The Investor

The Contradiction at the Heart of the $900B Mark

Anthropic is, by revenue growth, the fastest-scaling enterprise software company anyone has measured, and also a vendor that cannot show customers a usage dashboard. ServiceNow, which is approximately the most sophisticated enterprise software buyer on earth, blew through its full-year Claude budget by May 2026 because Anthropic offers no per-user, per-tool granularity and no SLAs worth the paper. National Life Group's CIO put it without ornament: Anthropic is 'great for consumer usage but not great for companies.'

The $900B mark in secondary markets is being priced off enterprise ARR quality. The revenue underneath looks like consumer quality. Those trade at different multiples in every cycle anyone here has lived through.

The Pricing Change That Broke Wrapper Economics

On May 12-13, Anthropic converted every Claude subscription into a dollar-matched API credit pool. Two hundred dollars a month now buys exactly two hundred dollars of programmatic tokens, which closes the seventy to ninety percent arbitrage Cline, OpenCode, and the rest of the coding-agent harness layer had been running against subscription tiers.

The timing is not accidental. It lands alongside a new CFO hire and a likely October IPO. This is margin recovery dressed as policy, timed for pre-IPO diligence. OpenAI answered within hours with two months of free Codex for enterprise switchers, which is a two-sided squeeze on the harness layer.

Every Claude-dependent portfolio company's gross margin changed last Friday. Most founders haven't flagged it yet because the change is four days old.

The Observability Gap Is a Category, Not a Bug

Four firms (Google, OpenAI via Bain DeployCo, Salesforce, ServiceNow) have independently concluded that deployment services, not model capability, is where the margin lives. They are all hiring hundreds of forward-deployed engineers and running the Palantir playbook. ServiceNow, in the most elegant version of this, is selling AI Control Tower to the same customers panicking about their Anthropic bills.

This is probably wrong, but the AI observability and FinOps category looks like it forms over the next six to twelve months with no incumbent. Modal at $4.5B is the closest private comp. After that window, lock-in.

What This Means for the October IPO

Anthropic is doing what every company does before going public: recovering margin, hiring a CFO, cleaning up pricing arbitrage. The revenue number will look extraordinary. The revenue quality questions (net revenue retention without SLAs, churn risk without contractual lock-in, budget overruns without telemetry) are what the S-1 has to answer.

The Category Is Being Named From Above

In a single week SAP committed €100M to an Autonomous Enterprise partner fund, ServiceNow shipped Action Fabric as headless APIs for agents via MCP servers, Notion launched a developer platform hosting Claude and Codex as teammates, and Airtable parked a $10M Hyperagent credit war chest on the table. That is not gradual adoption. That is the incumbents drawing the category boundary before the pure-play agent startups get to.

The cloud analogy runs the other way. AWS defined the infra and the incumbents arrived late. Here the ERPs and workflow platforms are moving first, which means the agent-layer value accrual story looks materially different from the one most decks are still pitching.

Vercel's Production Data Reveals Two Businesses Inside 'Foundation Models'

Vercel published the first production-grade AI Gateway index across 200K+ teams. The findings, in order of how much they should bother you:

• 59% of token volume is agentic, not chat completions, not copilot calls
• Anthropic captures 61% of spend via Opus as premium reasoning
• Google captures 38% of volume via Flash as commodity throughput

Two different businesses are now visible inside what we have been calling 'foundation models.' One charges premium for complex agent orchestration. The other absorbs cheap high-throughput work. Pricing to one as if it were the other is how term sheets go wrong.

a16z's System of Intelligence Thesis Has Already Written Checks

a16z's public thesis is that the system of intelligence layer captures the majority of next-decade GTM software value, not the system of record, and the Stitch investment is the receipt. The proof point everyone is forwarding is Jason Lemkin's SaaStr anecdote: cutting Salesforce from 10+ human seats to 2 humans plus 1 API seat, with spend rising 83% from $12K to $22K and 20+ agents running underneath.

This is probably wrong in places, but the moat has moved from data gravity to orchestration gravity. Institutional context, workflow memory, and multi-system synthesis are the new lock-in, and the investable window is 12-18 months before incumbents close it or consensus forms around it.

The June 15 Cliff

Anthropic's third-party credit unbundling takes effect June 15. Any portfolio company whose COGS assumed subsidized Claude access through a third-party integration takes a structural margin hit in 30 days. That is not a future risk. It is a dated event with a known impact, which is the worst kind to be unprepared for.

The MongoDB Ransomware Moment for AI Infrastructure

LiteLLM, the LLM-routing control plane sitting inside thousands of enterprise AI deployments, was added to CISA's Known Exploited Vulnerabilities catalog this week. It is the first piece of AI-specific infrastructure to be federally flagged as actively exploited. Stack that against Ollama's GGUF model-loader bug at CVSS 9.1 with data exfiltration through malicious model files, plus OpenClaw shipping six critical CVEs in a single cycle, and the AI stack has now reached the security maturity that enterprise SaaS occupied in 2014.

This is the regulatory event that converts "AI security" from pitch slide to budget line. Or rather, the more interesting version: it tells procurement teams which line to put it on.

Mythos + NSA = AI Cyber Is Now a Procurement Category

Anthropic's Mythos is the first model to clear both UK AISI simulated attack ranges. The reward is closed-door Congressional briefings, with NSA — not CISA — winning access. That routing tells you the buyer is the IC, not the civilian defensive distribution channel, which is a different procurement cycle and a meaningfully different budget.

AI cyber task performance also doubled the prior 'doubling every few months' trendline. The attacker curve is now compounding faster than the defender curve. DepthFirst's Open Defense Initiative claims 10x cost efficiency over Mythos on vulnerability discovery, citing $1,000 versus $10,000 for equivalent FFmpeg results. The claim could be benchmark-gaming, and probably is in part, but the order of magnitude is the point.

The EDR Moat Is Cracking From Two Directions

TrustedSec pointed LLMs at all five major commercial EDRs and found identical architectural furniture across them, with YARA rules, Lua engines and local ML classifiers extractable in days rather than weeks. In the same week OpenAI launched Daybreak with Cloudflare, Cisco, CrowdStrike, PANW, Oracle, Zscaler, Akamai, and Fortinet as 'partners'. That is the classic pre-disintermediation lineup.

When a platform launches with every major incumbent listed as a friend, the relationship is pre-disintermediation rather than partnership. Cloudflare cutting twenty percent of staff for "the agentic AI era" the same week suggests at least one name on the list reads the wind.

The Investable Categories Are Now Distinct

The market has split into three things that should not be priced as one. AI-native offensive discovery is DepthFirst and XBOW competing on specialized-harness economics. AI infrastructure defense covers LLMjacking, model-artifact scanning and MCP firewalls, greenfield with no defender category leader yet. AI-speed enterprise defense means autonomous patching and continuous red-team. The buyers don't overlap and the exit math isn't comparable.

The Healthcare AI Category Just Got Its Winner

Abridge raised $550M in 2025 alone ($250M early year, $300M at a $5.3B cap in June), services 250 large US health systems, and processes 80M+ patient-clinician conversations annually. That is not a Series B story. That is a data moat no foundation model or new entrant can replicate — 100M+ proprietary medical conversations with edit data and clinician preferences across 28 languages and 50 specialties.

The signal for capital allocation: the ambient clinical documentation wedge is closed. The alpha is in adjacent workflow layers.

The Three-Act Playbook Is Now Visible

Abridge's expansion path — save time → save money → save lives — maps precisely to buyer personas within the same account:

• Act 1 (Won): Clinical documentation → CMIO buyer
• Act 2 (Active): Prior authorization, revenue cycle → CFO buyer
• Act 3 (Unlocked): Clinical decision support → enabled by January 2026 FDA CDS guidance

Each act monetizes the same proprietary conversation asset against a different stakeholder. This is the vertical AI underwriting template going forward.

What Makes It Investable as a Framework

Health systems — historically the slowest enterprise buyers — compressed release cadence from quarterly/biannual to monthly for Abridge. That's unprecedented velocity in a regulated buyer and should rewrite the sales-cycle assumptions in every healthcare SaaS DCF.

Three criteria should now be non-negotiable for any vertical AI deal:

1. Per-use data flywheel — every interaction accrues proprietary training data competitors can't access
2. ≥2 monetization vectors beyond the initial wedge, targeting different buyer personas
3. Explicit non-compete posture with the dominant system-of-record (EHR, CRM, ERP)

Abridge proved vertical AI moats are built from proprietary interaction data, not model quality. The window for adjacent categories — payer-side prior auth, nursing workflows, specialty verticals — is open now.