Open Source AI — September 22, 2026
Four stories today, and they pull in different directions. New figures show America now spends more building data centers than building homes - and a ratings agency warns the six biggest AI spenders could run cash-flow negative through 2027.
In Plain English: Four stories today, and they pull in different directions. New figures show America now spends more building data centers than building homes - and a ratings agency warns the six biggest AI spenders could run cash-flow negative through 2027. Meanwhile actual compute prices are going up, not down: GPU rental and memory rates rose this week even as some AI infrastructure debt trades below face value, a classic stress signal. A US government lab published its first public benchmark of a Chinese open-weight model and found real progress, though it still trails on the hardest tasks. And independent reviewers got their hands on Apple's new M5 Ultra Mac Studio and called it "the dream machine" for running big AI models at home. Today's Apple Angle ties that last point back to the first three.
US Data-Center Spending Overtakes Home Building, and S&P Warns Hyperscaler Cash Flow Could Go Negative Through 2027
New national accounts data shows real spending on information-processing equipment hit $752 billion in Q2 2026 - up 51% from the prior comparison period - while real residential housing investment fell to $748 billion, down 18% from its 2021 peak, meaning data-center-driven capex has now overtaken home building in the US economy. The same reporting cites S&P Global's warning that "the industry's capex is growing faster than revenue," with the ratings firm projecting that the six major hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX) will collectively post negative operating cash flow in both 2026 and 2027, with combined capex projected to reach $1.3 trillion in 2027 alone. S&P pegs 2028 as the likely inflection point where revenue growth finally catches up to capex. For Lucien's thesis, this is prediction #1 given a macro anchor: it's no longer just one company's balance sheet under strain, but a reallocation of national capital large enough to show up next to housing in the GDP accounts, with a credit-rating agency now attaching a two-year timeline to when the math is supposed to turn around.
Infrastructure & Economics · Fortune · Sep 21, 2026
Compute Just Got More Expensive: GPU and Memory Prices Rise This Week Even as Some AI Infrastructure Debt Trades Below Face Value
A weekly infrastructure-markets report finds cloud provider Nebius raising GPU rental rates 17-21% effective October 1, with memory pricing up roughly 41% - a direct contradiction of the "AI compute keeps getting cheaper" narrative. The same report flags stress signals elsewhere in the financing chain: Oracle's $18 billion Project Jupiter debt is trading at just 89-91 cents on the dollar amid distribution and permitting delays, and AI cloud company Nscale has filed for an IPO showing $103 billion in contracted value against only $140.6 million of first-half revenue. CoreWeave, meanwhile, priced $3.7 billion in new convertible notes this week to keep funding buildout. For Lucien's thesis, this is the sharpest data point yet for prediction #1's specific claim - that per-token and per-GPU-hour costs rise rather than fall - landing in the same week that financing markets are starting to price in real doubt about whether the underlying contracts will pay out.
Infrastructure & Economics · North American AI Compute Infra Report (Substack) · Sep 21, 2026
A US Government Lab Benchmarked a Chinese Open-Weight Model for the First Time - and Found Real, if Partial, Progress
NIST's Center for AI Standards and Innovation (CAISI) published its first public cyber-capability assessment of a Chinese open-weight model, concluding that Z.ai's GLM-5.3 is "the most cyber-capable open-weight model released to date." On CAISI's own benchmarks it still trails current US frontier systems by roughly four months in aggregate - scoring 40.4% versus 90.2% on SEC-Bench Pro and 61.1% versus 100% on ExploitBench, for instance - so the gap in this specific, security-sensitive domain remains real. But the fact that a US standards body is now running head-to-head evaluations of a freely downloadable Chinese model at all is itself notable. For Lucien's thesis, this is prediction #2 getting an unusually credible referee: not a lab's own marketing claim about how close open weights have gotten, but a government assessment putting a specific, if sobering, number on both the progress and the remaining distance.
Open-Weight Models · NIST/CAISI · Sep 17, 2026
First Reviews of Apple's M5 Ultra Mac Studio Call It "The Dream Machine" for Running AI Models at Home
Independent reviewers published the first hands-on tests of Apple's new M5 Ultra Mac Studio this week, and the verdict leans heavily on its fitness for local AI: prompt processing jumped to 2,057-2,771 tokens per second versus the prior M3 Ultra's 861-1,112 tokens per second (roughly 150% faster), on a chip with memory bandwidth up 50% to 1.2 TB/s. One reviewer directly compared it to an Nvidia RTX 5090: the Nvidia card processes prompts faster in isolation, but its 32GB of VRAM can't hold the large contexts the Mac's unified memory handles comfortably, making the Mac Studio the more usable machine for sustained, real-world local AI work. The tested unit carried 256GB of memory; a 512GB configuration - the one this brief's standing Apple Angle section keeps referencing - ships in October 2026. For Lucien's thesis, this is prediction #3 with hands-on numbers attached, from reviewers with no stake in the outcome, rather than Apple's own marketing copy.
On-Prem/On-Device Shift · MacStories · Sep 21, 2026
The Apple Angle (Standing Perspective)
Every story above is really about the same tension: the industry's centralized, rented model of AI compute is looking more expensive and more fragile by the week, while the tools to own a capable model outright keep getting better and cheaper to use. 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 a single desktop machine, immune to the GPU rate hikes and financing stress showing up elsewhere in this brief.
Reaching that same memory footprint on Nvidia's own workstation hardware takes five to six RTX Pro 6000 cards, somewhere around $60,000 to $75,000 combined, while drawing on the order of ten times the power. New frontier-capable open-weight models keep shipping every few weeks, which keeps nudging the real competitive question away from "who trained the smartest model" and toward "who can actually afford to own the box that runs a nearly-as-good one, locally, with no metering and no cloud provider in the loop." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory into one larger effective machine, Nvidia has stripped NVLink out of its own consumer workstation cards, and its DGX Spark borrows more from Apple's unified-memory playbook than the reverse. The honest caveat still applies 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 at once - 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 cut against the comfortable version of the AI story from two directions at once - national capex now outrunning home building with cash flow projected negative through 2027, and actual compute prices rising rather than falling this week - while open-weight models keep closing ground (even in a hard domain like cybersecurity) and independently-reviewed local hardware keeps getting more capable. That combination of rising centralized cost and improving cheap alternatives is exactly the scissors this brief's thesis expects to eventually squeeze commercial API margins, and it's the same scissors the standing Apple Angle keeps pointing back to.
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.