Jassy predicts 6-8 evenly matched AI models: SMBs will have to choose

During Amazon's third-quarter earnings call, its CEO, Andy Jassy, left an idea worth noting even if your company has nothing to do with AWS: the future of AI does not point to a single winning model, but to a handful of them competing neck and neck. Specifically, Jassy anticipated that in two or three years there will be between six and eight models "roughly as good as each other, or at least good enough for more than 95% of workloads."

The prediction arose from a question by Goldman Sachs analyst Eric Sheridan about the potential of Amazon Bedrock, the service that provides access to more than a hundred foundation models from providers such as Anthropic, OpenAI, Meta, Mistral, DeepSeek, Moonshot AI or xAI through a single API. Jassy's answer is not just a defense of his product: it is a thesis on how the AI market will be organized in the coming years.

The argument: no model does everything well

Jassy drew on the recent history of technology to explain why he does not believe in a single dominant player. He recalled that over the past two decades he has seen many technology leaders lose their initial advantage, and he cited databases, analytics tools or the case of TensorFlow, which seemed destined to become the standard framework before PyTorch overtook it. With AI models he expects something similar: "there is no one tool that rules the world."

The reasoning is simple and very practical. Some models will be better at reasoning, others will stand out in programming, others in visual tasks, and some will win on speed or cost per token. That specialization, according to Jassy, will push most customers to use several models, and many will do so even within the same generative AI application. For a development team, that translates into choosing the model according to the specific task instead of committing to a single provider for everything.

Switching models is not free

The idea of a market with several models trading the lead fits perfectly with the Bedrock proposition, which offers a common access layer. Its Converse API allows working with models from different providers without rewriting integrations each time. But do not be fooled: switching models is not as simple as pressing a button.

Each model has different context limits, different tool capabilities and features that do not always coincide. If a team has built its workflow around a specific provider's agent system, migrating can become quite complicated. Jassy acknowledged that customers want to "swap them back and forth as they leapfrog each other," but that requires testing and validation by the developer. Add to that security: according to AWS, Bedrock does not give model providers access to customers' prompts or responses, and that information is not used to train the underlying models.

What this means for an SMB

Beyond the battle between hyperscalers —Google with Vertex AI and Microsoft with Azure AI Foundry offer very similar multi-provider catalogs—, Jassy's prediction has very concrete consequences for a small or medium-sized company that is starting with AI:

• Don't marry a single provider. If your architecture allows you to change models without redoing everything, you will gain room to maneuver when a cheaper or more capable option appears for your use case.
• Choose by task, not by brand. A model can be excellent at writing emails and mediocre at extracting data from invoices. Measure before deciding.
• Watch the cost per token. In recurring workloads, the difference between models can matter more than a few tenths of quality.
• Test before migrating. Context limits and tool capabilities change between models; what works in one may fail in another.

At ForgeNEX we see it every week in SMB projects: the temptation to standardize everything on a single model is strong because it simplifies, but in the medium term it makes adaptation more expensive. As we already discussed when talking about what to measure when the GPU is at 100%, what matters is not raw power but whether the system solves the problem at a reasonable cost.

The business behind the forecast

The prediction should be read with context: Jassy is not a neutral observer, he is the CEO of the company that wants to be the access point to all those models. His thesis is that Amazon does not need to have the best model as long as it is the place developers go to choose. And the numbers he handles suggest that this strategy works: according to his data, Fortune 500 companies now spend almost seven times more on Bedrock than a year ago, token volume has multiplied by almost 17 and the service is, in his words, the fastest-growing multibillion-dollar product in AWS history. All this while AWS grows 39.3% in the quarter to 46 billion dollars, with an annualized rate of 184 billion.

Jassy also noted that most of Bedrock's traffic already runs on Amazon's Trainium chips, which allows it to contain prices and protect margins. That is, the business is not only in the model, but in the infrastructure and the access layer.

Our recommendation

For an SMB, the practical reading is this: treat AI models as interchangeable parts of an assembly line, not as a single bet. Design your integrations with an abstraction layer that allows you to change providers without rewriting the flow, document which model you use for each task and review that choice from time to time. And do not forget security: check what each provider does with your data before sending it sensitive information. If you need help setting up that abstraction layer or evaluating which model fits best in each process, at ForgeNEX we can give you a hand.

Source: The New Stack. Analysis and adaptation: ForgeNEX.