AI NewsModels & agentsReported
Jassy predicts six to eight AI models will do most of the work
Andy Jassy told Q3 analysts to expect six to eight AI models good enough for 95 percent of workloads. He said teams will use several models in one app.

Image: The New Stack
Why it mattersThe person who commits a company to a single model provider now bets against the Amazon CEO. Keeping a model-swapping layer looks safer than it did a quarter ago.
Six to eight AI models will cover 95 percent of workloads within two to three years. That is Andy Jassy's prediction, delivered on Amazon's Q3 2026 earnings call. The CEO answered a Goldman Sachs question about Amazon Bedrock's market opportunity. The New Stack reported on the exchange on 31 October.
What Jassy said
Jassy told Goldman Sachs analyst Eric Sheridan that the model race will not produce one winner. The analyst asked about the market for Bedrock. Bedrock is Amazon's platform for calling many AI models through one API. Jassy said there will be six to eight models about as good as each other. They would be good enough for 95 percent of the workloads.
He said there is not one tool to rule the world. He pointed to databases, analytics and TensorFlow as past examples. In each case, an early leader lost ground to several competitors.
Jassy said most customers are going to use multiple models. Many will use multiple models in the same generative AI application. Teams will pick a model for each task, rather than one provider for every task.
Why Amazon talks it up
Jassy's prediction is also a description of how Bedrock sells. Bedrock reaches more than 100 foundation models through one API, by Amazon's count. The list includes models from Anthropic, OpenAI, Meta, Mistral, DeepSeek, Moonshot AI and xAI. Amazon brought OpenAI's models to Bedrock recently, Jassy said. It also added managed versions of open models from Moonshot and Zhipu.
Jassy told analysts customers want to swap models back and forth as they leapfrog each other. The New Stack puts the strategy this way. Amazon does not need to release the best model. It only needs to own the place developers go to pick one.
The numbers Amazon put on the record
Fortune 500 companies now spend close to seven times more on Bedrock year over year, Jassy said. Token volume on the service is up nearly 17 times. He called Bedrock the fastest-growing multi-billion-dollar product in AWS history. He compared the opportunity to EC2, which is still one of AWS's biggest businesses.
AWS as a whole grew 39.3 percent in the quarter to 46 billion dollars. The segment reached a 184 billion dollar annualised run rate. Most Bedrock traffic runs on Amazon's own Trainium chips, Jassy said. The company says that helps hold prices down.
The caveats Jassy stated
Switching models through Bedrock's Converse API is not free, Jassy acknowledged. Models have different context limits, tool capabilities and supported features. Teams still need to test how applications behave on each one. The cost rises when a team built its workflow around one vendor's agent harness.
Jassy said Bedrock does not give model providers access to customer prompts or responses. The data is not used for training, he said. That is Amazon's statement, repeated to analysts.
What this changes for a buyer
The person who signs a one-provider deal in Q4 now bets against Amazon's CEO. Jassy did not say the leaders will keep leading. He said the lead will keep rotating. For a buyer, that reads as a reason to keep a model-swapping layer. The one-week cost of adding a second provider is small. It is small compared to being locked in to the first provider when the price or the model changes.
Source
The New Stack, on Andy Jassy's remarks on Amazon's Q3 2026 earnings call.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


