Open Source AI — August 23, 2026
Two stories today, one on each side of the thesis.
In Plain English: Two stories today, one on each side of the thesis. A UBS forecast reported by 24/7 Wall St shows Amazon, Alphabet, and Microsoft are on track to spend about 102% of their combined cloud revenue on AI infrastructure this year — meaning virtually every dollar cloud computing brings in gets plowed straight back into new data centers, with total hyperscaler capex projected to triple over the next three years. Meanwhile, Alibaba's tiny 27-billion-parameter Qwen3.8 model is matching GPT-5.6 and beating Claude Opus 4.8 on agentic tasks despite being small enough to run on a personal computer — another data point that "good enough" open models keep getting cheaper to run locally, just as commercial infrastructure spending gets harder to justify.
UBS: Hyperscalers Set to Spend 102% of Cloud Revenue on AI Capex This Year
UBS projects that Amazon, Alphabet, and Microsoft will collectively spend roughly 102% of their 2026 cloud revenue on capital expenditures — essentially recycling all cloud income, and then some, straight back into AI infrastructure. The ratio is expected to ease only slightly, to 99% in 2027 and 94% in 2028, even as absolute spending keeps climbing: UBS puts cumulative hyperscaler capex at roughly $4.1 trillion from 2026-2028, more than triple the $1.3 trillion spent over the prior six years combined. The article's core caution is blunt — infrastructure alone doesn't guarantee returns; capacity has to stay utilized and AI revenue has to eventually cover depreciation, financing, and power costs, and some of today's spending may take years to reach full utilization.
24/7 Wall St · Infrastructure & Economics · Aug 22, 2026
Alibaba's 27-Billion-Parameter Qwen3.8 Matches GPT-5.6, Beats Claude Opus 4.8 — and Fits on a Personal Computer
Alibaba's Qwen3.8-27B, a comparatively tiny 27-billion-parameter open-weight model, is performing on par with OpenAI's GPT-5.6 Luna and outperforming Anthropic's Claude Opus 4.8 on the Agentic Index, a benchmark for AI agent workflows. It also comes close to much larger open-weight rivals — DeepSeek-V4-Pro (1.7 trillion parameters) and Zhipu's GLM-5.2 (753 billion parameters) — despite being a fraction of their size. Its small footprint means it's practical to run entirely on a personal computer through tools like LM Studio, rather than requiring a data center or an API subscription — a concrete example of the capability-per-parameter gains that keep making local, self-hosted inference more viable for ordinary hardware.
South China Morning Post · Open-Weight Models · Aug 18, 2026
The Apple Angle (Standing Perspective)
Set against today's UBS numbers, it's worth restating the case some commentators keep making about who's actually best positioned once companies start running large models themselves instead of renting them by the token — and their answer isn't Nvidia, it's Apple. The mechanism is unified memory: a single Mac Studio, priced around $9,500, can be configured with up to 512GB, enough to hold a trillion-parameter open-weight model on one machine without spreading it across a server rack. Matching that with Nvidia's RTX Pro 6000 workstation cards would take five or six of them — on the order of $60,000-$75,000 combined — while drawing roughly ten times the electricity.
The argument keeps compounding with each open-weight release, today's Qwen3.8-27B item included: as frontier-caliber open models from Alibaba, DeepSeek, GLM, and Kimi keep shipping every few weeks, the limiting factor stops being "who has the smartest model" and becomes "who can afford the hardware to run it locally." Multi-Mac clusters using MLX and Thunderbolt 5's RDMA networking can already pool memory across several Studios to handle even larger models, and Nvidia's own DGX Spark essentially concedes the point by adopting a unified-memory design of its own — notably after Nvidia stripped NVLink pooling out of its consumer and workstation cards. The fair caveat still applies: a Mac Studio serves one user running one model locally, while a data-center GPU cluster serves many people at once — so this looks like Apple leading a specific, valuable segment, not the entire AI compute market.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items sit on opposite sides of the same coin. UBS's 102%-of-cloud-revenue capex figure is exactly the kind of ratio that raises the profitability question at the heart of Lucien's thesis — infrastructure spending that outpaces the revenue meant to justify it, with returns pushed years into the future. At the same time, a 27-billion-parameter open-weight model matching or beating frontier closed models on agentic tasks, while fitting comfortably on a personal computer, shows the capability gap continuing to close from the bottom up — not just at the trillion-parameter frontier, but at sizes ordinary hardware can already handle. Set against the standing Apple argument, the throughline holds: as commercial AI infrastructure spending gets harder to square with returns, open-weight models keep getting more capable at smaller sizes — and the hardware to run them locally keeps looking more attractive by comparison.
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.