Open Source AI — September 27, 2026
Three stories today cut straight to the thesis. Anthropic just signed an $11.6 billion infrastructure deal with Akamai instead of staying dependent on hyperscaler GPU providers—a vote of confidence in third-party clouds.
In Plain English
Three stories today cut straight to the thesis. Anthropic just signed an $11.6 billion infrastructure deal with Akamai instead of staying dependent on hyperscaler GPU providers—a vote of confidence in third-party clouds. DeepSeek crossed $1 billion in annualized revenue this month, proving open-weight models can generate serious money. And token pricing continues its downward march: Anthropic's new Claude Opus 5.5 costs roughly 40% less to run than its predecessor. The pattern repeats: rising cost pressure on the centralized model, widening viability of alternatives.
Anthropic Inks $11.6 Billion Seven-Year Infrastructure Deal with Akamai
Anthropic secured a major cloud infrastructure contract with Akamai, potentially extending to $20 billion and including warrant provisions that give Anthropic up to 5% equity in Akamai. The deal signals a strategic pivot away from pure hyperscaler lock-in, allowing Anthropic to diversify its compute suppliers and reduce dependence on Nvidia-heavy data centers. The warrant structure is notable: it ties Anthropic's long-term incentive to Akamai's success, not just short-term GPU rental arbitrage.
Infrastructure & Economics · AI Adapted · September 26, 2026
DeepSeek Reaches $1 Billion in Annualized Revenue Following Price Increase
The Chinese open-weight model company DeepSeek has crossed $1 billion in annualized revenue following a 2.3x to 4.5x increase in API pricing. The company is now planning a fundraise around $7.5 billion by the end of October. This milestone is significant for the thesis: an open-weight model—freely downloadable and runnable on commodity hardware—is now generating enterprise-scale revenue, validating that the competitive advantage doesn't require proprietary lock-in or exclusive cloud access.
Open-Weight Models · AI Weekly · September 26, 2026
Claude Opus 5.5 Launches at 40% Lower Costs, Further Compressing Margins
Anthropic released Claude Opus 5.5 with input token costs approximately 40% lower than the prior Opus 5 release, while maintaining or improving performance on most benchmarks. OpenAI has moved in a similar direction: its latest models cut pricing by roughly 50%. Lower per-token costs are spurring greater AI usage, which should increase total volumes—but the denominator (cost per inference) keeps falling faster than denominator multiplication can recover, tightening margins for frontier model providers and strengthening the case for local, open-weight alternatives where inference cost approaches zero after a one-time hardware purchase.
Infrastructure & Economics · NewsCord · September 22, 2026
Model Release Pace Accelerates: Claude Opus 5.5, GPT-6 Variants, and Xiaomi MiMo Released in Single Week
Multiple frontier and near-frontier model releases landed in the week of September 22-26, including Anthropic Claude Opus 5.5, OpenAI GPT-6 Luna and GPT-6 Sol, and Xiaomi MiMo-V2.6 open-source variants. The release cadence—multiple capable models per week across three different organizations—reflects deepening competition and commoditization. Each new release applies downward pressure on pricing and capabilities gaps, raising the question of whether frontier model access remains a competitive differentiator or has become a commodity play.
Open-Weight Models · LLM Stats · September 22-26, 2026
The Apple Angle (Standing Perspective)
The Akamai deal and DeepSeek's milestone underline the same industrial shift: centralized cloud GPU provisioning—the model that justifies trillion-dollar data-center buildouts—is looking less like the inevitable endgame and more like one option among several, each with rising operational costs and narrowing margins. Meanwhile, a single Mac Studio with 512GB unified memory (~$9,500) can run an entire open-weight trillion-parameter model at zero per-inference cost, immune to API pricing wars and monthly GPU rental bills. Apple's unified memory architecture, Thunderbolt 5 RDMA for multi-machine networking, and MLX software keep improving quietly. Nvidia's consumer workstation cards no longer have NVLink; its own DGX Spark borrows more from Apple's unified-memory design than the reverse. The honest caveat remains: Mac Studios excel at single-user, high-memory local inference; the gigawatt-scale facilities elsewhere in this brief serve millions concurrently. But as frontier-capable open-weight models keep shipping every few weeks, the competitive bottleneck—"who can run a nearly-as-good model affordably and locally"—may be drifting in Apple's direction almost by accident.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's three items—a major cloud provider stepping into the picture alongside hyperscalers, an open-weight model crossing revenue inflection, and continued per-token price compression—are the industrial mechanics Lucien's thesis expects to see. Centralized capex is straining margins and balance sheets; open-weight model quality keeps improving and generating real revenue; and the cost of inference keeps dropping. The scissors are closing. Whether through Akamai-style third-party clouds, open-weight models on local hardware, or some hybrid, the likelihood of a permanent shift away from pure cloud API dependency keeps rising each week these trends continue.
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.