The Silicon Trap: Why AMD’s Bet on 'Etched' AI Models Might Be a Capex Nightmare
AI-generated image · US National Wire
Opinion: AMD's acquisition of Taalas promises blistering inference speeds, but in a world of monthly model updates, hard-wiring weights into silicon is a gamble on stability that the market may not support.
Let's be clear: I am a skeptic by trade, and right now, AMD is asking the market to believe in a level of model stability that has never existed in the history of generative AI.
As first reported by The Register, AMD has acquired Toronto-based AI chip startup Taalas. The pitch is seductive: instead of relying on High Bandwidth Memory (HBM) to store model weights, Taalas uses a process that etches those weights directly into the silicon, creating 'model-specific integrated circuits' (MSICs).
On paper, the performance is staggering. The Register reports that Taalas' first test chip, the HC1 (fabbed on TSMC’s 6nm process), served Meta’s Llama 3.1 8B at 16,960 tokens per second—roughly 48 times faster than Nvidia's GPUs and 8.5 times faster than Cerebras' accelerators.
But here is the P&L problem: the moment you etch a model into silicon, you have effectively frozen that model in time. In a field defined by rapid obsolescence, AMD is betting on a hardware architecture where any significant change to the model requires a re-spin of the chips. While Taalas claims that changing only two layers of metal makes re-spins cheaper, it is still a costly ordeal compared to a software update.
AMD’s strategy is to create a disaggregated architecture, pairing Instinct-based Helios racks with Taalas accelerators to offload token generation. Vamsi Boppana, AMD’s SVP of AI, describes this as a 'full-stack AI platform' to deploy the 'right compute solutions for every AI workload.'
For this to work, customers like Meta, OpenAI, and Anthropic must commit to a hardware cycle that cannot pivot on a dime. While Taalas told The Next Platform that etching weights is 100 times less expensive than training a frontier model, the real risk is the opportunity cost of being locked into an inferior model.
AMD plans to scale this with the second-gen HC2 chip, aiming for 20 billion parameters. The Register notes that 50 of these could support a trillion-parameter model, offering better power efficiency than Nvidia’s LPX systems.
Still, I remain unconvinced. AMD is attempting to move the AI industry from a software-defined era back to a hardware-defined one. In a race where the finish line moves every few weeks, etching your bets into silicon isn't a strategy—it's a prayer.