Open Source AI — September 1, 2026
Three stories today, none of them about a flashy new model — all about the plumbing underneath.
In Plain English: Three stories today, none of them about a flashy new model — all about the plumbing underneath. SpaceX is now making its own jet-engine turbine blades because the portable gas generators AI data centers depend on for power have a waitlist stretching to 2030, and normal suppliers can't keep up. A pro-AI advocacy group just put $50 million behind a political campaign to fight local backlash against data centers in three states, showing that community pushback is becoming a real cost of the buildout, not just noise. And the EU committed €388 million to build its own AI supercomputer in Finland, explicitly to reduce Europe's dependence on renting compute from U.S. cloud providers. None of today's news is about open-weight models directly, but it's a good snapshot of how strained and political the rented, centralized side of AI infrastructure has become — which is exactly the backdrop the "run it yourself" argument needs.
SpaceX Builds Its Own Turbine-Blade Factory to Break an AI Power Bottleneck
SpaceX is bringing turbine blade and vane casting in-house at a new Texas foundry rather than relying on outside jet-engine suppliers, after those suppliers proved unable to keep up with demand for the portable natural-gas turbines that AI data centers — xAI's in particular — depend on for power. Normal turbine order queues were reportedly stretching out to 2030; Elon Musk says the in-house foundry cuts generator delivery times by about 18 months, on top of the roughly $1 billion he has already spent building a portable turbine fleet. For Lucien's thesis, it's a vivid data point on the cost side of the profitability question: when a rocket company has to start manufacturing jet turbine parts just to keep its data centers powered, that is real, unglamorous capital being burned on infrastructure bottlenecks rather than on the AI itself.
Tom's Hardware · Infrastructure & Economics · Aug 30, 2026
Pro-AI Group Puts $50 Million Behind Political Fight to Defend Data Centers
Build American AI, a nonprofit tied to the pro-AI super PAC Leading the Future, has launched a roughly $50 million campaign to build political support for data center construction, starting in Kansas, Ohio, and Wisconsin ahead of the 2027 state legislative sessions. The push is a direct response to a bipartisan backlash that has made data centers a live midterm-election issue, with local communities objecting to electricity price increases, water consumption, grid strain, and land use. For Lucien's thesis, this is a cost the capex math doesn't usually include: the industry is now spending real money just to keep the political and social license to keep building, on top of the hardware and power bills already stacking up.
Axios · Infrastructure & Economics · Aug 31, 2026
EU Commits €387.8 Million to Its Own Sovereign AI Supercomputer, Built by Europe's Bull
The EuroHPC Joint Undertaking and a six-country consortium (Finland, Czechia, Denmark, Estonia, Norway, and Poland) have selected French vendor Bull to build LUMI-AI, a €387.8 million supercomputer using AMD Instinct MI430X GPUs that will deliver roughly ten times the AI capacity of the current LUMI system when it goes live in Finland in the second half of 2027. Officials framed the project explicitly around "European technological autonomy" — reducing reliance on renting frontier AI compute from U.S. hyperscalers. For Lucien's thesis, it's the government-scale version of the on-prem shift: rather than paying by the token to American cloud providers, a bloc of European governments is choosing to build and own its own AI infrastructure outright.
GlobeNewswire / LUMI Consortium · On-Prem/On-Device Shift · Aug 31, 2026
The Apple Angle (Standing Perspective)
Today's LUMI-AI story is a useful contrast for the standing argument here. A consortium of six European governments needed €388 million, a dedicated liquid-cooling system, and until late 2027 to stand up sovereign AI compute at supercomputer scale. The case for Apple is that a version of that same instinct — own your compute instead of renting someone else's — is available today, on a desk, for a small fraction of the price: 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 machine. Matching that memory footprint with Nvidia's RTX Pro 6000 workstation cards takes five or six of them, roughly $60,000–$75,000 combined, while drawing close to ten times the power.
As frontier-capable open-weight models keep shipping every few weeks, the argument goes that the competition stops being about who trains the smartest model and becomes about who can afford — and physically power — a machine to run it locally, whether that's a nation-state, a company, or an individual. Apple's MLX framework and Thunderbolt 5's RDMA networking already let several 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, notably after Nvidia stripped NVLink pooling out of its consumer and workstation cards. The fair caveat still holds: a Mac Studio serves one user running one model locally, while a data-center-scale system like LUMI-AI is built to serve many researchers and workloads at once — so this is Apple leading a specific, high-value segment of AI compute (sovereign, single-user, high-memory local inference), not replacing large shared infrastructure outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items all sit on the infrastructure and economics side of Lucien's thesis rather than the model-capability side, and together they paint a strained picture of the rented, centralized status quo: SpaceX manufacturing its own turbine parts because power supply can't keep pace with data center growth, a $50 million political campaign needed just to defend the industry's social license to keep building, and a bloc of European governments deciding it would rather spend €388 million building its own AI infrastructure than keep renting compute from U.S. clouds. None of this is proof the capex cycle is unprofitable, but it is evidence of how much friction, cost, and political risk now sits underneath it. The LUMI-AI story in particular sharpens the on-prem half of the thesis at the top of the market — if governments are choosing to own their AI compute rather than rent it, the same logic scales down to the Mac Studio on someone's desk, which is exactly the case the standing Apple Angle keeps making.
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.