Competing Silicon, and How a Non-Engineer Should Read a Chip Roadmap

Accelerator supply splits into three kinds of supplier. Merchant vendors of graphics processing units, or GPUs, design chips and sell them to anyone: NVIDIA and Advanced Micro Devices, or AMD. Hyperscalers design chips for their own clouds and their own models: Google, Amazon, Microsoft and Meta. And a group of smaller specialist designers compete for particular workloads. The first kind sells chips. The second kind sells computing time, and designs chips to lower its own costs.

Google is the longest-running case. It announced its tensor processing unit, or TPU, in May 2016, after more than a year of internal use. In April 2025, at its Cloud Next conference, it announced the seventh generation, Ironwood, describing it as its first TPU designed specifically for inference, the work of running trained models. Google says an Ironwood pod joins up to 9,216 chips. In November 2025 it said Ironwood would become generally available to cloud customers within weeks. The customer list matters as much as the specification. In October 2025, Anthropic announced it would expand its use of Google Cloud TPUs to up to one million chips.

Amazon Web Services, or AWS, is the second case. It announced its Trainium training chip in 2020. Trainium2 became generally available in December 2024. At its re:Invent conference on 2 December 2025, AWS made Trainium3 UltraServers generally available, each joining up to 144 Trainium3 chips, and previewed Trainium4, which it said would be designed to support NVIDIA's NVLink Fusion interconnect. Note the last detail. A hyperscaler building its own chip still plans to connect to the market leader's link technology.

Microsoft announced its first in-house accelerator, Maia 100, at its Ignite conference in November 2023. Meta has built a family of accelerators it calls the Meta Training and Inference Accelerator, or MTIA, and in March 2026 announced four further generations.

Why do hyperscalers build their own? The structural reasons are cost and control. A company that runs enough of one workload can design a chip that does only that workload, avoid paying a merchant vendor's margin and set its own schedule. The constraint is software. A merchant GPU comes with years of tools and libraries. A custom chip must earn its software support, usually by serving the company's own models first. Faraday's read is that custom silicon is a lever on price negotiation as much as a replacement: it caps what a hyperscaler will pay for merchant chips, even when the hyperscaler keeps buying them.

AMD sits between the two camps. It sells merchant accelerators, the Instinct line, and it competes by matching NVIDIA's cadence and by backing open standards for links and racks, including the UALink scale-up specification and the Open Compute Project, or OCP, rack designs.

Now the reader's method. A non-engineer does not need to judge transistor counts. Five questions are enough to read any chip announcement with some discipline.

First, which stage is this? Announced, sampling, in full production, available from partners or deployed at scale. The headline often blurs them.

Second, what number format is the performance quoted in? A figure at 4-bit precision cannot be compared with an older figure at 16-bit precision. If the release does not say, the footnotes will.

Third, what about memory? Look for the capacity and bandwidth of the high-bandwidth memory, or HBM, per chip. Those numbers usually say more about real performance than the compute figure.

Fourth, how big is the scale-up domain? Seventy-two GPUs in an NVIDIA NVL72 rack, 144 chips in a Trainium3 UltraServer, 9,216 chips in an Ironwood pod. The largest group that can work as one shapes which models run well.

Fifth, what does it ask of the building? Look for rack power, cooling method and distribution voltage. If the announcement does not mention them, the integrator's documentation will, as Hewlett Packard Enterprise's does for its GB200 NVL72 build.

Those five questions turn a press release into a checklist. They will not tell you which vendor wins. Nobody reliably knows that. They will tell you what has actually been delivered, what the figures mean, and what the building must do to host it. For an investor, an operator or an engineer starting in this field, that is the useful part of any roadmap.

The chips will keep changing every year. The questions will not, and that is what makes them worth learning.