Open Source AI — November 1, 2026 — Lucien Engelen
Home About Blog Books Booking Lucien
Blog · 1 November 2026

Open Source AI — November 1, 2026

Alphabet fell 12% in one day, its worst session on record and about $500B of market value, after reporting a second straight negative free-cash-flow quarter (-$9.1B) on $51.4B of quarterly capex (TechCrunch, Oct 30).

Hook of the Day

Alphabet fell 12% in one day, its worst session on record and about $500B of market value, after reporting a second straight negative free-cash-flow quarter (-$9.1B) on $51.4B of quarterly capex (TechCrunch, Oct 30).

In Plain English

Big Tech is spending more than $740B this year on AI data centers, and investors have started to push back because cash is running out faster than new revenue arrives. At the same time, companies that pay for AI by the token are looking for cheaper models, and small open models that run on a single machine keep getting better. Both trends push in the direction of running more AI yourself.

Section A: Open-Weight Models

Today's items show open weights still trailing the best closed models on third-party indices, but at a gap measured in months, and small open models are being tuned for local use. Both points matter for the thesis: the gap is real, yet "good enough to run yourself" keeps getting cheaper to reach.

Open-vs-closed gap: best open model scores 32 vs 58 for the best closed (secondary analysis)

A consultancy roundup updated Oct 28 cites the Artificial Analysis Intelligence Index v5.0: the best open-weight models (DeepSeek V4.1-Flash and GLM-5.3) score 32, against 58 for the best closed model (Claude Opus 5.5). It also cites Epoch AI for an average lag of about four months, and says Kimi K3.1 Pro reached its API on Oct 23 with weights announced for Nov 4. This is a secondary source summarizing other sites, so verify the figures against the originals before relying on them; the Nov 4 weights date is unconfirmed.

Continue Reading →

Slash Digital Lab · Analysis (secondary) · Updated Oct 28, 2026

NVIDIA Nemotron 3.5 Bolt E4B: a 4 GB open model for local voice agents

An NVIDIA employee's Hugging Face post (Oct 30) describes building a voice assistant around Nemotron 3.5 Bolt E4B, an open small language model "purpose-built for local conversational agents" with a 4-bit release of about 4 GB. The author reports (vendor-side, not independent) a median 1.70 s from end of speech to audio reply, and says the model has a smaller footprint than Qwen3 30B-A3B and Qwen 3.6 35B-A3B. The post is a project write-up rather than a launch announcement, and it does not state licenses.

Continue Reading →

Hugging Face blog (NVIDIA author) · Open-weight model · Oct 30, 2026

Section B: Infrastructure & Token Economics

Today's items show the capex side of the thesis getting stress-tested by markets, while the token-price side is more mixed: reported prices are falling, not rising, but enterprises say costs still bite. That complicates the "rising price per token" leg of the thesis without removing the profitability pressure on the builders.

Cracks in the AI spending boom: Nasdaq -2.9% as capex tops $740B

TechCrunch reports that Alphabet, Amazon, Meta and Microsoft outlined over $740B of capex this year. On Friday Oct 30 the Nasdaq fell 2.9% and Nvidia fell 4%. Free cash flow was negative at Alphabet (-$9.1B), Meta (-$2.1B) and Amazon (-$42.4B trailing twelve months); only Microsoft was positive ($18.6B). The piece also cites S&P Global Market Intelligence for $326B of debt raised in 2026 by Microsoft, Alphabet, Amazon, Meta and Oracle, triple last year, and relays an FT report that Blackstone warned Anthropic may not get an investment-grade rating after its IPO. Follow-up to earlier earnings coverage, with new market-reaction and debt figures.

Continue Reading →

TechCrunch · Infrastructure economics · Oct 30, 2026

Token costs push enterprises to cheaper models: SAP, Cloudera survey (secondhand)

A newsletter roundup dated Oct 29 cites Reuters (Oct 28) saying SAP is swapping costly AI models for cheaper ones where quality is equal, targeting more than EUR 2 billion in efficiency gains across more than 70 models. It also cites a Cloudera survey (via ITPro) in which 89% of 1,500 IT leaders limited or redesigned AI deployments because of token costs, and an Axios/Ramp index putting corporate token prices near $0.59 per million, about 35% lower since June. Note that falling prices cut against the thesis's "rising token price" premise, while the survey supports its "cost pressure" premise. The roundup is a vendor newsletter; I did not read the Reuters, ITPro or Axios originals, so verify before relying on the figures.

Continue Reading →

jCodeMunch Token Cost Radar (aggregator) · Token economics · Oct 29, 2026

Section C: On-Prem & On-Device Shift

Today's only new item is small but on-theme: local models reaching users through the browser rather than a cloud API. Hardware angles are covered in The Apple Angle below.

Opera One adds a preview that lets Macs run Apple's on-device model in the browser

9to5Mac reports that Opera One users can access a preview feature using local AI models through WebAI, a proposed web standard, including Apple's foundation model on Mac. I could only read the feed abstract, not the full article. It is a small sign of on-device inference becoming a default web capability.

Continue Reading →

9to5Mac · On-device · Oct 27, 2026

The Apple Angle

Some commentators argue that Apple, not Nvidia, is best placed for the on-prem and on-device shift. A single Mac Studio (about $9,500) with unified memory can hold very large open-weight models that would otherwise need five or six Nvidia RTX Pro 6000 cards (about $60,000-$75,000, with roughly ten times the power draw). As capable open-weight models keep shipping, the bottleneck moves from "who has the best model" to "who has affordable hardware to run it locally". Mac Studio memory configurations change; press coverage in mid-October described M5 Ultra machines with up to 512 GB each, but I cannot verify availability today. The counterpoint: Mac Studios suit single-user, high-memory inference, while datacenter GPU clusters still serve many users at once.

Continue Reading →

Limited Edition Jonathan (Substack) · Perspective

Why This Matters

Markets are now questioning whether $740B+ of capex earns a return, and enterprises are already trading down to cheaper models, which supports the profitability and shift-away-from-APIs legs of the thesis. The open-weight gap is still large on one third-party index (32 vs 58) and token prices have fallen rather than risen, so legs one and two are not settled.

Book Lucien for your next event

Keynotes, masterclasses, panels and board-room sessions.

Get in touch