Google rations Gemini for Meta: when compute becomes the rarest resource in the AI planet

Google rations Gemini for Meta: when compute becomes the rarest resource on the AI planet

🔎 Two giants, one same wall

Two of the world's richest companies find themselves in an unprecedented situation: Google had to limit Meta's access to its Gemini models simply because it didn't have enough chips to meet demand. The information, revealed by the Financial Times on June 28, 2026, does not concern a startup desperate for GPUs. It's Meta — with a market capitalization of $1.5 trillion — being throttled by a competitor over a resource that neither of the two can manufacture at will.

The signal is stark. In 2026, the AI bottleneck is no longer the quality of the models, nor the talent of the researchers, nor even money. It's raw compute. The chips, the data centers, the electricity to run them. Everything else — transformer architectures, alignment techniques, fine-tuning — becomes secondary as long as there is no silicon available to execute inference at the required scale.

This event marks a structural turning point. The era where anyone with a good model and an API key could scale freely is over. Computing capacity has become a geopolitical resource, and those who possess it dictate the pace.

The essentials

• Google imposed limits on Meta's use of Gemini models after the latter requested more computing power than Google could provide, according to the Financial Times relayed by CNBC.
• The restrictions disrupted Meta's internal AI projects and pushed the company to ask its employees to save AI tokens, reports Yahoo Finance.
• Google Cloud's order backlog reaches $460 billion, but this paper wealth does not translate into instant capacity, according to TechTimes.
• Google signs a $920 million per month deal with SpaceX to rent ~110,000 Nvidia GPUs in bridging capacity, from October 2026 to June 2029, according to the WSJ and CNBC.
• Half of US data center constructions are delayed or canceled due to a lack of electricity and components, according to Tom's Hardware.

Recommended tools

What exactly happened — The facts

Around March 2026, Meta placed an order with Google Cloud for massive inference capacity on Gemini models. The requested volume exceeded what Google could deliver immediately. Rather than rejecting the contract, Google accepted but imposed usage caps — contractual throttling.

According to TNW, Google informed Meta that it could not provide the total amount of compute capacity desired. Google Cloud's $460 billion backlog, cited by TechTimes, illustrates the scale of the imbalance between supply and demand.

The direct consequence at Meta: some internal AI projects were delayed. The company sent a note to its employees asking them to use AI tokens more sparingly, as reported by Yahoo Finance. When a company burning billions in AI CAPEX has to ask its engineers to save tokens, the situation speaks for itself.

For Wedbush, cited by Seeking Alpha, this limitation is "a clear signal that compute demand exceeds supply." It is probably the most polite understatement of the year.

Why compute is the real bottleneck in 2026

The argument is simple but its implications are massive. The AI compute demand curve is growing exponentially. The supply curve — constrained by chip manufacturing (18-24 month cycles), data center construction (3-5 years), and grid connection (5-10 years) — is growing linearly.

Up to 70% of all memory chips produced globally in 2026 will be destined for AI data centers, according to Accuristech. AI data center spending will exceed $600 billion this year. Alphabet, Amazon, Meta, and Microsoft alone are expected to spend over $650 billion to expand their capacity, according to Tom's Hardware.

But this money isn't enough. The "AI Power Wall" has become the main bottleneck. According to EnkiAI, access to the electrical grid now dictates the future of AI infrastructure. You can have 110,000 Nvidia GPUs on hand — if the data center doesn't have the gigawatts required to power them, those chips are just office decorations.

The result: half of planned US data center constructions are delayed or canceled. The causes are twofold — electrical infrastructure shortages and component supply chain breaks from China. The power wall is not a future problem. It is here, now, and it explains why Google throttled Meta rather than simply ordering more machines.

The SpaceX deal — $920M/month for 110,000 GPUs

Faced with this constraint, Google made a decision that would have been unthinkable two years ago: outsourcing some of its compute to SpaceX. The deal, revealed in early June 2026 by the WSJ, CNBC and TechCrunch, provides for $920 million per month for 32 months — from October 2026 to June 2029.

The volume: approximately 110,000 Nvidia GPUs. The total cost over the contract's duration approaches $30 billion. This is "bridge capacity" — temporary capacity to hold things over until Google's own data centers are built and connected to the grid.

This deal speaks volumes about the structural desperation of the market. Google, which has its own TPU chip line and one of the world's largest data center networks, is forced to rent Nvidia compute at a premium price from a space company that has diversified into the cloud. The lesson: even Google's vertical integration is no longer enough to cover internal + customer demand.

The parallel with Meta's situation is illuminating. Meta uses Google's Gemini models for certain projects. Google does not have enough compute to serve Meta and its own needs. So Google pays SpaceX nearly a billion a month to unlock capacity — but this capacity goes to Google's internal projects first. External customers, even one the size of Meta, remain left on the sidelines.

Vertical integration as a competitive weapon

The Gemini/Meta episode reveals Google's number one structural advantage in 2026: complete vertical integration. Google controls the stack end-to-end — models (Gemini 3.1 Pro, Gemini 3 Pro Deep Think), cloud infrastructure (Google Cloud), custom chips (TPU), and proprietary data centers.

When capacity tightens, this integration becomes a massive defensive moat. Google's internal projects — search, Android integration, consumer products — get priority. External customers, even those paying hundreds of millions, are on a lower priority tier.

This is exactly what happened with Meta. Despite a commercial contract, Google chose to ration rather than deprioritize its own products. For a client relying on Gemini for critical projects, this is an existential risk. No SLA compensates for the fact that your provider can unilaterally reduce your access because its own needs have increased.

This dynamic explains why every hyperscaler is now building its own custom chip. Meta is developing its Iris chip to reduce its dependence on Nvidia and third-party clouds. OpenAI is working on Jalapeño, its inference chip with Broadcom. The goal isn't necessarily to outperform Nvidia — it's to no longer depend on a spot market where availability is zero and prices are exploding.

The custom silicon war — Iris, Jalapeño, and what's next

Gemini rationing for Meta accelerates a trend that was already underway: the flight to custom silicon. Every major AI lab is now investing massively in its own chips, not out of technological ambition but out of a need for survival.

Meta is simultaneously building a $10 billion data center in Alberta, Canada, and developing its Iris chip. The choice of Canada is not insignificant — access to low-cost hydroelectric power partially bypasses the AI Power Wall blocking projects in the United States. By owning the data center AND the chip, Meta eliminates two layers of dependency.

OpenAI, for its part, is pushing Jalapeno as an inference chip optimized for its own models. The idea is to reduce the cost per token by 50% on GPT-specific workloads. If you control the model AND the silicon, you optimize both at the same time — an advantage that no one can replicate from the outside.

On the Nvidia side, the announcement of Vera Rubin and N1X at GTC Taipei shows that the GPU leader is not standing still. The transition to ARM for GPUs is a signal that even Nvidia recognizes that the x86 architecture has become a thermal and energy bottleneck. But these chips will not be available in volume until 2027-2028. Until then, rationing is the norm.

Where models fit in — Rankings and availability

Compute rationing has a direct impact on end users. A theoretically superior model becomes useless if it isn't available when you need it. The June 2025 general-purpose LLM rankings provide a glimpse of the current hierarchy:

In agentic, the picture changes slightly:

But this table doesn't tell the whole story. When Google rations access to Gemini 3.1 Pro for a client like Meta, what does that mean for a developer calling the API? Rate limits become more aggressive. Latency increases during peak hours. The most powerful models are degraded in practice by infrastructure constraints. The Gemini vs ChatGPT vs Claude comparison takes on a new dimension when you realize that the effective availability of each model varies drastically depending on the provider's load.

For developers, the lesson is clear: redundancy is no longer a luxury, but a necessity. Relying on a single model provider — especially if that provider is also your competitor — is a calculated risk that can turn into a blind risk overnight.

What this means for developers and businesses

Compute rationing has concrete and immediate consequences for anyone building on AI.

Firstly, prices will go up. When supply is constrained and demand is exploding, pricing follows. The fixed-price contracts some negotiated in 2024-2025 will be renegotiated upward at every renewal. Google has no incentive to lower prices when clients are fighting to get capacity.

Secondly, multi-cloud becomes mandatory, not optional. If your product relies on Gemini for 80% of its inference, you are one rationing incident away from an outage. Distributing the load across multiple providers — Google, OpenAI, Anthropic, or even free LLMs for non-critical tasks — is now basic hygiene.

Thirdly, token optimization becomes a competitive advantage. Meta asking its employees to save tokens is not a sign of weakness — it's a sign of maturity. Every unnecessary token is wasted compute, and compute is the company's most expensive resource. Frameworks for prompt engineering, semantic caching, and routing models to the right capacity level will become core competencies.

Fourthly, free AI APIs or low-cost ones (Groq, OpenRouter) are no longer prototyping tools — they are safety nets for tasks where the premium model isn't strictly necessary. Intelligent routing between a free model for simple queries and a premium model for complex cases becomes a standard production architecture.

The backlog paradox — 460 billion promises

The figure is dizzying: 460 billion dollars of backlog for Google Cloud, according to TechTimes. That is more than the GDP of many countries. It is irrefutable proof that AI compute demand is structurally higher than supply.

But this backlog is also a problem. These are signed contracts, made commitments, capacity promised to clients who are already paying. Every day that Google cannot deliver this capacity is a day that trust erodes. Meta's rationing is not an isolated incident — it is the mathematical consequence of a backlog that cannot be absorbed by existing infrastructure.

Building data centers takes 3 to 5 years. Electrical grid connection can take even longer. Components — transformers, cooling systems, memory chips — are subject to globalized supply chains that include China, whose exports of certain components are now restricted. According to Tom's Hardware, this combination of factors has delayed or canceled half of US data center projects.

The paradox is as follows: the more companies sign large compute contracts, the more the backlog lengthens, the more delivery timelines stretch, the more rationing worsens. It is a vicious cycle fueled by the fear of missing out — every client over-orders hoping to get at least half of what they need, which artificially inflates demand and worsens the shortage.

The next two years — What is going to happen

The 2026-2028 period will be defined by one word: triage. Hyperscalers will have to choose — and are already choosing — which projects, which clients, which models get compute and which ones wait.

Google will continue to prioritize its internal products. The integration of Gemini into Android, Google Workspace, and Search is non-negotiable. External clients, even enterprise ones, will be in a second tier of priority. The $920M/month SpaceX deal buys time, but 110,000 GPUs are not enough to close a gap of hundreds of thousands of chips.

Meta will accelerate its independence plan. The Alberta data center, the Iris chip, the expansion of its own infrastructure — all of this takes on new meaning in light of Gemini rationing. Meta no longer wants to be in the position of begging a competitor for compute. The goal is self-sufficiency, even if it costs 10-20 billion dollars per year in additional CAPEX.

OpenAI and Anthropic will face the same constraints on Nvidia compute. Their agreements with Microsoft and AWS respectively give them priority access, but not unlimited access. The ranking of the best LLMs for coding will become unstable — a model can be number one one month and unavailable the next.

For end users and developers, the era of unlimited abundance is over. The architectures that will work will be those that optimize every token, that route intelligently between models, and that maintain redundancy as a core principle. The ChatGPT vs Gemini comparison will have to incorporate a new dimension: actual availability, not just raw performance.

❌ Common mistakes

Mistake 1: Thinking the problem is temporary

This is not a temporary hype linked to a model launch. It is a structural imbalance between demand that doubles every quarter and supply that grows by 20-30% per year. The AI Power Wall, documented by EnkiAI, is not resolved with more money — it is resolved with years of building electrical infrastructure.

Mistake 2: Believing a large contract protects you from rationing

Meta had a contract with Google Cloud. Google throttled them anyway. An SLA does not create chips that don't exist. Penalty clauses may exist, but they do not compensate for an AI project delayed by six months. The only real protection is to diversify your providers and, ideally, to own your own compute.

Mistake 3: Ignoring the energy cost in your calculations

920 million dollars per month for 110,000 GPUs is about 8,360 dollars per GPU per month. This price includes energy, cooling, and SpaceX's margin. If you calculate your inference costs by only looking at the price per token, you are ignoring the energy component which is becoming predominant. More efficient models — Claude Sonnet 4.6, DeepSeek V4 Pro — aren't just cheaper, they consume fewer watts per token, making them more resilient to rationing.

Mistake 4: Underestimating the speed of migration required

When your main provider throttles you, you cannot migrate in a week. Integrating a new model, running regression tests, adjusting prompts, setting up monitoring — all of this takes months. If you don't already have a multi-provider architecture in production, you are behind.

❓ Frequently Asked Questions

Why doesn't Google just buy more Nvidia GPUs?

Nvidia doesn't fabricate GPUs on demand. Production is limited by TSMC, which is itself constrained by co-processor capacity, HBM memory (70% of which goes to data centers according to Accuristech), and various other components. Even with a blank check, Google cannot get more chips than the global supply chain can produce.

Could Meta simply use its own Llama models instead of Gemini?

Meta does indeed use Llama internally, but certain projects require Gemini's specific capabilities — particularly in multi-layer reasoning with Gemini 3 Pro Deep Think or in mobile integration via the Google ecosystem. The model comparison shows that each model has distinct strengths. The reliance on Gemini is not a default choice, but a technical one.

Does the SpaceX deal mean anyone can rent compute from them?

No. These are 110,000 GPUs dedicated to Google in a 32-month contract worth nearly $30 billion. SpaceX is building this capacity specifically for this client. It is not a public cloud service — it is dedicated infrastructure that illustrates how far major players are willing to go to secure compute.

Will less powerful models become more popular because of rationing?

Yes, and it is already happening. Claude Sonnet 4.6 (score 83) or DeepSeek V4 Pro High (score 84) consume significantly less compute than Gemini 3.1 Pro (92) or GPT-5.5 (91) while remaining excellent for the majority of use cases. Automatically routing to the lightest model that can handle a given query will become standard practice.

✅ Conclusion

Gemini's rationing for Meta is not a contractual incident — it is the moment the AI industry collectively realized that compute, not models, is the resource that defines who wins and who waits. Labs that own their own silicon, their own data centers, and their own power connections will dictate the pace for the next two years. The others — even the wealthiest — will be rationed. If you are building on AI today, your number one priority is no longer finding the best model, but ensuring you will actually have the compute to run it when you need it.