Open Source AI — September 11, 2026
Four stories today, and cost is the thread running through all of them. DeepSeek made its cheaper new model permanent and cut prices further, and the news knocked down shares of two rival Chinese AI labs.
In Plain English: Four stories today, and cost is the thread running through all of them. DeepSeek made its cheaper new model permanent and cut prices further, and the news knocked down shares of two rival Chinese AI labs. A startup called Cognition released a coding AI built on a free Chinese model that nearly matches the priciest options for a quarter of the price. France's Mistral partnered with data firm Cloudera so companies can run Mistral's free models entirely on their own servers, even fully disconnected from the internet. And a Deloitte survey found most corporate finance chiefs expect AI costs to keep climbing through 2027 even as they keep funding it. Today's Apple Angle ties the rising-cost side and the cheaper-open-model side together.
DeepSeek Makes V4.1 Flash Permanent, Cuts Prices Again, and Rattles Chinese AI Stocks
(Follow-up: yesterday's brief covered V4.1 Flash as a one-day test beta; today DeepSeek made it official.)
After testing it for a single day, DeepSeek made "V4.1 Flash" permanent: a free, MIT-licensed, 552-billion-parameter model that will fully replace the older V4-Pro starting September 14, when all V4-Pro traffic gets automatically rerouted to the cheaper Flash tier. Off-peak output pricing dropped as much as 32% to $0.60 per million tokens, reversing an August price increase, with cached input priced at a fraction of a cent. The launch knocked more than 8% off Hong Kong-listed shares of rivals MiniMax and Z.ai and over 2% off Alibaba, with the timing coinciding with DeepSeek's own preparations for a Shanghai stock listing. For Lucien's thesis, this is prediction #2 compounding in real time: a free, downloadable model got cheaper and more capable in the same week, and the market treated it as bad news for two other Chinese AI labs and priced it in immediately.
The Next Web (via Bloomberg reporting) · Open-Weight Models · Sep 10, 2026
Startup's New Coding AI, Built on a Free Chinese Model, Matches Frontier Performance at a Quarter of the Cost
AI coding startup Cognition released "SWE-2," a new coding agent built on top of Moonshot's open-weight Kimi K3 — a 2.8-trillion-parameter model that had already been heavily trained for agentic coding — rather than on a closed frontier model. On Cognition's own benchmark, SWE-2 scores within one point of Anthropic's Claude Fable 5.1 while costing 64% less to run, comes within a few points of OpenAI's GPT-6 Astra at roughly a quarter of the price, and cuts the number of steps needed per coding task by more than half compared to its predecessor. For Lucien's thesis, this is predictions #2 and #3 combining directly inside a commercial product: a well-funded AI application company chose to build its flagship product on an open-weight foundation instead of a closed API, and is using that choice as its main pitch to customers on cost.
Cognition (company blog) · Open-Weight Models · Sep 10, 2026
Mistral Partners With Cloudera So Enterprises Can Run Its Open Models Fully On-Prem — Even Air-Gapped
French AI lab Mistral, fresh off its $3.6 billion sovereign-AI funding round, announced a partnership with data platform Cloudera that lets enterprises deploy and fine-tune Mistral's open-weight models directly inside Cloudera's hybrid data environment — spanning private cloud, on-premises servers, and fully air-gapped networks with no external connection at all. Cloudera, which manages roughly 30 exabytes of customer data across regulated industries, is framing the deal around letting companies "own" the resulting intelligence outright rather than sending their data to an outside API. For Lucien's thesis, this is prediction #3 in close to its purest form: an open-weight model maker and an enterprise data platform jointly building the plumbing for customers who want to run frontier-adjacent AI entirely inside their own walls, with no cloud dependency required.
Mistral (company announcement) · On-Prem/On-Device Shift · Sep 10, 2026
Deloitte Survey: 60% of Finance Chiefs Expect AI Costs to Keep Rising Through 2027 — But They're Funding It Anyway
A Deloitte survey of 1,434 finance leaders found that 60% expect AI costs and operational complexity to rise substantially through 2027 and say they'll need more sophisticated cost-management practices just to keep up, while a smaller 35% believe expenses will stay modest and are sticking with their current approach. Despite that expected cost pressure, 43% of the same finance leaders named embedding AI into operations as one of their top three organizational priorities. For Lucien's thesis, this is a clean read on prediction #1 from the buyer's side rather than the vendor's: the people who actually approve AI budgets expect the bills to keep climbing rather than falling as usage scales — the opposite of what would need to happen for today's rented-compute economics to comfortably work out.
ESG Dive (Deloitte survey) · Infrastructure & Economics · Sep 10, 2026
The Apple Angle (Standing Perspective)
Today's four stories all lean on the same fault line: the cost of renting AI keeps climbing, according to the finance chiefs Deloitte surveyed, at exactly the moment the free alternative keeps getting stronger and cheaper, according to DeepSeek, Cognition, and Mistral. That's the setup this recurring perspective keeps returning to. Apple's Mac Studio line, built around unified memory shared between CPU and GPU, offers a fixed-cost way to sidestep both problems at once: a maxed-out configuration with up to 512GB of that memory costs roughly $9,500 and can hold an entire trillion-parameter open-weight model at once — the same class of model DeepSeek, GLM, and Kimi now ship every few weeks. Matching that memory capacity with Nvidia's RTX Pro 6000 workstation cards takes five or six of them, somewhere in the $60,000–$75,000 range, while pulling roughly ten times the electricity.
Today's items sharpen why that matters. Cognition just built its flagship commercial product on an open-weight base specifically to escape closed-model pricing, and Mistral is now selling the ability to run its own open weights fully air-gapped — both companies making the same bet that the real bottleneck is affordable hardware to run open weights, not the weights themselves. Apple's MLX software and Thunderbolt 5's RDMA networking already let several Mac Studios pool memory together for even larger models, which is part of why the competitive question keeps drifting from "whose model is smartest" toward "who can actually afford to run a nearly-as-smart model without a data center behind it." The usual caveat still holds: a Mac Studio serves one person running one model at a time, while the infrastructure behind a commitment like a $50 billion compute deal is built for enormous concurrent demand — Apple looks well positioned to own a specific, high-value corner of AI compute, not to replace the data center wholesale.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's four items land on both sides of Lucien's thesis at once. DeepSeek's price cut and Cognition's cost-focused coding agent both show the open-weight side of the ledger getting cheaper and more capable in the same week, while Deloitte's finance-leader survey shows the buyers of AI compute bracing for costs to rise rather than fall — exactly the scissors that would make renting frontier AI less attractive over time. Mistral's air-gapped partnership with Cloudera shows a serious enterprise vendor building the on-prem plumbing for that shift today, not hypothetically. None of this alone confirms the transition to on-prem and on-device AI has arrived, but together the four stories describe the same pressure building from multiple directions — and the standing Apple Angle argues that whoever needs an affordable, fixed-cost way to run tomorrow's open-weight models locally will keep finding Apple's Mac Studio line quietly ahead of the curve.
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.