HWNEWS
Hyperscalers (the cloud giants running the largest AI workloads) are shifting from buying chips to building them. Nvidia's full stack rack is the bet that system sales can outrun chip share loss.
Three years ago, Nvidia held 90% of the world's installed AI compute capacity. Today the share is 70%, and the 20-point slide is now the number every AI infrastructure buyer has to plan around. The custom silicon filling the gap is mostly Google's Tensor Processing Units and Amazon's Trainium and Inferentia accelerators, custom AI chips built in-house, with OpenAI's first in-house AI chip now joining the queue. Nvidia is now selling whole racks, not just faster chips.
The share figures come from a dataset maintained by Epoch AI, an independent research group tracking AI compute capacity, and were aggregated in an August 2026 analysis from Semiconductor Engineering. The dataset counts installed capacity by manufacturer, not by dollars or revenue, and it excludes Meta, Microsoft, and Cerebras systems. Even with that scope, the trajectory is clear: 20 points of share moved from general-purpose Nvidia GPUs to custom AI silicon in three years, while total installed AI compute capacity grew 17-fold. The chip market is not shrinking for Nvidia. The mix is shifting under it.
The shift is driven by the same buyers who once fed Nvidia's order book. Google has run its own TPUs, processors optimized for the matrix math at the heart of neural-network training and inference, alongside Nvidia hardware for nearly a decade. Amazon built Trainium for training and Inferentia for inference workloads, the latter aimed at cost-per-query rather than raw throughput. OpenAI is now in late development on its first in-house AI chip, known internally as Jalapeño, a name the company has not used in public filings. AWS has also committed to buy 2 million Nvidia GPUs alongside millions of Nvidia's new Vera CPUs, the kind of multi-year, multi-product order that binds a buyer to a vendor long after any single chip cycle.
Nvidia is not standing still. On August 26, 2026, the company reported earnings and CEO Jensen Huang projected roughly 70% revenue growth for the coming fiscal year, capped by supply rather than demand. Two weeks later, at the Hot Chips 2026 symposium, Nvidia presented five papers, one-fifth of the program. The papers covered the Vera CPU, the NVL72 rack-scale networking fabric, the Groq 3 LPX low-latency inference system, the BlueField data-center networking chip, and the Spectrum-6 SPX scale-out switch. Read together, the five papers describe a stack rather than a chip: compute, memory, networking, and the orchestration software that ties them into a single unit a hyperscaler can buy with one purchase order.
The strategic logic is the shift in the unit of competition. When the unit of comparison was a single chip, Nvidia won on raw throughput and a software stack, CUDA, that became its own moat. Now the unit is a rack, and buyers ask which vendor can deliver predictable performance at cluster scale, with networking, memory bandwidth, and inference economics that match the workload. The fastest-chip benchmark still matters, but it matters less than the system around it. Rack-level contracts are also harder to break than chip-level ones, which is why a 2-million-GPU AWS deal matters more to Nvidia's defense than its headline share suggests.
The exposure runs the other way too. The hyperscalers, the cloud companies whose training and inference workloads dwarf everyone else's, do not buy full stacks out of brand loyalty. They buy them when the math works. Google's TPUs, Amazon's Trainium and Inferentia, and OpenAI's forthcoming chip are not vanity projects. Each exists to keep Nvidia's pricing power in check by giving the largest workloads a credible second source. If those in-house chips keep improving, the 70% number can keep sliding, and Nvidia's defense becomes a race between its rack business and the hyperscalers' willingness to standardize on its full stack.
The next strategic question is which workloads go to which stack, and at what point the unit of competition shifts again. Training the largest models still favors Nvidia's full rack. Inference at scale, where cost per token matters more than peak throughput, is the workload most likely to keep migrating to custom silicon. If OpenAI's Jalapeño chip moves inference economics enough to lower per-token costs at scale, it could pull other large labs toward in-house silicon too.