APREX · EST. 2026
Nvidia has long dominated the AI hardware market almost without competition. Now, the company itself signals that the picture is changing — through a strategic shift in how revenues are presented to investors and analysts.
According to Stratechery, Nvidia will henceforth report hyperscaler sales separately from sales to other customers. This is not a random accounting adjustment. It is an admission that the two market segments function fundamentally differently.
Hyperscalers Building Their Way Out of Nvidia Dependence
The four dominant cloud providers — Google, Amazon Web Services, Microsoft, and Meta — all have ongoing programs to develop their own AI accelerators. The goal is twofold: lower costs and less vulnerability to a single external vendor.
Google is the furthest along. The company launched its first Tensor Processing Unit (TPU) back in 2015, and the seventh generation — codenamed Ironwood — was announced in April 2025. According to available specification information, each TPU v7 chip delivers 4,614 FP8 TFLOPS with 192 GB of HBM3E memory. Google holds an estimated 58 percent of the market for proprietary cloud AI chips and invests around $8 billion annually just in its collaboration with Broadcom.
AWS has been rolling out Inferentia chips for inference and Trainium for training since 2019. The third generation Trainium is expected, according to Amazon, to deliver over four times the performance of its predecessor with significantly lower energy consumption.
Microsoft presented Maia 200 — manufactured on TSMC's 3nm process — and claims it provides 30 percent better performance per dollar than existing hardware in the company's own data center network. Meta, for its part, unveiled four new MTIA generations in March 2026 and has already rolled out hundreds of thousands of its own inference chips across Facebook and Instagram.
Two Different Battlegrounds for Nvidia
This is where Nvidia's new reporting structure becomes meaningful. With hyperscalers, Nvidia is in a competitive situation where customers are actively working to replace — or at least supplement — GPUs with their own solutions. Margin pressure and volume negotiations characterize this segment.
For all other customers — businesses, research institutions, the public sector, mid-sized cloud providers — the situation is different. Here, Nvidia sells not just GPUs, but a complete technology stack: hardware, network infrastructure via Mellanox/InfiniBand, and software through the CUDA ecosystem. This is far more difficult to replace and gives Nvidia a structurally stronger position.
By reporting these two segments separately, Nvidia makes it easier for the market to understand where the pressure actually lies — and, crucially, where growth potential is more protected.
Source and Disclaimer
The analysis is based on Stratechery's review of Nvidia's quarterly figures and the new reporting structure. Information on hyperscalers' own chips comes from public announcements and industry research. It is worth noting that performance claims from the companies themselves — especially comparisons across chips — are marketing statements that are not always independently verified.
The overall picture is nevertheless clear: the AI hardware market is stratifying. The top of the pyramid — where the very largest players operate — is in motion. The rest of the market is currently Nvidia territory.