Open Source AI — August 31, 2026
Today's most on-the-nose story: OpenAI has quietly bought tens of thousands of Apple Mac minis and Mac Studios to run parts of its AI training, and Anthropic is renting similar Mac capacity through Amazon — both because Apple's shared-memory chip design turns out to fit this kind of workload well.
In Plain English: Today's most on-the-nose story: OpenAI has quietly bought tens of thousands of Apple Mac minis and Mac Studios to run parts of its AI training, and Anthropic is renting similar Mac capacity through Amazon — both because Apple's shared-memory chip design turns out to fit this kind of workload well. Nvidia itself has reportedly started treating Apple as a real rival in local AI hardware, not a hobbyist sideline. Separately, Tencent gave away the weights to Hy4, a giant 770-billion-parameter model that edged out rival open models from GLM and Kimi K3 in blind tests by outside engineers — one more sign free, downloadable models keep closing in on the best proprietary ones. Two different stories, same thread: the tools for running serious AI outside a rented data center keep getting stronger, and even the companies most invested in renting cloud AI are starting to use them.
OpenAI Buys Tens of Thousands of Mac Minis and Studios for AI Training — Nvidia Reportedly Sees Apple as a Local-AI Rival
OpenAI has purchased tens of thousands of Apple Mac minis and Mac Studios to run reinforcement-learning and computer-use-agent training workloads, according to The Information's Aaron Tilley — work that is "memory-bound and parallelism-light" compared to standard transformer pretraining, which makes Apple's unified memory architecture a surprisingly practical fit. Anthropic is taking a similar approach but renting Mac mini capacity through AWS rather than buying outright. The demand reportedly caught Apple off guard: delivery times for high-memory Mac Studio and Mac mini configurations have stretched to weeks or months, and Nvidia has, per the report, begun viewing Apple's silicon as a meaningful competitor in local AI processing rather than a niche platform. For Lucien's thesis, it's about as direct a confirmation as this brief is likely to see: the labs spending the most on rented cloud GPUs are themselves buying and renting Apple hardware for pieces of their own pipeline.
The Information (via Free Press Journal) · On-Prem/On-Device Shift · Aug 31, 2026
Tencent Open-Sources Hy4, a 770-Billion-Parameter Model That Edges Out Rival Open Models in Blind Tests
Tencent released the weights for Hy4-preview on Hugging Face under an Apache 2.0 license: 770 billion total parameters with 49 billion active per token, a mixture-of-experts design, and a context window past 1 million tokens. In blind evaluations run by 163 outside engineers across 203 real coding and research tasks, Hy4 narrowly outscored fellow open-weight models GLM-5.3 and Kimi K3. Tencent also said the model helped optimize its own training and inference systems, lifting throughput by nearly 32% against a baseline. It's another data point in the pattern this brief keeps tracking: as Western labs increasingly keep their strongest models closed, Chinese developers keep giving away weights that are not just competitive with, but in this case ahead of, other leading open alternatives.
TechNode · Open-Weight Models · Aug 28, 2026
The Apple Angle (Standing Perspective)
Most days this section deals in argument and projection; today it gets a literal, real-world data point instead. The recurring case here is that Apple — not Nvidia — may be quietly best positioned for AI's move away from rented, by-the-token compute, thanks to its unified memory design: a Mac Studio built out to 512GB of memory shared between CPU and GPU costs around $9,500 and can hold a trillion-parameter open-weight model on a single desktop. Reaching that same memory ceiling with Nvidia's RTX Pro 6000 workstation cards takes five or six of them — roughly $60,000–$75,000 combined — while pulling close to ten times the power. Today's Mac-buying story above is that argument playing out inside the very labs renting the most cloud GPU capacity, rather than in a hobbyist's home lab.
As frontier-capable open-weight models keep shipping every few weeks — Tencent's Hy4 above is just the latest — the argument goes that the competitive edge shifts from who has the smartest model to who can afford, and physically power, the hardware to run it locally. Apple's MLX framework and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory into a cluster for even bigger models, and Nvidia's own DGX Spark — a small unified-memory box of its own — quietly concedes the same point, notably after Nvidia stripped NVLink pooling out of its consumer and workstation cards. The fair caveat still holds: a Mac Studio serves one person running one model locally, while a data-center GPU cluster still serves many people at once — so this looks like Apple leading a specific, valuable segment of AI compute, not replacing the data center outright. But a report that Nvidia itself now treats Apple as a genuine local-AI rival is a sign this argument has moved past Substack essays and into how the industry's biggest players are actually behaving.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items both point toward the on-prem/on-device half of Lucien's thesis, from opposite ends of the pipeline. OpenAI buying Mac hardware and Anthropic renting it — with Nvidia reportedly taking Apple seriously as a local-AI competitor — is the demand side: even the biggest buyers of rented cloud compute are carving out workloads to run on unified-memory silicon instead. Tencent's Hy4 is the supply side: another frontier-capable open-weight model, free to download and now arguably ahead of its closest open rivals, giving anyone with the right hardware more reason to run models locally rather than pay per token. Neither story proves the profitability half of the thesis (nothing today addressed capex or token pricing directly), but together they sharpen the on-prem case: the models keep getting better, and the framing of Apple silicon as a genuine local-AI competitor — rather than a hobbyist curiosity — is starting to show up in how the biggest labs actually spend their money.
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.