Two of the three companies that matter in AI silicon have now bought opposing answers to a single question, eight months apart. Nvidia put $20 billion behind a license to Groq’s design in December. AMD then agreed on August 6 to buy Taalas, the Toronto startup we covered Saturday. Writing on August 8, the hardware blog Kernel argued these are not two chip deals but one shift.
Start with what physically travels. Inside a GPU, the thing in motion is the model itself. Parameters are parked in high-bandwidth memory beside the die, and emitting one token drags them across that link once per layer, then the trip repeats for the next token. The arithmetic is cheap. The commute costs the money, which is why memory outprices logic in an AI server. Both startups asked what would happen if the model held still and the prompt moved.
Taalas answered by turning the parameters into wiring. Its numbers are laid down as read-only structures while the wafer is patterned at the foundry, and each transistor pulls double duty, holding four bits and doing the arithmetic they govern. Activations fall through the chip’s layers in one direction, retrieving nothing from outside the package. A separate rewritable zone of ordinary SRAM sits alongside, the part most AMD coverage skipped. It carries the KV cache plus LoRA adapters, so a customer can trade a legal adapter for a clinical one while the part is live. Adapters bend the frozen model toward a vocabulary. They cannot grant it an ability the etched version never had.
Groq declined that last step. Its LPU also does without external memory, but the parameters land in conventional static RAM spread across the die, with their placement and the timing of every operation resolved by software before the job starts rather than while it runs. The model therefore stays swappable. Loading a different one means compiling again and reloading, a matter of minutes with no new hardware. The penalty lands on headroom. Kernel puts the first LPU generation near 230 MB per part, too little for a 7B model at 8-bit precision, so a 70B deployment sprawls across 576 of them. One Taalas die swallows an 8B model whole.
Taalas wagers that one set of numbers stays useful long enough to repay its masks. Groq wagers only that the transformer endures while the numbers inside keep turning over. One is costly per model, the other per rack.
Kernel’s larger thesis is that specialization always resolves the same way. Whenever a workload has been pulled off general-purpose hardware, the author writes, it has permanently rewritten which companies could afford to run it. The examples run from bitcoin mining’s slide into ASICs to the fixed-function video decoder in every phone. The pattern holds. The fit to inference does not. In each precedent a published standard had already frozen the workload before anyone cut a mask, and SHA-256 was ratified in 2002. Weights are nobody’s standard. Kernel supplies the damaging detail itself: the first Taalas part reached customers in February carrying Llama 3.1 8B, a nineteen-month-old release. The affordability half runs backwards too, since Taalas reached silicon with 24 staff and $30 million.
The number neither press release mentions is the cost of reversing a decision. When inference lives in fixed silicon, replacing a model stops being a configuration change and becomes a procurement cycle. Kernel relays an outside estimate near $3 million in masks per chip variant, before fabrication, validation, and the labour of pulling cards from racks. A switching cost that large rewards anyone who has already chosen correctly.
That favours the biggest inference vendors, whose volume on one settled model absorbs a mask set, and whichever lab’s parameters get cast into a customer’s fleet. Hardware lock-in outlasts any contract. The losers are the mobile: teams that re-shop providers quarterly on price, open-weight shops treating a checkpoint bump as free, and internal eval groups, whose leverage dies once a model change needs a purchase order. Narrow, settled jobs invert the logic. Transcription, OCR, translation, retrieval embeddings, and moderation have hardly moved in years and would clear a mask bill.
Both purchases share one exposure: each assumes text arrives a word at a time. Diffusion models generate a whole span and then polish it, and Kernel points to Google’s Gemini Diffusion and Inception’s Mercury as shipping evidence.
Anyone signing 2027 inference capacity should negotiate the switch, not the token. Ask whether changing models costs minutes or a new mask set.
Analysis published by Kernel on August 8, 2026.
