AMD and Cerebras join forces against Nvidia’s Groq LPUs

AMD teamed with Cerebras Systems to build a disaggregated AI inference platform pairing AMD Instinct GPUs with Cerebras wafer-scale SRAM accelerators, targeting ultra-low-latency agentic workloads. The setup is expected to improve efficiency—tokens per second per watt—by up to 5x, while leveraging SRAM (instead of HBM4) to exceed 2,000 tokens/sec in Cerebras’ inference performance. The offering will be available via Cerebras Cloud later this year, and AMD signaled more workload-disaggregation partnerships.

Analysis

This is more important as a portfolio-completeness signal for AMD than as a near-term revenue event. Inference buyers are buying latency and tokens per watt, so a credible heterogeneous stack can improve AMD’s win rate in agentic workloads where raw GPU FLOPS are no longer the bottleneck. The second-order effect is that AMD may start showing up in procurement conversations it previously lost before a workload even got benchmarked, which can expand its addressable share without needing a wholesale architecture win.

For Nvidia, the threat is narrative before it is economics. The market has been paying up for the assumption that one vendor can own both training and inference; disaggregated designs weaken that thesis and could compress the multiple if customers begin segmenting spend by workload rather than platform. The more durable spillover is to the memory stack: if inference traffic migrates toward SRAM-heavy accelerators, incremental HBM attach becomes less linear, though this is a months-to-years issue because training still dominates capacity demand.

The key catalyst is whether this becomes a real cloud product with repeatable benchmarks and named customers over the next 1-3 months. Without that, it is mostly ecosystem theater and the earnings impact for AMD this year remains modest. Contrarian view: the market may be overestimating the immediate share shift in NVDA while underestimating how valuable this is for AMD’s strategic positioning in enterprise inference bake-offs over the next 6-18 months.

Key Decisions for Investors

• Initiate a small AMD/NVDA relative-value long AMD / short NVDA pair over the next 1-3 months; use weakness to enter and target 3-5% spread capture, with a stop if NVDA commentary or backlog data re-accelerates inference demand more broadly.
• Do not chase this as an outright long NVDA short until there is evidence of lost platform share; the current setup is a narrative headwind, not a fundamental break, and shorting Nvidia here has poor asymmetry.
• Buy AMD 3-month call spreads on a pullback only if the market gives back the keynote premium; the trade is a cheap way to own ecosystem optionality without paying for a full rerating.
• Keep MU and the HBM complex on watch rather than shorting immediately; only turn bearish if public benchmarks and customer disclosures show inference mix shifting materially toward SRAM-heavy systems over the next 2-3 quarters.