AI inference.
Cheap tokens, big bills.

Google processed 9.7 trillion tokens a month in 2024. This May, it said the number was past 3.2 quadrillion. Over the same stretch, the price of any given level of AI fell about tenfold a year.

A token is the unit all that AI output gets counted in: a small chunk of text that costs compute, memory, and power to produce. Producing them is what inference means. Training builds a model once. Inference runs it every time someone asks a chatbot a question, an agent takes a step, or an assistant writes a line of code. The model reads your whole prompt in one gulp, then writes the answer one token at a time, and that second half is limited less by raw compute than by how fast memory can feed the chip.

So volume is exploding and the unit price is collapsing. That leaves one question hanging over AI inference stocks in 2027: is serving AI a commodity business or a toll road?

Right now, the hardware side is winning that argument. Micron kept 87 cents of every revenue dollar as gross profit last quarter. Cerebras, which sells much of its capacity as tokens from its own cloud, kept 14. Tokens are getting cheap. The parts that make them aren't.

These AI inference stocks run from the companies keeping the most margin to the ones keeping the least: the toll collectors, the giants designing their own chips, the price fighters, and the companies selling tokens directly.

These two own the parts nobody can serve tokens without, and so far they keep the margin no matter who wins downstream.

Nvidia NASDAQ: NVDA

Last December, Nvidia agreed to pay about $17 billion to license the technology of Groq, a startup whose whole pitch was that GPUs are the wrong tool for generating tokens fast. In August it shipped the result: Groq 3 LPX, an accelerator that sits beside its flagship Vera Rubin systems and handles the latency-sensitive part of writing answers.

That's the tell. The company with a 75% gross margin decided the threat from specialized inference chips was real enough to build one itself. Nvidia would rather cannibalize its own GPU sales than let someone else set the price of a fast token.

Micron Technology NASDAQ: MU

Micron is the purest bet that memory, not compute, is what inference runs short of. Every token a model writes means pulling its weights and the running memory of the conversation back through the chip. Longer contexts and longer agent sessions make that pile bigger.

Micron did $54 billion of revenue last quarter, nearly five times a year earlier. It's now shipping data-center SSDs aimed specifically at holding that conversation memory when it spills out of HBM. Even Nvidia's purchase commitments more than doubled last quarter, mostly to lock up memory.

The biggest token producers don't want to pay a toll forever, so they design their own silicon and hire these companies to help build it.

Alphabet NASDAQ: GOOGL

Alphabet owns the model, the chip and the cloud, so it hands less of each token's cost to outsiders than anyone else. Its Gemini API alone now handles 22 billion tokens a minute, about triple the rate of nine months earlier.

The bigger shift is that Google now sells its TPUs to others. It split its latest generation into a training chip and a dedicated inference chip, and it sells TPU systems as hardware. Anthropic has committed to multiple gigawatts of them starting in 2027. A chip built for one company's cost structure is now a product, and Google has every reason to price it below the merchant alternative.

Broadcom NASDAQ: AVGO

Broadcom is the contractor behind Google's TPUs and a handful of other custom AI chips. When a big buyer decides it's done paying Nvidia's margin, Broadcom is usually who it calls. AI chip revenue hit $16.7 billion last quarter, more than triple the year before, and management guided to $21.7 billion for the next one.

Marvell Technology NASDAQ: MRVL

Marvell just signed the most interesting contract on this list. In July it agreed to build custom chips for Google's TPU ecosystem, including inference accelerators, and gave Google a warrant that vests in slices as that custom business grows.

The structure tells you the ambition. The warrant vests in 240 tranches, one for every $500 million of custom revenue through fiscal 2033. Vest all of them and that's about $120 billion of chips.

These two sell merchant chips into the same market as Nvidia, and their pitch is simple: more tokens for the dollar.

AMD NASDAQ: AMD

AMD is the credible second source, and in July it got the customer that proves it. Anthropic agreed to deploy up to two gigawatts of AMD's next-generation MI450 chips, with the first gigawatt starting in the first half of 2027. AMD is also putting up to $5 billion into Anthropic, so it has skin in the outcome.

AMD's own marketing leads with "inference tokens per dollar," not peak speed. It's also willing to share the job. In a new partnership, its Helios racks read the prompt and Cerebras writes the answer, which is not something a company chasing a monopoly would sign up for.

Qualcomm NASDAQ: QCOM

Qualcomm is a pivot story. The company that spent two decades squeezing performance out of phone batteries is now selling inference racks, on the theory that tokens per watt matters more than peak speed in a power-starved data center.

It has an anchor customer in Saudi Arabia's HUMAIN, a second-generation rack due in 2027, and a target of more than $15 billion in data-center revenue by fiscal 2029. It also bought Modular, a software company focused on inference, to make its chips easier to program. None of that is proven yet. Qualcomm is the cheap option on the idea that the inference market gets big enough for a power-efficiency specialist.

These companies sell tokens directly, so they'd be the first to feel it if inference turns into a commodity, and the first AI inference stocks to show it.

Cerebras Systems NASDAQ: CBRS

Cerebras builds chips the size of a dinner plate and sells what they produce as some of the fastest tokens on the market. Its design uses no HBM and no CoWoS packaging, the two parts everyone else is fighting over.

OpenAI committed to 750 megawatts of Cerebras inference capacity through 2028, and Cerebras now has $25.4 billion of contracted future revenue against $180 million in quarterly sales. One Abu Dhabi university still accounted for 34% of last quarter's revenue, and gross margin was 14% by standard accounting, 41% on the company's adjusted measure. If speed earns a premium, Cerebras grows into the backlog. If it doesn't, the company is selling commodity tokens on very expensive hardware.

Fireworks AI (Private)

Fireworks AI doesn't make chips. It runs open models on other people's hardware and sells the output, more than 40 trillion tokens a day. It passed $1 billion in annual recurring revenue and raised at a $17.5 billion valuation in July. Fireworks is the cleanest test of whether software that makes tokens cheaper can keep any of the savings for itself.

Today a token is a token. Prices are quoted per million, and the cheapest provider wins.

That won't hold through 2027, because the hardware is already splitting. Nvidia sells a separate accelerator just for fast output. Google split its TPUs into training and inference chips. AMD hands answer-writing to Cerebras. All of it exists to make some tokens come out faster than others, at a higher cost.

Agents are what make that speed worth paying for. A chatbot reply that takes three seconds is fine. An agent chaining a thousand model calls turns a one-second delay on each into a quarter-hour wait.

So inference becomes two markets. Bulk tokens get priced like freight, and fast tokens get priced like overnight shipping. The bulk side keeps getting cheaper and fights over pennies. The margin moves to the fast lane, and to whoever owns the memory bandwidth that sets its speed limit. Those are the AI inference stocks worth owning.