Nvidia Needs to Watch Out for This Custom Chip Designer (Hint: It's Not Broadcom)

Amazon says its upcoming Trainium chip capacity is already sold out months in advance, and management claims the current generation offers about a 30% cost-performance improvement versus GPU-based training. The article argues AWS is the bigger profit engine, contributing 50% of operating profits in Q4 and 66% in Q3, while Amazon plans $200 billion of capex this year, mostly for AWS infrastructure. Overall tone is constructive on Amazon shares, though it also highlights a competitive threat to Nvidia rather than a near-term disruptive event.

Analysis

AMZN is starting to look less like a retail compounder and more like a vertical integration story: if AWS can materially improve training economics with proprietary silicon, the profit pool shifts from selling compute to controlling the stack. The second-order effect is that every customer choosing Trainium reduces NVIDIA attach rate, but it also pressures OEMs, networking vendors, and software layers that have been priced off a “GPU-everywhere” future. Near term, the market will likely treat this as a share-gain narrative for AMZN and a margin-defense narrative for NVDA, but the bigger implication is that custom silicon adoption can broaden faster once a few large workloads clear the performance risk hurdle.

The key catalyst is not chip substitution in the abstract; it is capacity conversion. Sold-out training capacity suggests demand visibility already exists, which means the next 2-4 quarters are about monetizing capex rather than proving a concept. That matters because AWS profit leverage can inflect faster than consolidated revenue, so AMZN can re-rate even if e-commerce remains sluggish. The main risk is execution: if the next chip generation slips, software compatibility lags, or customers refuse to replatform beyond pilot workloads, the “NVDA displacement” thesis becomes a long-dated optionality story rather than a near-term earnings driver.

Consensus is probably underestimating how this impacts NVDA at the margin but overestimating how quickly it becomes existential. Most hyperscalers will stay multi-sourced for years, which caps outright displacement; however, even a 5-10% share shift in internal training spend can compress NVIDIA’s pricing power and slow incremental upside. On the other side, AMZN may be undervalued relative to the operating leverage embedded in AWS capex, especially if investors still anchor on retail-multiple logic instead of cloud earnings power. The cleanest read-through is that this is a structural AWS winner with a contained but real competitive threat to NVDA, not a binary winner-take-all transition.

Key Decisions for Investors

• Long AMZN vs short NVDA into the next 1-2 earnings cycles: own the beneficiary of custom-silicon adoption and capex monetization while fading some of the premium for GPU scarcity; target a 6-12 month horizon, with the trade working if AWS revenue re-accelerates faster than NVDA can offset incremental share pressure.
• Buy AMZN call spreads 3-6 months out to express upside from AWS operating leverage without paying for full-multiple expansion; best if entered on any post-news consolidation, with risk capped to premium.
• Trim or hedge NVDA on strength rather than chasing downside: use downside puts or a partial hedge against long semis exposure, since the displacement thesis is gradual but can compress forward multiples before earnings actually deteriorate.
• Monitor AVGO as a secondary beneficiary and relative winner versus NVDA on custom-accelerator adoption; pair long AVGO / short NVDA for a cleaner expression of 'purpose-built silicon gains share' over the next 2-4 quarters.
• Avoid interpreting this as an immediate INTC-style collapse for NVDA; if the market overreacts to the narrative, use weakness in NVDA to rebuild only if data shows Trainium adoption stalls or AWS capital efficiency disappoints.