The Microsoft fiscal year ended on June 30, like it always does. For people outside the organization, such a milestone doesn’t matter much. For the markets, it’s just another quarter with the familiar set of outputs released as part of the press release and webcast for FY26 Q4.

I’m not going to try and deliver the kind of generic analysis of the numbers that news sources and analysts focusing on Wall Street and the tech sector do for a living. This is just my perspective on things, based on what I write about and what MS technology I use and care about the most.

As the money spent on AI infrastructure has kept growing, investors have grown concerned about what is a sustainable level to set money on fire without proof of actual end-customer revenue from AI products. For example, Alphabet posted its first-ever negative free cash flow in their quarterly report last week. Their CFO projected the full-year capex to grow to around ~$200 billion, which today is no longer something the markets cheer at just for the sake of “investing in cool new stuff”.

Satya Nadella and Amy Hood knew they had to avoid a similar report at all costs — especially given the poor performance of MSFT stock compared to its peers in 2026. Looks like they found a way to tweak the numbers in a way that made the company look good:

Due to an accounting change, the capex guidance went down from $190 billion to $175 billion. Because, effective FY27, Microsoft extends the estimated useful life of datacenters and office buildings from 15 to 25 years. Magic!🪄 The obligations don’t disappear, they just move to a different place in the books.

Fellow MS watcher Paul Thurrott wrote in his analysis that the depreciation practice changes would also impact GPUs and CPUs. I tried to find evidence to back this interpretation and couldn’t find it. Meaning: I fed all the docs to Opus 5 and Gemini Pro, asking them to find it (since I ain’t no financial analyst). Since Paul has a lot more professional experience on this front, I’m not going to say he’s got it wrong — just demonstrating how tricky it is to decipher the marketing comms and financial statements coming from one of the biggest corporations in the world.

In any case, Microsoft is walking a fine line here by trying not to be the one who sets off a selling spree by announcing less spend on AI data centers and GPUs. At the same time, they want to appear to be more in control of the situation than competing hyperscalers, making sure they’re not doing the same negative cash flow and bonds issuing moves as others. This may buy enough time for another quarter’s worth of speculation and adjustments on the inflation rate of the LLM tech valuation balloon and resulting hardware spend.

As a fun lil’ anecdote from MSFT FY26 Q4, the most profitable thing Microsoft did last quarter was own a piece of Claude. This can be seen from the other income part of the financial statement and the verbal statement from Satya: “these include a $3.2 billion gain from our investment in Anthropic.” There’s also a few billion annual boost from OpenAI ownership included in the numbers.

That’s how money is manufactured in the circular financing world of today’s tech biz. When frontier AI labs sell fixed fee monthly services that burn tokens at an unprofitable rate, their company valuations go up and the owners can book that as profit. Seeing this in action again made me revisit my earlier newsletter issue about AI profits from May 2025:

So, how have things changed since the last time?

Microsoft as a company remains very profitable. Its existing SaaS and platform businesses have been so amazing that this allowed MS to sit alongside Google and Amazon, throwing money in the endless bonfire to build more inference capacity. If the cloud businesses of Big Tech had not been such cash cows, everyone would have run out of money much sooner and we’d have a smaller data center environmental catastrophe to look forward to in the coming years.

Are AI services bringing in new profits from end customers in 2026? That’s still not so clear. The fact that we have to keep guessing on how specifically the numbers in press releases are designed to show the intended metrics to the public is in itself a big warning sign. Not about the usage or usefulness of AI in general, but rather about the stories we’ve been told by Microsoft & co. for the past three years now.

Just like AGI, the profits are always just around the corner. We’re forever stuck in the “early innings” and asked to patiently wait for the future to arrive and prove the marketing pitch was true. In summer 2026, things aren’t materially different from summer 2025, the way I see the numbers and reports that aren’t created by someone with a vested interest in making fetch happen.

Back to more realistic scenarios: is Microsoft 365 Copilot still the commercial failure that I said it was in October? The most accurate answer I can think of, based on the FY26 Q4 data, is another meme pic:

Like with the profits discussion, the number of paid M365 Copilot seats is growing. In fact, it’s been growing pretty nicely since the shockingly low numbers leaked from last Fall. Between January 2026 and now, the reported paid seats figure went from 15 million to 30 million. That’s great progress! Helped by the fact that when you start from single digits in a 400M+ addressable market of M365 users, doubling the number of subscribers is relatively easier than with a broadly adopted product category. But still.