AnthropicEdgesOpenAIonRampasMeteringBecomestheTrade

◆ DEEP DIVES

1. 01

   Enterprise AI Revenue Is Consumer-Grade Plumbing at Enterprise Prices — And the Fix Is a Category

   The Problem Nobody Priced

   ServiceNow, which is approximately the most sophisticated enterprise buyer on earth, exhausted its full-year Anthropic budget by May 2026. Not because Claude underdelivered, but because Anthropic ships no per-user, per-tool usage telemetry and no SLAs worth the name. National Life Group's CIO said it without ornament: 'great for consumer usage but not great for companies.'

   This is the company the market is being asked to value at over nine hundred billion dollars on the premise that enterprise revenue justifies the number. The revenue is real. The revenue quality is not enterprise-grade.

   Every dollar of enterprise AI ARR is being booked on infrastructure that would have embarrassed a mid-tier SaaS vendor in 2014.

   The Market Response: FDEs Become Consensus

   Four firms reached the same conclusion this month, independently, which is the kind of coincidence that usually isn't one: deployment is the bottleneck, not model capability. Google Cloud is hiring hundreds of forward-deployed engineers. OpenAI stood up DeployCo with Bain Capital and acquired a consulting firm for its 150-FDE starting roster. Salesforce and ServiceNow are staffing the same function. When four firms independently decide the margin is in deployment services rather than the model, the margin is probably in deployment services.

   The Category That Forms From This

   ServiceNow is already selling AI Control Tower to the same customers panicking about their Anthropic bills, which is the CDIO watching the category form inside her own P&L and deciding to be the vendor rather than the line item. Combined with Modal at four and a half billion dollars and the 'tokenmaxxing' vocabulary entering enterprise procurement, the category has shape:

   • Token-level cost attribution (per-user, per-tool spend caps)
   • SLA monitoring across model APIs
   • Usage observability of the kind that would have prevented the ServiceNow budget-exhaustion pattern
   • Multi-model routing governance

   No independent winner exists yet. The window is six to twelve months before ServiceNow or an incumbent locks it down.

   The Portfolio Implication

   Any LLM-layer portco whose revenue looks like ARR is, on closer reading, carrying a reversibility discount the mark doesn't reflect. No SLAs, no telemetry, no contractual lock-in adds up to enterprise AI spend that reverses the moment cost-efficient alternatives land. The honest answer is a twenty to forty percent haircut where these features are absent.

   ---

   The counter-thesis is that Anthropic fixes this before the October IPO; they hired a CFO this month and the credit-matching pricing is step one. Possible. The deployment-services layer gets built regardless of whether one lab fixes its plumbing, because every model lab has the same gap.

   Action items

   • Launch sourcing sprint on AI observability/FinOps at Seed-Series A — target token attribution, SLA monitoring, multi-model governance
   • Demand SLA and usage-telemetry roadmap from every model-layer portco claiming enterprise ARR — make it a board agenda item
   • Apply 20-40% 'reversibility discount' to any LLM-layer mark lacking SLAs, telemetry, or contractual switching costs
   • Map Palantir alumni network for FDE-layer investment targets and deployment-services diligence advisors

   Sources:Laura Bratton · The Pragmatic Engineer · Martin Peers · Bloomberg Technology · The Information AM

2. 02

   Anthropic's Pricing Trap: Wrapper Economics Collapse in a Week

   The Two-Sided Squeeze

   On May 12-13, Anthropic and OpenAI repriced the coding-agent market within hours of each other, which is either a coincidence or — to use the more honest description — two firms watching the same calendar. Anthropic converted every Claude subscription into a dollar-matched API credit pool, so a $200 plan now buys $200 of programmatic tokens and not a token more. That kills the 70-90% arbitrage that Cline, OpenCode and OpenClaw were quietly running against subscription-tier usage. OpenAI answered the same afternoon with two months of free Codex for enterprise switchers.

   One layer up, Notion's External Agents API now hosts Claude, Codex, Cursor, Decagon, Warp and Devin inside the same workspace, which is what commoditization of the interface layer looks like when it happens in public. The wrapper business got squeezed from both ends in under a week.

   Any portfolio company running COGS against subscription tokens lost 20-40% of runway since Friday. The change is four days old. Most founders haven't flagged it yet.

   The IPO Clock Explains the Timing

   Anthropic hired a CFO and is targeting an October IPO, and the credit-matching move is margin recovery wearing the costume of a policy update. Ramp's April data put Anthropic at 34.4% of business spend vs OpenAI at 32.3%, which is the first documented lead change and not the sort of number you leave lying around before a roadshow. The pricing offensive is timed to pre-book cleanup. OpenAI's free-Codex counter is the obvious move; both firms are buying developer lock-in ahead of their respective events, and neither is pretending otherwise.

   Who Gets Hurt, Who Benefits

   Category	Impact	Action
   Claude-wrapper dev tools	COGS up 3-5x overnight	Request updated cohort gross margin by month-end
   Multi-model routing infra	Beneficiary — enables cost arbitrage	Priority sourcing category
   Open-source harnesses (Cline)	OSS backlash play validates	Watch for enterprise traction signal
   Cursor/bounded-execution tools	Security-as-moat thesis strengthens	Mark up if holding

   The Forward View

   Anthropic's June 15 unbundling of third-party credits is the next shoe, and post-October the subsidy economics stop being economics at all. The three scenarios for app-layer holdings: the duopoly absorbs the layer above and multiples compress, the duopoly keeps subsidizing and the layer above gets a free input-cost cut, or the IPO prices the whole thing and the subsidy stops the week after. This is probably wrong, but the third case is the one that actually resolves the argument. It is also five months away.

   Action items

   • Contact every Claude-dependent portco CEO by Friday — request updated gross margin model assuming API-rate billing, not subscription-tier
   • Model Anthropic secondary entry at sub-$700B before the October IPO book-building firms up pricing
   • Downgrade any thin-wrapper AI deal in pipeline that can't demonstrate proprietary workflow data, OSS distribution, or enterprise-grade bounded execution
   • Add June 15 Anthropic credit unbundling as a portfolio monitoring trigger — model second-order margin impact across all Claude-exposed positions

   Sources:AINews · TLDR AI · ben's bites · Daily Dose of DS · Laura Bratton

3. 03

   GTM Software Value Migrates to the Intelligence Layer — a16z Deploys the Thesis

   The Deployed Thesis

   a16z published its GTM software map and wrote a check into Stitch in the same week, which is the kind of thesis that tells you what a firm actually believes rather than what it says at conferences. The claim is that the system of intelligence layer captures the majority of the next decade's GTM enterprise value, leaving the system of record (Salesforce at one hundred forty billion dollars, HubSpot at nine billion) as plumbing. The moat migrates from data gravity, where the customer record lived in Salesforce, to orchestration gravity, where the workflows and reasoning state live in the AI layer. That is a real claim. It has also been made about every middleware layer of the last twenty years, with mixed results.

   The Proof Point

   Lemkin's SaaStr anecdote is doing most of the work here: Salesforce went from 10+ human seats to 2 humans + 1 API seat, spend rose 83% ($12K → $22K), and twenty-plus agents now run on top. Seats collapsed and the bill went up. That is the consumption pitch in one customer.

   CRM usage is rising, not falling, since AI adoption began. The CRM becomes infrastructure consumed at the API layer — still valuable, still growing in usage, just no longer the value-capture point.

   Incumbents Moving First

   The cloud analogue everyone wants to draw is wrong in one important respect. AWS and Azure defined infra while incumbents argued about it for five years. This time the ERPs and workflow platforms are moving first, which is either evidence the incumbents finally learned something or evidence the threat is overstated:

   • SAP: a one hundred million euro Autonomous Enterprise partner fund, NVIDIA and Microsoft wired into the platform
   • ServiceNow: Action Fabric decouples logic from UI and exposes workflows as headless APIs for agents
   • Salesforce: API-first counter-punch to protect the record layer

   This is validation and compression in the same press release. Every agent-infra startup acquired potential buyers and lost moat on the same morning. Those are not symmetric events, and the second one usually wins.

   Where the Alpha Sits

   The investable window is roughly 12-18 months before consensus and incumbents close it, which is also the window in which one is most likely to be wrong in interesting ways:

   1. Vertical orchestration with institutional-context accumulation, by which I mean narrow high-frequency workflows with measurable outputs rather than horizontal toys
   2. Consumption-pricing GTM plays that can show the Lemkin template (10x agent deployment, 80% seat reduction, 80%+ spend increase)
   3. Agent-native security, where 81% bot-detection bypass by AI agents is a structural break requiring new detection primitives

   What to avoid: horizontal AI CRM wrappers, thin copilots without workflow moats, any pitch that confuses model quality with defensibility. Salesforce's API-first counter-punch hits those first.

   Vercel Production Data Confirms

   Agentic workloads are now 59% of production token volume across two hundred thousand-plus teams. Anthropic captures sixty-one percent of spend while Google captures thirty-eight percent of volume, which are two different businesses sitting inside one category called foundation models. Multi-model routing is the enterprise default. The single-model moat was a 2023 thesis.

   Action items

   • Rerank AI-GTM pipeline by orchestration moat depth — prioritize narrow workflows with measurable outputs over horizontal copilots
   • Request agent-to-seat ratio, API call growth, and NRR on existing logos in all active AI-GTM diligence
   • Source 3-5 agent infrastructure deals in MCP tooling, agent identity, and agent observability before SAP's corp dev activates
   • Stress-test every seat-based SaaS position using Levie's 'multiplicative seat value' thesis — model seat value at +30% and -30% scenarios

   Sources:a16z · TLDR IT · ben's bites · Laura Bratton · TLDR AI

4. 04

   Vertical AI Moats: Abridge Proves the Data Flywheel Template

   The Category Winner Print

   Abridge raised $550M in 2025 — two hundred and fifty million early in the year, three hundred million in June at a five point three billion dollar mark — and is now servicing 250 large US health systems with 80M+ projected patient-clinician conversations this year. That is not a Series B story. It is a corpus that compounds every time a doctor opens their mouth.

   The genuinely interesting signal, which most of the cap-table commentary is burying, is that health systems — the slowest enterprise buyers on earth, by a margin that has held for two decades — compressed release cadence from quarterly or biannual to monthly for Abridge. Every healthcare SaaS DCF written before this year is using the wrong velocity assumption.

   The Template for Vertical AI Investing

   Abridge's three-act structure is now the investable pattern, or rather the pattern people will claim to have seen coming:

   Act	Wedge	Buyer	Status
   1. Save Time	Clinical documentation	CMIO	Won by Abridge
   2. Save Money	Prior auth, billing	CFO	Active expansion
   3. Save Lives	Clinical decision support	CMIO/CIO	Unlocked by Jan 2026 FDA CDS guidance

   Vertical AI moats are built from proprietary interaction data, not model quality. 80M conversations × 250 systems × 28 languages × 50 specialties is a corpus no foundation model or new entrant can buy or scrape.

   The Non-Negotiable Criteria

   Three requirements for any vertical AI deal going forward, and we will be wrong about at least one of them within a year:

   1. Per-use data flywheel — every customer interaction accrues proprietary training data competitors cannot 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)

   Abridge satisfies all three by framing itself as a 'clinical intelligence layer' rather than an Epic or Cerner replacement. That positioning is table stakes now for any regulated-vertical AI deal worth the diligence.

   Adjacent Opportunities Still Open

   The ambient scribe category inside US large health systems is closed. What this means for allocation is that capital chasing it from here is buying a position someone else already owns. The remaining alpha sits in:

   • Payer-side prior authorization — Abridge attacks from the provider side; the payer counterparty is greenfield
   • Nursing and short-form clinical workflows — a 30-second visit is a different product surface, not a feature flag
   • Specialty verticals outside US — vet, dental, behavioral health, international
   • Clinical decision support post-FDA guidance — Jan 2026 CDS guidance unlocks the 'save lives' tier

   Action items

   • Kill or de-prioritize any active deals competing directly with Abridge in US large health system ambient scribing — redirect to payer-side prior auth and specialty verticals
   • Build thesis memo on payer-side prior-auth automation with 8-12 startup target list within 30 days
   • Update vertical AI underwriting framework to require per-use data flywheel, ≥2 monetization vectors, and non-compete posture vs. dominant system-of-record
   • Track Abridge secondary market activity — pre-empt any tertiary-round opportunity at $8-12B given likely IPO re-rate of entire vertical AI book

   Sources:Latent.Space · Laura Bratton