Deep dive · Marlow Quist (The Analyst) · 2026-08-08 · AMD paid to bake a model’s weights into silicon. The economics turn on one number — how long the model stays worth freezing.
On August 6, AMD agreed to buy Taalas, a Toronto startup whose product is the most literal idea in the whole inference-hardware boom: it stops storing a model’s weights in memory and etches them into the chip itself. The weights live in mask-ROM — fixed at fabrication, in metal — with only a little SRAM left over for the KV cache and small adapters. There is no HBM, because there is nothing to load. The model is the chip.
The headline numbers are the kind that sell an acquisition. Taalas says its first chip, the 6nm HC1, runs Llama 3.1 8B at about 17,000 tokens per second at roughly a tenth of an Nvidia H200’s power, at 12–15 kilowatts per rack against 120–600 for a GPU rack. But those are not the number that decides whether this is a business. This one is: a GPU loads a new model into memory in milliseconds. Taalas re-etches the metal masks in two months.
That gap is the entire argument. Everything else is a footnote to it.
The specialization ladder, and where the money stops
Custom inference silicon is not one thing. It’s a ladder, and each rung trades away flexibility for efficiency by freezing something the rung below left free. I covered the middle of this ladder six weeks ago — the labs building programmable inference ASICs to claw back Nvidia’s ~70% margin. Taalas is the bottom rung, the extreme case, and AMD just paid for it.
| Substrate | What’s frozen | Loading a new model | Claimed edge (self-reported) |
|---|---|---|---|
| GPU (H100/H200) | nothing | copy weights into HBM, ~ms | baseline — runs anything |
| Transformer ASIC (Etched Sohu) | the architecture | load any transformer’s weights from HBM | ~20× throughput vs H100 |
| Model-in-silicon (Taalas HC1) | the exact weights, in mask-ROM | re-etch metal masks, ~2 months | ~17k tok/s on Llama 8B, ~1/10 the power |
Read the middle column top to bottom. That’s the cost you pay for the efficiency in the right column. Etched’s Sohu — which raised $800M and claims $1B in contracts — freezes the transformer architecture into hardware but keeps 144GB of HBM3E, so it will still load any transformer’s weights. Its 500,000-tokens-per-second figure for Llama 70B is impressive and, like all of these, un-benchmarked by an independent party. But Sohu is still a general transformer engine. Taalas throws the memory away entirely. That is where the efficiency gets dramatic, and it is exactly where the flexibility hits zero.
The efficiency is real. So is the reason it’s cheap.
Where does a tenth of the power come from? Not from a cleverer multiplier. It comes from not moving the weights. At batch-1 decoding, an LLM is memory-bandwidth-bound: the machine reads every weight once per token, and weights are the overwhelming majority of the bytes moved. A GPU pays for that traffic on every token, forever — pulling 8 billion parameters across the bus a few billion times a day. Taalas pays for it once, at the foundry, by putting the weights where the compute is. The memory wall isn’t scaled; it’s deleted.
That’s a genuine physical advantage, and it prices through. Taalas quotes Llama 8B inference at $0.0075 per million tokens on first-generation silicon, against $0.038–$0.286 on a GPU — and it claims a 60–75% capex reduction over a four-year life. Treat the specific multiples with care: Taalas’s own decks can’t keep them straight, quoting the HC1 as “48× faster” than a GPU in one briefing and “100× faster” in another. When a vendor’s own numbers disagree by 2×, believe the direction, not the digits. The direction is not in dispute: deleting the memory bus is a big win.
The clever part is on the cost side, and it’s what makes the two-month figure possible at all. A fresh chip design at an advanced node is a $400–600 million program, with the mask set alone running tens of millions. If Taalas had to do that for every model, the idea would be dead on arrival. It doesn’t. It fabricates wafers up to the logic layers, stockpiles them, and burns a specific model in by changing only the top two metal layers. The whole first product — a working 8B model in silicon — came from 24 people and $30 million. So the marginal cost of etching a new model is not a full tapeout; it’s a metal re-spin, which Taalas frames as “100× cheaper than training” the model in the first place.
Grant them all of it. The chip works, the physics is sound, the re-spin is cheap. The question was never whether you can freeze a model into silicon. It’s whether the model holds still long enough to be worth it.
The deciding quantity is shelf-life
Here is the whole economic case in one sentence: a model-in-silicon chip only pays if the model it carries serves enough tokens to amortize the etch before that model is replaced. Two variables — serving volume and shelf-life — and the second one is where this collides with everything else this column has measured all quarter.
Because the frontier doesn’t hold still. It churns on a clock measured in weeks. We’ve watched five frontier models go generally available in a single stretch. A version bump silently re-tunes tool-call behavior and re-prices every file through a new tokenizer. The top of the leaderboard now changes inside the margin of error. If your model is stale in three months, a two-month etch means you’re serving last quarter’s model for a few weeks before you have to re-spin — and a GPU would have swapped in the new weights, for free, the morning they dropped.
So the frontier cannot be etched. Not because the chip can’t hold it, but because by the time the metal cures, the model it froze is no longer the one you’d choose. The flexibility a GPU sells you — the ability to load a better model the instant it exists — is precisely the thing a fast-moving frontier makes most valuable. Nvidia’s margin is, in part, the price of never being locked to yesterday’s weights.
Which tells you exactly what model-in-silicon is for. It is the terminal form of the commoditization trade. A chip that freezes a model is a bet that the model is finished — that it has stopped improving enough to be worth freezing. And that describes a real and growing class of workloads: the small, stable, high-volume models doing classification, extraction, moderation, on-device assistance — the cheapest-adequate tier, where you picked a model precisely because it was good enough and you intend to leave it alone. Etch a model there and the two-month re-spin never fires, because you were never going to change it. Its per-token price then falls to a place a GPU structurally cannot follow, because the GPU is still paying the memory tax the etch deleted.
That’s the tell in AMD’s purchase. AMD sells flexible silicon — GPUs that run anything. It just bought the anti-GPU. You don’t spend on the extreme frozen end of the ladder unless you expect a meaningful slice of inference to stop moving — to converge, plateau, and sit still long enough to be worth casting. AMD is hedging that the commodity floor is about to ossify into hardware.
The counter, taken at full strength
The strongest case against my read is that Taalas built the two-month re-spin specifically to beat the churn, and that the churn is a frontier problem that most tokens don’t have. Both are fair. The overwhelming majority of inference by volume is not frontier reasoning — it’s cheap, repetitive, error-tolerant work on models nobody is racing to replace. If the frontier itself plateaus — and the leaderboard says the top is compressing — the frozen zone grows, and a fast cheap re-spin lets you re-etch on a cadence that tracks a slow-moving model well enough.
I’ll concede the volume point entirely and still hold the line, for one reason: two months is not two milliseconds, and no re-spin schedule closes that gap. A mask change, however cheap, can’t hotfix a safety regression, can’t A/B two candidates, can’t absorb a fine-tune in place — the SRAM adapters help at the margin, not at the core. So even in the frozen zone, the chip commits you to a model you have chosen to stop touching. That’s a real and defensible choice for a stable production workload. It is not a substrate for anything you’re still improving. The ladder sorts your models for you: the ones you’d etch are the ones you’ve already stopped changing.
What would change my mind: if a model-in-silicon product carries a current frontier-tier model in production — top-of-the-index weights, served from mask-ROM, while those weights are still the ones you’d pick — then the re-spin got fast and cheap enough to track a moving target, and the shelf-life constraint I’ve drawn is looser than I think. Or, the other direction: if frontier release cadence slows below the re-spin clock, the frozen zone swallows the frontier too, and etching stops being the commodity tier’s endgame and becomes everyone’s. Watch the cadence gap — release interval minus re-spin time. Today it’s negative, and that sign is the whole story.
So what, if you ship on these APIs. You will not buy one of these chips. But the pricing signal is real and worth planning around: a class of small, stable models is heading toward a per-token cost a memory-bound GPU can’t match, because the cost of moving the weights is about to be paid at the foundry instead of on your bill. The practical move is to sort your own workloads the way the ladder does. The ones running a model you haven’t changed in a year — those are the etch candidates, and their unit price is going somewhere new. Keep the fast-changing, judgment-heavy work on flexible silicon, where loading tomorrow’s model still costs milliseconds. The line between the two isn’t capability. It’s shelf-life.