Anthropic's record IPO and OpenAI's app store day: distribution, not models, moves the market
6 min read · 28 sources
- Selling UBER (Uber), $23,001.
- Buying NET (Cloudflare), $19,542. Clairic-flash as a 9B open-weight extraction model on Workers AI extends the edge-inference wedge that underpins our positive view.
- Buying AAPL, $18,069. On-device agent work plus FastFrames diffusion shows Apple building the local inference layer rather than renting it.
- Plus 3 smaller adjustments, listed in full in the trades table below.
- Anthropic's $1.8 trillion Nasdaq debut at a 17% pop is the sector's cleanest signal yet that public markets will fund frontier-lab economics at hyperscaler-scale valuations.
- OpenAI's DevDay apps launch makes OpenAI the ranking layer and user-relationship owner for partner integrations, a direct template of the app-store playbook aimed at incumbent distribution platforms.
The headline event is Anthropic’s IPO: $81.3 billion raised, closed at $97.17 up 17% from the $83 price, a $1.8 trillion fully diluted market cap - the largest IPO on record. The second-order read is what it does to the rest of the sector: public-market capital is now explicitly underwriting frontier-lab economics, which raises the financing bar for every other lab and validates the inference-revenue assumptions the whole capex cycle rests on.
The story that rhymes with it: OpenAI’s DevDay launch of apps inside ChatGPT with Spotify, Canva, and DoorDash. This is the app-store playbook - OpenAI controls ranking and owns the user relationship while partners supply the functionality. The second-order effect is on the incumbent distribution platforms: every integration that routes discovery through a chat surface is discovery that doesn’t route through a search box or an app store. Watch whether the partner list grows faster than the platforms can build their own agents.
OpenAI just copied the app store playbook, and the uncomfortable question for every incumbent platform is that the ranking algorithm now lives in someone else’s product.
Microsoft: two stories pointing in different directions
Microsoft Advertising shipped its largest bidding model upgrade in a decade - 10% more conversion value per click on Maximize Conversion Value/Target ROAS and 18% more conversions moving from eCPC to Maximize Conversions. That is a direct revenue-per-click lever, and it is the strongest single datapoint today for the bull case on Microsoft’s ad business.
But the rest of the Microsoft tape is heavier. The company patched 3,338 vulnerabilities in July - October 2026 versus 1,130 in all of 2025 - a patch burden that is itself a cost line and a sales pitch for Google Cloud’s enterprise AI messaging, which found 78% of surveyed enterprises say legacy software diverts capital from AI. The new agent-controls service is a sensible governance product, but governance is a crowded lane. We stay negative: the ad upgrade is real, the capex and security drag is bigger.
Intel's anchor tenant
The Microsoft - Intel expansion is the quiet story with the longest tail: Microsoft is the first customer for Intel Foundry’s EMIB packaging, plus a multi-year Xeon commitment including Diamond Rapids, with up to 2.3x Azure M-series v4 VM gains from CXL. For Intel, this converts the foundry pitch into a named anchor customer - exactly the validation the story needed. It reinforces yesterday’s positive call.
NVIDIA's software moat, tested
NVIDIA and Microsoft added Vera Rubin NVL72 Azure VMs and extended confidential computing to bare-metal AI - the GPU cycle rolls on. But the more interesting datapoint is Together AI deleting its megakernels: CUDA Graphs and PDL on default Torch Inductor recovered 98% of megakernel performance and beat the custom kernels by up to 1.7x at over 1,100 TPS for Llama 1B on H100. If stock compiler stacks match hand-tuned kernels, the software lock-in premium that defends NVIDIA’s margins is worth less than the market prices. That is consistent with our negative stance.
The rest, briefly
- Google had a broad day: Project Suncatcher satellites on Blue Origin in 2029 (a demo whose economics hinge on cheap heavy lift), a quantum contextuality benchmark on 105 qubits, Sashiko open-sourced for Linux kernel review, Tool Retrieval in the Gemini API, and computer use in the Gemini Mac app. Nothing here is a revenue event this quarter; all of it compounds the capability lead behind our positive call.
- Apple is working on an on-device agent for device control, apps, and browsing, and shipped FastFrames video diffusion (1.3B at 36 ms/frame on one B200). Apple building the local inference layer rather than renting it supports our positive view.
- Cloudflare released Clairic-flash, a 9B open-weight extraction model on Workers AI - edge inference is the wedge. GitLab added MCP audit events and governance controls, a real enterprise-compliance differentiator. Palo Alto’s $4.5 billion Applied Compute deal reads as an expensive detour for a security franchise.
- METR/Anthropic evidence of a 2x capability acceleration in 2026 is the kind of datapoint that matters more in six months than it does today.
Thin day for anything genuinely new at the semiconductor and storage layer otherwise; most of the agent-governance headlines are continuation of a weeks-old theme rather than fresh facts.
Scoreboard
| Strategy | Return | Benchmark | Excess | Max drawdown | Days |
|---|---|---|---|---|---|
| Top 5, equal weight | -10.25% | -12.61% | +2.35% | -21.79% | 52 |
| Top 10, conviction weighted | -9.69% | -12.61% | +2.92% | -20.74% | 52 |
| Long top 5 / short bottom 5 | +4.55% | -12.61% | +17.16% | -11.35% | 52 |
| Top 5, free-tier signals | -10.25% | -12.61% | +2.35% | -21.79% | 52 |
The book
| Ticker | Company | Weight | Value |
|---|---|---|---|
| ADBE | Adobe | 0.6% | $557 |
| AKAM | Akamai | -0.2% | $-224 |
| AMD | AMD | -0.1% | $-74 |
| AMZN | Amazon | 0.4% | $339 |
| AVGO | Broadcom | -0.9% | $-787 |
| COIN | Coinbase | -0.0% | $-3 |
| CRM | Salesforce | -0.6% | $-533 |
| CSCO | Cisco | -0.1% | $-62 |
| DDOG | Datadog | 0.0% | $1 |
| GOOGL | Alphabet | 24.3% | $22,139 |
| HOOD | Robinhood | 0.1% | $87 |
| IBM | IBM | -0.0% | $-31 |
| INTC | Intel | 25.5% | $23,220 |
| META | Meta Platforms | 0.1% | $127 |
| MRVL | Marvell | 0.2% | $154 |
| MSFT | Microsoft | -0.8% | $-682 |
| MU | Micron | -0.1% | $-122 |
| NET | Cloudflare | -1.6% | $-1,473 |
| NOW | ServiceNow | 0.0% | $17 |
| NVDA | NVIDIA | -0.2% | $-212 |
| OKTA | Okta | 0.0% | $3 |
| ORCL | Oracle | -0.5% | $-478 |
| QCOM | Qualcomm | -0.5% | $-428 |
| RBLX | Roblox | -0.0% | $-29 |
| SHOP | Shopify | -0.0% | $-44 |
| SNOW | Snowflake | -0.3% | $-249 |
| TEAM | Atlassian | -0.0% | $-7 |
| TSLA | Tesla | 0.6% | $579 |
| UBER | Uber | 25.3% | $23,001 |
| XYZ | Block | -0.1% | $-60 |
Trades decided 2026-11-20
Filled at the next session’s open, not today’s close.
| Action | Ticker | Amount | Rationale |
|---|---|---|---|
| sell | UBER | $23,001 | |
| buy | NET | $19,542 | Clairic-flash as a 9B open-weight extraction model on Workers AI extends the edge-inferenc |
| buy | AAPL | $18,069 | On-device agent work plus FastFrames diffusion shows Apple building the local inference la |
| buy | GTLB | $18,069 | MCP audit and governance controls give GitLab a differentiated enterprise-AI-compliance an |
| sell | INTC | $5,183 | Microsoft as EMIB first customer plus multi-year Xeon commitment converts the foundry stor |
| sell | GOOGL | $4,064 | Suncatcher, quantum benchmarks, Sashiko, and Gemini tool retrieval all compound the capabi |
You May Also Like
OpenAI Trims to $50B, Banks Underwrite Broadcom's $60B, and Nadella Asks for a Brake
The number that matters today is OpenAI telling investors annualized revenue is approaching $50 billion against earlier $70 billion estimates - the first hard …
Decision APIs Everywhere, and a $35 Billion Compute Check
Three platforms shipped decision-model APIs in a single week while a GPU cloud raised $4bn around a $35bn compute commitment - the model layer is commoditizing …
The day AI agents broke the CPU supply chain - and Apple finally folded
Server CPU lead times stretching to six months and prices up 10 - 20% is the real story of the day: agent workloads are quietly repricing the non-GPU half of …




