Open Source AI — September 21, 2026
Two stories today, both about the price of AI's engine room.
In Plain English: Two stories today, both about the price of AI's engine room. A chip-industry executive argued AI compute has to get radically cheaper - the way electricity and broadband once did - before it can become a true mass-market product, which is really an admission that it's still priced like a luxury today. Separately, Raspberry Pi's famously unimpressed CEO said in a new interview that AI hype is overblown - even as his own company profits from real, growing demand for cheap, on-device AI hardware. Today's Apple Angle argues the destination both stories are circling - affordable, locally-owned AI hardware - is something Apple already sells.
A Chip Executive Says AI Inference Has to Get Radically Cheaper - or the Industry Stalls
In a new op-ed, Marshall Choy of inference-chip maker Rebellions Inc. argues AI compute needs to follow the same path electricity, broadband, and cloud computing took: from expensive and rationed to cheap and abundant. He points out that even well-resourced teams - he names Microsoft - currently throttle API calls and cap token usage to keep cloud bills under control, which he calls the wrong long-term model. Choy's vision of "ambient," always-on AI assistants only works once per-token costs fall sharply, and he argues the industry should judge itself by task completion and money saved rather than benchmark scores. For Lucien's thesis, the piece is notable less for its prescription than its concession: a chip-industry insider is publicly admitting inference is still priced like a scarce resource today - exactly the pressure point prediction #1 rests on - even though Choy's own proposed fix, industry-wide cost cuts, is a bet against the "prices keep climbing" version of that same argument.
Infrastructure & Economics · SiliconANGLE · Sep 20, 2026
Raspberry Pi's CEO Calls AI Hype Overblown - While His Own Boards Sell Into the Local-AI Boom
In a new Financial Times interview, Raspberry Pi CEO Eben Upton pushed back hard on AI industry hype, saying people are "very inclined to overestimate what these tools can do" and warning that overselling AI's abilities could distort young people's career and skills choices for the worse. The irony: Upton delivered that skepticism while running a company whose growth increasingly rests on the exact trend he's questioning - Raspberry Pi now pairs its boards with roughly $100 neural accelerators built specifically for local, on-device AI inference, and shipped more than 4 million units in the first half of 2026 alone. For Lucien's thesis, the interview matters precisely because of who's saying it: a hardware CEO with every incentive to stay quiet about AI's limits is nonetheless describing real, paying demand for affordable local inference hardware - separate from, and skeptical of, the cloud-API model this brief keeps tracking.
On-Prem/On-Device Shift · Financial Times (via Techmeme) · Sep 20, 2026
The Apple Angle (Standing Perspective)
Today's two stories sit on opposite sides of the same coin: one says AI compute has to get radically cheaper before it can go mainstream, the other shows a skeptical hardware CEO nonetheless building a real business on affordable, local AI inference. Apple's answer to both problems is already on shelves. Configure a Mac Studio with its full 512GB of unified memory - about $9,500 - and it holds an entire trillion-parameter open-weight model in memory on one machine, with no metering and no throttled API calls to worry about.
Matching that memory footprint with Nvidia's own workstation cards takes five to six RTX Pro 6000s, roughly $60,000 to $75,000 combined, pulling something like ten times the power. As frontier-capable open-weight models keep landing every few weeks, the more interesting question keeps drifting away from which lab trained the smartest model and toward who can simply afford to own the box that runs a nearly-as-good one, locally, without asking a cloud provider's permission. Apple's MLX framework and Thunderbolt 5's RDMA networking already let several Mac Studios pool memory into one larger effective machine, Nvidia has quietly dropped NVLink from its own consumer workstation cards, and its DGX Spark borrows more from Apple's unified-memory approach than the reverse. The honest caveat holds every day this section runs: a Mac Studio is built to serve one person running one model at a time, while the debt-financed, gigawatt-scale facilities covered elsewhere in this brief exist to serve millions of concurrent users - Apple looks well positioned to quietly own a valuable corner of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's items pull from opposite directions but land in the same place: a chip insider admits inference is still priced too high for mass adoption, while a skeptic's own sales numbers show real demand for cheap, local AI is already here. Both sharpen prediction #1's central question of who gets to make AI affordable, and reinforce why the standing Apple Angle keeps pointing at the same shelf-available answer.
The Daily: Open Source AI - a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.