8 min read · 33 sources

Most of today’s feed was product notes. Two things were actual evidence: an adoption dataset and a price list. US business AI adoption plateaued at 55.9% in September, and within that saturated base, enterprise share has flipped to Anthropic at 43.8% versus OpenAI’s 40.6%. Separately, Google published pricing for its Argon TPU at $1.99 per task with 80% more tokens than comparable GPU runs. Between them, they reframe the quarter: the model layer is no longer a land-grab story, it is a share-and-unit-economics story, and the second-order consequences land on Microsoft, Nvidia, and the SaaS front-end.

A published price list is stronger evidence than any benchmark: at $1.99 per task, custom silicon stops being a roadmap slide and becomes a procurement option.

The plateau makes it a share fight, and Microsoft is holding OpenAI's coat

A plateau at 55.9% means the marginal adopter is gone; from here, gains are share gains. Anthropic holding 43.8% against OpenAI’s 40.6% is the number that matters, and it cuts two ways. For Microsoft, the negative read is structural: its enterprise AI distribution is heavily routed through OpenAI, and its partner is now losing the share race it was expected to dominate. Microsoft’s response is visible in the same feed - Maia 300 and Jalapeño silicon in the pipeline and MAI-Base-2 pre-training complete at 80B active parameters. Decoupling from OpenAI is the right strategy, but it is expensive and mid-flight, which is why we keep Microsoft negative. For Alphabet, the same datapoint is a double win: Anthropic’s enterprise strength routes through Google’s cloud and capitol, and Google’s own TPU economics are now a published advantage rather than a claim. Add two Workspace rollouts - the AI-driven Inbox and Gmail Live - both default-on for paid tiers, and Google is monetizing distribution it already owns. We stay positive at the same conviction as the past week.

Two Anthropic side-items are worth reading as sales motion rather than features: OpenTelemetry tracing for Managed Agents and third-party evaluator access to its METR system are enterprise-procurement artifacts - observability and auditability are what get a model vendor onto a regulated buyer’s shortlist. One staleness flag: the researcher survey on extinction risk and the Hinton item are a 2024 survey being recirculated; it is not news and we did not trade on it.

Custom silicon publishes a price list

The Argon pricing is the day’s most consequential number for hardware. Benchmarks are arguable; a price list is not. At $1.99 per task with 80% more tokens than GPU runs, Google has turned the custom-ASIC threat into procurement arithmetic, and that is why we re-tighten Nvidia from neutral back to negative - what changed is that the threat now has a price attached, not just a roadmap. The software side points the same direction: TokenSpeed claims 685 tokens/s on GB300 with up to 90% cost reduction, and XERJ cuts token usage 2.7× by indexing locally. Note the nuance - efficiency tooling that runs on Nvidia’s own hardware sells more boxes near-term while structurally deflating per-token economics. Nvidia shipping Nemotron 3.5 Super VL (120B parameters, 12B active) reads as a hedge into the model layer. The clean second-order winner is Broadcom: hyperscaler ASIC programs - Google’s TPUs, Microsoft’s Maia line - shift design value toward the ASIC ecosystem regardless of which model vendor wins, which is why we hold our highest-conviction positive there. The TSMC item, by contrast, is a retrospective profile of the foundry model, not new information; we treat it as background, not a signal.

OpenAI moves up the stack; the SaaS front-end is the exposed layer

GPT-6’s launch is the new fact; the Codex retrospective is a year-old recap of the summer 2025 shift from cloud IDE to autonomous agent, and we treat it as such. The new part is the “Intelligent UI” generating interactive frontend components and “dots” agents operating across 4,000 apps. That is an attack on the interface layer - the place where subscription software prices its moat. The second-order exposure is Salesforce and Adobe: if the UI is generated per-task rather than owned per-seat, the front-end ownership premium erodes from below. Neither name got new negative evidence today; the existing short cases were reinforced, not changed. xAI’s Grok Voice 3.0 at flat pricing versus 2.0 is supporting color - voice is becoming a priced commodity at the edge of the stack - but xAI is not a name we can act on.

Meta's open weights meet Apple's idle silicon

Meta releasing open weights for Muse Spark Open - 816B parameters, 42B active, optimized for Apple’s M5 Ultra - continues the strategy we flagged earlier this week: commoditize the model layer, monetize the infrastructure around it. The Apple side of the pairing is the more interesting new fact. Beyond the foldable - which was yesterday’s story, and the $1,999 pricing adds nothing new - the leadership change to John Ternus, an engineering veteran, plus the “Minions” system pooling idle Macs into a distributed GPU cloud, clarifies a hardware-first strategy that treats the installed base as compute. We move Apple from neutral-low to neutral with slightly more conviction: strategy clarity, but no change to near-term revenue arithmetic.

Edge plumbing and the long tail

Cloudflare shipped three same-day GAs - Logpush Transformers on all self-serve plans, Hyperdrive for MySQL with distributed pooling and CDN caching, and programmable Secure Web Gateway and DLP actions via Workers. The last is the quiet displacement story: customers can replace hand-rolled NGINX proxies with edge-executed logic. We keep our positive stance, with one honest caveat - this is the fourth consecutive week of Cloudflare headlines, and at some point cadence becomes noise. The SWG/DLP item is the one with displacement economics behind it. Elsewhere: Amazon S3 landing on a developer marketplace is a small distribution win that does not change our Amazon view, and the memo-culture piece is a culture critique without a number - worth tracking for decision rigor, not for the book.

Agentic commerce gets its first measured dataset

Chainalysis counting 138 million AI-agent payments on Coinbase’s Base L2 is the first large-scale measured usage of agentic commerce rather than another projection. It builds directly on the Base identity work we covered earlier this month, and it is why we raise Coinbase’s conviction a notch: the thesis now has a denominator.

What was thin

The Willow processor’s ~40× error reduction via magic-state cultivation is genuine roadmap progress with no revenue line for years. The multi-language Agent Development Kit, the Claude-generated scanner driver, the GitHub-stars critique, and the validator-proof proposal are color. The Anthropic Scion monorepo story is worth one careful thought - AI-generated merge volume becoming an infrastructure problem at the labs is a signal about code volume - but one lab moving off Git is a data point, not a trend, and we will not over-read it for GitLab. Most of today was plumbing; the two numbers that mattered were the adoption plateau and the price list.

Scoreboard

The book

Trades decided 2026-10-16

Filled at the next session’s open, not today’s close.