AI chip companies in 2026: who makes AI chips and how they differ
Who makes AI chips? A plain-English guide to Nvidia, AMD, Google, Amazon, Intel, Qualcomm, Cerebras and TSMC, and what an AI chip actually is.

The big AI chip companies are Nvidia, which still sells most of the accelerators used to train and run large models, and AMD, its closest rival in data center GPUs. Behind them sit the cloud giants that design their own chips (Google, Amazon and Microsoft), challengers like Intel, Qualcomm and Cerebras, and TSMC, the Taiwanese foundry that actually manufactures nearly all of them. If you only remember one thing, I'd make it this: designing an AI chip and making one are separate jobs, and very few firms do both.
Here's a fact that surprised me. AMD now says inference, the work of running models rather than training them, takes about 60% of global AI compute this year. That shift explains a lot about who's building what, and why the list of AI chip makers keeps getting longer.
Short answer: who makes AI chips
An AI chip is a processor built to do the huge piles of matrix maths that neural networks need, very fast and in parallel. Nvidia and AMD sell GPUs for this; Google, Amazon and Microsoft build custom accelerators for their own clouds; and TSMC fabricates most of them. Intel, Qualcomm and Cerebras are smaller players with their own designs.
The background: what makes a chip an "AI chip"
A regular CPU is a generalist. It's great at doing a few complicated things one after another, which is what your operating system and browser need. Neural networks want something different: millions of simple multiply-and-add sums done at once, over and over, on numbers that don't need much precision.
That's why graphics cards ended up running AI. A GPU was already built to shade millions of pixels in parallel, and the same layout happens to suit the maths inside a model (I explain that maths in more detail in what compute means in AI). Over the past decade the chips drifted away from graphics entirely. Modern accelerators add tensor or matrix units, support tiny number formats like FP8 and FP4, and sit next to stacks of high-bandwidth memory (HBM) so the cores aren't left waiting for data.
An everyday comparison helps here, and it's the only one I'll use. A CPU is a skilled chef cooking one complicated dish at a time; an AI accelerator is a kitchen of 10,000 line cooks who can each only chop onions, but all at the same moment. For a model, it's mostly onions.
So what are AI chips made of? Mostly silicon, like any processor, but the packaging matters as much as the die. AMD's MI400 series uses chiplets with 3D hybrid-bonded compute dies, and its MI455X carries 432 GB of HBM4 memory. The leading-edge parts are made on TSMC's newest processes; its 2nm (N2) technology started volume production in the fourth quarter of 2025.
One more shift is worth knowing. The unit of sale is no longer really a chip. Nvidia and AMD both sell whole racks of 72 GPUs wired together, and Google talks about pods of thousands of chips. If you want a feel for the physical scale, our piece on how big AI data centers are covers it.
Real examples: the main AI chip companies
I've sorted these by how most people will meet them: chips you can buy (or rent), chips you can only rent from one cloud, and the factory behind both.
| Company | Main AI chip | Type | How you get it |
|---|---|---|---|
| Nvidia | Vera Rubin (Rubin GPU, Vera CPU) | GPU | Servers from OEMs, most clouds |
| AMD | Instinct MI455X / MI450X, Helios rack | GPU | OEMs and clouds |
| TPU 8t, TPU 8i | Custom accelerator | Google Cloud only | |
| Amazon (AWS) | Trainium3 | Custom accelerator | AWS only |
| Intel | Crescent Island | Inference GPU | Upcoming, sampling to customers |
| Qualcomm | AI200, AI250 | Inference accelerator | Cards and racks |
| Cerebras | WSE-3 Turbo, CS-4 system | Wafer-scale chip | Systems and Cerebras Cloud |
| TSMC | Manufactures most of the above | Foundry | Doesn't sell chips under its own name |
Nvidia
Nvidia is the answer most people expect when they ask what company makes AI chips, and they're right to. Its current platform, Vera Rubin, pairs the Rubin GPU with the Vera CPU. The flagship NVL72 rack holds 72 Rubin GPUs and 36 Vera CPUs. Nvidia confirmed full production in July and named OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta and Dell among the first customers. It also folded in a Groq 3 LPU for low-latency inference, which I read as a sign that even Nvidia thinks GPUs alone aren't the whole answer for serving models.
Its real moat probably isn't silicon, though. It's CUDA, the software layer that most AI code has been written against for well over a decade.
AMD
AMD is the clearest second source. Its Helios rack links 72 Instinct MI455X or MI450X GPUs and is rated at up to 2.9 exaFLOPS of FP4 compute with 31 TB of HBM4. Helios went into full production with first shipments set for the end of the third quarter. OpenAI and Meta each have six-gigawatt agreements starting with MI450 hardware, and Anthropic plans up to two gigawatts of Helios capacity from 2027. Those are big numbers, probably the biggest, for a company that was an afterthought in AI five years ago. AMD's weak spot has been software; its ROCm stack is open, but it's younger than CUDA.
Google, Amazon and the custom-chip crowd
The hyperscalers would rather not pay Nvidia's margins forever. Google's TPUs run Gemini, Search and Maps, and the eighth generation splits into TPU 8t for training (9,600 chips per superpod) and TPU 8i for inference, which Google says gives an 80% performance-per-dollar improvement over earlier generations. Amazon's Trainium3 UltraServers scale to 144 chips each, and AWS lists Anthropic, Databricks and OpenAI among its Trainium customers. Microsoft has its own Maia line as well, though it seems further behind.
The catch for you is simple: you can't buy any of these. You rent them inside one cloud.
Intel, Qualcomm and Cerebras
Intel's next data center accelerator, Crescent Island, is an inference GPU with 32 Xe3P cores and 160 GB of LPDDR5X memory, cheaper memory than HBM, in a 350W air-cooled card. Qualcomm took a similar bet: its AI200 card carries 768 GB of LPDDR, with commercial availability planned for 2026 and the AI250 for 2027. Cerebras goes the other way entirely, building one chip from a whole silicon wafer; its WSE-3 Turbo has four trillion transistors.
I find Cerebras the most interesting of the three, if only because it's so different. Whether a niche design can win at scale is another question, and I suspect it'll stay a specialist tool.
Benefits and limitations
The upside of specialised chips is efficiency. A part that only has to do low-precision matrix maths can drop circuitry a general processor needs, so it does more AI work per watt and per dollar. That's the whole pitch behind TPUs, Trainium and the inference cards from Intel and Qualcomm, and it's why AMD split its range into the AI-focused MI450X and the MI430X for traditional supercomputing.
The limitations are just as real. These chips are scarce and costly, they lean on a single foundry in Taiwan for the leading-edge parts, and they eat power: the vendors now quote deals in gigawatts, which tells you something. Software lock-in is the other big one. Code tuned for CUDA doesn't always move cleanly to ROCm, TPUs or Trainium, although PyTorch, JAX and vLLM support across vendors has made that much easier than it was. Custom cloud chips add a second kind of lock-in, since you can't take a TPU with you when you leave Google Cloud.
There's also the question of whether today's spending is sustainable. I don't know the answer, and I'm wary of anyone who claims to; our explainer on the AI bubble debate lays out both sides.
Practical takeaways for you
For most people, very little changes day to day. You'll meet AI chips indirectly, through faster chatbots and lower prices on AI tools as inference gets cheaper. If AMD's 60% inference figure holds, I'd expect most of the competition to show up as cheaper tokens rather than flashier training runs.
If you run AI workloads at work, the practical choices look like this:
- Renting GPUs in a public cloud is still the default, and Nvidia hardware is the easiest to find software support for.
- AMD instances are worth pricing out if your stack runs on PyTorch or vLLM without custom CUDA kernels.
- TPUs or Trainium can be cheaper for large, steady jobs, provided you're happy staying in that one cloud.
- Buying hardware outright usually only makes sense at real scale; our list of AI data center servers shows what that involves.
And if you're asking about chips for investing reasons, I'd stay out of it here: this piece explains the technology and the companies, not where their share prices are going.
The road ahead for AI chips
The pace is the first thing to watch. Nvidia and AMD have moved to yearly releases, and AMD has said its MI500 series follows in 2027. Google launched four eighth-generation TPU systems in a single year, which it calls a first. Qualcomm says it's on an annual cadence too.
The second trend is that inference is splitting off from training. Intel and Qualcomm are using cheaper LPDDR memory to chase it, AMD is pairing Helios with Cerebras for the token-generation step, and Nvidia bolted on Groq LPUs. My guess is that by 2028 "AI chip" will cover several quite different designs depending on whether you're training a model or serving one. That's a guess, not a forecast anyone can promise.
Last, keep an eye on manufacturing. TSMC's roadmap already lists A16 and A14 nodes after 2nm, and whoever gets capacity on them first will likely ship the fastest parts.
FAQ
What is an AI chip?
An AI chip is a processor designed to run neural networks efficiently. It packs thousands of simple maths units that work in parallel, supports low-precision number formats like FP8 and FP4, and sits next to fast high-bandwidth memory. GPUs from Nvidia and AMD and custom accelerators like Google's TPUs are the main kinds you'll hear about.
What are AI chips?
AI chips are the family of processors built for machine learning: data center GPUs, custom cloud accelerators (TPU, Trainium, Maia), inference cards from firms like Qualcomm, and the neural processing units (NPUs) inside phones and laptops. They differ in size and power. All of them trade general-purpose flexibility for fast, parallel matrix maths.
Who makes AI chips?
Nvidia and AMD design the most widely sold data center AI chips. Google, Amazon and Microsoft design their own for their clouds, while Intel, Qualcomm and Cerebras offer alternatives. TSMC in Taiwan builds most of them.
What are AI chips made of?
They're made mostly of silicon, patterned with billions of transistors on advanced processes such as TSMC's 3nm and 2nm nodes. Around the main die sit stacked memory and advanced packaging. AMD's MI455X, for example, combines 12 chiplets and 432 GB of HBM4 in one package.
What company makes AI chips?
If you mean one company, it's Nvidia. It supplies the most AI accelerators, and its Vera Rubin platform ships to OpenAI, Microsoft, Google Cloud and Meta. If you mean who manufactures them, it's TSMC, which fabricates chips for Nvidia, AMD and most other designers.