7 min read · 31 sources

Two documents carried the day: Anthropic’s S-1 and Microsoft’s quarter. Everything else - and there was a lot of it, with agentic coding hitting general availability on three platforms at once - orbits those. The tail of the news list (a Meta pet app, an Apple display refresh, a transcription model) was genuinely thin, and we say so rather than manufacture views.

Spawn’s $3.1M monthly token bill is the first hard proof that falling inference prices expand spend rather than shrink it.

The S-1 is the document: Anthropic's revenue is substantially a cloud-channel story

Anthropic filed to list as “ANT” at a potential $100bn raise against a ~$2trn valuation, on $19.7bn of FY2026 revenue with Q2 at $11.6bn, up 1,368% year over year. We covered the draft prospectus and the debt backstop recently; today is the filing itself with hard numbers, so this narrative is building, not stale.

The disclosure that matters is not the valuation - it is that 47% of 2025 revenue flows through Amazon and Google cloud partners, with the two largest customers at 12% of 2024 revenue each. The obvious name in this story is Anthropic, which you cannot own until the listing. The less obvious beneficiaries are the channels: the S-1 hands AMZN and GOOGL an audited, third-party-verified demand anchor for their AI consumption narratives. Cloud capex stops being a leap of faith when a private lab’s prospectus quantifies who resells the compute. The concentration cuts both ways - Anthropic depends on two channels - but for the clouds it is booked revenue.

The conspicuous absence is Oracle. ORCL spent the week courting AI workloads - Vera Rubin NVL72 Superclusters, the Acceleron RDMA fabric, first-party Snowflake - and now lands Google TPU v8 on OCI from 2027. That is a silicon-arbitrage strategy: rent everyone’s constrained compute to everyone. It is a real business, and we move positive on it, but the S-1 shows Oracle has not yet cracked the model-lab consumption channel that Amazon and Google own. SNOW’s first-party AI Data Cloud on OCI is a distribution win inside that same strategy; it does not answer the consumption-growth question, so it stays neutral for us.

Agentic software delivery goes GA, and the token bill arrives

GitHub shipped the Copilot app and made Copilot Fusion generally available: multi-model orchestration with three-tier automated routing, claiming ~50% token-cost reduction and 40% faster task completion, plus self-improving meta-automations. Microsoft exposed the same multi-model harness through Foundry with MCP support. Slack made the Code API generally available (two items in today’s feed, one story), opening dedicated multi-user code channels to any agent vendor.

The demand side is no longer hypothetical. Ramp’s Inspect agent generates 74% of merged pull requests at 3.4× PRs per person. Spawn burns $3.1M per month on Anthropic tokens to push ~77k production changes. Pair that with Anthropic shipping enterprise reporting, in-app surveys and pooled group budgets, and the picture is coherent: token spend is becoming an audited ROI line item, and the labs are building the finops layer themselves.

The second-order read sits above the models. When the flagship coding platform’s headline feature is routing between models to cut token cost 50%, the neutral gateway layer - policy, caching, routing, observability outside any single vendor - becomes infrastructure. That is the category NET sells through AI Gateway, and it is why we stay positive: every enterprise adopting routed multi-model inference needs someone neutral holding the routing table. And note the elasticity: Gemini 4 Carbon just cut prices while Spawn’s bill grows. Volume is outrunning price, which is the correct bear-case rebuttal for inference margins.

Google's pincer: cut inference prices, rent the silicon to rivals' clouds

Gemini 4 Carbon prices at $2/$10 per million tokens with a 95% discount on cached inputs - a structure that specifically favors agentic workloads, which are read-heavy - and leads security benchmarks at 88.2% on SEC-bench Pro. The Connected Apps Platform extends Gemini distribution through MCP servers into third-party apps.

The bolder move is TPU v8 on Oracle Cloud from 2027: 12.6 PFLOPS FP4 per chip, >97% goodput. Google is externalizing TPU distribution into a rival’s cloud. For NVDA, that seeds a credible second silicon channel at hyperscaler scale, and it is why our negative stance holds - though the Vera Rubin NVL72 expansion on OCI is the partial offset that keeps conviction low.

Microsoft: the print that broke our bear case

We reverse the MSFT negative call today. The thesis rested on capex squeezing free cash flow; the quarter printed $18.6bn of FCF, roughly $10bn above estimates, with Azure at 48% and Q2 guided to 49 - 50% against 47.6% consensus. What changed is the numbers, not the narrative - if Q2 confirms acceleration with FCF intact, we go further positive. Separately, Entra retires SMS and voice MFA in February 2027 with passkeys defaulting in September: the passkey transition is now enforced by the largest identity provider, which validates the category while raising the default-installed competitive bar for standalone vendors like Okta.

Crypto rails: settlement, market microstructure, and an agent with a wallet

Ripple partnered with HOOD to run RLUSD through institutional settlement, treasury automation and financing via Ripple Prime and Bitstamp; Fynd’s “last-mover” AMM went live on Robinhood, Base, Arbitrum and Ethereum with a 2.7× LP-fee claim we treat as unverified. Slack open-sourced the Centaur multi-agent framework and the Reth 2.0 high-throughput Ethereum client. The thread tying it together is Instinct’s demo: an agent converted stablecoins via Stripe, moved USD to a Lead Bank prepaid account, and paid with a Visa virtual card - $700, end to end. Agents need programmable money, and stablecoin rails are becoming that substrate, which is the quiet case for HOOD’s infrastructure drift.

Quick hits, honestly thin

Apple’s touchscreen MacBook Pro with the “PROLED XDR” panel (2,000 nits, 120Hz, 3 lb) is a decade-late concession at a revenue scale where display refreshes move nothing. Meta’s “Waddle World” AI pet in the Philippines is a consumer toy, irrelevant to our enterprise-cloud negative. NVIDIA’s Nemotron 3.5 Transcribe is solid vertical speech work, not a thesis mover. OpenAI’s Dot agents over iMessage/RCS extend the agent-surface pattern to a private company - watch the channel, not the ticker. The one tail item with teeth: Claude autonomously found five critical vulnerability chains including a one-click Chrome RCE, producing 62 patched CVEs in Chrome 156. That is the strongest public evidence yet that agentic security is production reality, and the security vendors’ turn to monetize it is coming.

Scoreboard

The book

Trades decided 2026-10-29

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