Computational capacity has become the Achilles heel of the artificial intelligence industry — and not even the biggest players in the world are escaping it.
In March 2025, Google told Meta that it simply could not supply all the Gemini capacity the company wanted to purchase.
Sounds unreal, right?
Two of the biggest tech companies on the planet, and one could not meet the other’s demand.
But this episode is more than just behind-the-scenes gossip. It exposes something much bigger: AI infrastructure is hitting its limits, even with billions being poured into chips, data centers, and electricity every year. The demand for computation grew so fast that even the companies building the models are struggling to keep up. And when that happens, projects stall — including internal ones at a giant like Meta. 👇
What happened between Google and Meta
According to people familiar with the matter, Google notified Meta around March 2025 that it would not be able to provide all the Gemini capacity that Mark Zuckerberg’s company intended to buy. Meta‘s plan was straightforward: supplement its artificial intelligence infrastructure by using Google’s models across various internal operations. But the plan hit an unexpected wall — Google did not have the headroom to handle the volume Meta needed. The answer was blunt: we cannot deliver all of that right now.
Those restrictions, which remain in place, ended up disrupting and delaying several of Meta‘s internal AI projects. As a result, combined with a broader effort to trim AI costs, the company began encouraging its employees to be more efficient with so-called AI tokens — the units that measure model consumption. In other words, the directive became: save processing power whenever possible.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
And Meta was not the only one affected. Several other Google customers also felt the impact of these restrictions, though on a smaller scale. The reason Meta was hit so hard comes down to the exceptionally high volume of demand the company had for Google’s models. When you need a lot and the line is long, the fallout is bigger.
What stands out here is not just the incident itself, but what it represents for the industry. If Google — the company that built Gemini, operates some of the largest data centers in the world, and invests tens of billions a year in infrastructure — cannot keep up with all the demand, the problem is structural. It is not a one-off glitch. It is a sign that AI growth is outpacing the physical capacity to support it, and the consequences reach far beyond a negotiation between two tech giants. 🔍
The compute crisis nobody saw coming
For years, the tech market ran on a fairly predictable logic: demand grew, companies invested in infrastructure, and the cycle repeated at a manageable pace. The arrival of large language models — massive-scale language models like Gemini, GPT, and others — changed all of that. Training these models requires an absurd amount of processing power, and inference — the moment a model answers a question or generates content — also consumes resources at enormous scale. When these models were made available to the public en masse, demand simply exploded at a rate no forecast could accurately predict.
This surge in inference workloads — all the tasks required to run models after they have already been trained — has become one of the sector’s biggest challenges. The core issue is chips. NVIDIA’s GPUs, which power the vast majority of AI systems worldwide, have become a scarce and fiercely contested resource. The supply chain cannot expand at the same speed demand is growing. Data centers need to be built from scratch, which takes time, money, and — increasingly — a massive amount of electricity.
Because of all this pressure, especially from large enterprise clients like Meta, Google itself rushed to secure additional capacity. In early December 2025, the company signed a staggering deal worth 920 million dollars per month to lease computational capacity from SpaceX, Elon Musk’s company. In other words, even Google is not relying solely on its own infrastructure — it is also seeking outside reinforcements to avoid falling behind.
What happened between Google and Meta is a direct reflection of this environment. When the available computational capacity is less than the total demand — whether from external clients or internal projects — companies have to make choices. And those choices carry enormous strategic weight. In the world of AI, having the money to pay is not enough — you still need someone with the infrastructure to actually serve you. 💡
The numbers that explain how big this problem really is
To grasp the scale of this race, it is worth looking at what Google itself revealed. In its first-quarter 2025 results, released in April, CEO Sundar Pichai shared that the company’s cloud revenue surpassed 20 billion dollars for the first time. On top of that, the backlog of cloud contracts already signed but not yet delivered nearly doubled from one quarter to the next, reaching more than 460 billion dollars.
Pichai himself was candid about the situation. According to him, the company is capacity-constrained in the short term, and cloud revenue would have been even higher if they could meet all the demand. When the CEO of one of the largest tech companies in the world publicly admits he is leaving money on the table due to lack of infrastructure, it is clear the bottleneck is real and massive.
Demand for AI compute skyrocketed as companies began deploying chatbots, coding assistants, and AI agents into their operations. And Google is far from the only one scrambling. Anthropic, the company behind the popular chatbot Claude, closed a deal with SpaceX in November 2025 similar to its existing arrangement with Google. Everyone is chasing the same thing: more capacity to keep growing. 🚀
Why Meta relied so heavily on Gemini
These restrictions reveal just how much Meta had been leaning on rival models like Gemini, even while spending aggressively to become a leader in AI and improve its own models. Mark Zuckerberg has been pouring billions of dollars into attracting talent and securing infrastructure, all in the name of what he calls personal superintelligence.
Unlike Google, Meta does not have a cloud business and is racing to build its own fleet of data centers, geared toward both training and inference for its models. As part of that effort, the company has committed to investing no less than 600 billion dollars in the United States through 2028. It is a jaw-dropping number, but it shows the size of the bet.
Gemini had been used internally by Meta across several areas. One of them was automating safety processes, such as identifying scams and removing harmful content from its platforms. The model also powered customer service chatbots and ad-support tools, and it was used in some internal workflows and coding — alongside other models like Claude from Anthropic.
And why did Meta pick Gemini in the first place? Simple: it performed better than the company’s own open-source models, the Llama family. But things are changing. More recently, Meta started prioritizing its new model Muse Spark, which is seen as more competitive against Gemini and capable of reducing dependence on external models in certain applications. In other words, the company is learning to stand on its own two feet — and fast.
What this changes for the future of AI
This episode opens up an important discussion about how the artificial intelligence market will organize itself in the coming years. If even the largest companies in the sector are facing computational capacity limitations, what is left for startups, mid-sized companies, and independent developers who depend on APIs and cloud services to run their products? The answer, for now, is uncertain — and that creates an instability the market is still trying to absorb. Anyone who relies on external providers to access models like Gemini is subject to limitations that do not always show up in the contract but definitely show up in practice when demand gets tight.
On the other hand, this scenario is accelerating moves that were already underway. Meta, by prioritizing Muse Spark and Llama, is demonstrating exactly this push toward autonomy. Other companies are also heading in similar directions, seeking greater independence over their AI infrastructure. But building that independence comes at an extremely high cost, and not everyone has the scale to foot the bill. The result is a market increasingly concentrated in the hands of those with the resources to build and maintain their own computational capacity — which could create significant imbalances in terms of competition and innovation.
Google, for its part, knows that infrastructure scarcity is both a problem and an opportunity. Whoever can scale their computational capacity fastest — whether through new chips, strategic partnerships like the SpaceX deal, or breakthroughs in energy efficiency — will come out ahead. The road is long, the investments are massive, and results do not always arrive at the pace the market demands. What is clear is that the race for AI has also become a race for infrastructure — and that part of the competition happens far from the spotlight but has an enormous impact on what we see on our screens.
The limitation Google imposed on Meta is not an isolated incident. It is a barometer of where the artificial intelligence industry stands right now: growth accelerating too fast for the current infrastructure to keep up, with real consequences for companies of all sizes.
The AI market will keep growing — that much is practically certain. But how it grows, and who will be able to keep pace, depends heavily on how the computational capacity question gets resolved in the coming years. And if an episode like this between Google and Meta serves any purpose, it is to remind us that behind every response generated by a model like Gemini, there is a massive physical infrastructure — and it is finite.
