16,960 tokens a second, from a single chip drawing about 200 watts.

That was the number Taalas put on the table in February, when it pulled its HC1 processor out of stealth. Running Meta's Llama 3.1 8B, the chip finished roughly 48 times faster than Nvidia's GPUs and about 8.5 times faster than Cerebras's wafer-scale engine. It earned that gap by refusing to do the one thing every other AI chip does: keep the model in memory.

Taalas burns the model's weights straight into the silicon. No HBM stack to stream them from. No runtime to schedule them. No software layer in between. The model is not loaded onto the chip. The model is the chip.

On August 6, AMD said it would buy the company, terms undisclosed, the deal expected to close in the fourth quarter subject to regulatory approval. The catch sits on the same datasheet. A Taalas chip runs one model. Change more than a LoRA adapter and you do not retrain anything — you fabricate a new chip. In a field where leading models ship every few months, AMD just paid for silicon that starts aging the day it works. Whether that is a masterstroke or an expensive mistake hinges on one unanswered question. Does inference belong to hardware built for a single model, or to hardware built to survive all of them?

The case for silicon that forgets

Every GPU-based inference stack pays a tax. During token generation, the chip spends most of its energy shuttling weights from memory to compute and back. The arithmetic is cheap. The movement is not. Taalas deletes the movement by deleting the distance: each chip splits into a mask-ROM region where the weights are etched as transistors, and an SRAM region that holds the KV cache and fine-tuning adapters. Nothing streams. Nothing waits.

AMD is not alone in reading it that way. As we wrote in October, Nvidia is funding the chip startup built to beat it — a direction it committed to when it licensed Groq's inference technology at the end of 2025 in a deal reported around $20 billion.

The economics follow the physics. SDxCentral put the HC1 at roughly 14,000 tokens per second, with under 200 milliseconds of latency. It draws about 200 watts; a comparable GPU rack pulls 120 to 600 kilowatts. Train the model once, etch it, and the cost per token falls by orders of magnitude. Taalas's own framing, from an interview with The Next Platform, was blunt: baking a model's weights into silicon is around 100 times cheaper than training a frontier model.

📊
One year, two incumbents, one conclusion
Nvidia licensed Groq's inference tech in December 2025. AMD bought Taalas in August 2026. Both now treat decode as a separate problem from training, and both are willing to bolt non-native silicon into their own racks to solve it.
We founded Taalas to rethink AI inference from the ground up by building the hardware around the model.— Ljubisa Bajic, co-founder and CEO, Taalas

For a model that will not change, that is close to free money. The hardware exists, the throughput is measured, and the power bill collapses.

The case for silicon that stays flexible

Start with what Taalas gave up. Its first chip runs Llama 3.1 8B and nothing else — a model that shipped in mid-2024 and is, in AI years, a fossil. Switching to a bigger or newer model means redesigning the chip. Taalas says it has compressed that to about two months by changing only two metal layers at TSMC, against up to six months to fabricate a processor like Nvidia's Blackwell. Fast, but not instant, and not free.

The scale ceiling is sharper. Because the design is entirely SRAM-based, a bigger model needs more chips. DeepSeek's 671-billion-parameter model would take somewhere near 30 separate tape-outs. Taalas's answer is parallelism. Its second-generation HC2, due around the turn of the year, moves to 4-bit MXFP4 weights and targets roughly 20 billion parameters per chip, which pulls a trillion-parameter model inside about 50 accelerators. That fits neatly into AMD's Helios rack platform — and it also means a fleet of chips frozen to one model, replaced when that model moves on.

Then there is the moat that never appears on a die shot. Nvidia's advantage is not only FLOPS. It is CUDA, a software ecosystem that is hard to leave and easy to depend on. AMD has spent years answering it with ROCm, with mixed results. A chip that runs exactly one model does not compete with that ecosystem. It ignores it. That is a strength in a narrow, stable deployment, and a liability in a market where model turnover is measured in weeks.

Where the numbers actually land

🎯
Hard-wired inference silicon wins on cost per token in production.

It loses on flexibility, and model turnover punishes that.

The likely end state is hybrid: GPUs for prompt processing and change, dedicated silicon for decode.
MetricTaalas HC1 (hard-wired)Nvidia GPU (H200/B200 class)
Throughput (Llama 3.1 8B) ✔ ~16,960 tokens/sec ✗ 1× (baseline)
Power draw ✔ ~200 W per chip ◐ 120–600 kW per rack
Model flexibility ✗ One model per chip ✔ Any model, anytime
Model switch cost ✗ New masks, ~2 months ✔ Software update

Sources: Taalas / The Register / SDxCentral, 2026. Throughput and power are vendor claims.

The table is not a verdict. It is the shape of the trade. One column optimizes a single number until it is nearly free; the other keeps the whole board playable.

What AMD is actually buying

The deal is expected to close this quarter, which puts the real test close at hand. Taalas is the third inference bet AMD has placed in about nine months, after MK1 in software and Mext in memory optimization. Read together, the pattern looks deliberate. AMD is not trying to beat Nvidia's GPUs at their own game. It wants GPUs to be the flexible half of a two-part system and to move the expensive half — generating tokens at scale — onto silicon that does one thing perfectly.

That logic holds only if the models hold still. Deploy a hard-wired accelerator and you are signing a fixed contract: the model you bet on today is the model you run until the next fabrication run. For a narrow, high-volume workload — a coding assistant, a fixed enterprise agent — that is a reasonable trade. For anything chasing the frontier, it is a trap with a two-month floor.

The two poles will not be reconciled by a press release. They get settled by the thing the deal is really betting on: whether the next two years bring a stable, standard model that everyone deploys, or the same monthly churn that defined the last four. AMD has now put real money behind one answer. Nvidia, thanks to Groq, is hedged against the other. Taalas's silicon runs fast because it forgot how to change. The market will decide whether that was wisdom or nostalgia.

AMD Acquires Taalas to Advance Compute Solutions for the AI Inference Market
AMD's own announcement of the definitive agreement, the technology rationale, and the integration plan with Instinct GPUs and Helios racks.
The primary source — what AMD says it is buying, in AMD's words.
AMD acquires Taalas to boost inference by etching models into silicon
The clearest technical breakdown of the model-specific chip: mask-ROM weights, SRAM caches, the HC2 roadmap, and the disaggregated GPU-decode architecture.
Where the throughput numbers and the architecture details actually come from.
AI Chip Startup Taalas Acquired by AMD
EE Times on the founders, the two-month tape-out tool flow, and why a 671B model would need dozens of separate tape-outs.
The clearest account of the design trade-off — and its scale ceiling.
Chip designer Taalas bets on hard-wired AI chips
Background on the HC1's ~200 W draw against a 120–600 kW GPU rack, the $169M round, and the frontier-chip target for late 2026.
The power and cost framing that makes the hard-wired bet legible.