, an average near $11 per GPU-hour gives readers an order of magnitude to hold against public cloud list prices and against the rental figures in our H100 guide.
What “Soaring Compute Rates” Does and Does Not Mean
The phrase “soaring compute rates” comes from the Barron’s headline. It does not appear in the filing as a defined term, and the filing excerpts we reviewed give no earlier rate to compare against. The one confirmed figure is the $920 million monthly payment.
So the honest reading is narrower than the headline. The contract shows a very large recurring payment for a very large block of GPUs. It does not show that SpaceX raised prices, that rates climbed over time or that Google overpaid relative to rivals. Those claims need a second data point, such as a prior SpaceX compute contract or a disclosed price list, and neither is on the record.
The headline still points at something real. A single customer committing $11.04 billion a year for GPU capacity tells the market that compute is scarce enough for a non-traditional supplier to land a hyperscaler as a buyer. That scarcity story has run through chip supply, memory pricing and data center power all year, and the filing adds a hard contract number to it.
Delivery Deadlines and Exit Ramps: The Fine Print
Big headline numbers often hide soft commitments. This contract has both a hard delivery test and an early exit. The filing states the delivery remedy in plain terms: “If we fail to deliver access to the committed amount of GPUs by September 30, 2026, then following a one-month grace period, Google may immediately terminate the agreement or accept the number of GPUs provided, with a corresponding pro rata reduction in the monthly fees.”
Read that carefully. If SpaceX delivers only part of the 110,000 GPUs, Google can keep the smaller allocation and pay proportionally less. The $920 million figure therefore depends on SpaceX actually standing up the full fleet, and the ramp period through September was the window to do it.
The second clause matters even more for forecasting. The filing says: “After December 31, 2026, the agreement may be terminated by either party upon 90 days’ notice.” That language cuts both ways. Google can leave, and so can SpaceX. A schedule that runs to June 2029 on paper can shrink to a few quarters of revenue in practice, which is why analysts who model the full $30.36 billion are making an assumption the filing does not guarantee.
Tesla Ties: From $11 Million to $147 Million
The second half of the Barron’s framing is the Tesla relationship. SpaceX disclosed commercial transactions with Tesla in four periods, and the pattern is anything but smooth.
| Period | Commercial transactions with Tesla | Change vs. prior period | Share of one month of the Google payment |
|---|---|---|---|
| 2023 | $11 million | Baseline | About 1.2% |
| 2024 | $4 million | Down about 64% | About 0.4% |
| 2025 | $147 million | About 37x the 2024 level | About 16% |
| Jan. 1 to Apr. 30, 2026 | $4.0 million | Four months only | About 0.4% |
The last two columns are our calculations from the filed figures. Across 2023 to 2025 the filing shows $162 million in transactions, and $147 million of that landed in 2025 alone. The first four months of 2026 dropped back to $4.0 million, which would annualize to roughly $12 million if the pace held. A simple extrapolation like that says little, because the filing gives no breakdown of what Tesla bought or sold.
Tesla has its own compute and energy appetite, and our coverage of Megapack demand from AI data centers shows how tightly the two companies’ hardware stories already overlap.
Reading the Tesla Numbers Without Overreach
The “deep Tesla ties” label is Barron’s framing. The fact sheet behind this story confirms the four dollar figures and nothing else about the relationship. It does not confirm what the 2025 transactions covered, whether they were one-time or recurring, or whether any of them tie into SpaceX’s compute business.
That limit matters because the temptation here is obvious. Elon Musk is listed in the filing as Chief Executive Officer, Chief Technical Officer and Chairman of the Board of SpaceX, and his name sits on the other side of many Tesla storylines. A reader can easily stack those facts into a theory of coordinated strategy. The filing supports none of it. It supports a spike in 2025 transactions, followed by a quiet start to 2026.
What the figures do support is a modest claim. Tesla is a real, disclosed commercial counterparty, and the relationship was large enough in 2025 to equal about 16% of a single monthly Google payment. Investors who track Musk-linked companies will want the next filing to say what changed.
The 2028 Orbital Compute Plan
The Google contract covers GPUs on the ground. The filing also lays out a second, far more speculative compute story. SpaceX expects to begin deploying orbital AI compute satellites as early as 2028. The company calls moving power-intensive AI workloads into orbit “the only truly scalable solution” to rising compute demand set against terrestrial energy constraints.
Two phrases deserve attention. “As early as” means the date is a target and not a commitment. And the company frames the case as an energy argument, not a cost argument. The quoted language frames the case around energy supply and not around cost.
Launch capacity is the other half of that bet. Putting meaningful compute into orbit takes a very high launch cadence, which ties the plan to the ground infrastructure we covered in Musk’s Starbase Louisiana ambitions for 10,000 Starship launches a year. The filing excerpts we verified do not put a cost or launch count on the orbital program, so those links remain context and not confirmed plan details.
Why Energy Limits Drive the Compute Pitch
The “terrestrial energy constraints” argument is not unique to SpaceX. Hyperscalers, neoclouds and chip designers all report the same bottleneck. GPUs arrive faster than the grid can power them, and local opposition to new data centers keeps growing. Our reporting on why data centers use over 2 million gallons of water daily covers the resource side of the same pressure.
SpaceX’s answer is to leave the grid entirely. If the orbital plan works, power becomes a satellite design problem and not a utility interconnect queue. If it fails, the company still holds the terrestrial GPU contract that pays $920 million a month. The two programs hedge each other, at least in the narrative the filing builds.
That structure also explains why the Google deal reads as a bridge. Ground capacity funds and proves the compute business while the orbital hardware remains a 2028 target. Whether SpaceX can run a GPU fleet to Google’s service standards is the nearer test, and the September 30 delivery clause turns that test into a contract term.
Market Impact: Who Gains and Who Feels Pressure
The immediate winner is NVIDIA. A block of roughly 110,000 GPUs, bought or leased by a company that is not one of the usual hyperscalers, adds another large buyer to a market that already struggles with supply. The filing does not name the GPU model, so the revenue split across product lines stays unknown.
Neocloud providers face the sharper question. SpaceX now sells GPU capacity to Google, a company with its own giant cloud. If a hyperscaler will rent capacity from an outsider at this scale, small GPU clouds lose some of their claim to being the only flexible overflow supply. Our look at the AI funding crunch and neocloud squeeze shows how little cushion some of those providers have.
Financing is the other pressure point. Large GPU fleets need large debt and equity raises. We reported earlier that Musk said SpaceX closed a $40 billion GPU debt deal, and a $920 million monthly receivable gives lenders something concrete to underwrite. The 90-day termination clause, however, caps how much comfort they can take.
Competitive Comparison: SpaceX Against Hyperscalers and Neoclouds
SpaceX enters compute from an unusual direction. Hyperscalers own data centers, networking and enterprise sales channels. Neoclouds own GPU inventory but depend on a few anchor customers. SpaceX brings launch capability and Starlink, but the filing excerpts we verified name only one compute customer.
That concentration is the biggest weakness. One contract worth $11.04 billion a year, cancellable on 90 days’ notice after December 31, 2026, makes SpaceX’s compute revenue as dependent on Google’s goodwill as a neocloud is on its largest tenant. Hyperscalers spread that risk across thousands of customers.
The upside is structural. No hyperscaler can offer a credible plan for compute in orbit by 2028, and no neocloud owns a launch business. If the orbital program slips, the comparison collapses back to GPU rental economics, where scale and financing decide the winner. For the broader spending picture across the largest buyers, see our preview of how four AI spenders face their earnings test.
Historical Context: From Rocket Company to Compute Seller
SpaceX built its reputation on reusable launch vehicles, then added Starlink as a recurring-revenue business. Selling GPU capacity to a hyperscaler is a third act that would have sounded odd a few years ago. The Verge covered the filing under the headline “SpaceX just filed for what could be the biggest IPO in human history,” and the compute contract shows why the story reaches beyond launch economics.
Tesla’s path offers a parallel. It started as a car maker and now sells batteries and energy storage to data center operators. Both companies took hardware expertise built for one market and pointed it at AI infrastructure demand. The timing matches a period when capital flows toward anyone who can deliver power and silicon at scale.
The filings themselves are public. SpaceX hosts its SEC filing details on its investor relations site, and readers who want to check our reading of the Google agreement can compare it against the primary document.
Five Predictions for SpaceX Compute
These are our forecasts, built from the filed terms and not from any company guidance or outside expert.
- The September 30 delivery test becomes the first public checkpoint. Expect analysts to ask whether SpaceX hit the full 110,000-GPU commitment or accepted a pro rata cut.
- The 90-day termination clause drives the stock narrative after January 1, 2027. Any notice from either party would move more sentiment than a launch milestone.
- Compute revenue gets its own line in analyst models. A $11.04 billion annual run rate is too large to bury inside a Starlink or “other” bucket.
- The Tesla line stays lumpy. The 2025 spike to $147 million looks like a project and not a trend, so later filings will likely show wide swings again.
- Orbital compute remains a slide through 2028. The “as early as 2028” target depends on launch cadence and satellite hardware that the filing does not yet price.
What to Watch Next
Four items will settle most open questions. First, whether SpaceX names the GPU generation behind the 110,000 units. Second, whether later filings explain the 2025 Tesla spike. Third, whether a second compute customer appears. Fourth, whether anyone invokes the termination clause once it opens after December 31, 2026.
For readers tracking the hardware side, the hardware section on Tech Insider follows GPU supply, pricing and AI chips as the story develops. The filing gives investors one firm number and several open questions. That mix is why this story will keep drawing attention long after the IPO headlines fade.
Frequently Asked Questions
What did SpaceX disclose about its Google deal?
SpaceX entered a Cloud Service Agreement with Google LLC on June 5, 2026. It covers approximately 110,000 NVIDIA GPUs, plus CPUs, memory and related components, and pays $920 million per month from October 2026 through June 2029.
How much is the SpaceX Google compute deal worth in total?
At the full monthly rate, 33 months of payments come to $30.36 billion. That is our calculation, and it assumes neither side terminates the agreement. The filing allows either party to exit on 90 days’ notice after December 31, 2026.
Did SpaceX raise compute rates?
The filing does not say so. “Soaring compute rates” is a headline characterization from Barron’s, and the only confirmed figure is the $920 million monthly payment. No earlier rate is on the record for comparison.
How much did SpaceX transact with Tesla?
The filing reports $11 million in 2023, $4 million in 2024, $147 million in 2025 and $4.0 million from January 1 through April 30, 2026. It does not break those totals down by product or service in the excerpts we verified.
When will SpaceX launch orbital AI compute satellites?
SpaceX expects to begin deploying them as early as 2028. That is a target and not a guaranteed date, and the filing excerpts we reviewed give no cost or satellite count.
Which NVIDIA GPUs are in the Google agreement?
The verified excerpts name NVIDIA but not the generation. Treat any claim that the fleet is Blackwell or Rubin as unconfirmed until SpaceX or Google says so.
What happens if SpaceX misses the delivery deadline?
If SpaceX does not deliver the committed GPUs by September 30, 2026, and a one-month grace period passes, Google may terminate immediately or accept the GPUs provided with a pro rata cut to the monthly fee.
