The Investor

The Revenue Quality Alarm

AI coding tools just delivered the fastest value creation event in software history: approximately $6.5 billion in combined ARR across Claude Code (~$2.5B), OpenAI Codex (~$2B), and Cursor (~$2B) — built in roughly twelve months. But a phenomenon called "tokenmaxxing" is materially inflating these numbers, and the market hasn't priced it in.

At Meta, an internal leaderboard dubbed "Claudeonomics" tracked AI token consumption across 85,000+ employees. Top users earned titles like "Token Legend" and "Session Immortal." The result: 60.2 trillion tokens consumed in 30 days, estimated at $100M+ even at heavy volume discounts. At Salesforce, leadership set minimum weekly spend targets — $100/week on Claude Code, $70/week on Cursor — with a Mac widget updating every 15 minutes and a web tool to browse colleagues' spend. Microsoft has maintained a similar leaderboard since January 2026, where VP-level executives who rarely write code appear in the top 20.

Enterprise AI usage metrics are the new vanity metrics: mandated consumption floors create guaranteed vendor revenue that looks like organic PMF but isn't.

Why This Matters for Your Portfolio

The implications cascade across every AI coding investment. Anthropic, Cursor, and GitHub Copilot are all benefiting from demand that is partly manufactured by corporate mandates, not organic developer preference. Cursor is the only player demonstrating unsubsidized user preference for in-house models — and even Cursor benefits from Salesforce's $70/week minimums. OpenAI and Google are heavily subsidizing usage, further distorting true demand signals.

Meanwhile, Anthropic's rationing of individual users while prioritizing enterprise accounts — and GitHub freezing Copilot signups because costs doubled YTD — reveals a supply-demand tension that tokenmaxxing makes worse. The vendors are capacity-constrained serving waste.

The Contrarian Angle

One long-tenured Meta engineer suspects the leaderboard was deliberate data generation strategy. If Meta uses 60.2T tokens/month of real-world coding traces to train its next-gen coding model, the $100M+ monthly cost is an R&D investment in proprietary training data no competitor can replicate. If true, Meta's waste is actually a moat — and the resulting model could challenge Codex and Claude Code directly.

The Investable Category: AI FinOps

Shopify offers the counter-model: renamed leaderboards to "usage dashboards," implemented circuit breakers for runaway agents, and conducts per-token cost analysis. This is exactly the infrastructure every enterprise will need. The parallel to cloud cost optimization is exact — that market spawned $2B+ in value (CloudHealth for $500M, Spot.io for $450M). AI cost governance is at the same inflection point right now.

The Distribution Innovation

OpenAI is investing up to $1.5 billion into a PE joint venture called DeployCo, valued at $10 billion, with TPG, Bain Capital, Advent International, Brookfield, and Goanna Capital contributing an additional $4 billion. Anthropic is running a parallel play with Blackstone and Hellman & Friedman. This is the most consequential enterprise AI distribution shift since cloud marketplaces.

The mechanism is elegant: PE firms mandate AI tool adoption across their portfolio companies, AI labs get distribution without building enterprise sales teams, and the JV captures the economics. OpenAI's initial $500M commitment — with an option for $1B more at the same $10B valuation — is capital-efficient customer acquisition dressed as an investment. If even 20% of the combined PE portfolios adopt these tools, we're talking about thousands of enterprise accounts acquired through a single channel.

AI labs partnering with PE firms to mandate adoption across portfolio companies is the most consequential enterprise distribution innovation since cloud marketplaces.

The SaaS Repricing Is Already Happening

ServiceNow just posted 22% revenue growth to $3.77B, raised full-year subscription guidance to ~$15.76B, and its stock still cratered 13% after-hours on top of a 30%+ YTD decline — a combined 43%+ drawdown. The $7.8B Armis acquisition compressed operating margin to 31.5%. But the real story is the 18-point gap between non-GAAP operating margin (31.5%) and GAAP margin including stock compensation (~13.5%). The market is pricing AI disruption risk into enterprise software incumbents regardless of current fundamentals.

This creates a pincer movement: DeployCo pushes AI adoption into PE-owned enterprises from the top down, while AI agent platforms (OpenAI Workspace Agents, Google Gemini Enterprise) attack incumbent workflow tools from the bottom up. Horizontal SaaS companies without deep integration moats face structural demand erosion on both fronts.

What the Market Isn't Pricing

Google's $750M consulting fund and $1B Merck deployment deal confirm the enterprise AI implementation gap is widening, not closing. Merck's CIO put it bluntly: "The gap between what companies are able to do and what the technology allows is getting bigger and bigger." This creates a paradox: DeployCo will generate demand, but enterprises still can't deploy what they buy. The companies that bridge this gap — AI services enablement platforms with SI partnerships — capture pull-through demand from both the PE adoption wave and the implementation gap.

AmEx Validates the Acquisition Playbook

American Express acquiring Altman-backed Hyper for corporate expense automation proves the pattern is repeatable: build AI automation for a specific financial workflow, secure AI-leader backing, sell to an incumbent. Block's Goose framework powering dual-sided agents (MoneyBot for consumers, ManagerBot for merchants) shows the vertically integrated approach. The 12-18 month window to build or buy an agent position is closing.

Agents Are Now the Primary Software Buyer

A paradigm shift in software distribution is forming across multiple independent data points, and most investors are still evaluating companies as if humans are the customer. 60% of Vercel's admin traffic is now bots. Claude recommends Resend ~70% of the time as the default email provider. Supabase is the default agent-recommended Postgres database. When AI agents become the primary evaluators of software products, they bypass UI, onboarding, brand, and traditional GTM entirely.

The implications are structural: companies that agents don't recommend by default will see customer acquisition costs approach infinity. The window to get into AI training data is closing — this is a now-or-never moat for developer tools and infrastructure companies. A new category called AEO (Agent Engine Optimization) is emerging, analogous to early SEO. WordPress.com is already positioning around it with server-rendered pages and semantic markup specifically for AI crawlers.

Products that win the next cycle won't be the most beautiful or the best-branded — they'll be the most machine-readable, API-first, and protocol-compatible.

Fintech's Agent Layer Is Forming Simultaneously

Five major fintech incumbents launched AI agent strategies in the same news cycle, confirming this is a category birth:

The pattern across both dev tools and fintech is identical: proprietary data + agent framework = compounding moat. Startups building app-layer products without proprietary agent frameworks face a rapidly closing window. Block's Goose, Alipay's OpenClaw, and Coinbase's agentic wallets create infrastructure that compounds with usage.

The Enterprise Agent Platform War

OpenAI, Google, and Microsoft all launched enterprise agent platforms in the same week — Workspace Agents, Gemini Enterprise Agent Platform, and token-based Copilot billing respectively. This simultaneous launch signals the enterprise agent market is entering its land-grab phase. OpenAI's pivot from the failed GPT Store to Workspace Agents is a tacit admission that value accrues at the workflow layer, not the distribution layer — bearish for horizontal AI agent marketplaces.

Microsoft's shift to token-based Copilot billing ($19/user with $30 credits for Business, $39/user with $70 credits for Enterprise) is the first major SaaS pricing model break from per-seat to consumption. This creates a new category need: enterprise AI spend management — the Datadog moment for AI consumption.

Where to Invest

The investable whitespace is in three layers: (1) Agent security and identity management — who manages credential lifecycle and audit trails for autonomous agents? Nobody yet. (2) Agent discoverability infrastructure — the AEO toolchain for making products agent-discoverable. (3) Vertical agent specialists with proprietary data — the only startup-accessible wedge as horizontal platforms get bundled by incumbents.