The Investor

Three Concurrent Tests on One Day

Thursday is going to produce more usable pricing information about the AI stack than any single day this quarter, which sounds hyperbolic until you look at the calendar. Cerebras prices at $35B, the loudest independent AI-silicon IPO in a market where two direct peers already gave up the ghost (Groq → Nvidia license + talent; Graphcore → Meta acqui-hire). Figma reports $316M of revenue, up 38.5%, with 75% of paying customers consuming AI credits every week — the first clean public test of whether usage-based AI pricing survives enterprise procurement. And FactSet's -8% print from earlier this week keeps repricing as Anthropic's ten finance agents wired into Microsoft 365 and Moody's start to look like credible workflow unbundling.

The Cerebras Outcome Tree

The bull case is easy to see. CoreWeave is up 185% from its $40 IPO and Bloomberg says the book is strong. The bear case is structural, or rather, the more honest version is structural: every hyperscaler is building its own silicon (Trainium, TPU, MTIA, Maia, OpenAI+Broadcom), and the independent merchant-chip thesis has exactly zero successful precedents in this cycle. Price above range, trade up twenty percent or more, and late-stage AI-silicon secondaries open for roughly thirty days. Break issue and every active chip deal collapses into acqui-hire math: $2-5B to a hyperscaler as the base case.

The Figma Margin Question

The 75% weekly AI-credit number is impressive and also hides a demand-elasticity problem. Usage-based pricing converts gross margin into COGS while the customer spreadsheet tells procurement to push back. A miss resets every AI-native SaaS comp in the portfolio. If consumption converts to expansion revenue at current margins the model works, which is what the bulls are paying for. If consumption outpaces willingness-to-pay, Figma becomes the warning label on every AI-attach pricing strategy shipped in the last eighteen months.

The FactSet Domino

FactSet down eight percent on day one is a data point, not a conclusion. The more interesting signal is Viceroy Research's public pivot from fraud shorts to shorting high-margin clean-balance-sheet businesses sitting in the path of AI. When the sharpest activists start targeting the quality factor, the compounders that have anchored long-only allocation for a decade need re-underwriting. MSCI, SPGI's analytics segment, and Bloomberg's non-terminal revenue are the second-derivative names, which is a polite way of saying the list is long.

Thursday answers whether independent AI silicon is a fundable category or a hyperscaler talent pool. It also answers, separately, whether usage-based AI pricing survives at enterprise scale, and whether financial-data moats hold under agent pressure. Pre-position, or pay for the information after it prints.

The Bifurcation Is Real

Sierra raised nine hundred and fifty million dollars at $15B post-money on $150M ARR with Fortune 50 penetration north of forty percent. DeepSeek doubled from twenty billion to $45B in weeks on Chinese state capital, with Big Fund leading and Tencent and Alibaba alongside. Zhipu shipped GLM-5.1 under MIT license scoring 58.4 on SWE-Bench Pro, which is, narrowly, ahead of GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. The same week, xAI priced Grok 4.3 at $1.25/$2.50 per million tokens, undercutting Anthropic and OpenAI by forty to sixty percent.

These are not three AI stories. They are three asset classes wearing similar labels, and a fund that runs them through one framework will mismark the vintage.

Underwriting Logic #1: Applied Agents (Sierra, Moonshot/Kimi)

Enterprise CX agents at roughly one hundred times ARR with named Fortune 50 customers and renewal cycles. This is SaaS math applied to agents, defensible if retention behaves and embarrassing if it doesn't. The alpha sits in vertical agent plays at Series A — insurance claims, banking back-office, healthcare RCM, field service — where the Sierra comp just reset valuation expectations upward. Or rather, the more interesting version of that trade is the one priced before the reset propagates.

Underwriting Logic #2: Frontier Labs as Geopolitical Options (DeepSeek)

DeepSeek at forty-five billion dollars is a sovereign strategic auction, not a valuation. The marginal buyer is a state actor who cannot purchase the Western alternative. Exposure at this layer is a macro position wearing a software ticker. Diligence on anything with China revenue or Chinese-sourced models needs explicit geopolitical and export-control clauses, because the downside is not a down round.

Underwriting Logic #3: Open-Weight Commoditization (GLM-5.1, Grok 4.3)

The frontier-model premium is compressing faster than secondary valuations assume. GLM-5.1 at MIT-zero and Grok at $1.25/M tokens reset the enterprise floor. Every portfolio company with >30% COGS tied to OpenAI/Anthropic APIs should model a migration scenario this quarter. The thesis, which has been approximately right for several quarters and could still be wrong: the moat has migrated up-stack to orchestration, vertical data, and inference optimization. Stanford AI Index 2026, Netflix's metadata architecture, and DeepMind's 3x Gemma 4 inference speedup via speculative decoding each say the same thing, independently.

Sierra at fifteen billion and DeepSeek at forty-five billion are not the same trade. The funds that run both through one framework will mismark the vintage.

SAP Just Fired the First Shot

SAP quietly retooled its API access to block third-party agents like OpenClaw while whitelisting its own Joule and Nvidia's NemoClaw. In the same cycle it bought Prior Labs all-cash with €1B committed over four years for tabular foundation models, which makes the gating architectural rather than merely contractual. Call it an enterprise-data moat, paid for in advance.

The base case, and this is probably wrong in the particulars but right in shape, is that Salesforce, Workday, ServiceNow, and Oracle follow within two to three quarters, because they always do. If they do, horizontal agent platforms lose 30-50% of addressable distribution in a quarter. The agents that survive are endorsed partners, agents sitting on proprietary data substrates, or workflow layers above the SaaS tier.

The Portfolio Audit

Any agent portco whose pitch opens with "we read your SAP/Salesforce/Workday data" needs a 2026 Plan B on the desk this week. The specific questions:

• Does the portco have a formal partnership or whitelisting agreement with the platform it depends on?
• Can the product function on exported data rather than live API access?
• Is there a proprietary data layer that the platform cannot replicate?
• Does the agent sit above the SaaS tier (orchestrating across platforms) rather than within it?

Four noes is not a slow fade. Four noes is a guide-down.

Where the Survivors Live

The walled-garden thesis actually increases conviction on three archetypes, or rather, the more interesting version of those three:

1. Cross-platform orchestration agents that route across SaaS vendors rather than reading any single one — immune to single-platform gating
2. Vertical agents with proprietary domain data that Joule and Copilot cannot replicate — healthcare claims adjudication, insurance underwriting, legal contract analysis
3. Endorsed partners who invested in platform relationships early, analogous to the Salesforce AppExchange winners of the prior cycle

The FactSet repricing runs in parallel. Anthropic's finance agents inside Microsoft 365 and Moody's suggest platform-endorsed agents from frontier labs will displace independent middleware. The incumbent data providers have three paths: license data to agent builders and accept margin compression, get routed around, or ship their own agents and reconstitute the moat one layer up. FactSet's silence on which it is picking is what the 8% is pricing.

The agent plays that survive enterprise walled gardens are endorsed partners, proprietary-data holders, or cross-platform orchestrators. Everything else just became a feature of Joule.

The Meta Case Changes the Math

Five publishers and Scott Turow allege that Meta torrented 267TB of pirated books to train Llama after walking away from a $200M licensing deal, with Zuckerberg personally authorizing the call. That last detail is the one doing the work. This is the first CEO-authorized AI piracy case that arrives with a quantified alternative already sitting in the record, which hands the court a clean damages anchor prior training-data cases never had.

The investment read here is structural, not idiosyncratic. Every frontier-model position now carries a mispriced tail risk that lives inside indemnification language rather than the P&L, which is a less comforting place for it to live. Labs that paid for data look overpriced today and defensible after judgments land. Labs that didn't are carrying a contingent liability dressed up as an asset.

Where the Liability Sits Determines Which Layer Gets Repriced

The practical move, or rather the boring version of it, is to read the indemnity clauses before the valuation notes. The clauses are shorter and they tell you more. A portco that cannot document training-data provenance is not holding an asset. It is holding a lawsuit waiting for a plaintiff.

The Counter-Thesis

Courts move slowly, settlements get structured, and enterprise buyers writing checks do not care as long as their own indemnities hold. This has been approximately true for eighteen months and the market has arguably already absorbed it. The counter-counter is narrower and more interesting: what has actually changed is the willingness of counterparties to put indemnification language into contracts they used to sign without it, which matters to procurement teams today and to valuation multiples the day it shows up in a filing.

If the Meta case survives motion to dismiss, a training-data licensing marketplace boom becomes the base case rather than the speculative one. Scale AI, Shutterstock-style licensing, and provenance-tracking infrastructure move from interesting to must-own. This is probably wrong in its timing and roughly right in its direction.

An LLM portfolio that cannot document training-data provenance is not holding an asset; it is holding a lawsuit waiting for a plaintiff.