The financial markets are getting increasingly nervous about the AI capex spend of tech giants. Especially now that “the Don is gone” — meaning Jensen Huang has left Nvidia and is no longer there to position the company as the backstop for massive investments made by neoclouds and basically any tech company smaller than Microsoft. The hyperscalers still seem relatively safe, even though their war chests are looking empty compared to a couple of years ago.

It feels as if the blissful days of waiting for AGI are over and now the pressure is on for all the CEOs to show they’ve got a more concrete plan in their back pocket to make money, not just spend it. Microsoft sure knows how to make money when looking at the business as a whole. But merely growing the quarterly revenue figures steadily isn’t sufficient anymore. To offset the unprecedented amount of capital tied up in increasingly expensive hardware, investors are expecting to see more than just a steady climb in Azure revenue. They want to also see cost cuts.

This week I’ll look at recent Microsoft moves and reports from current and former employees for signs of that pressure. The capital commitments stretch years into the future. The patience for research, employees and products seems to be running out much sooner.

In a recent blog post titled The End of AI Frontiers, longtime Microsoft researcher John Langford explained why he left the company after the AI Frontiers group he helped establish in 2023 was absorbed into Mustafa Suleyman's Microsoft AI organization. His story offers an insider perspective on the clash between two very different approaches to AI development at Microsoft:

While nothing will be formally announced by Microsoft, the future direction we can infer from these organizational conflicts and personnel changes is not hard to predict. Launching new MAI models that can be immediately put into use inside Microsoft’s products and reducing dependence on GPT and Claude are the primary goals the MAI org is pursuing. Humanist AI may be the branding used for distinguishing the MAI approach from other labs. The practical side seems to be “we do less crazy stuff than Anthropic and OpenAI, plus we ship efficient models that fit inside monthly SaaS bills of MS products.”

Microsoft Research has been around since 1991, with a long tradition of letting researchers explore ideas without having to justify their immediate commercial value. That's quite different from focusing resources on scaling existing AI models. Given how much Microsoft is now spending on AI infrastructure, I can see why predictable results are becoming a priority. But it does raise questions about how much room there will be for research that challenges the current direction, and what kind of talent MS will be able to attract and keep for its AI team.

Meanwhile, a bigger wave of uncertainty is hitting not just the AI or Research orgs but practically all of Microsoft. Employees across the US, UK, Canada and India have been directed to record their decision on this VEO process screen:

“Should I stay or should I go?” Compared to the earlier VRP that was scoped to the long-time MSFT’ies possibly looking to officially retire, this VEO does not discriminate by age or tenure. Everyone is on the chopping board and the question everyone affected will be thinking is: “if I don’t take the offer now, how soon will I get hit with the layoffs?”