Open Source AI — August 19, 2026
Three stories today all point to strain hiding underneath the AI buildout.
In Plain English: Three stories today all point to strain hiding underneath the AI buildout. A Guardian investigation found Microsoft has only 2.2 million AI chips actually installed — well short of what its $280 billion in spending implied — because it doesn't have enough power-ready data-center space to plug them in; Microsoft's own CEO admitted as much. Separately, the computer memory chips that AI data centers depend on have jumped 500-1,000% in price over the past year, because AI companies have locked up almost the world's entire 2027 supply, squeezing ordinary PC buyers in the process. And a closer look at Nvidia's reworked $105 billion financing deal for OpenAI's Ohio data center — already down $145 billion from the figure first reported weeks ago — shows investors increasingly worried the AI industry may be financing its own demand rather than responding to real customer need. None of this proves the buildout won't pay off, but the gap between money spent and value actually delivered got harder to explain away today.
Guardian Investigation: Microsoft's AI Chips Barely Moved Despite $280 Billion Spend
A Guardian investigation (reported by Aisha Down and Ed Zitron) found Microsoft targeted 1.8 million AI chips installed by the end of 2024 but has only about 2.2 million running nearly two years later — a shortfall well below what outside experts expected given roughly $280 billion in AI infrastructure spending through mid-2027. CEO Satya Nadella confirmed the underlying problem isn't chip supply but power: "you may actually have a bunch of chips sitting in inventory that I can't plug in... I don't have warm shells to plug into." It's one of the most concrete public admissions yet that AI capex isn't converting into usable capacity as fast as headline spending figures suggest.
The Guardian, via AOL · Infrastructure & Economics · Aug 17, 2026
AI Data Centers Have Bought Up the World's Memory Supply — Prices Are Up to 10x
Tom's Hardware reports that computer memory (DRAM) prices have climbed 500% over the past year, with some 128GB DDR5 kits up nearly 1,000% and now costing more per kilogram than half the price of solid gold. The cause: hyperscale AI data-center operators have pre-purchased almost all of the world's 2027 DRAM production capacity, diverting factories toward high-margin AI memory (HBM) and away from consumer chips. SK Hynix's CEO called 2027 set to be "the worst year for memory supply in the industry's history." It's a cost the AI buildout is exporting directly onto everyone else's hardware bills — and a reminder that the infrastructure squeeze runs well beyond GPUs.
Tom's Hardware · Infrastructure & Economics · Aug 17, 2026
Nvidia's OpenAI Financing Keeps Shrinking, Reviving "Circular Financing" Fears (Follow-up)
Fortune reports that Nvidia's guarantee backing OpenAI's 8-gigawatt Pike County, Ohio data center — the deal covered here on Aug 17 at $105 billion — was cut down from an initial $250 billion in stages over just a few weeks, a $145 billion reduction. The repeated downsizing has revived investor unease that Nvidia is essentially financing demand for its own chips: when the $250 billion figure first surfaced, Nvidia's stock dropped roughly 4.5% intraday. CEO Jensen Huang disputes the "circular financing" framing, saying the deal is about "securing long-lived infrastructure for Nvidia compute so OpenAI can deploy the most productive AI factories" — but the scale of the walk-back is itself a data point on how fluid these numbers still are.
Fortune · Infrastructure & Economics · Aug 18, 2026 — follow-up to Aug 17 item
The Apple Angle (Standing Perspective)
A perspective worth carrying forward every day: while attention stays fixed on Nvidia, some observers think Apple has quietly assembled the more practical toolkit for running big AI models outside the cloud. The reason is memory architecture — Apple Silicon's unified memory lets a single Mac Studio, around $9,500, be configured with up to 512GB in one machine, enough to hold trillion-parameter open-weight models without spreading them across a rack. Nvidia's workstation cards would need five or six RTX Pro 6000s — roughly $60,000-$75,000 combined, and about ten times the power draw — to match that capacity.
That gap matters more each week a new frontier-capable open-weight model ships from labs like DeepSeek, GLM, or Kimi: once a model good enough to matter exists for free, the constraint shifts from "who built the smartest model" to "who can afford to run it locally" — and on hardware economics, Apple's Mac Studio line looks unusually well positioned, largely without anyone declaring it so. MLX and Thunderbolt 5's fast networking already let multiple Mac Studios pool memory to split a model between machines, and even Nvidia's own DGX Spark borrows the unified-memory approach Apple popularized. The fair caveat: Mac Studios are built for one user running a large model locally, while data-center GPU clusters still serve many concurrent users at once — so this may be Apple winning a valuable slice of AI rather than the whole contest.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items all sit on the "commercial AI spending isn't converting cleanly into value" side of Lucien's thesis: Microsoft's own CEO admitting chips are sitting unplugged for lack of power, memory prices spiking 500-1,000% as AI data centers buy up global DRAM supply, and Nvidia's OpenAI financing shrinking by $145 billion amid circular-financing concerns all describe an infrastructure build-out that keeps costing more and delivering capacity more slowly than advertised. That's exactly the kind of friction that makes the standing Apple argument above more relevant, not less — if a single unified-memory machine can already run a trillion-parameter open model today, without waiting on power buildouts, chip allocation, or DRAM supply chains that are themselves under visible strain, the case for local inference gets stronger as the commercial cloud AI supply chain gets more visibly congested.
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.