Amazon has lagged OpenAI and Anthropic, but AI chief sees path to catch up in 'coming year'

Amazon said it expects to compete with OpenAI and Anthropic on frontier AI models in the coming year, though its current models are not yet at the very frontier for the largest workloads. The company highlighted Nova2, launched in December, with about 50,000 customers, and said its AI strategy remains split between Bedrock model access and in-house model development. Amazon also emphasized custom chip efforts under Trainium and Graviton, with no timeline set for selling those chips externally.

Analysis

The key read-through is not that Amazon is “catching up,” but that it is trying to compress two different gaps at once: frontier-model credibility and inference-cost leadership. That combination matters because the market will ultimately reward whichever platform can make model quality and unit economics reinforce each other; if Amazon can pair acceptable frontier performance with materially lower deployment costs via custom silicon, it can defend AWS workloads even without being the absolute benchmark model vendor. The more important second-order effect is that this is a margin strategy disguised as an AI product strategy.

For Nvidia, the threat is less near-term displacement than gradual erosion of bargaining power in cloud negotiations. If major hyperscalers keep proving they can train and serve useful models on in-house accelerators, Nvidia’s mix at the margin becomes more exposed to price competition, longer sales cycles, and customer demand for portability across chips. The overhang is likely measured in quarters to years, not days, but the first derivatives show up sooner in hyperscaler capex commentary and custom-silicon allocation shifts.

The contrarian point: the market may be underappreciating how little a “frontier” model matters for Amazon’s economics if Bedrock continues to aggregate third-party models while Nova becomes “good enough” for enterprise workloads. In that world, Amazon wins by monetizing distribution and infrastructure rather than chasing the best benchmark score. The risk to the bullish thesis is execution slippage—if Nova does not materially narrow the quality gap within 2-4 quarters, AWS could still participate in AI demand but surrender the highest-value workloads to other clouds and model vendors.

Key Decisions for Investors

• Add to AMZN on 3-6 month weakness: thesis is asymmetric if AWS custom silicon plus model distribution drives higher AI attach; target is multiple expansion from AI infrastructure monetization rather than model leadership.
• Initiate a medium-term NVDA underweight versus AMZN (or short NVDA / long AMZN pair) with a 6-12 month horizon: downside comes from hyperscaler silicon substitution and procurement leverage, while upside in AMZN comes from internal AI stack monetization.
• Sell NVDA upside against AMZN strength via a call spread pair trade: long AMZN Jan-2027 calls, short NVDA Jan-2027 calls, to express the view that cloud-custom silicon adoption is a slow-burn headwind for Nvidia margins.
• If NVDA rallies on near-term AI capex headlines, use that strength to trim and rotate into AMZN: the risk/reward is better where AI spend can be converted into recurring cloud usage and chip economics.
• For event-driven positioning, watch the next AWS capex / guidance cycle; if management increases Trainium-related allocation, add to the AMZN/NVDA pair as that would validate the substitution path over the next 2-3 quarters.