Open Source AI — September 17, 2026
Three stories today, all circling the same question: who pays for AI's electricity? Meta revealed a new in-house chip it says will beat Nvidia on cost and power when it ships in 2027.
In Plain English: Three stories today, all circling the same question: who pays for AI's electricity? Meta revealed a new in-house chip it says will beat Nvidia on cost and power when it ships in 2027. Anthropic signed a $32 billion deal for one of Australia's largest data centers - built purely to answer Claude questions, not train new models. And in Washington, the House voted 417-3 to stop AI data centers from quietly pushing their power costs onto ordinary households' electric bills. No open-weight model releases or on-prem adoption stories cleared the bar for "genuinely new" today - so today's brief stays entirely on the infrastructure and economics side of the thesis. Today's Apple Angle asks why a technology now straining national power grids is exactly the problem Apple's approach was built to sidestep.
Meta's Next In-House AI Chip Is Built to Beat Nvidia on Cost and Power, Not Just Speed
Meta disclosed plans to deploy its MTIA 450 chip, code-named Arke, across its data centers starting in the first half of 2027, with the company saying it will deliver better performance per watt and per dollar than Nvidia's current lineup for running AI models (as opposed to training them). It's the latest move in a broader hyperscaler pattern - Meta alongside Google, Amazon, and Microsoft - of building proprietary inference silicon rather than continuing to pay Nvidia's margins on every GPU. Meta hasn't published independent benchmarks or pricing yet, so the cost claim is still the company's own framing rather than a verified figure. For Lucien's thesis, this is a hedge against prediction #1 coming from inside the industry itself: one of the biggest AI spenders is betting that owning its own inference chips, not just renting Nvidia's, is how it keeps per-token costs from spiraling as usage scales.
Infrastructure & Economics · Quartz · Sep 16, 2026
Anthropic Signs $32 Billion Deal for a Data Center the Size of a Small Australian Town
Anthropic will build a roughly AU$32 billion (about US$21 billion) data center on 725 hectares near Dalby in Queensland, Australia, with construction expected to take four to six years and first use targeted for 2027. Unlike many mega-facilities announced this year, this one is explicitly for inference - running Claude to answer everyday user questions - rather than training new models, and its power draw is expected to rival 1.5 million average Australian households. Queensland's premier called it a "major win" for jobs and the local grid, while a local opposition group raised concerns the facility could run on gas-fired power. For Lucien's thesis, this puts a concrete price tag on the abstraction behind Anthropic's $517 billion global compute commitment (covered in this brief on September 16): even one inference-only facility, in one mid-sized country, now costs more than the GDP of many nations - a follow-up data point on the same story, not a new one.
Infrastructure & Economics · ABC News (Australia) · Sep 16, 2026 (follow-up to Sep 16 coverage of Anthropic's $517B compute commitment)
Washington and Silicon Valley Both Moved on AI's Power Problem in the Same 24 Hours
The US House passed the Ratepayer Protection Act 417-3 - a rare near-unanimous vote - requiring state regulators to make large data centers (100+ megawatts) cover the full cost of grid upgrades built to serve them, so ordinary households don't quietly absorb the bill; Rep. Bob Latta put it plainly: "American families should not have to pay higher electricity bills so large technology companies can build and operate data centers." The same day, Nvidia, Google, Anthropic, and several utilities launched the AI Energy Management Alliance, aimed at making data centers flexible grid participants that can shift or shed load at peak times instead of running as fixed, always-on draws. Both moves treat electricity, not chips, as the binding constraint on how much AI infrastructure can actually get built. For Lucien's thesis, this is the power side of prediction #1 made literal: industry and regulators are now openly negotiating over which power-hungry AI hardware gets to exist at all - exactly the friction a much lower-power, single-machine alternative sidesteps entirely.
Infrastructure & Economics · Unite.AI / NVIDIA Blog · Sep 16, 2026
The Apple Angle (Standing Perspective)
Today's items make the power argument literal rather than abstract. Congress just voted 417-3 to stop data centers from pushing their electricity bills onto ordinary households, Nvidia and Google are now co-founding an alliance just to get data centers to negotiate with the power grid rather than simply draw from it, and Anthropic's newest single facility in Australia will draw as much power as 1.5 million homes. Apple's answer to all of that friction is to not need it in the first place: a Mac Studio maxed out with 512GB of unified memory, at around $9,500, can hold an entire trillion-parameter open-weight model in memory on its own, drawing something like 200-270 watts - closer to a couple of hair dryers than a small city.
Matching that memory capacity on Nvidia's own workstation-grade hardware takes five to six RTX Pro 6000 cards, on the order of $60,000-$75,000 combined, while drawing roughly ten times the power - the same order of power draw now being fought over in Congress and pooled into flexible grid alliances. As frontier-capable open-weight models keep shipping every few weeks, the practical question keeps drifting from "who trained the smartest model" toward "who can afford hardware simple enough that it never has to show up in a grid-capacity negotiation." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for bigger jobs, Nvidia has stripped NVLink from its own consumer workstation cards, and its DGX Spark increasingly borrows Apple's unified-memory approach rather than the reverse. The honest caveat still holds: a Mac Studio serves one person running one model at a time, while the gigawatt-scale facilities behind today's stories are built for enormous concurrent demand across millions of users - Apple looks positioned to quietly own a valuable, low-friction corner of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items - Meta's cost-driven push into its own chips, Anthropic's $32 billion single facility, and Washington's near-unanimous vote to keep AI's power costs off ordinary electric bills - all sharpen prediction #1's central tension: the industry is now openly negotiating, in public, over who pays for AI's electricity, even as it keeps spending at a pace that assumes the answer will sort itself out. That exact negotiation is the friction the standing Apple Angle argues a low-power, locally-owned alternative was built to avoid.
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.