Open Source AI — August 21, 2026
Two stories today, and both are about real-world limits catching up with the AI buildout.
In Plain English: Two stories today, and both are about real-world limits catching up with the AI buildout. SK Hynix — the top maker of the memory chips feeding Nvidia's GPUs — announced its biggest-ever stock buyback after investors sold off its shares even though chip demand is strong, a sign markets are getting nervous about whether AI infrastructure spending keeps paying off. Separately, new data shows European AI data centers are now being built nearly four times farther from major cities than a few years ago, simply because there's no cheap power left near London, Paris, or Frankfurt. No open-weight model news broke today, but the standing argument below — that Apple's memory-heavy Macs sidestep exactly these power and land constraints — gets a little more relevant with each story like these.
SK Hynix Launches Record $28.6 Billion Buyback After Investors Sold Off the Stock Despite Strong AI Chip Demand
SK Hynix — the dominant supplier of the HBM memory chips that feed Nvidia's GPUs — announced South Korea's largest-ever share buyback and cancellation after its stock fell nearly 10% in Seoul, even as the company reported record cash flow and robust chip demand. The disconnect between strong operating results and a falling share price is a concrete signal that investors are growing warier about whether current AI infrastructure spending levels are sustainable long-term, separate from any doubt about near-term chip demand itself.
Cryptopolitan (Bloomberg) · Infrastructure & Economics · Aug 19, 2026
Europe's AI Data Centers Are Moving Nearly 4x Farther From Cities as Power, Not Chips, Becomes the Bottleneck
New data shows AI data centers are now sited an average of 175km from major European hubs like London, Frankfurt, and Paris — up from just 46km a few years ago — because available "powered land" near those cities has effectively run out. Powered land in established hubs now costs roughly 4-10x more (about €2.36 million per megawatt) than in secondary regions (as low as €200,000), pushing the entire buildout toward cheaper, more remote power rather than proximity to customers.
Finimize · Infrastructure & Economics · Aug 19, 2026
The Apple Angle (Standing Perspective)
Set against today's stories about scarce power and jittery infrastructure investors, it's worth restating a case some commentators have been making about who's actually best positioned for the on-prem AI shift — and it isn't Nvidia. The trick is Apple Silicon's unified memory: a single Mac Studio, about $9,500, can be configured with up to 512GB, enough headroom to load a trillion-parameter open-weight model on one machine. Matching that on Nvidia's workstation line would take five or six RTX Pro 6000 cards — roughly $60,000-$75,000 combined and about ten times the electricity draw, the same kind of power constraint now pushing Europe's data centers into the countryside, just playing out at desktop scale.
The case sharpens with every open-weight release: as frontier-caliber models from labs like DeepSeek, GLM, and Kimi keep shipping every few weeks, the bottleneck stops being "who has the smartest model" and starts being "who can afford the box to run it locally." Multi-Mac clusters using MLX and Thunderbolt 5's RDMA networking can already pool memory across several Studios to handle even bigger models, and Nvidia's own DGX Spark essentially concedes the point by borrowing Apple's unified-memory approach — notably after Nvidia stripped NVLink pooling out of its consumer and workstation cards. The fair caveat: a Mac Studio is built to serve one user running one model locally, while a data-center GPU cluster still serves thousands of people at once — so this looks like Apple winning a specific, valuable lane, not the whole race.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items sit squarely on the infrastructure-strain side of the thesis: SK Hynix's buyback-despite-strong-demand shows investor confidence in AI capex decoupling from the underlying operating numbers, and Europe's data-center migration shows that even when capital is available, physical power is becoming the real constraint on scaling commercial AI. Neither story is about open-weight models directly, but both make the standing Apple argument above more relevant by the day — if the bottleneck on the commercial side is increasingly land, power, and investor patience, a self-contained machine that sidesteps all three by running a frontier-caliber open model locally looks more attractive with each headline like these, not less.
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.