How are AMD and Broadcom challenging Nvidia in AI chips?

There are two ways to loosen Nvidia's hold on AI chips. One is to sell a competing product. The other is to help your customer stop needing one. AMD is the main example of the first approach, and Broadcom of the second.

The current scale, from the latest quarterly reports: Broadcom's AI chip revenue was $16.7 billion, up 221% from a year earlier, and it expects $21.7 billion the next quarter. AMD's data centre segment, which also includes server processors, was $6.7 billion, up 107%. Nvidia's data centre revenue was $89.0 billion. So both challengers are growing quickly from a much smaller base.

Two different businesses

A custom chip, sometimes called an XPU or an ASIC (application-specific integrated circuit), gives up flexibility. It only has to run the customer's models, so it can leave out what they do not need. Its owner also keeps the margin that would otherwise go to Nvidia. In exchange, it takes on years of design work and the risk of guessing wrong about what future models will need.

Google's TPU is the longest-running example. Broadcom's April 2026 filing says the two signed a long-term contract under which Broadcom builds future custom TPU generations, along with networking and other rack components through up to 2031.

What the big deals actually say

Press headlines give the largest number. The documents are more careful.

A gigawatt (GW) measures electrical power, not chips: it is roughly what a very large power plant produces. Labs now size their chip deals in power because power is often the limit on what they can build. For the details, see how a gigawatt of AI is counted.

Three patterns stand out in this table:

• "Up to" is a ceiling. AMD's filing describes what each customer "intends" to deploy. Actual revenue arrives as chips ship.
• Money flows in both directions. A warrant lets its holder buy shares for almost nothing ($0.01 each, per AMD's filing) once milestones are met. So AMD pays part of the cost of winning OpenAI and Meta in its own stock, and it is investing in Anthropic too. Critics see this kind of arrangement as circular financing, since a supplier helps fund its own customer. Supporters see a sensible way to share risk on an unproven product.
• Customers are spreading their bets. Anthropic says it uses AWS Trainium, Google TPUs and Nvidia GPUs, and has now added AMD. OpenAI has signed with both AMD and Broadcom. Having several suppliers gives a lab bargaining power and protection against shortages, at the cost of running its models well on more than one kind of chip.

What could stop them

Both challengers share Nvidia's bottlenecks. AMD's filing says it relies on TSMC for all its GPU wafers at 7 nanometres or smaller, and warns that rack systems depend on third-party parts that may be in short supply. A rival GPU also needs a software path away from CUDA, which is improving but not finished.

Custom chips carry a different risk: concentration. Broadcom's risk language flags the loss of significant customers, and a business built on a few enormous clients can swing when one changes plans. And these are forward-looking agreements about hardware that, in many cases, has not shipped in volume. None of this is a judgment on the companies as investments. It is simply what to check before treating a headline number as a done deal.

Will more chip competition lower what you pay for AI?

It should help over time, but do not expect a direct line from a chip deal to your subscription price. Chips are a large part of what it costs to run a model, and a buyer with real alternatives can negotiate harder. If Nvidia's 75% gross margin narrows, or if in-house chips let labs serve requests more cheaply, some of that saving could reach customers.

Against that, demand is currently outrunning supply. Memory, factory capacity and electricity are all tight, and memory makers describe their output as effectively sold out for years, so competition may mostly mean more total capacity at first, not lower prices. Labs may also spend savings on larger models instead of cutting prices. What you are more likely to notice first is fewer usage limits and more choice, if the extra capacity arrives on schedule.