AMD takes a shot at Nvidia by betting on AI's next big shift

AMD is partnering with Cerebras on disaggregated AI inference, moving Helios into Cerebras data centers later this year. At Advancing AI, AMD unveiled Helios and claims it delivers up to 30% more inference tokens per dollar versus Nvidia’s Vera Rubin NVL72 rack, as customers shift from training to deploying models. The move aligns with UBS’s view that architectural limits are pushing the industry toward disaggregated inference, though orchestration across chips remains a key challenge.

Analysis

The market should not treat this as a simple GPU-share headline; the more important mechanism is that inference is becoming a systems business, where the winner is whoever can optimize cost per token across compute, networking, memory, and orchestration. That tends to expand the wallet for AMD, AVGO, and hyperscalers that can arbitrage lower inference cost into higher utilization, but it also raises the bar for standalone chip claims because the bottleneck may move from silicon performance to software coordination.

AMD’s upside is real if Helios gets embedded in production clusters, but the trade is less about one product launch and more about whether the company can convert architecture talk into repeatable design wins over the next 1-3 quarters. The near-term loser is not necessarily NVDA revenue, but NVDA’s multiple if investors start to believe inference becomes more modular and price-competitive; still, NVDA is better insulated than a pure loser call because it can participate in the orchestration layer and will likely respond aggressively on pricing.

The contrarian point is that the headline efficiency claim may be overstated until workloads are live at scale: heterogeneous stacks often fail on integration, not FLOPS. If the software layer is the real moat, the biggest beneficiaries could be AMZN and MSFT, which can pass lower inference costs through to customers while keeping the margin spread, with META and ORCL seeing faster product iteration rather than direct chip economics. Watch for design-win disclosures and any AMD guidance improvement; absent that, this is a narrative trade, not an earnings re-rate.

Key Decisions for Investors

• Long AMD / short NVDA as a 1-3 month relative-value trade, but only on confirmation of additional production design wins; thesis breaks if NVDA maintains pricing discipline and AMD fails to raise AI revenue guidance in the next two quarters.