Amazon CEO Andy Jassy expects the frontier AI model race to produce several competitive models rather than a single winner, predicting during the company's Q3 2026 earnings call Thursday that within two to three years there will be "6- 8 models that are about as good as each other, or at least good enough for 95%+ of the workloads." Jassy told Goldman Sachs analyst Eric Sheridan that the challenge ahead will be selecting among models rather than committing to one provider, arguing that the AI landscape will mirror what happened with databases, analytics, and frameworks like TensorFlow, which lost ground to PyTorch after seeming poised to dominate.
The early numbers for Amazon's Bedrock service lend weight to Jassy's multi-model thesis, with Fortune 500 companies now spending close to seven times more on the platform year over year and token volume up nearly 17 times. Jassy called Bedrock the fastest-growing multibillion-dollar product in AWS history and suggested it could grow into a service "as large as EC2," still one of the company's biggest businesses. The optimism arrives as AWS reaccelerates, with the segment growing 39.3% in the quarter to $46 billion and reaching a $184 billion annualized run rate. Most Bedrock traffic already runs on Amazon's own Trainium chips, which lets the company hold prices down while protecting its margins.
Jassy said he expects "most customers are going to use multiple models, and many will use multiple models in the same generative AI application," pointing to different strengths across models in reasoning, coding, visual tasks, speed, and cost per token as reasons companies will adopt more than one. According to Jassy, customers want to "swap them back and forth as they leapfrog each other, as different constraints of their applications require them." The report notes that Bedrock reaches more than 100 foundation models from providers including Anthropic, OpenAI, Meta, Mistral, DeepSeek, Moonshot AI, and xAI through one API, and has recently added managed versions of open models such as Moonshot's Kimi and Zhipu's GLM.
The strategy beneath Jassy's prediction is that Amazon doesn't need to ship the best model as long as it owns the place developers go to choose one, a position it shares with Google's Vertex AI and Microsoft's Azure AI Foundry, both of which pitch similar multi-provider catalogs and unified access. A market where six or eight models keep trading the lead and teams constantly rotate between them is exactly what a neutral access layer monetizes, and Jassy called the opportunity "gigantic." The report explains that Bedrock's Converse API gives developers a common way to work with models from different providers without rewriting their integrations each time, though models have different context limits, tool capabilities, and supported features, so developers still need to test how their applications behave. According to AWS, Bedrock doesn't give model providers access to customer prompts or responses, and that information isn't used to train the underlying models, protections Jassy pointed to as another reason enterprises choose the service. The basic bet is simple: if no single model wins, the platform that makes switching easiest captures the value—and Amazon is racing Google and Microsoft to own that layer. Whether enterprises will tolerate the operational complexity of constant model rotation, or eventually consolidate around fewer providers for simplicity's sake, will determine if Jassy's prediction becomes a self-fulfilling prophecy or just expensive infrastructure hedging.