The funny thing about the AI boom is that nobody actually agrees on what "booming" even means anymore. Companies keep announcing datacenters, models keep shipping, and every other post on my feed is convinced the end is nigh. So when the Guardian dug into Microsoft's numbers this week and found the company running roughly 2.2 million AI chips — a number that sounds massive until you remember Microsoft once said it wanted 1.8 million by the end of 2024 — my first reaction wasn't about capacity. It was about how little any of us actually knows.
Here's the short version. Microsoft has poured somewhere around $280 billion into AI infrastructure since 2022, including $41 billion in the past quarter alone. CEO Satya Nadella has been promising to double the company's datacenter footprint by mid-2027. Yet the internal documents the Guardian got hold of suggest the newest datacenters either aren't fully switched on or don't have the chips to fill them. That's not a tiny gap. That's the kind of discrepancy that turns analyst models into astrology.
To be fair, this is barely even a Microsoft problem. It's an industry problem wearing a Microsoft logo. Nvidia won't say how many chips it sells or to whom. The big cloud providers won't say how many they've actually deployed. So "is AI really scaling, or are we all just… pretending?" is a question nobody can answer with confidence. The richest companies on Earth are guessing in the dark right alongside the rest of us. I find that both reassuring and unsettling, in roughly equal measure.
The Watermark Fight Nobody Asked For
Meanwhile, Anthropic decided this week was a good time to roll out an AI watermark, and let me tell you — the Claude crowd did not take it well. The feature is meant to satisfy updated European transparency rules, which on paper sounds reasonable. In practice, users are worried that even lightly edited AI-assisted text will get flagged, turning it into a digital scarlet letter that follows your writing around.
I get the EU's intent here. I do. Transparency around machine-generated content has a place. But the implementation detail that keeps nagging me is how blunt these tools tend to be. A watermark that can't tell "I wrote this whole thing myself" from "I asked for a comma fix" is not transparency, it's a blunt instrument. A lot of people are wondering why the people who actually write for a living keep getting treated like the problem.
Nuclear Energy Is the New GPU
Off the beaten path, the SK Group and Bill Gates' TerraPower met again in Seoul this week, and the subtext was pretty loud: AI's electricity appetite is now big enough that small modular reactors are getting the full courtship treatment. SK already holds a strategic stake — it put $250 million into TerraPower back in 2022 and is now the second-largest shareholder. The two sides have reached a preliminary agreement on SMR cooperation, which would let SK plug into TerraPower's Natrium reactor project in Wyoming, a fourth-gen sodium-cooled design that got its construction permit from the U.S. NRC back in March.
Honestly, I find this one of the more quietly important stories of the week. Nobody's throwing confetti over it, but when industrial conglomerates start treating nuclear power as the long-term power source for AI datacenters, that's the market telling you something. This isn't a hobby anymore. It's the power grid rewiring itself for inference.
RAMapocalypse Hits the Wallet
Now the part that actually stings for regular people. DDR5 memory prices in Germany have climbed to nearly 5x what they were in July 2025, with the price index jumping from 445% to 486% in a single month. The pattern is identical elsewhere in the world. The AI boom keeps swallowing DRAM supply, and the people left holding the bag are budget PC builders — the same folks who were told two years ago that AI would make everything cheaper and better.
From my perspective, this is where the "AI for everyone" story starts to wobble. You don't need a GPU cluster to feel the boom. You just need to want 32GB of RAM at a sane price. I'd love to tell you when this normalizes, but every month that passes the answer looks the same: not yet, and probably not soon. Quick add-on note — if you're planning a build right now, it might be worth holding off on the RAM purchase unless you absolutely need it.
Your Vacation Photos Are a Geolocation Database
Last one, and it's a creepy one. McAfee ran a study feeding travel photos to two open-weight models — Google's Gemma 3 27B and Alibaba's Qwen 3 VL 30B — and found they could identify the correct city and country from the images alone about 87% and 91% of the time respectively. No GPS data. No EXIF. No captions. Just architecture, signage, skylines, street markings, and the occasional food stall. Tourist spots were easy; even generic beaches and hotel rooms often gave away the country.
The models are open-weight and free to run, which is the uncomfortable part. Privacy tools that used to require a sophisticated attacker are now sitting in any hobbyist's home lab. I'm not here to scare anyone, but keep this in mind the next time you post that beach photo with the distinctive palm-tree skyline while on vacation. The internet's memory is long, and now it's got better eyes.
Also floating around this week: a native Postgres GUI for macOS that runs Apple's on-device model for text-to-SQL (Widen), and a small open-source tool that cryptographically proves which checks actually ran against your code (ProofRun). Both are worth a look if you live in the terminal.
That's the week. The boom is real, the costs are real, and somewhere in between, the industry is still figuring out what any of it actually means. If you're curious about smarter planning around all this — from budgets to builds — I keep a small calculator collection that has saved me more than once: PayCalc. See you tomorrow.
Top comments (0)