AI + ML
OpenAI and Anthropic have built AI Swiss Army Knives, but the future may be smaller built-for-purpose tools
To cater to the broadest possible market, OpenAI and Anthropic build ever-larger models capable of making a brute-force attempt to tackle almost any task.
These models are the Swiss Army Knives of the AI world. When used with sufficient force, they can do almost any job … but nobody needs a frontier class model to summarize emails, draft replies, or summarize meeting notes.
It’s cheaper and easier to train a small domain-specific model that can run dozens of instances of on a single accelerator. Plus, building your own means you don’t have to worry about your apps rambling about goblins when OpenAI replaces an aging but still beloved model with a new one — or Uncle Sam decides your model of choice is a bit too dangerous for general consumption.
Microsoft appears to be embracing the belief that bigger isn’t always better, and has quietly built small army of domain-specific models. Detailed at its Build developer conference in June, Microsoft’s MAI family covers a broad range of use cases from general purpose reasoning and coding to image generation, editing, and voice models.
According to a recent Bloomberg report, these models are now slowly but surely replacing OpenAI’s models as the power behind the AI features in Microsoft products.
When Microsoft was trying shoehorn generative AI into every facet of our digital lives, a Swiss Army Knife like GPT-5 was useful. But now the software giant knows how its customers are actually going to use AI, it can replace a frontier model with more precise tools that do the same job as quickly and cheaply as possible.
Cheap is the key word here, because while AI has proven itself useful in certain areas, bean-and-token-counters aren’t yet sure if it’s possible to sell AI at a profit. For hyperscalers like Microsoft, that means smaller models may be necessary.
Redmond describes MAI-Thinking-1 as a “medium-sized model that stands among the strongest models in its weight class” and says it “matches leading models on key software engineering benchmarks, demonstrates advanced mathematical reasoning capabilities, and is preferred to Sonnet 4.6 in our blind human side-by-side evaluations.”
What Microsoft considers to be a medium model we can’t say, but when it comes to gen AI, size matters. The bigger the model, the better and more reliable it tends to be, but the more expensive it is to run. That’s because when a model uses fewer parameters, it frees up memory and improves hardware utilization.
Smaller models also mean Microsoft can deploy the right AI for the right job at the right time. If Redmond sees a surge in speech-to-text traffic, the company can spin up more instances of the best model for that function while keeping costs controlled.
In addition, Microsoft now designs and builds its own AI accelerators, as do Amazon and Google. Its Maia 200-series parts announced in January promise to deliver performance comparable to Nvidia’s Blackwell parts. Custom chips mean operators can optimize the entire AI stack – software, hardware, and models – for greater efficiency.
Microsoft isn’t the only hyperscaler thinking about smaller models. Google has been playing this game from the beginning with its Gemini and Gemma families of models built around its custom TPU architecture.
However, the closest parallel to Microsoft is probably Amazon.
At the dawn of the AI boom, Microsoft hitched its cart to OpenAI’s horse, while Amazon placed its bet on rival Anthropic. And just like Microsoft, Amazon has been investing heavily in its own Nova family of models and applications and coding assistants powered by them.
General purpose frontier models still have their place, and someone still needs to drive innovation forward. Refining existing tools is much easier than inventing never-before-seen ones, which means OpenAI and Anthropic are still valuable to its hyperscale partners. And that’s why they’re willing to invest billions to keep them afloat.
The cloud titans still need the great model houses, but the less they rely on them the greater their chances of finally turning AI into a profitable business line. ®
