Reading Room

You've been thinking about your own AI tool spend. Revolut, a banking app with more than 80 million customers, just gave a real-world playbook. It cut its AI coding bill by 90 percent. Its engineers also started shipping code 70 percent faster. It did this not by cutting back on AI, but by changing how it paid for it.

David Abitbol leads AI operations at Revolut. He described the change at a conference called AI Engineer New York. Revolut had already given about 2,000 engineers its best coding tools: Claude Code and Codex. Those are AI assistants that write and edit code. Revolut paid for them per use, through Anthropic and OpenAI. But giving everyone the best tool did not speed up delivery on its own. The real jump came when Revolut switched engineers to a model it runs itself. It is an open-weight model (the underlying files are public, so anyone can run a copy) called GLM, built by a company named Z.ai. Revolut hosts it on 56 Nvidia B200 chips (specialized hardware built to run AI models fast), rented through a company called Together AI.

The blended cost works out to $0.30 for every million tokens (chunks of text, roughly three quarters of a word each) the model processes. That is about one tenth of what Revolut paid before. After the switch, the time from an engineer picking up a feature to having code ready to merge fell by about 70 percent. Revolut also tuned the setup, rather than just buying more chips. The same 56 GPUs ended up handling 76 percent more work than when they started. Abitbol has a blunt label for one wasteful default setting in this kind of system: he calls it 'a scam.'

The numbers carry extra weight because of who is reporting them. Revolut sold shares in July 2026 at a valuation of $115 billion. That made it Europe's most valuable private company. It reported more than $6 billion of revenue in 2025, up 46 percent, from over 80 million customers. It also built its own banking-specific AI model, called PRAGMA, with Nvidia, and runs an in-house research lab called Revolut Research. A company with that much of its own AI expertise choosing to move off the big paid AI providers is a signal worth watching, not just a cost-cutting story. One gap: the newsletter is itself a teaser for a longer paid report, so some promised detail, like exactly how a European Union-regulated bank cleared a Chinese-developed model through compliance, is not included here.

A report on AI spending by the investor a16z found that 69 percent of S&P 500 companies run AI in production. Only 2 percent of them track any AI metric over time. Revolut is in that 2 percent, with nearly two years of data across 2,000 engineers. That is part of why Abitbol's numbers are worth more than the usual conference claim.

An opinion piece in Crain's Chicago Business argues that Anthropic's own warnings about AI risk are really a case for taking control of AI away from a handful of Silicon Valley companies, not just a safety statement. The piece frames the debate as one about who holds power over AI, not only whether it is dangerous.

Worth tracking because rules shaped over the next year or two, on who gets to build and control AI, could affect how he and others build with it.

Why it is here: a useful counterweight to the bullish AI infrastructure stories he keeps running into, even though this is one person's opinion on reddit, not research backed by data.

A poster on r slash ArtificialInteligence argues that today's drop in chip and PC hardware maker stock prices is just the first crack. They claim the huge sums poured into AI data centers and chips (the computers that run AI models) are no longer backed by real financial returns, so a much bigger price collapse is coming. The poster is specific about timing: they do not expect the crash to hit full force until late 2027, because long term corporate supply orders already placed, and new chip production lines already scheduled, are propping prices up for now.

Once those props run out, the poster expects two problems to hit at once: the current pricing cannot be sustained, and looming energy shortages plus the high cost of power needed to run data centers will force prices down hard, potentially below where they stood before the current chip shortage years. They call the whole boom hype mixed with institutional ignorance, and predict that ordinary PC builders and gamers, the market big manufacturers deprioritized to chase corporate AI orders, become the main buyers again once prices fall.

The caveat: this is one person's prediction with no data, model, or named analyst behind it, posted to a general AI subreddit. There is no report or firm backing the 2027 date. Worth tracking as a contrarian marker against the bullish AI spending stories, not as a forecast to act on.

Why it is here: a sharp counterpoint to the idea that only desk jobs are exposed to AI. This poster, a tradesperson, argues manual labor is not the safe harbor people assume.

Posting to r slash ArtificialInteligence, the poster says blue collar trade work is at risk from AI, and not for the reason people expect. It is not that robots will physically replace plumbers and electricians. It is that most everyday trade calls, a dripping tap, a boiler that will not fire, moving a plug socket six feet, putting up a few shelves, are really knowledge and diagnosis problems wrapped around simple physical work. AI can now supply the diagnosis and the how-to instantly, which removes the main reason someone used to need to call a professional for a small job.

The poster estimates, from their own country's trade market, that these small, low ticket jobs make up around 14 percent of total trade revenue but 25 to 30 percent of total trade jobs. That gap matters: it is a large share of the entry level work in the trades, even though it is a small share of the money. AI, plus a bit of a do it yourself mindset, removes the real barrier to doing this work yourself, which was never the physical skill. It was knowing what is wrong and which tool or part fixes it. YouTube partly solved this before, the poster says, but when a tap is actively leaking, people want the fix now, and AI gives an instant answer rather than a video to search through.

The poster draws a direct comparison to how AI is already squeezing entry level software engineering jobs: cutting out the low ticket work removes the bottom rung new workers use to enter a trade and build experience, which over time thins out the whole profession from the bottom up.

Outside your usual reading: ASN Bank, a Dutch bank, blocks its app from working on phones running privacy focused Android alternatives like GrapheneOS, even when the phone itself works fine, by requiring Google's own verification service to log in. It is a small, clear example of how software gatekeeping can lock people out of devices they fully own, a pattern worth recognizing in other apps.

Outside your usual reading: Kugelblitz is a new Mac app that renders the Earth and its neighboring bodies as a smooth 3D globe you can fly over like a pilot, built entirely from free, openly published map data such as OpenStreetMap, with no account needed. It runs smoothly even on an older M1 Mac and can stream the view straight into a Vision Pro headset. A nice example of a small, well built tool doing something big companies usually charge for.