The Chinese startup DeepSeek shook up the world of AI last week after showing its supercheap R1 model could compete directly with OpenAI’s o1. While it wiped nearly $600 billion off Nvidia’s market value, Microsoft engineers were quietly working at pace to embrace the partially open- source R1 model and get it ready for Azure customers. It was a decision that came from the very top of Microsoft.
Sources familiar with Microsoft’s DeepSeek R1 deployment tell me that the company’s senior leadership team and CEO Satya Nadella moved with haste to get engineers to test and deploy R1 on Azure AI Foundry and GitHub over the past 10 days. For a corporation the size of Microsoft, it was an unusually quick turnaround, but there are plenty of signs that Nadella was ready and waiting for this exact moment.
While the open-source model has upended Wall Street’s idea of how much AI costs, Nadella seemed to know that something like DeepSeek was coming eventually. Appearing on the BG2 podcast in early December, he warned of the exact thing DeepSeek went on to achieve weeks later: an algorithmic breakthrough that results in compute efficiency.
How this breakthrough has been achieved is still up for debate. Microsoft and OpenAI have reportedly been investigating whether the Chinese rival used OpenAI’s API to train DeepSeek’s models using a technique called distillation. That, too, was a threat Nadella warned about.
“It’s impossible to control distillation,” Nadella said on December 12th. “You don’t even have to do anything. You just … reverse engineer that capability, and you do it in a more compute efficient way.” He even joked that this approach was “kind of like piracy.”
On Christmas Day, DeepSeek released its V3 reasoning model, the foundation for the R1 release early last week. DeepSeek’s progress might have looked like it came out of nowhere to Wall Street, but anyone following AI closely, like Nadella, will have witnessed the progress the Chinese AI lab has made with its consistent releases throughout 2024.
DeepSeek claims its final training run cost $5.6 million, and AI labs in the US are currently replicating the R1 recipe to see if DeepSeek’s numbers are accurate. It looks like Microsoft is happy with the quality of the model either way, as it’s not just Azure AI Foundry and GitHub where the software maker is looking to deploy R1.
Distilled R1 models can now run locally on Copilot Plus PCs, starting with Qualcomm Snapdragon X first and Intel chips later. This brings a lot more AI capabilities to Windows, and it’s something Microsoft was already working on with its Phi Silica language models.
Sources tell me Microsoft is also looking at the prospect of bringing R1 to some of its Copilot tools for businesses. Microsoft is currently anticipating that more businesses will use its AI tools in the coming months, particularly its AI agent capabilities. Models like R1 could help Microsoft sell more access to Copilot inside business apps, low-code platforms, and other industry-specific tools at a lower cost to businesses.
R1 could present the turning point for bringing AI costs down: a focus on software and model improvements rather than constantly chasing costly hardware compute power. “I think there will be a governor on how much people will chase,” Nadella said in that same December podcast. “Right now it’s a little bit of everybody wants to be first. It’s great, but at some point, all the economic reality will set in on everyone.”
Nadella hinted at this new turning point in an early morning post on X on Monday, just hours before the US stock market opened and reacted to the buzz around DeepSeek. Nadella referenced Jevons paradox, an observation from 1865 by English economist William Stanley about how technological improvements in producing coal led to increased consumption. “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of,” Nadella wrote.
Microsoft isn’t trying to produce all of the coal itself, but it certainly wants to sell the shovels needed by AI application developers. Consumers and businesses have already shown that they’re not willing to pay extra for AI yet, so Microsoft is increasingly trying to find ways — like R1 — to drive the costs down and consumption up. We’re still waiting on Microsoft’s R1 pricing, but DeepSeek is already hosting its model and charging just $2.19 for 1 million output tokens, compared to $60 with OpenAI’s o1.
“Why would I want to spend a lot on some model capability when the network effects are all on the app layer?” Nadella asked in the December podcast. That sensitivity to spending more and more on model capability when new, more efficient models are coming might just explain why Microsoft was willing to renegotiate its OpenAI partnership.
Prior to the renegotiation announced last week, OpenAI exclusively used Microsoft’s Azure cloud infrastructure for its needs. Now, it can shop around as long as it gives Microsoft the right of first refusal to provide additional capacity. This should ease some of the apparent tension between Microsoft and OpenAI, with a report in October suggesting OpenAI leaders had become frustrated with Microsoft not supplying servers fast enough.
While OpenAI is now teaming up with SoftBank on Stargate, a potential $500 billion AI data center project, Microsoft still has a complex revenue-sharing agreement in place, and OpenAI’s APIs remain exclusive to Azure. That gives Microsoft the flexibility to experiment with rival models that can push costs down, while also getting access to OpenAI’s latest and greatest. We’ll see that play out again very soon, as I’m told Microsoft is also working on its own version of OpenAI’s new Operator AI agent that can perform tasks for you on the web.
In a world where the next big AI advancement could clearly come from anywhere, Microsoft is very much having its cake and eating it, too.
New Intel-powered Surface devices arrive next month
Microsoft announced today that it’s launching new Intel-powered Surface devices next month. The Surface Pro 11 and Surface Laptop 7 will soon both come in options for Intel’s Lunar Lake chips as well as the existing Qualcomm versions. These new Intel-powered variants will be available primarily to businesses, and they have NPUs powerful enough to make them Copilot Plus PCs.
Microsoft is also launching a new USB 4 docking station, which works with existing USB-C Surface devices. In another business-focused move, the Surface Hub 3 is also getting web apps, access to personal content, Edge browser support, and wireless Miracast projection in the coming months.
- Microsoft’s AI business is booming — Xbox, not so much. While Microsoft was keen to show Azure and AI growth in its latest earnings, the company’s gaming revenues didn’t fare so well. Gaming revenue declined 7 percent overall this quarter, with Xbox hardware down 29 percent. This was the all-important holiday quarter, with the big releases of Indiana Jones and Call of Duty: Black Ops 6 failing to move the needle on Xbox hardware. Microsoft was also banking on Call of Duty on Game Pass improving its subscription numbers, but Xbox content and services, which includes Game Pass, was only up 2 percent.
- Microsoft is bringing your iPhone to the Windows 11 Start menu. Windows Insiders can now try out quick access to an iPhone’s messages, calls, and more straight from the Start menu. Microsoft started rolling this out to Android phones last year, and it’s now expanding it to iOS in a new test with Beta and Dev Channel Insiders this week. You’ll be able to access calls and messages in a neat little sidebar that floats next to the Start menu that’s similar to Microsoft’s dedicated Phone Link app.
- Trump says Microsoft wants TikTok. Microsoft is in talks to acquire the US arm of TikTok owner ByteDance, according to President Donald Trump. Microsoft is reportedly among several investors, including Oracle, that are working on a joint bid. Microsoft previously tried to buy the social networking site in 2020, and Nadella called it the “strangest thing I’ve ever worked on.”
- Windows 11 is getting colorful new battery icons that are easier to understand at a glance. I’ve always thought the battery icons in Windows 11 needed to be a bit clearer, and Microsoft agrees. It’s now testing improved battery icons that clearly show your charge status at a quick glance. The battery icons will soon be green if you’re charging, yellow if you’re in energy-saving mode, and red if your battery is critically low.
- The Windows 11 File Explorer is getting shared content. It’s not just battery icons that Microsoft is looking to improve in Windows 11. It has now started testing quick access to files that have been shared with you inside the File Explorer window. This will include files shared over email or in a Microsoft Teams chat.
- Microsoft was the No. 1 game publisher in the world last month. Gaming revenue might have taken a hit in the recent quarter, but analysts at Ampere say Microsoft was ahead of Electronic Arts for game spending in December. That’s an early sign that Microsoft’s multi-platform approach is starting to pay off.
- Microsoft and AWS face more scrutiny from UK watchdog. The UK’s Competition and Markets Authority (CMA) has published provisional findings that show cloud competition isn’t working as well as it could be. “Microsoft is using its strong position in software to make it harder for AWS and Google to compete effectively for cloud customers that wish to use Microsoft software on the cloud,” says the CMA. A final decision from the CMA is expected later this year, but it looks like both Microsoft and AWS will face greater scrutiny under the UK’s Digital Markets Act.
- Neural rendering is coming to DirectX. Nvidia has been hyping up the potential of neural rendering techniques with its new RTX 50-series GPUs recently, and now Microsoft is committing to bringing these new 3D graphics techniques to its DirectX APIs. Neural rendering will use AI workloads for real-time rendering, allowing AI models to power shaders instead of them taking up the entire GPU. Microsoft’s DirectX team is working with AMD, Intel, Nvidia, and Qualcomm on bringing cross-vendor support for neural shaders to DirectX.
- Bill Gates reflects on cofounding Microsoft 50 years ago. Bill Gates founded Micro-Soft, which later became Microsoft, with his childhood friend Paul Allen nearly 50 years ago. It started as a consulting gig for microprocessor software, but two years later, Gates wrote to his Harvard professor to let him know he was taking the spring semester off to focus on Microsoft. He never returned, and the rest, as they say, is history. Gates’ first memoir about the early days of Microsoft, Source Code, arrives on February 4th. He also sat down with The New York Times this week to discuss his early life and other billionaires.
- Microsoft makes OpenAI’s o1 reasoning model free for all Copilot users. Microsoft first launched o1 inside Copilot as Think Deeper in October, but you had to have a Copilot Pro subscription to access it. You no longer need to pay $20 a month for Copilot Pro or ChatGPT Plus to get access to the o1 reasoning model. Microsoft just made it free to use for all Copilot users.
Thanks for subscribing and reading to the very end. If you’ve heard about any of Microsoft’s other secret projects, you can reach me via email at notepad@.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.
