A Shift in the AI Power Dynamic? (Premium)

Everyone reading Thurrott.com is familiar with the classic scene in The Empire Strikes Back in which Darth Vader reneges on his deal with the bounty hunter Boba Fett and Bespin administrator Lando Calrissian, later exclaiming to the latter, “I am altering the deal, pray I don’t alter it any further.” Even as a child of just 13 years old at the time, I immediately understood two things from this altercation. Even bad guys–Vader and Boba Fett, at least, were clearly bad guys–can’t trust a bad guy. And though this was inherently a partnership of sorts, Vader was exercising a power dynamic in which he was fully in control.

On that note, it’s difficult not to view the secretive Microsoft and OpenAI partnership in this same light all these years later. These two firms are both objectively evil in how they push AI in unethical ways, and their partnership was specifically designed to prevent any regulatory oversight and be as secretive as possible. But it’s never been clear which company was in the driver’s seat. That is, partnerships like this are ideally a win-win, meaning that both companies benefit in ways that are meaningful to each and further their respective corporate strategies. But business deals are rarely equally beneficial to all parties, with one perhaps operating from a position of weakness. And so the debate rages here. Which is Vader and which is Boba Fett, Microsoft or OpenAI?

Early on, it appeared that OpenAI had the upper hand: Microsoft’s AI strategy of 2023 was solely based on taking whatever OpenAI could give it and repackaging it for its consumer and commercial partners. Indeed, Yusuf Mehdi once described what was then called Bing Chat as a combination of OpenAI’s ChatGPT (GPT-4) and Microsoft’s “secret sauce,” which he then incorrectly described as its Prometheus model plus the Bing search algorithms and relevancy ranker. But that’s not the case: Prometheus is/was the secret sauce. That is, Prometheus is/was a model that Microsoft created by combining ChatGPT with Bing index, ranking, and answers. Put another way, Bing is the secret sauce.

And you can see the problem there: Bing is terrible. Terrible as a product and terrible as a brand, and it was only a matter of weeks before Microsoft starting backing away from the Bing-centric branding for its AI offerings. More to the point, OpenAI was free to release its own AI offerings and did so. And ChatGPT exploded in popularity, quickly eclipsing Bing from a usage perspective, in a major embarrassment to Microsoft. There are contradictory numbers out there, but the latest figures I could find place ChatGPT users in the hundreds of million and traffic at over 2 billion visits per year. (ChatGPT had over 200 million weekly active users as of last August.) ChatGPT is like Microsoft AI, hold the secret sauce.

But then November 2023 happened. When OpenAI CEO Sam Altman was unexpectedly fired in a boardroom coup, Microsoft CEO Satya Nadella exercised his own power dynamic quickly, forcing OpenAI to embarrass itself, bring back Altman, and replace much of its board of directors. Thanks to this episode, it seemed clear which company was in the driver’s seat. Microsoft–or Nadella, perhaps–was the Darth Vader to OpenAI’s Boba Fett. Yes, the two companies were partnering. But Microsoft had the upper hand.

Or did it?

The OpenAI drama triggered fears up and down Microsoft’s corporate leadership ladder. The company wasn’t just all-in on AI, it was literally betting the company on AI, and not just on AI generally, but on OpenAI and whatever it was that the startup could produce. And OpenAI had just shown itself to be toxic and unstable. This would not do. And so a plan was hatched internally to lessen Microsoft’s reliance on this one source for AI. This was likely met with relief by most inside Microsoft, and I suspect many had been leery of OpenAI, and certainly of the unilateral nature of this partnership, from the beginning. But Nadella’s next move was shocking, internally and externally.

In March 2024, Microsoft unexpected announced the creation of a new Microsoft AI “organization” that didn’t seem to have a direct place in the company’s corporate structure. Curiously, it would be led by outsiders, including DeepMind and Inflection co-founder Mustafa Suleyman, who would report directly to Nadella, bypassing the rest of the company’s leadership. Worse, it would be staffed largely by former executives and engineers from Suleyman’s previous employer, Inflection, a company Microsoft was not acquiring and thus would not need to seek regulatory approval. But Microsoft “acquired” Inflection in the same sense that it “acquired” Nokia: It got all the best people and all the best technology that the other company offered and took it as its own.

Microsoft AI is a curious beast. It consumed previous AI teams at Microsoft, including those responsible for Copilot, Bing, and Edge, and for generative AI. But the leadership is all from outside the company. And this was so unacceptable to the person who led the first of those two Microsoft teams that he quit. In the months since, we’ve seen only limited evidence of this new organization’s work in whatever AI advances Microsoft has added or will add to Windows 11, Copilot+, Bing, and other consumer products. Suffice to say it’s not particularly inspiring, and certainly nothing like the advanced computer science we see at OpenAI. For example, the Sora generative video service and the incredible advanced voice mode feature that debuted with GPT-4o.

Did this shift the power dynamic at all? Not exactly, but the ensuing year was chaos. OpenAI partnered with Apple for Apple Intelligence and then put an Apple executive on its board before removing him and the Microsoft representative. We learned that Microsoft partnered with OpenAI solely out of fear that it would lose to Google. OpenAI announced a search competitor and has added features to ChatGPT at an astonishing clip. Microsoft began partnering with other AI startups, like Mystral, and opened up GitHub Copilot to other models. Nadella was so desperate to distance Microsoft from OpenAI that he referred to its AI as “a commodity.” And there were rumors that Microsoft would soon stop using OpenAI models exclusively in its core AI offering, Microsoft 365 Copilot. And a report explaining that this relationship would only end if OpenAI somehow made over $100 billion on AGI, a technology that may in fact be impossible.

Years ago, a friend from Microsoft and I were discussing the contentious relationship that the Windows team, then the core of Microsoft, had with Dell, one of its biggest partners. Both were terrible to each other, overtly undermining each other despite how much each needed the other. “These companies deserve each other,” he declared in the perfect summary.

And that’s how I view Microsoft and OpenAI today. They need each other, and explicitly. They are actively working to undermine the other and cut ties as soon as possible, almost as explicitly. And yet, neither will be successful without the other. In short, they deserve each other.

And then this week happened.

On Tuesday, Microsoft and OpenAI said they were “thrilled” to announce that they would “evolve” their partnership. The base relationship remains unchanged: Microsoft gets everything OpenAI makes to use in its own products and services, OpenAI’s API are still exclusively hosted on Azure, and there are never-explained revenue sharing agreements in both direction. (Microsoft is rumored to collect 20 percent of whatever money OpenAI earns.) But there is one major change, and it’s raised many questions: OpenAI is free to get capacity, “primarily for research and model training,” so not exclusively, from others. There are only a few others: Amazon, Google, and maybe Oracle. But whatever.

The most obvious question is whether this changes the power dynamic and, if so, how. Is there some version of this story where OpenAI shifts into the Darth Vader role? This could happen if it was Microsoft that was more reliant on OpenAI than the reverse. And if you believe that to be true, then you will view this shift as something that is bad for Microsoft and good for OpenAI.

I do not view it that way at all.

OpenAI runs the most expensive and resource-intensive computer workloads on the planet. There are now reports that Microsoft cannot meet OpenAI’s demands for resources, despite spending $80 billion this fiscal year building out the required infrastructure as fast as it can. And then these demands, and Microsoft’s inability to meet them, has caused much friction. But I don’t need reports to know that’s true. As much as Microsoft benefits from OpenAI, its workloads are clearly a strain, a negative. It is bogging down everything else that Microsoft is trying to do. It’s impacting the user experience of its many other online offerings. Microsoft has more capacity than almost any company on earth. But there are still limits.

My take on this is that this shift is a relief for both companies. OpenAI will put workloads on other cloud infrastructures, lightening the burden on Microsoft. But OpenAI is, for your D&D fans, chaotic evil. And that means it will no doubt probe these competing clouds to see which are cheaper or more performant, and it will begin running business workloads elsewhere too. As I pointed out up front, even 13-year-old Paul understood that you can’t trust a bad guy. And that’s true if you’re a bad guy yourself. They’ll always betray each other.

And Microsoft and OpenAI will continue to betray each other as much as possible. That’s their nature.

To beat this Star Wars reference to death, it’s not difficult to see Microsoft as the Empire in this equation. But OpenAI is no rebel alliance. It’s just a bunch of bad guys going after other bad guys and now allying with even more bad guys. Because when it comes to Big Tech and the AI startup bros, there are no good guys, no Luke Skywalker. And, yes, they all deserve each other. But we, consumers and customers, deserve a better outcome. The problem is, that won’t happen.

After all, it’s a dark time for the rebellion.