Everyone wants their own AI chips

Oct 28, 2025

GPUs are powering the AI boom. If you are a big AI company, you do not want to be left behind should this AI thing turn into a big money-maker. You invest many tens of billions of dollars a year in data centers. GPUs comprise as much as 39% of the value of your data center investments (Business Insider report).

There is only one company - Nvidia - selling the GPUs you want. The hardware is top-of-the-line and it comes with a software toolkit (Wiki on CUDA) that your engineers like. You pay up to $40,000 per module even though it probably costs Nvidia only $6,000 to make it (Tom’s Hardware report). Because you are in a rush, you pay whatever is needed to secure your GPUs. But it still hurts. You wish you did not have to pay as much.

Nvidia designs its GPUs, and relies on TSMC to fabricate them (Nvidia blog). Others provide critical memory components (Korea Herald report). You buy billions of dollars worth of GPUs a year. If you went to TSMC with your own chip design, would they say no to you? Probably not. You have enough volume that you would make it worthwhile for them. Maybe they will charge you more than what they charge Nvidia but you think you will still save a ton.

If you pull this off, you can probably cut your AI compute costs in half. Maybe more. You can hopefully build better AI models. You can afford to charge your customers less than your competitors. Customers will like the cost-performance trade-off. They will use your products. Enterprises will build applications using your APIs and cloud offerings. You bet this will build enduring value for your company.

Google has been doing some version of this exercise with its TPUs since 2015. Estimates suggest they had one-fifth of the world’s AI compute in 2024 (AI 2027 has an analysis). Their compute costs are lower (Hedder thinks Google saves 80% by using its own hardware). Amazon has had its own chips for training (Trainium) and inference (Inferentia) for a few years now. Microsoft is a bit early in the game with its Maia chip but it is directionally on the same path as Google and Amazon (CNBC report).

Nvidia’s biggest customers are trying to wean off Nvidia. Its $100B investment in OpenAI is a bet to preserve a chunk of its compute market. Even so, OpenAI struck a deal with AMD - Nvidia’s only notable competitor, despite recent moves from Intel (Intel newsroom) and Qualcomm (CNBC report). However, playing competitors against each other may not be enough if you see yourself as the next multi-trillion dollar company. OpenAI has also teamed up with Broadcom to design its own chips.

If you are Nvidia, you still have a lot of things going for you. One, many of your biggest customers will still buy heavily from you. It might take a few years before they become self-reliant. Two, the market for GPUs is growing fast. There is a fat tail of companies out there for which designing their own chips is not practical. Three, you have helped build a new industry, neocloud, that rents out GPUs over the cloud. This widens your market access, and gets you customers who build using your GPUs and your software toolkit.

Now, what comes next? Maybe Nvidia will run faster than everybody else and preserve its lead. Maybe more credible competitors will decrease the premium investors are willing to pay for Nvidia’s and AMD’s stocks. Maybe Nvidia will be able to diversify and capture the fat tail of the market. Maybe we will get even better AI chips as competition heats up. Maybe we will enter an AI winter. Maybe.