Open Source AI — September 4, 2026
Three stories today, all touching different corners of the thesis.
In Plain English: Three stories today, all touching different corners of the thesis. Microsoft will finally start telling investors exactly how much money Azure — its cloud/AI business — brings in each quarter, instead of vague growth percentages, a rare bit of transparency demanded by investors nervous about whether AI spending is paying off. Meta released a new closed model that matches top rivals at a fraction of the price, and Mark Zuckerberg publicly promised an open-weight (freely downloadable) version is coming soon, plus another new model in the pipeline — direct evidence the "gap is closing" prediction keeps playing out, this time from an American lab. And Google cut prices again on its latest fast model while adding a cybersecurity-focused version, continuing the price war that's squeezing everyone's margins. Nothing today is about hardware directly, but the Apple Angle below explains why cheaper, more open models make the on-device case stronger, not weaker.
Microsoft to Disclose Azure Quarterly Revenue for the First Time
Microsoft is consolidating from three operating segments to two and will begin disclosing Azure's actual quarterly revenue — not just year-over-year growth percentages — starting with fiscal Q1 2027 results. Azure grew 42% to $29.4 billion last quarter under the old growth-only disclosure regime. The move follows years of Wall Street pressure for genuine cloud transparency and lands amid heightened investor scrutiny over whether AI infrastructure capex is translating into disclosed, verifiable returns. For Lucien's thesis, this is a crack in the opacity that has made the profitability question hard to check from the outside — once real Azure revenue is public every quarter, it becomes much easier to see whether AI infrastructure spending is actually paying for itself.
CNBC · Infrastructure & Economics · Sep 2, 2026
Meta Ships Muse Spark 1.3, Teases Open-Weight Release and New "Watermelon" Model
Meta released Muse Spark 1.3, a closed, API-only model that matches GPT-5.6 Sol and Claude Opus 5 on several benchmarks (88.8 on Terminal-Bench 2.1, 75.4 on DeepSWE) at aggressively low pricing ($0.10/$0.20 per million input/output tokens). Mark Zuckerberg posted that an open-weight version of Muse Spark is "coming soon," alongside a new frontier model codenamed "Watermelon," continuing Meta's stated strategy of keeping an American open-weight option competitive with China's DeepSeek and Moonshot's Kimi. For Lucien's thesis, this is a major Western lab signaling, in public, that it intends to keep releasing frontier-adjacent capability into the open-weight ecosystem — the exact gap-closing dynamic the thesis depends on, this time from Meta rather than a Chinese lab.
The Register · Open-Weight Models · Sep 2, 2026
Google Cuts Prices Again With Gemini 3.8 Flash, Adds Cybersecurity Variant
Google launched Gemini 3.8 Flash at the same aggressive introductory pricing as its predecessor ($0.75/$3.75 per million input/output tokens through year-end), positioned as beating larger, pricier frontier models on coding and agentic benchmarks at a fraction of the cost. A restricted variant, Gemini 3.8 Flash Cyber, went out to vetted security teams through a new "Fairwind" early-access program, reportedly producing 2.6x more correct vulnerability patches than commercial rivals. For Lucien's thesis, another major lab racing prices toward the floor is more evidence the commercial API business is being squeezed from both directions at once — cheaper open-weight competition from below, and internal price wars between the proprietary labs themselves.
Help Net Security · Infrastructure & Economics · Sep 3, 2026
The Apple Angle (Standing Perspective)
Today's stories point at the demand side of the equation — Meta promising an open-weight frontier model, Google racing prices toward zero, Microsoft finally opening its books on Azure — but the standing argument here has always been about the supply side: once a capable model is free to download, what do you actually run it on? Apple's quiet answer is the Mac Studio. Configured with up to 512GB of unified memory shared between CPU and GPU, a single machine costing around $9,500 can hold a trillion-parameter open-weight model entirely in memory — the kind of model Meta is now hinting it will give away for free. 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 on the order of ten times the power.
As frontier-capable open-weight models keep shipping every few weeks — soon, it seems, including one from Meta itself — the bottleneck shifts from "who trains the best model" to "who can afford, and physically power, a machine to run it locally," and that's a race Apple's unified-memory design is winning almost by default, largely unremarked upon. Apple's MLX framework and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for even larger models, and Nvidia's own DGX Spark — a small unified-memory box of its own — is a tacit admission that this is the right shape for local inference, notably after Nvidia pulled NVLink pooling from its consumer and workstation cards. The honest caveat remains: a Mac Studio serves one user running one model, while the data centers Microsoft is now disclosing revenue for serve many thousands of users at once — Apple looks positioned to dominate a specific, high-value slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items span both halves of Lucien's thesis. Meta's open-weight tease is a direct, first-party signal that the capability gap the thesis bets on keeps closing — this time from a major Western lab rather than only Chinese ones. Google's latest price cuts and Microsoft's new Azure transparency both bear on the profitability question: prices keep falling toward a floor that makes the rented-API business harder to sustain, and the industry's biggest cloud provider is now under enough investor pressure to start showing its real numbers. None of it proves the on-prem transition has arrived, but each piece nudges the underlying economics in the direction the thesis expects — and the standing Apple Angle argues that when it does arrive, the on-device winner may already be sitting on shelves as Mac Studios.
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.