The Investor

The New Data That Changes the Calculus

Three developments landed in the last 24 hours that transform Anthropic from a well-covered story into an actionable mispricing. First, a federal appeals court in D.C. upheld the Pentagon's 'supply-chain risk' designation, with judges explicitly citing 'an active military conflict' — language that makes reversal before 2027 unlikely. Second, Michael Burry disclosed a short position against Palantir, arguing Claude is 'eating Palantir's lunch' in enterprise AI. Third, a $600B secondary SPV attempted to form — falling through, but establishing demand at a 71% premium to last round.

The resulting profile is unprecedented: the same company is simultaneously locked out of all Pentagon contracts and cited by one of history's most famous short sellers as the reason to bet against a $150B+ government AI incumbent.

The Valuation Arbitrage Is Quantifiable

At its $350-380B valuation on $30B annualized revenue, Anthropic trades at roughly 11.7-12.7x revenue. OpenAI sits at $730B on ~$25B — a 29.2x multiple. That's a 2.5x gap between the two frontier AI leaders, with Anthropic growing faster (3x in one quarter), generating higher-quality enterprise revenue (1,000+ $1M+ customers), and locking in 3.5GW of TPU compute.

The market is pricing OpenAI's brand and consumer distribution at 2.5x the value of Anthropic's enterprise revenue machine. Burry is betting that equation is backwards.

Anthropic's new $200M commitment to a $1B joint venture with Blackstone, General Atlantic, and Hellman & Friedman is the distribution signal most investors will miss. This PE channel gives Claude embedded access to thousands of portfolio companies simultaneously — the same playbook OpenAI already runs. When both frontier labs independently converge on PE distribution, that's a structural go-to-market shift, not a partnership announcement.

The Government Risk Is Real But Bounded

The appeals court language matters: 'On one side is a relatively contained risk of financial harm to a single private company. On the other side is judicial management of how the Department of War secures vital AI technology during an active military conflict.' This framing subordinates corporate interests to national security in a way that creates precedent risk for every AI company — not just Anthropic.

But the government TAM Anthropic loses (~$15-20B addressable in defense) may be more than offset by concentrated enterprise focus. The contrarian thesis: Anthropic locked out of government becomes the most focused enterprise AI company in the world, and secondary market pricing drops 15-25% while commercial trajectory strengthens. May 19 oral arguments and the California litigation resolving are the catalyst dates.

What OpenAI's $102B Ad Revenue Forecast Tells You

Buried in this week's intelligence: OpenAI is now forecasting $102B in advertising revenue by 2030. If you're holding OpenAI secondary at 29.2x, ask what you're actually valuing — a SaaS business or a blended model-provider/ad-tech company. The ad revenue pivot fundamentally changes the comp set from pure SaaS multiples to a lower-multiple hybrid. Combined with the valuation premium versus Anthropic, this warrants a hard reassessment.

The Revenue Proof Point the Market Needed

Perplexity hit $450-500M ARR after a 50% monthly surge — driven specifically by its Computer agentic product launched in late February. The mechanism is critical: subscription credits for agent usage create a consumption-driven upgrade flywheel where heavier agent use triggers plan upgrades. This isn't chatbot revenue — it's usage-based pricing that scales with value delivered.

Combined with Cursor at $2B ARR in coding, we now have two independent proof points that vertical AI agents can build venture-scale businesses beneath the foundation model layer. The revenue stratification is clear:

The Commoditization Bomb From Above

In the same cycle, Anthropic launched Claude Managed Agents at $0.08/hr — containerized execution with sandboxing, checkpointing, and scoped permissions. Early adopters include Notion, Rakuten, Asana, and Sentry. Rakuten reportedly deployed agents across five departments in about a week each.

When a foundation model provider makes your startup's core value proposition a feature priced at eight cents an hour, the term sheet math changes fast.

This creates a barbell dynamic: agent orchestration middleware gets crushed from above, but vertical agents with proprietary workflow data and domain expertise appreciate. Cursor proves the pattern — it coexists with GitHub Copilot because its moat is workflow integration, not model capability. The same logic applies to agents in tax, legal, healthcare, and compliance — domains where accuracy requirements are highest and general chatbots fail catastrophically (TaxSlayer's test found $2,000+ average errors from general chatbots on tax scenarios).

Vertical AI Has Captured the Majority of New Company Formation

Vertical AI now represents 53% of 2025 VC deal volume, over half of exits, and a staggering 60% of earliest-stage startup formation. The breakout verticals: Manufacturing (strong deal growth, massive TAM), Legal (breakout financings, regulatory moat), and AEC (fastest-growing category, lowest AI penetration). The investable filter: does the company have a genuine orchestration harness generating proprietary workflow knowledge from execution data? Or is it a thin wrapper that Anthropic subsumes at $0.08/hr?

Beyond Mythos: The Threat Is Now Quantified and Commoditized

We covered Anthropic's Mythos capabilities earlier this week. What's genuinely new: Sequoia-backed Buzz demonstrated that an AI agent built from publicly available models (Anthropic + OpenAI + Google) autonomously exploited 103 of 122 known vulnerabilities (84.4%) in under an hour. The React2Shell vulnerability — one of 2025's most dangerous — fell in 22 minutes. No novel techniques. No restricted model access. Just commercial APIs orchestrated against CISA's Known Exploited Vulnerabilities catalog.

This is the proof point that changes procurement conversations: the offensive capability isn't locked behind Anthropic's consortium. Anyone with API access can build an autonomous exploit agent today. Chevron CISO Jon Raper confirmed the paradigm shift: 'Finding vulnerabilities isn't the problem — it's remediating them in time.'

Gartner Creates a New Budget Line Item

Separately, Gartner formalized IVIP (Identity Visibility and Intelligence Platforms) as a new framework — a continuous discovery layer above access management. The data is stark: 46% of enterprise identity activity occurs outside centralized IAM visibility, and orphaned accounts represent 40% of observed environments. Orchid Security published the quantifying analysis. When Gartner creates a category, enterprise budgets follow within 12-18 months.

Supply Chain Attacks Hit Sovereign Targets

The TeamPCP supply chain campaign weaponized Aqua Security's Trivy scanner to breach the European Commission — 340 GB exfiltrated, 52,000 email files exposed, 71 clients across 42 EC departments compromised. Wiz documented a 24-hour post-compromise operational tempo. Separately, the FBI confirmed US cybercrime losses hit a record $21B in 2025, with investment scams ($8.6B) and crypto theft ($6.2B) as dominant categories. Infostealers surged 59% YoY with 1M+ banking credentials circulating and 74% of compromised cards still valid.

The cost of finding vulnerabilities just went to near-zero. The cost of exploiting unfixed ones just went to infinity. Every cybersecurity dollar should be re-underwritten against this new equilibrium.

The Investable Categories Are Now Clear

The Risk No One Is Modeling

The most underappreciated signal from this week's HumanX conference wasn't a revenue number — it was a behavior pattern. 'Tokenmaxxing' — Meta employees competing on internal leaderboards to show how many AI tokens they consume — exposes a demand authenticity problem that could affect the entire AI revenue stack. Databricks CEO Ali Ghodsi called it partially 'performative' while noting coding use cases generate real productivity. Vinod Khosla countered with his 'massive gap between capability and deployment' thesis, arguing current valuations are justified by early adoption curves.

The tension between these views is the insight. If Khosla is right, the gap closes and valuations are cheap. If Ghodsi is right, a meaningful percentage of enterprise AI token consumption is driven by internal signaling rather than value creation, and the revenue base sits on softer ground than it appears.

How to Stress-Test Your Portfolio

Every AI company's revenue pitch depends on enterprise adoption curves that assume token consumption = value delivered. Tokenmaxxing breaks that assumption. It doesn't mean all demand is fake — Ghodsi explicitly validates coding use cases. But investors need to distinguish between two metrics:

• Consumption metrics: tokens processed, seats activated, API calls made
• Productivity metrics: time saved, code shipped, decisions automated, revenue generated

Companies that can demonstrate measurable workflow outcomes hold up in a correction. Those selling token volume face a reckoning.

If even 15% of enterprise AI token consumption is performative and gets curtailed, the entire AI revenue stack reprices — and the companies reporting 'consumption-driven growth' become the first casualties.

The Sector-Wide Valuation Check

Three AI application companies tripled valuations simultaneously in 12 months: ElevenLabs ($11B), Lovable ($7B), and Fireworks AI ($4B). This correlated re-rating across distinct verticals is a sector-wide phenomenon driven by capital inflows, not company-specific outperformance. The tokenmaxxing risk adds a new question to every valuation: how much of the usage metrics driving these multiples reflects genuine enterprise value versus organizational performance theater?

The practical framework: build a 'durable demand' scorecard for every AI holding. Weight retention and expansion revenue over raw token consumption. Companies with customer-reported productivity gains and measurable workflow outcomes get premium treatment. Companies selling 'engagement' and 'token volume' are your trim candidates before Q2 earnings.