Open Source AI — September 13, 2026
Four stories today, split cleanly along Lucien's thesis.
In Plain English: Four stories today, split cleanly along Lucien's thesis. On the centralized-buildout side: the Pentagon is reportedly in talks to lend AI cloud startup Fluidstack $5 billion to shore up America's data-center supply chain, and Nvidia signed eight Australian partners to build up to 2 gigawatts of new AI capacity by 2027. On the open-weight side: Cohere released a translation model that beats DeepL and Google Translate across 50 languages but gated commercial use behind a paid license despite marketing it around "sovereign AI," while Singapore's Agnes AI open-sourced a 33-billion-parameter model, under a fully permissive license, that runs on a single GPU even as its flagship hosted version already matches DeepSeek's pricing and speed. Today's Apple Angle ties the shrinking hardware footprint of new open models back to the Mac Studio case.
Pentagon in Talks to Lend AI Cloud Startup Fluidstack $5 Billion to Shore Up the US Data-Center Supply Chain
The Wall Street Journal reports the U.S. Department of Defense is negotiating a roughly $5 billion loan to Fluidstack, a young AI cloud-computing provider, aimed at strengthening the domestic data-center supply chain that underpins the current AI buildout. Reuters said it could not independently verify the report, but multiple outlets have since corroborated the talks, with the loan reportedly structured to help Fluidstack secure the chips and facilities needed to compete with better-capitalized rivals. The move would mark one of the more direct examples yet of a government stepping in to backstop private AI infrastructure financing. For Lucien's thesis, this is prediction #1 showing up in an unusually direct form: when a sovereign government has to consider lending billions to keep one data-center operator afloat, it raises the same question the thesis keeps asking about whether today's centralized buildout can stand on its own economics.
Infrastructure & Economics · Reuters (via Yahoo Finance) · Sep 10, 2026
Nvidia Signs Eight Australian Partners for Up to 2 Gigawatts of New AI Data-Center Capacity by 2027
Nvidia announced a coordinated buildout with eight Australian data-center and cloud partners - including Firmus, Sharon AI, IREN, NEXTDC, AirTrunk, CDC, ResetData, and Latitude.sh - to bring as much as 2 gigawatts of new AI computing capacity online by 2027, more than doubling the country's existing AI-related power load. Sharon AI alone plans to deploy up to 68,000 Nvidia GPUs, while IREN is building an 800-megawatt campus in South Australia, with the whole effort billed as a "sovereign AI" hub running on renewable power. No cost figures were disclosed, but a buildout of this scale implies billions of dollars in committed capital across the partner group. For Lucien's thesis, this is prediction #1's other side showing up again: a fresh multi-gigawatt commitment to centralized, GPU-dense infrastructure, arriving the same week open-weight models kept shrinking toward hardware an individual buyer can actually own.
Infrastructure & Economics · NVIDIA Newsroom (company announcement) · Sep 9, 2026
Cohere's New Translation Model Beats DeepL and Google - But Its "Open Weights" Are Gated Behind a Commercial License
Cohere released North Small Translate 1.0, a 218-billion-parameter (25B active) mixture-of-experts translation model covering more than 50 languages, which the company says outperforms both DeepL and Google Translate. The weights are posted on Hugging Face, but only under a non-commercial license - any company that wants to actually deploy it in production has to buy a commercial license and run it through Cohere's own hosted inference platform, a notable reversal from Cohere's earlier fully-open, Apache-licensed releases. Commentators have flagged the tension with Cohere's own "sovereign AI" pitch: a company can inspect the weights, but can't fork them, build a product on them, or keep using them if the terms change at renewal. For Lucien's thesis, this is a useful complication to predictions #2 and #3: capability keeps closing the gap with proprietary rivals, but the license terms attached to an "open" release increasingly decide whether that capability actually buys real independence from a vendor.
Open-Weight Models · The New Stack · Sep 11, 2026
Singapore's Agnes AI Open-Sources a 33B Model That Runs on One GPU, While Its Flagship Matches DeepSeek at a Fraction of the Price
Singapore-based Sapiens AI, trading as Agnes AI, released Agnes-3.0-Flash-Preview, a 33-billion-parameter multimodal checkpoint under a fully permissive Apache 2.0 license that scores 85.05 on GPQA Diamond and fits on a single Nvidia H100 or H200 GPU. Separately, the company's flagship hosted Agnes 3.0 Flash model matches DeepSeek V4 Pro's intelligence score on Artificial Analysis's index while generating responses more than three times faster and charging a small fraction of DeepSeek's per-token price, though it still trails on raw factual recall. For Lucien's thesis, this is prediction #2 in miniature: a well-funded but far-from-household-name lab is now shipping frontier-adjacent capability that fits on hardware a single workstation can hold, not a data center, at costs that undercut even other cost-focused rivals.
Open-Weight Models · Hugging Face / Artificial Analysis · Sep 12, 2026
The Apple Angle (Standing Perspective)
Today's clearest thread back to this recurring argument is Agnes AI's new 33-billion-parameter checkpoint, which needs nothing more exotic than a single Nvidia H100 to run - a small but telling reminder that the hardware required to hold a genuinely competent model keeps shrinking even as the model itself gets more capable. That is the same logic behind the case that Apple, not Nvidia, may be quietly best positioned for a shift toward local AI. Apple's Mac Studio uses a unified memory architecture shared between CPU and GPU, and a maxed-out configuration with 512GB of that memory - around $9,500 - can hold an entire trillion-parameter open-weight model at once, dozens of times larger than what today's Agnes checkpoint needs. Matching that capacity with Nvidia's RTX Pro 6000 workstation cards takes five to six of them, roughly $60,000-$75,000 combined, while drawing something like ten times the power.
Cohere's licensing reversal sharpens why that capacity matters on its own, not just on paper: if "open" weights can be gated, revoked, or repriced at a vendor's discretion, the appeal of hardware you outright own - capable of running whichever permissively-licensed model wins this month, with no renewal risk - only grows. As frontier-capable open-weight models keep shipping every few weeks, the competitive question keeps drifting away from "which model is smartest" and toward "who can afford the hardware to run a nearly-as-good one locally, on terms nobody else controls." Apple's MLX software and Thunderbolt 5's RDMA networking already let multiple Mac Studios pool memory for bigger jobs, while Nvidia has stripped NVLink from its consumer workstation cards and its own DGX Spark increasingly borrows Apple's unified-memory playbook. The honest caveat still applies: a Mac Studio serves one person running one model at a time, while the 2 gigawatts Nvidia just lined up in Australia - or the data center a $5 billion Pentagon loan would help keep standing - are built for enormous concurrent demand. Apple looks well 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 four items sit on both sides of Lucien's thesis at once. The Pentagon's reported willingness to lend $5 billion to a single AI cloud startup, and Nvidia's fresh 2-gigawatt commitment in Australia, both show enormous capital - public and private - still flowing toward the centralized, GPU-dense way of building AI, exactly the model prediction #1 expects to come under strain. But the open-weight side kept moving too: Agnes AI shipped a 33-billion-parameter model that runs on a single GPU while its flagship hosted version already undercuts DeepSeek on price and speed, showing real capability now fits on hardware far short of a data center. Cohere's licensing reversal is the necessary caveat to that story - "open weights" only deliver the independence prediction #3 expects if the license actually lets you keep them. None of this resolves the argument on its own, but it captures the exact tension the thesis is tracking: capital keeps committing to the current centralized buildout even as capable open models keep shrinking toward hardware individuals and enterprises can own outright - precisely the setup the standing Apple Angle argues Apple is already positioned to benefit from.
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.