Nvidia has led the AI chip race. Its GPUs power the biggest AI models and much of today’s AI infrastructure. According to Felix Prehn, founder of Goat Academy and former investment banker, that leadership faces a real test. The surprising challengers are Nvidia’s largest customers, who are now making their own chips.
Here is the key risk in simple terms. About 40% of Nvidia’s revenue comes from two customers. The market believes these are Microsoft and Amazon. When a big part of sales depends on a few buyers, small changes can have a large impact. If even one of these companies cuts orders, the stock could fall, not because Nvidia is weak, but because concentration increases volatility.
Why would these companies switch? Cost and control. Nvidia’s gross margins are around 75%. Gross margin means the percentage of revenue left after subtracting the direct cost to make the product. If a chip costs $25 to manufacture and sells for $100, the gross margin is 75%. Large buyers who spend billions see a strong incentive to design their own chips and keep that margin for themselves. They also gain control over supply, pricing, and features tailored to their own workloads. A workload is the type of computing task, like training a language model or running search results.
Google’s TPU effort shows how this plays out. TPUs (tensor processing units) are specialized chips for AI math. Think of Nvidia’s GPU as a Swiss Army knife—flexible and good at many tasks. A TPU is a chef’s knife—optimized for one cut, very efficient at it. Reports suggest TPUs can be two to three times more power efficient and up to four to ten times more cost effective than general-purpose GPUs for specific AI tasks. Google has already trained its top models on TPUs, proving they can handle the biggest jobs.
Other hyperscalers—Microsoft, Amazon, and Meta—are on a similar path. Hyperscalers are very large cloud providers that run massive data centers. They are designing chips such as Amazon’s Trainium and Microsoft’s in-house accelerators. There are also signs of cross-usage. For example, reports indicate Meta plans to rent and later purchase Google’s TPUs. If competitors are willing to use each other’s silicon, it validates that alternatives to Nvidia are real and improving.
What does this mean for investors? Prehn’s view is that 2025 and even 2026 can still be strong for Nvidia, but 2026–2027 is the window when the market may start to price in margin pressure and more competition. Markets usually look six to twelve months ahead. If custom chips scale, investors may expect lower margins for Nvidia even if unit sales and total compute demand rise. In valuation terms, that can translate into a lower price-to-earnings multiple. A multiple is how much investors are willing to pay for each dollar of earnings. If the multiple falls while earnings rise, the stock may trade flat instead of climbing.
There is still a strong bull case. Nvidia is an execution leader. Its software ecosystem, especially CUDA, is a moat because many AI developers write code that is optimized for Nvidia. Switching takes time and money. Also, demand for compute remains high. But customer concentration is a real risk, and the hyperscalers have the talent, capital, and motivation to push custom chips.
How can readers position themselves? First, assess concentration risk. If a portfolio is heavily tilted to Nvidia through direct shares or index exposure, consider risk controls like position sizing and stop-loss rules. A stop-loss is a preset price that triggers a sell to limit losses. Second, watch signals that the market cares about: hyperscaler silicon roadmaps, Nvidia’s gross margin trends, and news about large chip supply agreements in 2026–2027. Third, consider “picks-and-shovels” exposure in the AI chip supply chain. These are companies that provide components, manufacturing, assembly, and networking for many chipmakers, not just one brand. Examples cited by Prehn include Broadcom (ABGO/AVGO), Amphenol (APH), Taiwan Semiconductor (TSM), Amkor (AMKR), Jabil (JBL), Flex (FLEX), and other networking and packaging specialists. These businesses can benefit from the overall growth in AI compute, regardless of which chip brand wins share.
In simple terms, the story is shifting from one champion selling out every unit at premium margins to a more crowded field where the biggest buyers become competitors. If that happens at scale, pricing power often changes hands. Understanding this shift early can help investors balance risk and find opportunities across the ecosystem.
Learn more about Felix Prehn and Goat Academy here: Felix Prehn Goat Academy