-
Microsoft signs new AI deal with French startup Mistral
-
Hyperscalers across the board are looking to add more models to their roster
-
Analysts emphasized the importance not only of model selection but also of cloud platform tools to realize the model.
Be aware that Library of Alexandria, hyperscalers are building the next wonder of the world, an AI model library.
Case in point? Microsoft announced this week that it will host French AI startup Mistral’s flagship proprietary model, Mistral Large. This news builds on Microsoft’s move last year to host an open source model for Mistral, unlike competitors such as Amazon Web Services (AWS) and Google Cloud, which only offer open source options for Mistral. distinct, at least temporarily.
So why not stick with OpenAI? After all, Microsoft plans to hire OpenAI’s CEO at some point, and the AI company’s technology is deeply embedded in Microsoft’s Copilot.
Well, that card catalog is everything, so to speak.
As AvidThink’s Roy Chua said in an email to Silverlinings, “This announcement is primarily about Microsoft expanding its model library to include a variety of models, similar to AWS and GCP. We recognize that our model does not fit all customer needs (cost, tasks) – specificity, need for offline or disconnected operations, privacy).
That’s why you often hear about the proliferation of Meta’s Llama 2 model or Anthropic’s Claude, and why you read about hyperscalers’ partnerships with startups like Hugging Face and AlphaSense.
Jason Wong, Distinguished Vice President at Gartner, agrees: “Each industry has different needs, so it’s important to provide organizations with choice and options.”
Wong points out that Mistral is also based in France, and that its models are natively trained in various European languages and therefore could possibly perform better in those languages than native English models. He pointed out that there is. Additionally, he said, “we’re definitely going to see more regional regulations” regarding AI in the near future, so owning AI assets across a variety of regions could be a solid strategy. Stated.
“This is similar to how we think cloud providers need specific infrastructure in their region to support their clients and regulations,” Wong explained.
However, he pointed out that the focus should not just be on who has which model. The real differentiator will come down to which hyperscalers have the right tools on their platform to make the most of their models.
“If you take ‘this is Llama 2’ at face value, Llama 2 is open source, so everyone is hosting it. What’s the difference? …What determines my choice is… It’s a tool and a technology in the cloud provider’s arsenal,” Wong said.
Where is my phone company?
Telecommunications companies have been in a battle with cloud providers ever since they emerged. But by and large, they have kept AI models on the sidelines.
Wong said this could be because telecommunications is a completely different business than cloud in terms of intensive computing power. And given that we’re in the early stages of model training, compute is paramount.
Still, Wong said telecom companies could also get in on the action as models shrink and inference comes to the forefront. That’s because much of the inference happens at the edge, an area where carriers have some strength.
We’ll have to wait and see whether they’re willing to accept AI by then.
