AMD and Cerebras team up to take on Nvidia's Groq LPUs together

The enemy of my enemy is my friend. That’s the logic driving a new partnership between AMD and Cerebras, two companies that have been trying to chip away at Nvidia’s AI dominance from different angles. Now they’re joining forces with a specific target in mind: Groq’s LPUs, the inference-focused chips that have been turning heads in the hyperscaler market.

AMD brings the GPU muscle and the ROCm software stack — which, let’s be honest, has been playing catch-up to CUDA for years. Cerebras brings its wafer-scale engine, a chip the size of a dinner plate that’s been more of a science project than a revenue driver. Together, they’re pitching something that looks a lot like an integrated hardware-software play for inference workloads. The specifics are still fuzzy, but the direction is clear: they want to offer an alternative to Groq’s low-latency inference architecture without forcing customers to abandon the AMD ecosystem entirely.

Groq CEO Jonathan Ross has been vocal about his company’s performance advantages, claiming their LPUs can run large language models at speeds that make GPUs look sluggish. But Groq’s biggest weakness is availability — they’re still scaling up manufacturing while demand for inference compute is exploding. AMD and Cerebras are betting they can fill that gap with a combined offering that’s actually shippable at volume. Whether the performance numbers hold up is another question entirely.

The real story here is about consolidation among the Nvidia alternatives. For years, every startup with a novel chip architecture positioned itself as the David to Nvidia’s Goliath. That fragmentation made it easy for enterprises to just stick with CUDA. Pairings like this one suggest the also-rans are realizing they need to bundle their strengths to look like a real platform rather than a collection of point solutions. It’s a smart move, but the clock is ticking — Nvidia’s next-generation Rubin architecture is already on the horizon.