The Investor

The Revenue Quality Problem Nobody Is Pricing

Anthropic overtook OpenAI in enterprise billing share last month — 34.4% vs 32.3% on Ramp's April data — while quadrupling business adoption year-over-year against OpenAI's 0.3%. That is the headline a lot of people are quoting at each other this week. The footnote is more interesting: ServiceNow burned its entire annual Claude budget by May 2026, because Anthropic ships no per-user telemetry, no granular cost attribution, and no SLAs that survive a procurement review.

National Life Group's CIO put it plainly. Anthropic is 'great for consumer usage but not great for companies.' This is the firm the market is pricing at $900 billion on the explicit premise that enterprise is what justifies the number.

Enterprise AI ARR is not SaaS ARR. It reverses faster, has no contractual lock-in, and the buyer discovered this quarter that nobody was watching the meter.

Where the Money Actually Goes

The reaction from the platforms tells you where the next category forms, and what these firms are choosing not to spend the same dollar on:

• Google Cloud is hiring hundreds of forward-deployed engineers
• OpenAI/Bain stood up DeployCo and bought a 150-FDE consulting firm
• Salesforce and ServiceNow are staffing the same function
• ServiceNow is selling its own AI Control Tower to the customers panicking about Claude bills

When four firms independently conclude that the margin lives in deployment rather than the model, the margin probably lives in deployment. The Palantir playbook is now consensus — deployment is the bottleneck, not model capability.

The Category That Doesn't Exist Yet

AI observability and FinOps has the structural features of a Datadog-scale opportunity: token-level cost attribution, per-user spend caps, SLA monitoring across model APIs, anomaly detection on usage patterns. ServiceNow validated the need by building AI Control Tower in-house. No independent category winner exists. The window is 6-12 months before incumbent absorption closes it.

The June 15 Repricing Event

Anthropic's third-party credit unbundling on June 15 converts every Claude subscription into a dollar-matched API credit pool, ending the 70-90% arbitrage harness providers (Cline, OpenCode) were running on subscription-tier usage. OpenAI countered with 2 months of free Codex for enterprise switchers. Every Claude-dependent portfolio company's gross margin model is wrong in 30 days.

The Tension Worth Naming

This is probably wrong, but here is the view. Sources disagree on whether Anthropic's enterprise lead is durable. The bull case: quadrupling while OpenAI grew 0.3% means enterprise buying has structurally moved. The bear case: Ramp measures card billing, which skews SMB, the gap is only 2.1 points, and the next OpenAI release could flip it back — which has happened before in this category. The honest read is that both are true, and the zero-vendor-loyalty finding is the actual insight. Enterprise AI spend is reversible at a speed SaaS multiples do not reflect.

Production Data Arrives

Vercel just published the first production-grade AI Gateway index from more than 200,000 teams, and the numbers do something the thesis decks have been gesturing at for a year: they put a base rate on it. 59% of token volume is now agentic — tool-calling, multi-step, orchestrated. Chat completion, the thing most people still mean when they say 'AI use case,' is now the minority workload in production. That is the entire story, or rather, it is the boring half of the story.

The interesting half is the bifurcation. Anthropic captures 61% of spend on expensive Opus agentic calls. Google captures 38% of volume on cheap Flash throughput. Two different businesses are sitting inside the phrase 'foundation models,' and the enterprise default is to route between them. Single-vendor pitches are pricing in a world that already ended.

The Incumbents Are Defining the Category

In a single week:

The cloud transition had AWS and Azure defining infrastructure while incumbents caught up late. This time the ERPs and workflow platforms are moving first — the question of where agent-layer value accrues looks materially different as a result.

Where the Alpha Still Sits

This is probably wrong, but the picks-and-shovels layer still looks investable: MCP gateways, agent identity/auth, knowledge-graph tooling, agent observability, and governance. SAP's hundred million euros lands here eventually, which is the bull case and also the timer. The question is whether you are on the cap table at seed or Series A before corp dev activation compresses entry points to something less interesting.

The a16z GTM thesis reinforces this from a different angle. Lemkin's customer cut from ten human seats to two plus twenty agents, with spend rising 83%. The moat migrated from data gravity to orchestration gravity: institutional context, workflow memory, multi-system synthesis. That is the new lock-in, assuming the incumbents do not own it by default, which is the counter-thesis worth taking seriously.

What to Avoid

Standalone agent orchestration layers — the LangChain and CrewAI archetype — face the compression risk that always shows up when platforms absorb a category. Notion's developer platform launch could collapse horizontal agent-ops in two or three quarters. Vertical beats horizontal in agent ops from here. Also dead: any pitch where the moat is which model the company uses. The Vercel data says multi-model routing is standard and switching costs approach zero. That is not a moat. It is a rounding error.

Deals Worth Pricing This Week

AI security stopped being a narrative this cycle and started being a procurement line item, which is a less exciting framing and a more useful one.

1. LiteLLM's AI Gateway hit CISA's KEV catalog, the first time an LLM-routing control plane has been federally flagged as actively exploited. Call it the MongoDB ransomware moment for AI infrastructure, with the standard caveat that anything called a moment tends to age poorly.
2. Anthropic's Mythos cleared both UK AISI simulated attack ranges, the first model to do so, and the Congressional briefings are routing through NSA rather than CISA. That second detail is the whole story. Offensive and IC-led procurement is a different budget cycle than defensive, and a different sales motion entirely.
3. DepthFirst's Open Defense Initiative landed FFmpeg, Envoy, and Kata with a public claim of twelve memory corruption bugs for roughly a thousand dollars of compute against Anthropic's ten thousand across several hundred scans. A 10x cost efficiency headline against the frontier, or rather, the more honest version: a number that is either real or an artifact of which scans you choose to count, and we will not know which for two more quarters.

The EDR Moat Is Cracking

TrustedSec ran LLMs at five commercial EDRs and reported that reverse engineering that took weeks now takes days. All five products share the same architectural furniture: YARA rules and Lua engine plumbing, plus local ML classifiers trained on a malware corpus nobody publishes. In the same week OpenAI launched Daybreak with CrowdStrike, Palo Alto, Cisco, Zscaler, Akamai, Oracle, Cloudflare, and Fortinet listed as partners. When a platform launches with every incumbent as a friend, the friendship has a half-life.

The detection IP inside the security stack is commoditizing, and the platform layer is lining up to collect the value that leaves.

The Investable Segmentation

Sources Disagree on Timing

The bull case is that NSA access for Mythos routes first government dollars through IC budgets inside twelve months, and that LiteLLM on KEV creates defensible enterprise budget lines now. The bear case is that government procurement is slow and briefings are not contracts, while incumbent absorption (Daybreak absorbing its partners) may close before challengers scale distribution. The middle case, that both are true on different clocks, argues for positioning in the six to twelve month window before either resolves.