AMD Bought Taalas, Which Burns AI Models Into Chips; Nvidia Rose

On August 6, AMD announced it had bought Taalas, a Toronto company of roughly 25 engineers. No price was disclosed.

The next day the market delivered its verdict: AMD fell about 2%, while Nvidia rose 4%.

An acquisition where the buyer drops and the rival rallies looks backwards. The answer is what Taalas built: etching the model into the chip.

Nobody else can put an eight-billion-parameter model on one die and read nothing from external memory to run it.

What It Means to Etch Weights Into Silicon

A language model is billions of numbers called weights; everything it has learned lives in them.

On a conventional GPU the weights sit in HBM stacks beside the compute die — fast, but still a separate layer.

Taalas inverted that: the weights are etched into the transistors' wiring during fabrication — mask ROM, a fixed circuit printed once at the foundry. A second region of the die is SRAM, holding the KV cache and fine-tuning adapters.

The result is a model-specific integrated circuit, or MSIC: one model, run straight out of the silicon with no external fetch.

Bajic's Second Attempt

Taalas was founded in 2023 by Ljubisa Bajic, who started Tenstorrent in 2016 and designed processors at AMD and Nvidia before that. Co-founders Drago Ignjatovic and Lejla Bajic came from Tenstorrent's early ranks. AMD has just bought the second company its own alumni started.

The $50 million seed came from Pierre Lamond and Quiet Capital; Fidelity joined the $169 million round.

Taalas did not sell for lack of money: it held more than $170 million in unspent cash, and HC1's development reportedly cost roughly $30 million. A working chip for $30 million is unusual; ten times that is routine here.

The Memory Wall: What Actually Limits a GPU

AI chips are discussed in terms of compute, but text generation is limited by memory speed — the memory wall.

The arithmetic finishes quickly; the time goes to the second link. For every fragment of a word the model rereads all of its weights, while the compute units sit idle.

This is how Taalas quotes 16,960 tokens per second: not an engineering refinement but a relocation of the constraint. AMD's Helios systems push bandwidth to 19.6 TB/s per MI455X, six times an H100 — which moves the wall back rather than removing it.

On cost, the sources conflict. TrendForce reports 0.75 cents per million tokens against 3.79 cents for Nvidia's B200; another outlet gives the same figure as $0.75, inverting the comparison. Until an independent measurement exists, treat it with caution; every performance claim comes from the company.

Why the Market Rewarded Nvidia Instead of AMD

Friday, August 7 opened with a weak July employment report — nonfarm payrolls fell by 23,000, September rate-hike odds were priced out, indices rose. The interesting move was in the spread between two chip stocks.

Investors read Taalas not as a threat to Nvidia but as AMD's long-dated bet. William Blair's Sebastien Naji called the move "strategic" while noting the technology is less proven. Benchmark's Cody Acree said the near-term gap does not close until AMD names a product, a customer, a software integration and a shipment schedule.

That is also why Nvidia rallied: one model on one die does not change this quarter's standing of the market leader.

Nvidia's position does not rest on hardware alone; CUDA is the layer customers have written code against for years.

AMD's data center business grew 107% year over year in the second quarter and made up 58% of revenue while gaming contracted 31% — two narratives in one earnings statement.

Consolidation took two deals in eight months; independent inference-chip makers are disappearing fast.

The Other Side: What Is Etched Cannot Be Unetched

The cost comes from the same place as the benefit.

Model lock-in is the most serious constraint. A model change means respinning two metal layers and takes two months — not short where model lifespans are measured in months.

A second: once a security flaw is etched into silicon there is no patch to ship. Third, area — the same 4 GB of weights takes about 80 square millimeters in DRAM, a tenfold penalty that rules the architecture out for phones and laptops.

What Is Left

What AMD bought is a method and the 25 people who built it, not a product. The deal should close in the fourth quarter, after which the technology is meant to reach Instinct accelerators, Helios racks, EPYC processors and the ROCm stack — no timeline given. Which is why a valuation is impossible: no price, no revenue, no customers.

If the claim is right, GPU makers are not the only ones holding the bill: this architecture wants neither HBM nor TSMC's CoWoS line. HBM is the most profitable product Micron, SK Hynix and Samsung sell, and packaging is the tightest bottleneck; if it spreads, both weaken.

A decade-long question, not a quarterly one — but the asking is already part of the volatility in memory names.

The Price of Giving Up Flexibility

This piece draws on AMD's investor relations announcement, public statements from AMD and Taalas, and technology and financial press coverage dated August 6–7, 2026. Performance and cost claims originate with the company and have not been independently verified; conflicting figures are noted in the text. This is not investment advice.

Correction (September 13, 2026): prose tightened for rhythm; figures and sources unchanged.