Open Source AI — August 28, 2026
Two real developments today, and both are really about who controls the pipes AI runs through.
In Plain English: Two real developments today, and both are really about who controls the pipes AI runs through. Anthropic signed a $45 billion, six-year deal to rent computing power from Nscale, a two-year-old British infrastructure company — its latest in a fast-growing string of mega compute purchases, and a concrete reminder that even a leading AI lab still needs tens of billions of dollars in future hardware capacity long before anyone knows whether renting it out will ever turn a profit. Separately, European regulators are reportedly circling Nvidia's newly confirmed purchase of Hugging Face — the site that hosts most open-weight models — over concern that a chipmaker owning the model marketplace could squeeze out rivals. The standing Apple argument below lands with extra force today: both stories are about who controls centralized AI infrastructure, and the whole point of the on-prem/on-device shift is sidestepping that question entirely.
Anthropic Signs $45 Billion, Six-Year Compute Deal With British Infrastructure Startup Nscale
Anthropic has agreed to pay roughly $45 billion over six years to rent AI computing capacity from Nscale, a British infrastructure company founded in 2024, using Nvidia's newest Vera Rubin chip systems at a flagship data center in West Virginia. The capacity is expected to come online in late 2027. It's the latest in a rapid string of enormous compute commitments from Anthropic — alongside recent deals with Norway-based Volta ($10B), AMD ($5B), SpaceX, Google/Broadcom, and Amazon — and it's a concrete data point for the profitability question this thesis tracks: one of the industry's frontier labs still needs tens of billions of dollars in future compute just to keep pace, years before anyone can say whether inference revenue will cover those costs at today's — or tomorrow's — token prices.
Bloomberg / CNBC / TechCrunch · Infrastructure & Economics · Aug 26, 2026
European Regulators Reportedly Eyeing Nvidia's Hugging Face Acquisition on Antitrust Grounds (Follow-Up)
A day after Nvidia's $12.9 billion agreement to buy Hugging Face was confirmed (covered in this brief on Aug 27), reporting suggests European authorities may scrutinize the deal on competition grounds. The concern: letting the industry's dominant AI chipmaker also own the leading distribution hub for open-weight models could let Nvidia extend its market power into model distribution, developer tooling, and enterprise infrastructure — potentially disadvantaging rival chipmakers and cloud providers who also depend on Hugging Face to reach developers. No formal review has been announced, and outcomes could range from unconditional approval to imposed access commitments to an outright block. For Lucien's thesis, it's a reminder that even as open-weight models keep closing the capability gap, who controls the pipes those models travel through remains a live and contested question.
Dealroom.co / Forkast · Open-Weight Models (Follow-Up) · Aug 27, 2026
The Apple Angle (Standing Perspective)
Every mega compute deal like today's Anthropic-Nscale agreement is a reminder of the alternative path some commentators think is getting overlooked: running frontier-caliber models locally instead of renting them by the token. The case rests on Apple's unified memory architecture — a single Mac Studio, roughly $9,500 in a mid-tier build, can be configured with up to 512GB of memory shared between CPU and GPU, enough to hold a trillion-parameter open-weight model on one desktop. Reaching that same memory footprint with Nvidia's RTX Pro 6000 workstation cards takes five or six of them, on the order of $60,000-$75,000 combined, while pulling roughly ten times the power.
As frontier-capable open-weight models keep shipping every few weeks, the argument goes, the bottleneck stops being "who has the smartest model" and becomes "who can afford, and physically power, the box 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 larger models, and Nvidia's own DGX Spark — a small unified-memory box of its own — quietly concedes the same point, especially notable since Nvidia stripped NVLink pooling out of its consumer and workstation cards. The honest caveat still holds: a Mac Studio serves one person running one model locally, while a data-center GPU cluster serves many people at once — so this looks like Apple leading a specific, valuable segment of AI compute rather than replacing data-center AI outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Both of today's items are, at bottom, about control over centralized AI infrastructure — one financial, one regulatory. Anthropic's $45 billion Nscale deal shows a frontier lab still has to keep buying enormous amounts of rented compute capacity years in advance, with no guarantee that today's spending will look rational once the bill comes due — exactly the profitability tension this brief tracks. Meanwhile, European regulators reportedly circling Nvidia's Hugging Face purchase shows that even the "open" side of the AI ecosystem can end up gated by whoever owns the distribution pipes, a risk that doesn't exist if models are simply downloaded once and run locally. Set against the standing Apple argument, today's news reads as two more reasons the on-prem path looks more attractive with each passing week: one because renting compute keeps getting more expensive and uncertain, the other because owning your own hardware means no marketplace gatekeeper to worry about.
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.