The Investor

The Revenue Quality Problem Nobody Is Marking

ServiceNow, which is one of the most sophisticated enterprise software buyers on earth, exhausted its full-year Anthropic budget by May 2026. Not because Claude underperformed, but because there is no granular per-user, per-tool telemetry and no SLAs worth the name. National Life Group's CIO put it less politely: Anthropic is 'great for consumer usage but not great for companies.'

This matters because Anthropic just passed OpenAI on Ramp's billing data, 34.4% vs 32.3%, the first documented leadership change in enterprise AI, and over the trailing year Anthropic quadrupled business adoption while OpenAI grew 0.3%. The revenue underneath that lead is reversible at near-zero switching cost.

The Capacity Crisis Beneath the Revenue

Anthropic planned for 10x growth and got 80x, which is a forecasting error, or rather the more interesting version, a forecasting error that bends strategy. Hence the silent Claude Code nerfs, the mid-trial Pro revocations, the corporate account bans, and then the unthinkable: Anthropic leased xAI's entire Colossus 1 cluster (220K+ GPUs) from Elon Musk, the person who publicly called them 'misanthropic and evil.' When you rent from a declared enemy, you are not making a strategic choice. You are short capacity.

At the same time, every subscription became a dollar-matched API credit pool, so a $200 plan now buys $200 of programmatic tokens, which closes the 70-90% arbitrage that Cline and OpenCode had been running on the harness side. This is margin recovery timed to an October IPO, with a new CFO already in the chair.

Enterprise AI ARR is not SaaS ARR. It has no SLAs, no lock-in, no telemetry, and customers blow annual budgets by Q2. Price it accordingly.

The Forward-Deployed-Engineer Consensus

Four firms independently conceded the same point this week: deployment, not model quality, is the bottleneck. Google Cloud is hiring hundreds of forward-deployed engineers, OpenAI stood up DeployCo with Bain, and Salesforce and ServiceNow are staffing the same function. The Palantir playbook is now consensus, which means the interesting category is one layer over, in the AI observability and FinOps layer forming in real time.

This is probably wrong, but: ServiceNow is already selling an AI Control Tower to the same customers panicking about their Anthropic bills, and Modal is raising at $4.5B. The next Datadog-scale category lives wherever CDIOs are both buying and building the same tool.

The First Production Data

Vercel finally published the first production-grade AI Gateway index across more than two hundred thousand teams, and the headline that ought to rewrite a few AI memos is that 59% of all token volume is now agentic workloads. Chat completion, which is what most of these decks were sold on, is now the minority case in production. The agent thesis is not arriving. It is most of what is running.

The split between spend and volume is the more interesting puzzle. Anthropic captures 61% of spend on long, tool-heavy reasoning calls, while Google captures 38% of volume on the cheap, high-throughput commodity work. Two different businesses are sitting inside the phrase 'foundation models,' and they do not deserve the same multiple.

The Platform Land Grab

Three incumbent moves landed in the same quarter, all pointed roughly the same way:

• Anthropic's June 15 third-party credit unbundling — ends subsidized experimentation through downstream tools
• Notion's Developer Platform — Claude, Codex, Cursor, and Devin as hosted teammates inside the workspace
• SAP's €100M Autonomous Enterprise Fund — wired into NVIDIA and Microsoft at the platform layer

This is probably wrong, but the read is that standalone agent-orchestration startups of the LangChain or CrewAI class face absorption risk inside two to three quarters. The counter-thesis is that orchestration is hard enough that incumbents buy rather than build. Either way, the value migrates to neutral infrastructure doing multi-model routing, observability and eval, or to vertical integrators sitting on proprietary workflow data. Standalone orchestration is the layer that gets squeezed.

Agent workloads are 59% of production token volume. The winners are neutral infrastructure and vertical integrators, not standalone orchestration layers.

a16z's GTM Thesis Sets the Clock

a16z has now publicly declared that the 'system of intelligence' layer captures more value than the system of record beneath it, and put the Stitch cheque behind the claim. The Lemkin number makes the pitch concrete, or rather makes the consumption-based GTM legible in one customer: 2 human seats replacing 10+, spend up 83% from $12K to $22K, 20+ agents running on top. Fewer seats, higher ARPU, more compute. That is the deck.

What the actor is not doing, while spending here, is funding horizontal AI CRM wrappers, which is the trade that has already started compressing. The moat has moved from data gravity to orchestration gravity: institutional context, workflow memory, multi-system synthesis. The investable window is twelve to eighteen months before incumbents or consensus close it, and the thesis has been right for several quarters and could be wrong by the next earnings cycle. The alpha, if there is any left, sits in vertical-specific orchestration with measurable outputs.

The Category-Formation Event

LiteLLM is the first AI-infrastructure component to land on CISA's Known Exploited Vulnerabilities catalog, which is interesting less as a vulnerability story than as the regulatory-validation moment that converts 'AI security' from pitch deck to enterprise budget line. In the same cycle Ollama shipped a CVSS 9.1 model-loader bug enabling data exfiltration via malicious GGUF files, and PraisonAI's auth-bypass was weaponized within four hours of disclosure. Four hours is not a window. It is a confession.

The AI-infra stack is sitting roughly where enterprise SaaS sat in 2014 on security maturity, which is to say: not great. Every dollar of AI-infra ARR is creating latent demand for AI-sec ARR, and the Series A window for runtime-layer winners is open now. It will not stay open long.

The Economics Split

Two models are emerging and the middle gets squeezed, which is the usual story but worth watching anyway.

The Daybreak partner list — Cloudflare, Cisco, CrowdStrike, PANW, Oracle, Zscaler, Akamai, Fortinet — is the tell. When a platform launches with every major incumbent listed as a friend, what you are looking at is pre-disintermediation rather than partnership. Cloudflare cutting twenty percent in the same week is one of those partners already reading the wind.

Cyber-AI is pricing as an AI feature and transacting as a defense-tech category. The Series B window closes on whichever framing the market agrees to first.

The EDR Moat is Cracking

TrustedSec ran LLMs at five commercial EDRs. Reverse engineering that previously took weeks now takes days, and all five products rest on the same architectural furniture — YARA rules, Lua engines, local ML classifiers. The detection IP that justified CRWD and PANW multiples is commoditizing in public. Microsoft's MDASH, meanwhile, shipped 16 validated Windows CVEs in a single Patch Tuesday, which is the first production evidence that autonomous vuln discovery works at hyperscaler scale.

Mozilla found 271 bugs in Firefox with a custom agentic harness on top of Claude Opus 4.6. Anthropic's Mythos found 1 real CVE in curl with a generic scan. The moat is in the harness, not the model. Fund accordingly.

The Category Winner Template

Abridge raised $550M in 2025 ($250M early year, $300M at a $5.3B mark) with 250 US health systems live and 80M+ patient-clinician conversations annually. That is a data moat no foundation model or new entrant can buy or scrape. Health systems — historically the slowest enterprise buyers — compressed release cycles from quarterly to monthly for this product.

The three-act structure is the investable template: save time (clinical documentation, won) → save money (prior auth, active) → save lives (clinical decision support, unlocked by January 2026 FDA guidance). Each act monetizes the same proprietary conversation asset against a different buyer persona.

The Underwriting Framework Update

Abridge validates three non-negotiable criteria for any vertical AI deal:

1. Per-use data flywheel — every customer interaction accrues proprietary training data competitors can't access
2. ≥2 monetization vectors beyond the initial wedge, targeting different buyer personas in the same account
3. Explicit non-compete posture with the dominant system of record (EHR, CRM, ERP) — companies pitching 'replacement' of incumbent platforms in regulated verticals are largely uninvestable at scale

The critical positioning: Abridge frames itself as a 'clinical intelligence layer' rather than an EHR replacement, treating Epic/Cerner as a filesystem. That neutralizes the single biggest existential risk in healthcare AI.

Vertical AI moats are built from proprietary interaction data, not model quality. Back the data flywheels, fade the AI wrappers, pre-empt the adjacent workflow layers before the same playbook prices them in.

Where the Adjacent Alpha Lives

The category-leader print doesn't close the healthcare AI book — it opens adjacent chapters:

• Payer-side prior authorization — Abridge attacks from the provider side; the payer counterparty remains greenfield with state-fragmented policies creating a parsing moat
• Nursing and short-form workflows — visits can be 30 seconds, a fundamentally different product surface
• Specialty verticals outside US large systems — veterinary, dental, behavioral health, international markets Abridge won't prioritize for 24+ months
• Picks-and-shovels — Kafka, Temporal, CRDTs, event-driven orchestration, memory stores, LLM judges — infrastructure serving multiple vertical AI winners