Amazon Web Services' Bedrock artificial intelligence inference platform could grow as large as its foundational Elastic Compute Cloud offering as businesses embrace multiple models and resist vendor lock-in, according to Amazon CEO Andy Jassy during the company's third fiscal quarter earnings call. Jassy argued that customers are seeking access to competing models without giving up security, controls, or the flexibility to change providers. The CEO predicted that over the next two to three years, six to eight AI models could prove capable of managing at least 95 percent of workloads, with different models excelling in reasoning, coding, visual tasks, latency, and cost.

AWS reported dramatic growth across its AI portfolio, with Bedrock token volume surging nearly 17-fold year over year while spending by Fortune 500 customers climbed nearly sevenfold. Jassy called Bedrock the fastest-growing multibillion-dollar service in AWS history. Tens of thousands of customers now use Amazon Bedrock AgentCore, deploying twice as many agents during the third quarter as in the prior three quarters combined. Lambda, the company's flagship serverless offering, crossed a $1 billion annualized revenue run rate. AWS sales jumped 39 percent year over year to $46 billion with a $184 billion annualized revenue run rate, marking the division's sixth consecutive acceleration quarter and its fastest growth in nearly five years. Backlog grew triple digits year over year to $578 billion. Revenue commitments for Amazon's Graviton CPU and Trainium AI accelerators each increased more than 20 times year over year, with customers using Amazon's custom silicon saving more than $2 billion in the third quarter.

According to Jassy, "twenty-plus years in technology has taught" him "that there is not one tool to rule the world." He noted that Google's TensorFlow was initially expected to dominate as the AI framework, but now PyTorch is much bigger. "People always assume that whatever is in the lead early on is going to be the only thing or things that people use, and it's never the case," Jassy said. "Most customers are going to use multiple models, and many will use multiple models in the same generative AI application." The CEO stated that inference accounts for "well over" half of AWS' AI revenue even though enterprises remain very early in using inference at scale. Jassy described AI operating margins as "rapidly improving" and "well ahead of where they were on core at a similar time of evolution," despite the high cost of memory affecting the entire IT channel and depreciation from Amazon's capital expenditures.

Jassy explained that AWS benefits from a flywheel effect in which customers place applications and data in its cloud, run AI inference near that data, and then consume additional compute, storage, and networking for post-training, reinforcement learning, and agent tools. The CEO appeared to criticize neoclouds, saying their use is "about 3,500 basis points worse in operating margin" for cloud providers compared with long-leasing data centers, and noted that most of AWS' capacity is either owned or long-leased. Amazon's custom silicon strategy aims to bring down the cost of AI inference to encourage more AI usage, with the CEO stating the cost has dropped significantly over the last couple of years and "will come down meaningfully more." AWS still expects to spend about $220 billion in cash capital expenditures in 2026 and "meaningfully more in 2027, with a clear line of sight to attractive financial returns," with the company installing an AI accelerator about every 14 seconds. Jassy said AWS projects revenue growth rates of more than 40 percent in both 2027 and 2028, reaching $1 trillion in revenue in the next five to 10 years, calling it "the very early days of this inflection." The combination of massive infrastructure investment and differentiated chip economics positions AWS to maintain leadership while the enterprise cloud migration cycle still has 10 to 20 years left to run, given that roughly 85 percent of global IT spending remains on-premises. The real test will be whether multi-model flexibility proves more durable than raw compute scale as the basis for customer loyalty.