The Investor

The Flip Is Real. The Foundation Beneath It Is Not.

Ramp's April card-spend data puts Anthropic at 34.4% of US B2B spend versus OpenAI's 32.3%, the first documented lead change in any credible enterprise AI dataset. Over the trailing year Anthropic quadrupled business adoption while OpenAI grew three-tenths of a percent. The 0.3% deserves its own sentence. It is not a slowdown in a flat market. It is a stall while the category grew roughly four times around it.

Anthropic also admitted it planned for 10x growth and got 80x, then leased the entire Colossus 1 cluster (220K+ NVIDIA GPUs) from Elon Musk's xAI, the same Musk who publicly called Anthropic 'misanthropic and evil.' When a competitor rents compute from a sworn enemy, the supply situation is not a glut. It is a shortage that bends strategy, and what Anthropic is not doing with that capital is building the next cluster itself.

The Revenue Quality Problem Nobody Is Discussing

The share number obscures a structural fragility. ServiceNow, one of the most sophisticated enterprise software buyers on earth, blew through its full-year Anthropic budget by May 2026. Not because Claude underdelivered. Because Anthropic provides:

• Zero granular per-user, per-tool usage telemetry
• Zero enterprise SLAs worth the name
• No enterprise dashboard that would embarrass a mid-tier SaaS vendor from 2014

National Life Group's CIO put it without ornament: Anthropic is 'great for consumer usage but not great for companies.'

Enterprise AI ARR is not SaaS ARR. Switching costs are near zero, and the budget overruns are invisible until the invoice arrives.

The Deployment Gap Creates the Alpha

Every major player is now running the Palantir playbook. 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. The industry has quietly conceded that deployment is the bottleneck, not model capability.

Which means the $30B ARR figure is real but reversible. The quality of that revenue, measured by lock-in, telemetry, and contractual switching costs, is consumer-grade at best. There is a version where this gets cleaned up before a Q4 IPO, which is plausible and boring if it happens. There is a version where the gap persists long enough for an AI observability category to fill it, which is the more interesting one. This is probably wrong, but the second version is the base case. The bull case is not crazy. Read the next ServiceNow invoice before deciding which one you are underwriting.

What Happened

On May 12-13, Anthropic and OpenAI repriced the coding-agent market within hours of each other, which is the sort of coincidence that isn't. Anthropic converted every Claude subscription into a dollar-matched API credit pool, so a two hundred dollar plan now buys two hundred dollars of programmatic tokens, and the seventy to ninety percent arbitrage third-party harnesses had been running quietly disappears. OpenAI answered the same afternoon with two months of free Codex for enterprise switchers inside a thirty-day window.

Context: Ramp's April print showed Anthropic leading OpenAI for the first time, 34.4 against 32.3 percent, a new CFO is in the seat, and the IPO is probably October. The honest read is narrow. Margin recovery dressed as developer defense and timed to pre-IPO diligence.

Who Gets Hurt

Any coding tool running COGS against subscription tokens has lost twenty to forty percent of effective runway since Friday. The arbitrage was unsubtle: buy a two hundred dollar a month subscription, push thousands of API-equivalent calls through the UI, pay seventy to ninety percent below list. That trade is closed.

Meanwhile Claude Code shipped /goal and Auto Mode, which removes the human-in-the-loop steps entirely. The product surface that standalone autonomous agents (Cognition's Devin, Factory, Cosine) were selling against is now a free CLI primitive. The rounds those companies raised were priced on features that just became table stakes.

The Squeeze From Above and Below

This is probably wrong in the timing, but the cleanest framing is a two-sided squeeze. From above, Anthropic meters tokens and OpenAI subsidizes switches. From below, Notion's External Agents API now hosts Claude, Codex, Cursor, Decagon, Warp, and Devin inside the same workspace, which commoditizes the harness layer. Anyone thin-wrapping a frontier model now eats cost inflation and distribution compression at once.

The coding-agent thesis is now a duopoly subsidy fight with a commoditized harness layer beneath. Every portco priced on a Claude subscription arbitrage is worth less today than it was last Friday.

Three positions survive: proprietary workflow data, open-source distribution on Cline's model, or bounded enterprise execution on Cursor's security moat. The counter-thesis is that one of the harnesses quietly accumulates a real workflow datamoat before the subsidy ends, which is possible and not obvious. Anything that doesn't clear one of the three is a pass.

The Data Point That Changes the Thesis

Vercel published what is, to my knowledge, the first production-grade AI usage index across 200K+ teams, and the headline is that agentic workloads carry 59% of all token volume. This is not a lab benchmark or a self-reported ARR slide. It is downstream production data, which is rarer than it sounds. The chat-completion era is now a minority revenue line.

The interesting cut is underneath. Anthropic captures 61% of spend, mostly via Opus on long, tool-using agent calls, while Google takes 38% of volume via Flash on cheap, high-throughput commodity work. Two different businesses are now visible inside the thing we have been calling 'foundation models,' and they probably should not trade on the same multiple.

Incumbents Are Moving First This Time

The cloud transition was defined by startups, with the incumbents catching up late and overpaying. The pattern this round looks different. The ERPs and workflow platforms are moving first, which is worth flagging because it changes the resource math for anyone underwriting a greenfield platform play:

• SAP committed €100M to an Autonomous Enterprise partner fund wired into NVIDIA and Microsoft
• ServiceNow shipped Action Fabric, decoupling logic from UI and exposing workflows as headless APIs for agents to consume
• a16z published its system-of-intelligence thesis with a live Stitch investment, arguing the moat is shifting from data to orchestration

The Lemkin case makes it concrete. Seats fell from 10+ humans to 2 humans plus 1 API seat, while spend rose 83% ($12K→$22K) with 20+ agents running on top. Seat count collapsed. The bill went up.

Value in GTM software is migrating from data gravity to orchestration gravity, and the next 12-18 months is the last clean window to own the intelligence layer before incumbents or consensus close it.

Where the Alpha Sits

Three layers, three different capital dynamics, and only one of them is obviously defensible.

Abridge at $5.3B is the vertical template, or rather the most legible version of it: 250 health systems, 80M+ conversations a year, wedging from documentation into prior auth and clinical decision support. The moat is proprietary interaction data per use, not model quality. Health systems compressed release cycles from quarterly to monthly for them, which is faster than regulated buyers usually move.

This is probably wrong, but the timing question splits cleanly. a16z's thesis implies a 12-18 month window. ServiceNow's Action Fabric and SAP's fund suggest incumbents compress that to 6-9 months at the platform layer. The vertical layer holds longer, because incumbents will not prioritize narrow markets until someone embarrasses them in one.

AI security gets its first KEV entry

LiteLLM, the LLM-routing control plane, landed on CISA's Known Exploited Vulnerabilities catalog this week. That is the first time AI infrastructure has been federally flagged as actively exploited, which matters less for what it says about LiteLLM than for what it does to procurement budgeting. Separately, Anthropic's Mythos became the first model to clear both UK AISI simulated attack ranges, and the House Homeland Security Committee routed access through NSA, not CISA. That last detail is the tell. Offensive and IC-led procurement, not civilian defense.

Stack on Ollama's GGUF model-loader bug at CVSS 9.1, OpenClaw shipping six critical CVEs in a single cycle, and a PraisonAI auth-bypass weaponized within four hours of disclosure. The AI-infra stack is roughly where enterprise SaaS sat in 2014.

Two markets under one label

Sources describe a bifurcation in AI security economics that most investors are still pricing as one category:

OpenAI's Daybreak launch with 8 incumbent 'partners' (Cloudflare, Cisco, CrowdStrike, PANW, Oracle, Zscaler, Akamai, Fortinet) is the setup pattern of a platform preparing to disintermediate its partners. That move has played out in four prior waves of enterprise software. This is probably wrong, but the partners usually sign anyway, because the alternative is sitting outside the platform while a competitor sits inside it.

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

The Mozilla-vs-curl test

One quality signal worth keeping. Mozilla found 271 bugs in Firefox using a custom agentic harness on Mythos Preview + Claude Opus 4.6. Mythos alone found 1 real CVE in curl, and the maintainer called it "primarily marketing." Harness design beats model selection, or rather, the more honest version of that claim: any AI AppSec founder who cannot explain why their harness investment produces 271 bugs instead of 1 is selling the model, not the product. Use it as a diligence filter.