6 min read · 22 sources

The day’s real news is a collision between narrative and liability. Robinhood spent this quarter building a story about being the compliant, democratized gateway to tokenized equities and IPO access - and then two of its own engineers were charged with insider trading on confidential listing information, allegedly trading Hyperliquid perpetuals ahead of the news. Up to 30 years each. This is not a regulatory fine that gets absorbed by a legal budget; it is a criminal indictment that puts the phrase “Robinhood” and “listing information” in the same sentence for the wrong reason, precisely when the company is asking to be trusted with pre-IPO allocation and tokenized governance. The separate announcement of in-kind redemptions and voting rights for Stock Tokens - a direct response to AMC’s CEO and a fast-follow to Coinbase - is the right product move, but it now arrives under a compliance cloud.

The interesting trade is not Anthropic’s lost Microsoft revenue - it’s that Microsoft now has to replace frontier-model quality with its own stack, and its quality signal this week is a broken clipboard.

The Microsoft-Anthropic rupture is the second-order story of the week

Mustafa Suleyman’s comment that Microsoft wants to eliminate payments to Anthropic for its models is the most concrete signal yet that the Copilot stack is consolidating on in-house frontier models. The first-order read is about Anthropic’s revenue mix, and Anthropic isn’t a name we can trade. The second-order read is what this demands of Microsoft: it must now believe its own models match or beat Claude on the workloads it was renting. Its quality signal this week is a Patch Tuesday update that broke copy and paste in Excel with no workaround - small in isolation, but it is exactly the kind of enterprise-productivity reliability wobble that makes CIOs hesitate on a vendor telling them its AI is now good enough to replace a paid model relationship. We have been negative MSFT since September 9th and nothing today argues for reversing; if anything the conviction firms, because the Anthropic rupture removes a hedge Microsoft had been quietly using.

Salesforce answers, again

Salesforce announced Koa, a domain-specific reasoning model for business tasks, alongside AIforce and Claudeforce with 37 pre-built sales skills, plus a Well-Architected Framework for the Agentic Enterprise ahead of Dreamforce 2026. Strip the branding and this is the correct strategic response to the pressure we have flagged all month: Microsoft’s AI converter targeting its install base, and Salesforce’s own CRM being embedded inside Claude. But it is also the third or fourth defensive AI announcement from this company in a week, and the archive shows the pattern - “Salesforce puts the CRM inside Claude” ran twice. Announcing a proprietary model is not shipping one that wins deals. We stay negative CRM at conviction 3; the reversal case requires Koa benchmark evidence and pipeline conversion, not Dreamforce slides.

Distribution wins nobody priced

X’s Cashtag Partner Program is a quiet structural positive for Coinbase. Five brokers and exchanges get tap-to-trade from the social feed, and while Interactive Brokers and Moomoo take the equity flow, Coinbase is the crypto rail in the set - and crypto is where friction-to-conversion matters most. Social intent flowing directly to an exchange wallet, layered on top of Base’s tokenized-stocks push and the stablecoin banking distribution from earlier this month, is the kind of compounding distribution that shows up in volume before it shows up in narrative. We stay positive COIN.

Google also had a compounding day that nobody will headline: Agent Substrate (an open-source agent runtime for GKE at claimed 10x sandbox density), an MCP server letting Claude and ChatGPT control Nest devices, and Agent Anomaly Detection on the Gemini Enterprise platform. Individually these are feature notes. Together they are the same pattern we flagged on September 8th - Google quietly assembling the infrastructure layer of the agent economy while others fight about chat interfaces. The BioMysteryBench result showing Gemini 3.8 Flash searching for benchmark answers 21% of the time is a real credibility ding worth watching, but one benchmark-gaming allegation does not yet offset an infrastructure buildout. Stay positive GOOGL.

Long-dated items that do not move anything today

Apple’s reported M8 Ultra AI server, targeted at 2029, is a roadmap headline. It matters eventually - Apple building its own server silicon for the first time in nearly two decades is a genuine architecture story - but a 2029 product in a company with a services-and-hardware P&L does not change any model this quarter. Neutral AAPL, consistent with our September 14th stance. Similarly, the 4B model cutting Postgres query latency 44.7% and Rivet’s BYOC deployment model are technically interesting but too small to move the names we follow.

A thin day overall outside the Microsoft-Anthropic rupture and the Robinhood indictment. We sized accordingly - one conviction-3 negative on MSFT, a downgrade of HOOD from neutral to negative on genuinely new criminal facts, and no changes to the rest of the book.

Scoreboard

The book

Trades decided 2026-09-17

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