SpaceX becomes an AI infrastructure provider: Google pays $920 million per month for 110,000 NVIDIA GPUs
🔎 The most brutal reversal in the AI industry
In 2021, Google was providing compute to SpaceX for its internal needs. Five years later, Google is paying $920 million per month to SpaceX to rent NVIDIA GPUs. This complete role reversal, revealed in an SEC filing on June 5, 2026, by TechCrunch, marks a geopolitical shift in the artificial intelligence industry.
SpaceX is no longer just a space company. Following the acquisition of xAI in February 2026 and the recovery of its massive datacenters, Elon Musk's company has become one of the world's largest AI infrastructure providers. And this, the very same week it is preparing to go public with a valuation that could reach $1.75 trillion, according to analyses of the S-1 filed on May 20, 2026.
The irony is overwhelming: Google is literally financing the Musk empire while xAI, now a subsidiary of SpaceX, develops Grok 4.1 — a model that directly competes with Gemini 3 Pro in the general segment, with a score of 90 points compared to 92 for Google's model.
The essentials
- Google will pay SpaceX $920 million per month from October 2026 to June 2029, representing a total contract of around $30 billion for ~110,000 NVIDIA GPUs, CPUs, and memory, according to the SEC filing analyzed by TechCrunch.
- SpaceX acquired xAI in February 2026 via a triangular merger, taking over the Colossus datacenters and positioning AI as a key valuation driver for its IPO on the Nasdaq under the ticker SPCX.
- Anthropic pays $1.25 billion per month for the entirety of Colossus 1 (220,000 GPUs), a separate contract but one that confirms SpaceX as an industrial-scale AI infrastructure lessor.
- A pro-rata reduction clause protects Google if SpaceX fails to deliver the GPUs on time, suggesting significant execution risks.
Recommended tools
| Free AI APIs | Access to AI models without a credit card | Free (June 2026, check on ai-master.dev) | Testing and prototyping |
|---|---|---|---|
| Hostinger | Web hosting to deploy AI apps | Starting at 2.99€/month (June 2026, check on hostinger.com) | Production deployment |
The figures of the Google-SpaceX contract decoded
$920 million per month. That is the price Google agrees to pay to access approximately 110,000 NVIDIA GPUs in SpaceX's datacenters, according to details published by Euronews. The contract runs from October 2026 to June 2029, with a ramp-up period.
The total amount approaches $30 billion over 32 months. That is enough to fund several Starship launchers, but Google is not paying for space — it is paying for terrestrial compute.
The math behind the $920M/month
Breaking the contract down to the GPU level, we get about $8,360 per GPU per month. This figure includes CPUs, memory, cooling, electricity, and maintenance. It is above the bare rental price of an NVIDIA GPU in a traditional datacenter, but the AI compute market has been in a structural shortage since 2024.
Frandroid highlights this irony: Google, which owns its own TPU chip and datacenters all over the world, is forced to rent from a competitor because it cannot deploy enough NVIDIA GPUs internally.
The pro-rata reduction clause
The SEC filing reveals a crucial detail: a pro-rata reduction clause if SpaceX fails to deliver the GPUs on time. This clause protects Google against deployment delays, a real risk given that SpaceX is building its datacenters at an unprecedented pace.
This also indicates that the contract is structured around delivered capacity, not promised capacity. Google only pays for what is actually operational.
How SpaceX became an AI compute giant
SpaceX's transformation into an AI infrastructure provider does not date back to June 2026. It is the result of two major strategic moves.
The acquisition of xAI in February 2026
SpaceX absorbed xAI via a triangular merger — a legal structure that protects SpaceX from the risks associated with xAI's litigation, as explained by MEXC dans son analyse de l'IPO. xAI becomes a wholly-owned subsidiary of SpaceX, and its assets — notably the Colossus datacenters — come under the direct control of the space company.
It is this acquisition that gave SpaceX the 220,000+ GPUs needed to sign contracts of this magnitude. Without xAI, SpaceX did not have the compute capacity. Without SpaceX, xAI did not have the balance sheet to finance the expansion.
The space + AI + telecoms merger
TF1 Info note que SpaceX is strengthening its position well beyond space and Starlink. The company is now building three pillars of value creation: orbital launch, satellite telecommunications, and AI compute.
The SEC filing referenced by Blockspace even mentions an "AI compute satellite" — an orbital datacenter project with up to one million satellites. This is speculative at this stage, but the fact that SpaceX mentions it in a regulatory document shows the strategic direction.
Anthropic vs Google: two contracts, one same provider
The Google-SpaceX contract is not an isolated case. Anthropic signed a separate agreement for Colossus 1, with 220,000 GPUs and 300 MW of power, worth $1.25 billion per month according to Geo TV.
Comparison of the two contracts
| Criterion | Google-SpaceX Contract | Anthropic-SpaceX Contract |
|---|---|---|
| Allocated GPUs | ~110,000 | 220,000 |
| Monthly price | $920M | $1.25M |
| Price per GPU/month | ~$8,360 | ~$5,680 |
| Duration | 32 months (Oct. 2026 – June 2029) | Not publicly specified |
| Infrastructure | SpaceX datacenters (ex-xAI) | Dedicated Colossus 1 |
| Pro-rata clause | Yes | Not specified |
The Anthropic contract is more cost-effective per GPU, which is explained by the double volume and probably by a longer-term commitment. Anthropic uses this compute to power Claude Opus 4.7 Adaptive, its flagship agentic model which scores 94.3 points on reference benchmarks.
What these contracts reveal about the market
Two of the world's largest AI companies are renting their infrastructure from SpaceX, not from AWS, Azure or GCP. This speaks volumes about the NVIDIA compute shortage. Traditional hyperscalers are failing to meet demand, creating a secondary market dominated by non-traditional players like SpaceX.
Voice AI also illustrates this tension: companies like ElevenLabs are surpassing $500 million in ARR partly because compute has become the bottleneck for the entire industry.
The irony of Google funding Grok
This is the point that Frandroid highlights: Google pays $920 million a month to a company whose subsidiary xAI develops Grok 4.1, a direct competitor to Gemini 3 Pro.
Grok vs Gemini: a competition funded by the loser
Grok 4.1 reaches 90 points overall, closely trailing Gemini 3.1 Pro at 92 points. In agentic, the gap widens: Claude Opus 4.7 dominates at 94.3, but Grok 4.1 remains competitive at 79 points compared to OpenAI's GPT-5.4 Pro at 91.8.
Every dollar Google pays to SpaceX indirectly strengthens xAI's capacity to train models that threaten Google's position. It's a vicious cycle: Google needs compute to stay competitive, but the only compute available in sufficient quantities belongs to its competitor.
Why Google has no choice
Google has its own in-house TPUs, but the most performant models — Claude Opus 4.7, GPT-5.5, Gemini 3 Pro Deep Think — are all trained primarily on NVIDIA GPUs. The transition to in-house chips is slow, and the market isn't waiting.
According to TechTimes, the scarcity of AI compute is pushing tech giants into unnatural alliances. The Google-SpaceX contract is the most striking illustration of this.
The SpaceX IPO: AI compute as a valuation narrative
The timing is no coincidence. The Google contract was revealed on June 5, 2026. The SpaceX IPO is targeted for June 12, 2026, according to the analysis of the S-1 on LinkedIn. The company is aiming to raise $5 billion and achieve a valuation of up to $1.75 trillion.
Three pillars to justify $1.75T
SpaceX's valuation argument now rests on three distinct narratives:
- Space launch: Starship, Falcon 9, NASA contracts. The historical pillar.
- Telecoms: Starlink, which already generates billions in recurring revenue.
- AI compute: The ex-xAI datacenters, the Google and Anthropic contracts, the orbital datacenter project.
The third pillar is entirely new. It did not exist a year ago. And it could represent a significant portion of the IPO valuation if investors extrapolate the recurring revenues from the compute contracts.
What investors will take away
Coinfomania points out that this deal arrives "just ahead of SpaceX's historic IPO." The signal is clear: SpaceX wants to show the markets that it is not dependent on a single sector, and that AI is a growth driver in its own right.
With $920M/month from Google and $1.25M/month from Anthropic, SpaceX potentially generates over $2.1 billion per month in compute revenue — or more than $25 billion per year annualized. This is a figure that alone justifies a significant fraction of the target valuation.
What this means for AI developers and businesses
If you are building AI products, this deal has direct consequences for your business, even if you are not Google.
The compute shortage will last
When Google — which owns datacenters all over the world — has to rent from SpaceX, it is a sign that the NVIDIA compute shortage is structural. Models like GPT-5.5 (98.2 in agentic), Claude Opus 4.7 (94.3) and Gemini 3 Pro Deep Think (95.4) require colossal volumes of compute to train.
For developers, this means that API costs will not drop as much as hoped. Demand exceeds supply, and providers pass on the cost of compute in their pricing.
Alternatives become strategic
This is where APIs IA gratuites via providers like Groq, Google, or OpenRouter make all the sense in the world. For prototyping and low-volume use cases, these alternatives make it possible to partially bypass the cost constraint.
For production, the choice of hosting becomes critical. A hébergeur comme Hostinger at €2.99/month does not solve the GPU compute problem, but it allows you to deploy AI applications developed with external APIs without adding a prohibitive infrastructure cost layer.
Less resource-intensive models gain relevance
In a context of scarce compute, models like DeepSeek V4 Pro (88 points in general) or Claude Sonnet 4.6 (83 points) become rational choices for many use cases. Not everyone needs GPT-5.5 to generate content or automate simple tasks.
The risks of this deal for Google
Not everything is rosy on the Mountain View side. This deal carries significant risks that analysts seem to underestimate.
Dependence on a competitor
Google is relying on SpaceX for a critical part of its AI infrastructure. If the relationship deteriorates — and with Elon Musk, that's always a possibility — Google finds itself without compute at the worst possible time.
The pro-rata clause partially mitigates this risk, but it does not solve the underlying issue: in the event of a conflict, SpaceX's capacity to disrupt Google's AI infrastructure is considerable.
A price that could be overvalued
$8,360/month per GPU is expensive. Very expensive. If the compute market loosens — for example, if NVIDIA significantly increases its production or if Google's in-house chips (TPUs) reach performance parity — Google will find itself locked into an overvalued contract for 32 months.
The signal sent to the market
By signing this deal, Google is publicly admitting that its internal infrastructure is not enough. This is an admission of weakness that competitors — Microsoft with OpenAI, Amazon with Anthropic — will heavily leverage in their sales narratives.
The orbital datacenter project: speculation or reality?
SpaceX's SEC filing mentions an "AI compute satellite" and a "SpaceX Orbital Data Center System" capable of scaling up to a million satellites, according to SentiSight. This is the most speculative part of the whole story.
Why orbital compute makes sense in theory
Space offers a major theoretical advantage for datacenter cooling: the vacuum of space allows for almost unlimited thermal dissipation. Terrestrial datacenters devote a massive share of their energy to cooling. In space, this constraint disappears.
Solar energy is also continuously available in orbit, without terrestrial intermittencies. An orbital datacenter powered by solar panels could operate 24/7 without grid contribution.
Why it is unrealistic in the short term
Launch costs, even with Starship, make orbital computing prohibitive today. The communication latency between space and Earth is a problem for AI models that require real-time interactions. And the maintenance of electronic hardware in the space environment remains a major challenge.
This project likely serves more as a narrative for the IPO than as a short-term operational plan. But the fact that SpaceX positions it in an SEC filing shows they are seriously thinking about it on a 5-10 year horizon.
❌ Common mistakes
Mistake 1: Confusing SpaceX and xAI
xAI has no longer existed as an independent entity since February 2026. The triangular merger made it a subsidiary of SpaceX. Saying "Google leases from xAI" is technically wrong — it leases from SpaceX, which controls xAI's former assets. This distinction has a real legal and financial impact.
Mistake 2: Comparing the per-GPU price with public cloud rates
The $3,600/GPU/month from Google's contract includes much more than just the GPU: CPUs, memory, cooling, electricity, maintenance, and likely a scarcity premium. A direct comparison with the spot price of a GPU on AWS or Azure is misleading.
Mistake 3: Thinking Google is the only client
Anthropic pays $1.25M/month for Colossus 1. Other clients could follow. SpaceX is building an industrial-scale compute rental business, not a one-off deal with Google. The case of the company that blew up its Claude bill through incompetence shows that companies are burning through compute at unpredictable rates, creating structural demand.
Mistake 4: Downplaying the impact on the IPO
Some analysts treat this deal as a minor complement to SpaceX's space narrative. This is a mistake. At $25B/year in potential compute revenues, this business could represent a significant portion of the $1.75 trillion target valuation.
❓ Frequently Asked Questions
Why doesn't Google build its own NVIDIA datacenters?
Google is investing heavily in its in-house TPUs, but cutting-edge models are optimized for NVIDIA GPUs. The transition is slow, and current demand exceeds internal deployment capacity, even for a giant like Google.
Do Anthropic and Google share the same GPUs?
No. Anthropic has a dedicated Colossus 1 with 220,000 GPUs. The Google contract covers ~110,000 GPUs in a separate infrastructure. The two contracts are distinct, even though they rely on the same provider.
What happens if SpaceX doesn't deliver the GPUs?
The pro-rata reduction clause protects Google: it only pays for the capacity actually delivered. This shifts the delay risk onto SpaceX, which must therefore accelerate deployment to receive the full $920M/month.
Does this deal affect AI API prices?
Indirectly, yes. The high cost of compute is reflected in the pricing of models like GPT-5.5 or Claude Opus 4.7. Developers can mitigate this by using APIs IA gratuites for prototyping and lighter models like Claude Sonnet 4.6 for production.
Is SpaceX really going to put datacenters in space?
The "SpaceX Orbital Data Center System" is mentioned in the SEC filing, but it is a long-term project. The technical and financial challenges are considerable. In the short term, the vast majority of compute remains on Earth.
✅ Conclusion
The $920M/month Google-SpaceX deal is not just a GPU leasing contract — it is confirmation that AI infrastructure has become a market parallel to traditional cloud, with its own rules, its own players, and its own prices. By repurposing xAI datacenters after its February 2026 merger, SpaceX created a compute business in just a few months that could generate $25 billion a year. The fact that Google — a direct competitor of xAI via Grok — is forced to pay to access it says everything about the structural shortage hitting the industry. For developers, the lesson is simple: compute will remain expensive, and architectures that optimize model usage — starting with free AI APIs for prototyping — are no longer a luxury but a necessity.