For the better part of two years, Nvidia has operated less like a semiconductor company and more like the sole landlord of the artificial intelligence revolution. By controlling the hardware—the H100s and Blackwell chips—and the software ecosystem (CUDA) that runs them, the company has created a bottleneck that every major tech firm must pay to pass through.
But the relationship between Nvidia and its largest customers is beginning to look more like a strategic hedge than a simple partnership. During Amazon’s first-quarter earnings call, CEO Andy Jassy delivered a message that served as both a validation of Nvidia’s current dominance and a warning about its long-term vulnerability.
The paradox is stark: Amazon is committing to one of the largest hardware acquisitions in the history of cloud computing, while simultaneously building a parallel infrastructure designed to eventually make those incredibly purchases unnecessary. For Nvidia investors, the “good news” is a massive, guaranteed revenue stream; the “poor news” is that the world’s largest cloud provider is actively migrating its workloads away from Nvidia’s proprietary moat.
The Million-GPU Commitment
On the surface, the news is a windfall for Nvidia. Amazon and Nvidia have reached an agreement that will see Amazon Web Services (AWS) take delivery of 1 million Nvidia GPUs by the end of 2027. This deal, which includes ancillary networking equipment and specialized chips, ensures that Nvidia will see tens of billions of dollars in revenue over the next two years from a single client.
From Amazon’s perspective, this isn’t just about scaling; it’s about insurance. In the current AI arms race, compute capacity is the primary currency. By locking down supply and pricing now, AWS ensures that its enterprise customers—who are often locked into Nvidia’s CUDA software—don’t migrate to Microsoft Azure or Google Cloud simply because they can’t find available GPUs.
“We will always have customers who want to run Nvidia on AWS,” Jassy noted during the call. This admission highlights the strength of the CUDA ecosystem. Because so many developers have written their AI models specifically for Nvidia’s architecture, the cost of switching to a different chip is often too high for a company to justify in the short term.
The Rise of the ‘Trainium’ Alternative
However, the most critical part of Jassy’s update was a quiet pivot regarding Amazon’s internal hardware. While the 1-million-GPU deal makes headlines, Jassy revealed that Amazon is now bringing in a larger number of its own custom AI chips, known as Trainium, than it is bringing in Nvidia chips.
This shift is being facilitated by Amazon Bedrock, the service that allows developers to access various foundation models via a single API. By abstracting the hardware layer, AWS can move workloads from Nvidia GPUs to Trainium chips without the end-user ever knowing the difference. Jassy indicated that a majority of Bedrock workloads are already running on Trainium.
The financial incentive for this migration is staggering. By utilizing its own silicon, Amazon can bypass the “Nvidia tax”—the high margins the chipmaker commands—saving the company tens of billions of dollars in capital expenditures. Custom silicon allows AWS to optimize the hardware specifically for the types of workloads its customers run, potentially offering better price-performance than a general-purpose GPU.
The demand for this internal alternative is not just theoretical. Jassy reported a $225 billion backlog for Trainium compute alone, contributing to a total AWS backlog of $364 billion. This suggests that the market is becoming comfortable with non-Nvidia hardware, provided the cloud provider handles the complexity of the integration.
Custom Silicon Landscape: The Big Three
Amazon is not alone in this strategy. The “hyperscalers”—the massive cloud providers—are all racing to build their own AI accelerators to reduce their dependency on a single vendor.
| Company | Custom Chip | Primary Strategic Goal |
|---|---|---|
| Amazon | Trainium / Inferentia | Lowering Capex; Bedrock integration |
| Alphabet | TPU (Tensor Processing Unit) | Vertical integration; selling to select partners |
| Microsoft | Maia | Expanding capacity; reducing operational costs |
Alphabet has perhaps the longest track record with its Tensor Processing Units (TPUs), which have become so efficient that the company has begun selling access to them to select external partners. Recently, Anthropic entered a massive deal with Alphabet that includes both Google Cloud compute and the direct purchase of TPUs for its own data centers.
Microsoft is following a similar blueprint with its Maia chips. By designing its own silicon, Microsoft aims to expand its compute capacity while keeping the cost per token—the basic unit of AI generation—as low as possible.
What This Means for Nvidia Shareholders
For investors, the narrative is shifting from “how much can Nvidia grow” to “how long can Nvidia maintain its margins.” The threat is not that Nvidia’s chips are becoming obsolete—they are still the gold standard for training the most complex models—but that the most profitable part of the market (the hyperscalers) is diversifying.
When Amazon, Google, and Microsoft move their internal workloads to custom silicon, Nvidia loses its most reliable, high-volume buyers. While Nvidia continues to find new customers in sovereign AI (national governments building their own clusters) and smaller enterprises, those clients do not have the spending power of AWS.
This creates a ceiling for future earnings growth. Even with a relatively modest forward earnings multiple of 25x, the stock carries risk because the “moat” is being chipped away from the inside. The very companies that fueled Nvidia’s meteoric rise are now its most dangerous competitors.
The next major indicator of this trend will arrive during the next cycle of quarterly earnings reports, where analysts will be scrutinizing capital expenditure guidance and the specific breakdown of “AI-driven” revenue. Investors should watch for any further mentions of “custom silicon” or “internal accelerators” in the filings of the major cloud providers, as these are the leading indicators of Nvidia’s market share erosion.
Do you think custom silicon will eventually replace the GPU for most AI workloads, or is Nvidia’s software moat too deep to cross? Let us know in the comments or share this story with your network.
