AMD agreed to acquire Toronto startup Taalas, which builds AI chips hardwired for a single model, in a bid to undercut Nvidia's dominance in AI inference. The deal won't move AMD's revenue this quarter, but it targets a growing bottleneck for cloud companies: electricity, not chip supply.
AMD announced on August 6 that it agreed to acquire Taalas, a Toronto startup building chips for AI inference, the work of running a trained model to answer a user's prompt. Financial terms were not disclosed.
Nvidia controls roughly 80% to 90% of the data center AI chip market, and its CUDA software keeps developers locked into its hardware. AMD's stock has more than doubled this year as its data center business expanded, but the Taalas deal signals AMD no longer sees general-purpose graphics chips as the only path to winning AI workloads. Nvidia made a similar bet, paying about $20 billion for assets from inference startup Groq roughly seven months earlier.
Taalas etches models directly into silicon
Most AI chips, including Nvidia's, are general-purpose and shuttle billions of model weights between the processor and memory. Taalas instead etches a single model's weights directly into the chip.
Its first chip, the HC1, runs Meta's Llama 3.1 model and nothing else. Taalas says it can generate more tokens per second than Nvidia's H200 and B200 chips while using one-tenth of the power. The tradeoff: switching models means building new silicon, though Taalas says only a couple of metal layers need to change, taking about two months from design to finished chip.
Power, not chip supply, is the constraint
Cloud companies including Meta and Microsoft are hitting hard limits on how much electricity their data centers can draw. AMD confirmed it will fold Taalas into its roadmap alongside its Instinct GPUs, EPYC processors, Helios racks, and ROCm software. The plan looks like a division of labor, with GPUs handling the demanding work of processing a prompt while Taalas chips take over the higher-volume job of generating the response.
Inference is on track to make up about two-thirds of all AI compute spending, according to Silicon Analysts estimates cited in the report.
Market reaction stayed muted
AMD shares slipped about 2% on August 7 to around $479, though the stock was still up about 4% over the prior five days. The deal follows AMD's record second-quarter revenue of $11.5 billion, up 50% from a year earlier, with data center sales more than doubling. Nvidia's near-term business remains secure because general-purpose GPUs are still required for training and fast-changing models, and revenue from Taalas is likely quarters away.
The strategy carries risk. New AI models still arrive almost monthly, and if frontier labs move to entirely new architectures, chips hardwired for older models could lose much of their value. Whether major cloud customers commit to AMD-Taalas silicon, and whether AMD sets a shipping timeline inside its Helios racks, will show whether the bet pays off.
Source: TheStreet
Trading involves risk.
Risk Disclaimer
