Anthropic is discussing a new custom chip with Samsung

Anthropic is reportedly in talks with Samsung to explore a collaboration on pending AI chips, signaling it may move beyond “toying” with custom hardware as chip shortages persist. The company indicated it has not yet decided the chip’s purpose or server fit, while reiterating that a diversified compute stack using chips from Google, Amazon, and Nvidia remains central. The move also parallels OpenAI’s custom inference effort with Broadcom (“Jalapeño”), potentially intensifying competition around efficiency/performance-per-watt in AI accelerators.

Analysis

This reads more like procurement leverage than a near-term displacement of GPUs. The first-order winner is still the platform layer: AMZN and GOOGL can use custom silicon to improve unit economics and defend cloud margins, while AVGO benefits from the secular normalization of application-specific chips as a service line. NVDA is not facing an immediate demand shock; the software stack, qualification cycle, and training workload complexity keep GPUs entrenched for at least the next 2-4 quarters.

The second-order effect is on inference pricing, not training demand. If frontier-model customers keep pushing custom ASICs, cloud providers can reprice tokens downward and still preserve margin, which should increase usage and make the largest clouds stickier relative to standalone model vendors. The real medium-term risk for NVDA is mix: even a modest 10-15% shift of inference spend into custom silicon would cap upside in gross margin expansion and reduce pricing power, but that is a 6-18 month story, not a single-event trade.

Contrarian view: the market may overstate how quickly custom chips can erode NVIDIA’s moat, while underappreciating the valuation support for the hyperscalers that own both demand and infrastructure. The key falsifier is continued commentary from AMZN/GOOGL that GPU demand remains unconstrained and that custom silicon is incremental rather than substitutive; if that persists through the next earnings cycle, the bearish NVDA angle fades quickly. Conversely, if more model labs announce third-party ASIC partnerships within 1-2 months, the stock-to-stock rotation into vertically integrated platforms should accelerate.

Key Decisions for Investors

• Prefer long GOOGL and AMZN on any market-wide AI pullback over the next 1-3 months; the trade is that custom silicon improves cloud ROI before it meaningfully hurts their spend, with downside limited by existing TPU/Trainium optionality.
• Pair trade: long AVGO / short NVDA for a 6-12 month horizon if you want to express the custom-chip secular shift; AVGO benefits from design-win proliferation, while NVDA faces the most risk from inference mix erosion. Exit if NVDA commentary shows no slowdown in inference demand or if AVGO order conversion fails to show up in backlog.
• Use NVDA only as a tactical hold, not a fresh add, until the next earnings guide; the upside is still intact, but the risk/reward is less compelling if customers keep signaling silicon diversification. Falsifier: another quarter of strong GPU utilization and no evidence of pricing pressure.