One model per chip: AMD bought Taalas, whose test part hit 17,000 tokens a second

AMD is buying Taalas, a three-year-old Toronto startup that etches AI models into silicon. The deal landed on August 6 with no price attached. What Taalas sells is a chip that runs one model, forever, and cannot be reprogrammed to run another.

That sounds like a downgrade until you look at the number attached to it. Taalas’ HC1 test part served Meta’s Llama 3.1 8B at roughly 17,000 tokens per second per user, which the company claimed in February was 73 times an Nvidia H200 at one-tenth the power. Every AI company on earth is currently paying for inference serving by the token and buying power by the megawatt. A part that changes both figures by an order of magnitude is worth a hard constraint, and AMD just decided how hard a constraint it’s willing to accept.

What Taalas actually built

The company’s slogan is “the model is the computer,” and it means that literally. A conventional accelerator keeps model weights in memory and hauls them across a bus to compute units for every single token, which is why high-bandwidth memory costs what it does and why the memory bus, not the math, is usually the bottleneck. Taalas removes the trip. The weights and the dataflow are cast into the transistors themselves, so the model isn’t data the chip reads. It’s the wiring.

That design would normally be economic suicide, because a new chip per model means a new mask set per model. Taalas’ answer is structural: only two of the chip’s 100-plus layers change between one design and the next, per SiliconANGLE’s reporting, and its own tooling gets a design from weights to tape-out in about two months. The HC1 ran on TSMC’s 6nm process node, a deliberately unexotic choice, since the whole point is that you don’t need a leading process node when you’ve deleted the memory traffic. A second part, HC2, targets models around 20 billion parameters.

Ljubisa Bajic founded the company in 2023 after running Tenstorrent, and raised about $219 million, including a $169 million round in February 2026 from Quiet Capital, Fidelity and Pierre Lamond. “We founded Taalas to rethink AI inference from the ground up by building the hardware around the model,” Bajic said in AMD’s announcement. Read that sentence twice. Everyone else builds the model around the hardware, which is why quantization, kernel fusion and attention rewrites are entire engineering disciplines.

The trade you’re making

One model per chip is a brutal deal in a market where a frontier model has a useful life measured in months. Buy silicon for today’s model and you own a paperweight the week its successor lands. That objection is real, and it’s why nobody funded this idea seriously until recently.

Two things changed. First, open-weight models stopped churning as fast at the small end. An 8B or 20B open model that’s good at one job now stays deployed for a year or more, because the workload is classification, extraction, routing or retrieval rather than chat, and nobody upgrades a working router for fun. Second, inference volume exploded past the point where dedicated hardware pencils out. If you serve one model billions of times a day, a two-month tape-out is a rounding error against the power bill, and the ASIC math starts looking like the same math that moved Bitcoin mining off GPUs in 2013.

It helps to see where Taalas sits against the other companies attacking the same bottleneck. Groq and Cerebras also blame memory rather than math, and both answer with enormous on-chip SRAM so weights never leave the die. Crucially, their parts stay programmable: load a different model, run a different model. Taalas gives that up completely, which is why CNBC framed the pitch as thousands of times faster on specific models with zero flexibility in exchange. It’s the difference between a very fast general road and a railway. The railway wins on the route it was built for and is useless anywhere else.

The honest caveat is that the 73x claim is Taalas’ own, measured on its own test chip, against a two-generation-old Nvidia part, on a model small enough to fit the approach. Nobody outside the company has reproduced it. AMD paying real money for the team is a meaningful vote, though it isn’t an independent benchmark, and AMD hasn’t said when a Taalas-derived part ships. Quartz noted the deal is still subject to regulatory clearance, so even the corporate part isn’t finished.

Where this fits in AMD’s stack

AMD isn’t positioning this as a GPU replacement, which is the sane read. Taalas silicon goes into Helios racks next to Instinct accelerators and EPYC CPUs, programmed through ROCm, so a rack can put general-purpose parts on the models that keep changing and fixed-function parts on the ones that don’t. “AMD is building a full-stack AI platform that gives customers the flexibility to deploy the right compute solutions for every AI workload,” said Vamsi Boppana, who runs AMD’s AI group, in the announcement. Taalas’ technology, he added, strengthens the portfolio “by delivering differentiated inference performance and efficiency.” That’s the actual pitch: heterogeneity as a product, not a single hero chip.

The software story is the part worth watching, and AMD said nothing about it. ROCm has spent years catching up to CUDA on general GPU work. Adding a class of part that runs exactly one model means a new compilation path, a new deployment model, and a new question for every platform team about which of their models are stable enough to freeze into hardware. Nobody has shipped a good developer experience for that yet, because nobody has needed one.

It’s also AMD’s third AI acquisition in nine months, after MK1 in November 2025 and Mext in June 2026, with the FastFlowLM team joining in July. Nvidia made the equivalent move by licensing Groq’s technology in December 2025. Both companies have concluded that the general-purpose GPU can’t hold the whole inference market on its own, and both are buying their way to the specialized end rather than building it. AMD shares rose about 1.5% on the news, which is roughly the market saying “reasonable, small, tell us more.”

The pattern extends past the chip vendors. Anthropic is staffing its own silicon team for the same reason, and AMD’s own Helios rack with 72 GPUs and 31TB of memory exists because memory capacity became the axis of competition. Taalas attacks that axis by refusing to play on it.

What this means for you

Nothing you deploy this year changes. The deal hasn’t closed, no product has a name or a date, and the first AMD parts built on this will land in racks, not in anything you rent by the hour for a while yet.

What’s worth adjusting now is how you think about model choice. If you’re running a small open-weight model at high volume for a narrow job, you’re the customer this hardware is designed for, and the deciding variable becomes how long you can hold a model still. Teams that pin a model version for a year are about to get a cost curve that teams chasing every new release will never see. That’s a genuinely new tradeoff, and it runs opposite to how most AI teams have operated since 2023. Start measuring what a model upgrade actually buys you in production, because within two years the answer might have a power bill attached.