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

On August 6, AMD announced it had bought a small Toronto company. No price was disclosed. The company is called Taalas, it employs roughly 25 engineers, and it has raised $219 million to date. 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 at first glance. This piece explains why that happened. But the reaction only makes sense once you understand what Taalas actually built, because it is the most radical bet AI hardware has seen in some time: etching the model itself into the chip.

All four numbers collapse into a single sentence: an 815-square-millimeter die holds an eight-billion-parameter language model inside itself and reads nothing from external memory to run it. Nobody else in the industry does this.

What It Means to Etch Weights Into Silicon

A language model is billions of numbers. Those numbers are called weights, and everything the model has learned lives in them. On a conventional GPU, the weights sit in HBM stacks — high-bandwidth memory — lined up beside the compute die. HBM is fast and physically close to the tensor cores, but it is still a separate layer of memory.

Taalas inverted that. On its chip the weights are not written to memory at all; they are etched into the wiring pattern of the transistors during fabrication. The technique is called mask ROM: a fixed circuit printed once at the foundry and never changed afterward. A second region of the die is SRAM — fast on-chip memory — holding the KV cache that accumulates during a conversation, plus fine-tuning adapters.

The result is a model-specific integrated circuit, or MSIC in the company's own terminology. The chip runs exactly one model and nothing else. In exchange, it runs that model straight out of the silicon, with no external fetch at all.

Bajic's Second Attempt

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

The company came out of stealth in March 2024 with a $50 million round led by Pierre Lamond and Quiet Capital. In February 2026, alongside the unveiling of its first chip, the HC1, a second round of $169 million arrived, with Fidelity participating.

One detail deserves attention. Taalas did not sell because it ran out of money. At the time of the announcement it still held more than $170 million in unspent cash, and HC1's development is reported to have cost roughly $30 million. A hardware company producing a working chip for $30 million is unusual; ten times that figure is routine in this industry. This looks like a sale by choice, not by necessity.

The Memory Wall: What Actually Limits a GPU

Discussions of AI chips almost always center on compute. But when a model is generating text, compute is not the binding constraint — memory speed is. The industry calls this the memory wall.

The third link in that chain — the actual arithmetic — finishes very quickly. The time is spent on the second. For every single fragment of a word, the model must read all of its weights again, and then repeat the whole thing for the next token. The compute units sit idle waiting for that read to finish.

Doing the Arithmetic: How Many Milliseconds Per Token

The wall gets concrete once you put numbers on it.

This is how Taalas can quote a figure like 16,960 tokens per second. The gap is not an engineering refinement; it is a relocation of the constraint. AMD's own next-generation Helios systems push memory bandwidth to 19.6 TB/s per MI455X, roughly six times an H100 and a genuinely impressive jump. It still only moves the wall back rather than removing it.

On cost, the sources conflict, and there is no reason to paper over it. The comparison TrendForce reports is 0.75 cents per million tokens against 3.79 cents for Nvidia's B200 — about one-fifth the cost. Another outlet gives the same figure as $0.75, which would invert the comparison entirely. Both cannot be true. Until an independent measurement is published, the number deserves caution; every performance claim so far comes from the company itself.

Why the Market Rewarded Nvidia Instead of AMD

Friday, August 7 opened with an unexpectedly weak July employment report: nonfarm payrolls fell by 23,000. Odds of a September rate hike were priced out quickly and the indices rose. The more interesting move that day, though, was not in the index but in the spread between two chip stocks.

The distinction is this: investors read Taalas not as a threat to Nvidia's throne but as a long-dated bet by AMD. The analyst language ran the same way. William Blair's Sebastien Naji called the move "strategic" while noting the technology is less proven than its competitors'. Benchmark's Cody Acree was blunter: the near-term competitive gap does not close until AMD discloses a product, a customer, a software integration and a shipment schedule.

That is also why Nvidia rallied. An architecture that fits an eight-billion-parameter model onto a single die does not change this quarter's standing of the company that owns the data center AI chip market.

No one dislodges that split in a quarter. And Nvidia's position does not rest on hardware alone: CUDA is the software layer customers have been writing code against for years. Swapping a chip is easy; abandoning a software ecosystem is not. Nor has Nvidia left this flank open — in December 2025 it acquired the assets of inference specialist Groq for roughly $20 billion. The size gap between that deal and AMD's undisclosed one is itself a measure of what each company has committed to the category.

The three-month path in that chart carries a longer story than one deal day. AMD's data center business grew 107% year over year in the second quarter and made up 58% of total revenue, while gaming contracted 31% over the same period. Two different narratives are running through the same earnings statement.

Timeline

What the calendar shows is that consolidation on the inference side was largely completed by two deals inside eight months. The number of independent inference-chip makers is shrinking fast.

The Other Side: What Is Etched Cannot Be Unetched

The cost of the architecture comes from exactly the same place as its benefit. What is permanent is also what cannot be changed.

Model lock-in is the most serious constraint. The moment a die leaves the fab, the model it will run is fixed. The company says a model change requires respinning only two metal layers and takes roughly two months — genuinely cheap next to a full chip redesign. But in a field where model lifespans are measured in months, two months is not short.

A second constraint gets discussed less: if a security flaw is found in a model after it has been etched into silicon, there is no software patch to ship. The vulnerability is in the hardware.

Third is area efficiency. Etching 4 GB of weights into silicon takes roughly 800 square millimeters; holding the same data in DRAM takes about 80. That tenfold penalty rules the architecture out for phones and laptops.

What Is Left

What AMD bought is not a sellable product today; it is a method and the 25 people who built it. The deal is expected to close in the fourth quarter, after which the technology is meant to be integrated into Instinct accelerators, Helios rack-scale systems, EPYC processors and the ROCm software stack. No timeline was given for any of that.

Which is why a valuation is impossible right now: no price, no revenue, no customers. What exists is an architectural claim and an arithmetic account of what changes if it turns out to be right.

And if it is right, GPU makers are not the only ones holding the bill. This architecture wants neither HBM nor TSMC's CoWoS packaging line. HBM is the most profitable product the memory makers — Micron, SK Hynix, Samsung — currently sell, and packaging capacity is the industry's tightest bottleneck. An architecture that etches weights into silicon weakens both. This is a decade-long question rather than a quarterly one, but the fact that it is being asked at all is part of the recent volatility in memory names.

The index's overall direction does not yet reflect any of this. These are changes that will show up in supply contracts before they show up in quarterly reports.

Deals with undisclosed prices usually signal one of two things: either the number is immaterial to the buyer, or it is embarrassing to the seller. Given the $170 million still sitting in Taalas's account and the founders' track record, the first explanation looks more likely.

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.