Is Nvidia’s AI monopoly starting to crack?

Nvidia (NVDA) has built one of the most extraordinary monopolies ever.

In 2025, its AI chips (GPUs) accounted for roughly 80%–90% of the market for training AI models.

That dominance helped Nvidia rake in more than $200 billion in annual revenue. And its stock surged multiplied by 10X since the AI boom began in late 2022.

By any measure, Nvidia is one of the greatest businesses of our time. That is why Nvidia became the first $5 Trillion company.

But monopolies rarely stay intact forever. And Nvidia’s grip on the AI chip market is starting to weaken.

On December 11, 2024, Nvidia’s monopoly showed its first real crack.

It was the day Google unveiled Gemini 2, the first major AI model trained entirely on chips that did not come from Nvidia.

Google trained it on TPUs, short for Tensor Processing Units, Google’s proprietary AI chip.

TPUs already sit underneath much of Google’s AI stack. They power everything from Gemini to YouTube recommendations to real-time translation.

TPUs are not general-purpose chips like CPUs. And they are not as flexible as GPUs. A TPU is narrower than either. It’s designed from the ground up to accelerate machine-learning workloads inside Google’s own infrastructure.

That makes it especially valuable when Google wants to run large AI systems more efficiently and at larger scale across its own data centers.

Google is not doing this necessarily to beat Nvidia in performance. The main goal is to reduce dependence. TPUs give Google tighter control and the ability to scale AI on their own terms.

Most importantly, making AI chips in-house lower costs.

Back in the SaaS era, software had zero marginal cost of replication. But in the AI era, every single answer has a real, physical cost in electricity and compute. OpenAI burns through close to $700 million a month to keep its AI machine running.

He who can produce the lowest-cost AI has a huge advantage.

Google’s flagship TPU, the v7 “Ironwood”, for example, has a Total Cost of Ownership 44% lower than comparable Nvidia’s GB200 (Blackwell) systems. It’s saving Google tens of billions of dollars each year.

And Google is not alone in this. Every hyperscaler is switching to in-house AI chips.

Amazon has Trainium chip for training AI models and Inferentia chip for running them. Microsoft has Maia chips for Azure. Meta has MTIA chips for recommendations and AI inference.

Of course, it’s one thing for a hyperscaler proving its own custom silicon works for its own model.

But that’s changing too.

Independent labs are also diversifying away from Nvidia.

Anthropic is using Amazon’s Trainium chips through Project Rainier to train and run Claude, including future versions of the model.

Rainier is one of the biggest AI clusters ever built. One of its main hubs in New Carlisle, Indiana, is designed to consume up to 2.2 gigawatts of power. Roughly enough to power 1.5 million homes.

Rainier started on nearly 500,000 Trainium2 chips in 2025. It’s expected to scale beyond one million Trainium2 chips by the end of 2026.

Anthropic also added Google’s TPUs to the mix. Last week, the company signed a 3.5 gigawatt deal for access to Google’s chips starting in 2027.

A frontier AI lab that once would have been assumed to need Nvidia for everything is now spreading its biggest workloads across multiple custom-chip platforms.

OpenAI is making a similar move at even larger scale. It still has a huge Nvidia partnership, including plans for at least 10 gigawatts of Nvidia systems. But it has also signed a deal with AMD for up to 6 gigawatts of Instinct GPUs. And it has a separate Broadcom partnership to build 10 gigawatts of custom AI accelerators.

Nvidia could eventually lose up to half of its AI chip revenue.

Amazon, Google, Microsoft, Meta, and Oracle are expected to spend roughly $700 billion on AI infrastructure in 2026. So far, Nvidia has captured an enormous share of that wallet.

In its latest earnings commentary, Nvidia said large cloud service providers accounted for one out of every two dollars it makes from AI.

That puts Nvidia in a squeeze. Its largest customers are the buyers with the most money, the most scale, and the strongest incentive to build around it.

If Google trains more models on TPUs, that is Nvidia revenue it never sees. If Amazon routes more Claude workloads to Trainium, that is another chunk gone. If OpenAI splits future clusters between Nvidia, AMD, and custom Broadcom chips, Nvidia loses again.

I’m not saying hyperscalers will replace Nvidia overnight. But Nvidia is starting to lose first claim on the next wave of AI chip demand. A hyperscaler that used to send 50 cents of every AI infrastructure dollar to Nvidia might send 30 cents instead. Then 20. Maybe less for the workloads it controls most tightly.

Case in point. Last month, Broadcom reported it expects to make more than $100 billion by 2027, designing custom chips for companies like Google, OpenAI, Meta, Amazon, and Microsoft. That is the size of the leak already forming in Nvidia’s moat.

Even the contracts Nvidia keeps winning will become less valuable. Because once buyers have credible alternatives, Nvidia loses pricing power, one of the biggest benefits of being a monopoly.

What does this mean for Nvidia stock?

AI is still the biggest, fastest-growing megatrend of our time.

And Nvidia’s chips still win in performance and flexibility.

That is why hyperscalers and independent AI labs keep buying them, even while building their own alternatives.

Nvidia also has another ace in its sleeve. Software. CUDA, networking, developer tools, and full-stack systems make Nvidia hard to replace overnight. Once AI labs build around Nvidia, they do not switch easily.

That should be enough for Nvidia to keep making good money for years.

If you own Nvidia stock, hyperscalers moving into custom chips is not automatically a reason to sell.

But it is a reason to seriously recalibrate your expectations.

NVDA stock surged because investors treated Nvidia like a monopoly. If that monopoly is starting to crack, returns will be much more muted.

Nvidia (NVDA) has a Disruption Score of 4. Click here to learn how we calculate the Disruption Score.  

Nvidia is also part of the Disruption Aristocrats, our quarterly list of the world’s top disruptive stocks.