I’ll translate this article from Thai to English, preserving all MDX components and structure.
The blog article translation:
AMD acquired Taalas to get technology that “burns” AI models directly into chips, instead of running them on general-purpose GPUs like has always been done. The idea is that if a model doesn’t need to change often, you might as well fix it into an actual circuit — you get faster speed and lower power draw.
- Why burn it into silicon: General-purpose GPUs are designed to be flexible, able to run any model, but that flexibility trades away speed and power efficiency that isn’t strictly necessary.
- What AMD gets out of it: A boost to its inference-side arsenal to compete with Nvidia in the fastest-growing market right now — actually running models in production, not just training them.
- Does this really change the battlefield: I think we still need to wait and see, because AI models change versions extremely fast. “Burning” one into a chip means updates become a lot harder.
What this deal actually looks like
Put simply, Taalas builds ASICs that “burn” the structure of an AI model directly into silicon — not general-purpose chips that load a model and run it afterward. The concept is to cut out unnecessary steps and make the hardware and the model one and the same.
AMD bought this company specifically to strengthen its inference portfolio. Unlike general-purpose GPUs, which are flexible but power-hungry, this kind of chip is built for a single task but is far faster and more efficient.
The interesting part, in my view, is that AMD is betting the inference market will grow large enough to justify this kind of specialized investment — not just buying more GPUs and calling it a day.
Why this matters if you actually run AI in production
Every infra team runs into the same problem: the GPU cloud bill keeps climbing, but latency still isn’t stable enough for real production use.
General-purpose GPUs are built to be flexible and run any model, but that comes at the cost of power draw and per-token cost that’s higher than it needs to be. Once a model has settled and gets run millions of times a day, that flexibility turns into an unused luxury nobody actually needs.
Honestly, this is the pain point every ML team quietly complains about — but with no alternative besides buying more GPUs.
Fixed-function chips that etch the model directly into silicon address this exact problem. They cut away flexibility that isn’t needed, in exchange for a big jump in speed and efficiency. For workloads that run the same thing over and over, this is a more sensible path than continuously piling on more GPUs.
To be clear, this doesn’t replace AMD’s MI300/MI350 series — the two flagship lines still handle flexible training and inference workloads. Taalas fills in the gap general-purpose GPUs aren’t great at: inference for models that have already stabilized and don’t need frequent retraining.
What’s notable is that Taalas isn’t tied to the ROCm ecosystem the way traditional GPUs are, since it’s a chip that etches a specific model straight into silicon, cutting out the complex software stack layer entirely.
Compared to Nvidia, which is still fully committed to GPU-first plus the CUDA moat, I think AMD is opening up a different game entirely — not chasing Nvidia on the same field, but finding new ground where efficiency wins outright, starting at the silicon level.
Viewed as a portfolio play, this is risk diversification — no longer going all-in on GPUs alone.
What changes before and after the deal
In practice, these two approaches serve different jobs — one doesn’t replace the other.
The old approach (GPU-based) is highly flexible, can run any model, and can adjust workloads on the fly. It’s suited for training or workloads that haven’t stabilized yet.
The new approach from Taalas (etched silicon) takes a model that has already stabilized and embeds it directly into the chip, cutting away the power-hungry software layer. It’s suited for inference that gets deployed repeatedly at high volume.
| Factor | GPU-based (Traditional) | Etched Silicon (Taalas) |
|---|---|---|
| Flexibility | Runs any model | Fixed to 1 model |
| Power efficiency | Higher power draw | More efficient for inference |
| Time-to-deploy | Adjust and run instantly | Must re-etch every time the model changes |
| Best suited for | Training / unstable workloads | High-volume, repetitive inference |
I see this as reinforcing the arsenal, not switching lanes entirely — AMD still needs GPUs for training; this just adds an option for actual deployment.
What it’s like in practice
Picture a data center running LLM inference around the clock, over and over with the same model — that’s exactly what Taalas was designed for. Silicon that etches the model directly in makes inference faster and lower-power than running it on a general-purpose GPU.
It also suits edge inference well — for example, specialized devices that run a single model repeatedly without frequently switching workloads. The power cost per query has room to drop too, since the hardware is tuned to match the model exactly, with no overhead from being designed to do everything.
But the weakness is just as clear — once a new model comes out, you have to re-etch the whole batch, not just push a software update. I don’t think fast-moving models like frontier LLMs are a good fit for this approach; it suits workloads that have already “settled” much better.
How it compares to other options on the market
| Factor | Taalas (Etched Silicon) | Groq (LPU) | Nvidia (GPU) |
|---|---|---|---|
| Approach | Etches the model directly into the chip | Purpose-built chip for inference | General-purpose GPU |
| Inference speed | Highest (for that specific model) | Very high | Moderate-to-high |
| Flexibility | Low, must re-etch for every model | Adjustable via software | Most flexible, runs any model |
| Best suited for | Stabilized workloads that don't change often | Specialized inference tasks | Both training and inference |
Honestly, I don’t think these approaches compete head-on — Nvidia still dominates the training market, while Taalas and Groq are playing in the inference arena where extreme speed matters most.
Pros and cons you should know before cheering or criticizing
Pros
- +More power-efficient than general-purpose chips, since it's built for one specific model
- +Lower cost per inference, well-suited to high-volume repetitive workloads
- +Faster than general-purpose GPUs for the specialized task it was designed for
Cons
- −Low flexibility — if the model gets updated or its architecture changes, a new chip may be required
- −Risk of vendor lock-in, since it's tightly bound to a single model and a single piece of hardware
- −Not yet proven at real large scale — this is still an early-stage concept
Honestly, I think this is more of a long-term bet for AMD than a quick win, because AI models change so fast — etching them permanently into silicon risks becoming obsolete just as quickly.
The cost that isn’t on the original price tag
Something people overlook is the cost of “model updates.” If you etch into silicon, then every time a new model version comes out, you have to fabricate a whole new batch of chips — not just push a firmware update.
Another point is that AMD’s ROCm ecosystem is already trailing behind CUDA, and now it has to split R&D resources to also maintain this additional line of fixed-model hardware.
Honestly, I think the real risk isn’t the technology — it’s the commercialization timeline, which nobody has clearly stated will be ready when. In the meantime, AMD has to keep carrying the R&D cost without a large-scale proof of concept to back it up.
I’ll say it plainly: this looks great on paper, but the real cost lies in “speed of adaptation” — which has always been the weak point of fixed hardware.
Who this is for, and who it isn’t
Made for
- Hyperscalers running large, continuous inference workloads who are ready to lock in with specialized silicon
- Enterprise AI teams with a stabilized model architecture that doesn't change often, well-suited to etching into silicon
Think twice
- Companies still experimenting with multiple model types, who need to balance cost against flexibility
Skip this one
- Startups whose models change every quarter — better off with general-purpose GPUs that adapt faster, not fixed hardware
I think this project suits people who “already know what they’ll be running” more than those still figuring it out. If your model hasn’t stabilized yet, it’s better to wait and see how commercialization plays out first.
So what does this deal actually change
Honestly, the AMD-Taalas deal isn’t a direct fight with Nvidia — it opens up new ground that Nvidia hasn’t fully moved into yet: hardware that etches the model straight into silicon.
I think the key shift here is a change in mindset around inference chips — from “flexible, general-purpose” to “fixed-function, fastest possible for one job” — similar to how mining ASICs once changed the game for crypto.
Things worth watching going forward: 1) how far AMD will open this silicon-etching approach to third parties, 2) how hyperscalers like Google/Amazon, who already build their own custom chips, will respond, 3) whether the real cost-per-token once commercialized will be compelling enough for enterprises to migrate off their existing GPUs.
Long term, if model architectures keep trending toward stability, this approach could well become the standard for the inference layer.