-
Microsoft inked a fresh AI deal with French startup Mistral
-
Hyperscalers across the board are looking to add more models to their rosters
-
Analysts stressed the importance of model choice but also of cloud platform tools to bring the models to life
Watch out, Library of Alexandria, hyperscalers are building the next wonder of the world: AI model libraries.
Case in point? Microsoft’s announcement this week that it will host Mistral Large, the flagship proprietary model from French AI startup Mistral. The news builds on Microsoft’s move last year to host Mistral’s open source models and — at least temporarily — sets it apart from competitors like Amazon Web Services (AWS) and Google Cloud, which just offer Mistral’s open source options.
So, why isn’t it just sticking with OpenAI? After all, Microsoft did plan to hire OpenAI’s CEO at one point, and the AI company’s technology is deeply embedded in Microsoft’s Copilot.
Well, it’s all about that card catalog, so to speak.
As AvidThink’s Roy Chua put it in an email to Silverlinings, “This announcement is primarily Microsoft expanding its model library, just like AWS and GCP have, to include different models — recognizing that a single model may not fit all customer needs (costs, task-specificity, need for offline or disconnected operation, privacy).”
That’s why you’re hearing so much about the proliferation of Meta’s Llama 2 model and Anthropic’s Claude, and reading about partnerships between hyperscalers and upstarts like Hugging Face and AlphaSense.
Jason Wong, distinguished VP analyst at Gartner, agreed, “It is important to offer choice and options for organizations because different industries have different needs.”
Wong noted that Mistral is also based in France, which means its models are natively trained on an array of European languages and thus could perform better in those languages than perhaps English-native models. Additionally, “certainly there’s going to be more regional regulations” around AI in the near future, so having AI assets that cater to different geographies could be a solid strategy, he said.
“It’s kind of like how we see cloud providers needing to have specific infrastructure in those regions to support those clients and regulations,” Wong explained.
However, the focus shouldn’t solely be on who has which models, he noted. The real differentiator will be on which hyperscaler has the right tools on their platform to put those models to the best use.
“If you just take at face value ‘here’s Llama 2,’ well Llama 2 is hosted by everybody because it’s open source. What’s the difference? …It’s the tools, it’s the technology that’s in the cloud provider’s arsenal that’s going to drive my choice,” Wong said.
Where are telcos?
Telcos have been duking it out with cloud providers since the latter stepped onto the scene. But by and large, they’ve stayed on the sidelines of the AI model game.
Wong said that’s likely because telecom is a totally different business than the cloud when it comes to concentrated compute power. And given we’re still in the early days of model training, compute is the name of the game.
Still, Wong said telcos could have a shot to jump into play as models shrink and inferencing comes to the forefront. That’s because inferencing is largely done at the edge — an area where telcos have some strength.
Whether or not they have the stomach for AI by then, we’ll have to wait and see.