Nvidia’s Best Customers Have a Reason to Stop Buying So Many Nvidia Chips

Nvidia reports earnings on August 26. Most investors will be watching Blackwell demand, Rubin margins, and whether (NVDA) can beat expectations again. I am watching something else. Nvidia’s biggest customers are spending extraordinary amounts on its chips while also spending billions trying to reduce their need for them. That is not because Nvidia’s products are failing. It is because they are working too well.

Nvidia’s gross margins remain around 75%. That is fantastic if you own (NVDA). If you are (MSFT), (AMZN), (GOOG), or (META), paying those margins is also an incentive to build more of the technology yourself. Google has its TPUs, Amazon has Trainium, (MSFT) is developing Maia, and (META) is pushing further into custom silicon. None of them need to replace (NVDA). They only need to reduce the percentage of their next AI dollar that goes to Nvidia.

This matters because investors often treat rising AI capital spending as automatically bullish for (NVDA). It has been for several years, but I am not convinced every additional dollar of AI spending will continue flowing through (NVDA) at the same rate. AI spending can keep rising while Nvidia captures a smaller share of it.

Nvidia Created the Incentive

One of the strange things about exceptional businesses is that their economics eventually attract competition. (NVDA) has spent years building an extraordinary position in AI infrastructure. Its lead is not simply the GPU. CUDA, networking, software, and the architecture of the entire data center make (NVDA) difficult to replace. But extraordinary margins create extraordinary incentives. If your largest customers are spending tens of billions of dollars with you every year, they will eventually start asking whether some of that economic value should belong to them. Custom silicon is their answer.

I do not think this development means (NVDA) suddenly loses its leadership. That is too simplistic. The more compelling question is whether (NVDA) can remain dominant while its customers slowly take more workloads in-house. (NVDA) does not merely need AI to keep growing. Investors need (NVDA) to continue capturing an exceptional amount of the profit from that growth. Those are not the same thing.

There is another reason I watch the stock closely. Nvidia’s biggest customers are not weak buyers without alternatives. They are some of the best-capitalized technology companies in history. They can afford to spend billions developing their own chips, fail, learn from those failures, and try again. They also employ some of the best engineers in the world. Most suppliers lack this combination of capital, technical ability, and a strong financial incentive to reduce their dependence when facing customers.

The Threat Does Not Need to Kill Nvidia

This is where I think investors can make the wrong assumption. (NVDA) does not need to lose the AI market for its stock to struggle. A company can remain the dominant supplier, continue growing revenue, and still see its valuation compress if investors begin to believe customer-designed silicon will take a larger share of incremental spending. The competitive threat does not need to become existential. It only needs to become material.

That is important when you are dealing with a company valued at more than $5 trillion. The market is not pricing (NVDA) as a business fighting for survival. It is pricing an extraordinary business that is expected to continue producing extraordinary economic results. A small change in those expectations can matter far more to (NVDA) 

Inference May Be the First Pressure Point

Training the largest AI models still plays heavily to Nvidia’s strengths, but inference may be where the economics begin changing first. Once a model has been trained, the problem becomes serving enormous numbers of requests as cheaply as possible. At that point, saving even a small amount on every token starts to matter.

@MSFT, @AMZN, @GOOG, and (META) do not necessarily need to build the best general-purpose AI chip in the world. They need hardware that works well enough for specific workloads at a lower cost.

That creates a different competitive problem for (NVDA). This is not simply a chip race anymore. It is a fight over who keeps the economics.

(NVDA) wants to reduce the cost of AI computing quickly enough that customers continue buying its systems. Its customers want greater control, lower costs, and less dependence on one supplier. When your customers start designing around your margins, I pay attention.

What August 26 Needs to Show

That is why I am less interested in whether (NVDA) beats quarterly revenue estimates by another billion dollars when (NVDA) reports on August 26. At this point, another beat would surprise almost nobody.

I want to know whether the economics remain as strong as the growth. If (MSFT), (AMZN), (GOOG), and (META) are still expanding their (NVDA) purchases and (NVDA) can protect margins while its customers continue developing their own chips, then the custom-silicon threat may remain more theoretical than real.

If internally designed chips begin taking meaningful workloads away from Nvidia, the investment debate changes. The bear case does not require AI spending to collapse. It does not even require Nvidia’s revenue to decline. (NVDA) could continue growing quickly and still disappoint investors if the market starts believing (NVDA) will capture a smaller share of future AI spending.

That is the part I think matters most. For years the big (NVDA) question was whether AI demand was real. (NVDA) answered it. The next question is more difficult: who ultimately benefits from the economics? Nvidia’s biggest customers are buying more of its chips today while spending billions trying to make sure they do not have to buy quite as many tomorrow.

They have strong financial incentives to do so.