Satya Nadella: DeepSeek is the new bar for Microsoft’s AI success

Microsoft CEO Satya Nadella moved quickly to get DeepSeek’s R1 deployed on Azure in January. Nadella appeared to have anticipated the rise of a breakthrough like R1, and now he says a new bar has been set for the company’s own AI work.

Speaking at an employee-only town hall last month, Nadella and his senior leadership team discussed DeepSeek and the company’s own $80 billion investment into AI at length, answering questions from employees about the giant spend and its effect on Microsoft’s carbon-free goals, sources at the company tell me.

DeepSeek made headlines earlier this year after doing a bunch of low-level systems work to optimize below Nvidia’s CUDA layer with architecture changes that allowed its AI models to be more compute efficient. The efficiency, use of pretrained models, and small team required to pull it off all caught Nadella’s attention.

“What’s most impressive about DeepSeek is that it’s a great reminder of what 200 people can do when they come together with one thought and one play,” Nadella said in response to a question about how Microsoft will keep up in the ongoing AI battle. “Most importantly, not just leaving it there as a research project or an open source project, but to turn it into a product that was number one in the App Store. That’s the new bar to me.”

Microsoft hasn’t seen that kind of App Store success with Copilot yet. While ChatGPT is regularly in the top free app spot, Copilot is often outside even the top 100 apps. That result comes despite Microsoft having access to OpenAI’s latest models, spending big on a Super Bowl ad, and giving Copilot a design refresh that included voice and vision features. Nadella is now looking to Microsoft’s own AI work, and not just OpenAI’s, to improve its position.

Microsoft unveiled its own Muse model last month, which has been trained on the Xbox game Bleeding Edge to generate gameplay. Microsoft thinks that Muse will help Xbox developers prototype games or even use it to preserve games and optimize them for modern hardware.

“When I think about what we’re trying to do with Muse, that’s the bar,” said Nadella. “You go do fundamental research, you bring out a model, and then translate it into a real breakthrough feature in Copilot.

Some short interactive AI experiences will be part of Microsoft’s Copilot Labs soon, thanks to Muse, and it will be interesting to see if this work makes Copilot more appealing to the masses. Most of whom are currently playing around with ChatGPT’s improved image generation.

Microsoft’s AI challenge — beyond having a popular app — will be moving fast, just like it did with onboarding DeepSeek, to take advantage of new models or develop key advances itself. “Given what [DeepSeek have] done with 200 people… it does make a point for us in terms of how we work together internally where we have all of these organizational boundaries,” said Jay Parikh, head of Microsoft’s new CoreAI engineering group, during the same town hall meeting.

Parikh only recently joined Microsoft, and he’s already been tasked with overhauling Microsoft’s own engineering efforts in a similar way to how he was instrumental to Meta’s engineering work for more than a decade. “[DeepSeek] is just a reminder to me that we have to continue to push to go faster, because the innovation that’s happening in these smaller teams around the world is so impressive from a speed perspective.”

Beyond AI model improvements, Microsoft is also investing $80 billion into datacenters to power AI workloads this fiscal year. One employee wanted to know why Microsoft thinks this is a good investment, and how the company is ensuring it’s done in a sustainable way. Nadella believes it’s a key investment, especially as Microsoft reshapes its engineering groups to focus on an AI-first app stack.

”At some level, we want to position ourselves where every workload going forward will look like ChatGPT,” said Nadella. “If you look at ChatGPT it’s not just using an AI accelerator, it’s got a state in Cosmos DB, Azure search, and lots of other places. So there’s a ratio between the AI accelerator, storage, and compute, so that’s the other thing we have balanced. That’s where the spend is going, and even independent of AI we have cloud growth.”

That cloud growth requires a big chunk of its investment, in an effort to get customers ready to adopt Microsoft’s AI features. “We have nearly $300 billion of contracted revenue that we have to deliver on over the next year, two, or three, and so that amount of demand is what supports these numbers,” said Amy Hood, chief financial officer at Microsoft.

There are signs Microsoft might be holding back on some data center investments though, with analysts warning of a potential bubble in data center buildouts, where there might be an oversupply relative to demand for AI. That’s a big risk for Microsoft.

”What we’re focused on is making sure that $80 billion counts for our customers, and if they get real value I assure you we’ll feel great about spending $80 billion,” said Hood. “If they don’t get real value, or if others deliver ahead of us on innovation, then we also won’t feel great.”

Surprise the world in 2030

Five years ago Microsoft announced its goal to be carbon negative as a company by 2030, by reducing carbon emissions and buying more carbon removal than Microsoft emits every year. The fresh push into AI is hugely impacting Microsoft’s green energy efforts, and the company’s leadership team was ready to address employee concerns about this.

”It was all on track and then came this thing called generative AI,” said Brad Smith, Microsoft’s vice chair and president, at the town hall meeting. “In effect what we then saw was that the goal that we had was basically four times, or four and a half times as big.”

Instead of shying away from the goal, Smith says Microsoft is doubling down. “It’s as if the thing that everybody called our moonshot now had a moon that had moved. So what do you do when the moon moves? Journalists keep asking that all of the time, and our answer has been consistent: you build a more powerful rocket ship.”

Microsoft has lined up 34 gigawatts of carbon-free energy in 24 countries around the world in an effort to still meet its carbon negative goal, alongside efficiency improvements in energy and water use. But there’s still more to be done to hit the goal, including Microsoft’s use of steel, concrete, chips, and sustainable fuels.

Smith even thinks AI itself will help solve the problem somehow. “You use the power of AI to crack the code to make the further advances we’re going to need to make.” That’s a lot to ask of AI models over the next five years, especially as Microsoft’s original goal for 2023 already seemed big.

To many onlookers the carbon-free goal now seems unrealistic given the company’s AI push. Smith thinks Microsoft will prove the doubters wrong, though. “I still have enormous optimism that we’re going to surprise the world in 2030 and show them just what an extraordinary thing we’ve accomplished.”

I’m always keen to hear from readers, so please drop a comment here, or you can always reach me at notepad@theverge.com if you want to discuss anything else. If you’ve heard about any of Microsoft’s secret projects, you can reach me via email at notepad@theverge.com or speak to me confidentially on the Signal messaging app, where I’m tomwarren.01. I’m also tomwarren on Telegram, if you’d prefer to chat there.

Thanks for subscribing to Notepad.