The Board Room

Anthropic Now Owns Enterprise AI — And the Data Is Unambiguous

The enterprise AI market has undergone its most significant power shift since OpenAI launched ChatGPT. Anthropic now commands 40% of enterprise AI spending while OpenAI has cratered from roughly half to 27%. This isn't a temporary fluctuation — it reflects a structural failure in OpenAI's product strategy. Fidji Simo's internal memo acknowledging 'spreading our efforts across too many apps' (Sora, Atlas, Prism) is the rare corporate admission that amounts to: we lost our focus, and now we're losing the market.

The partnership that underpinned 80% of enterprise AI procurement decisions — Microsoft + OpenAI — is no longer a safe assumption.

Model Makers Are Eating the Tool Layer

The AI coding market has crossed $5.5B ARR across three players: Claude Code at $2.5B+, Cursor at $2.0B+, and Codex at $1.0B+. The critical insight isn't the revenue — it's that model makers are winning against tool builders. Notion migrated hundreds of engineers from Cursor to Claude Code and Codex because engineers increasingly argue that the companies who build the models are best positioned to build the harness around them. Junior engineers gravitate to Claude Code for intuitive task completion; senior engineers prefer Codex for 8-hour autonomous sessions running overnight.

Cursor's response — releasing Composer 2, built on Chinese startup Moonshot's open-source Kimi 2.5 — compounds its positioning problem. This is the platform-eats-the-app-layer dynamic that has played out in every prior technology cycle, happening faster than expected.

Meta's Revealed Preference Is the Strongest Signal

Perhaps the most devastating competitive signal this week: Meta's internal executive tools — MyClaw and Second Brain — run on Anthropic's Claude, not Meta's own LLaMA models. When one of the world's most sophisticated AI companies chooses a competitor's model for its own mission-critical agentic tools, that's a $2 billion data point for your vendor evaluation. Meanwhile, OpenAI's advertising model is failing badly — 0.91% CTR versus Google's 6.4% benchmark — revealing that conversational AI may not be an advertising medium at all, narrowing OpenAI's monetization path to subscriptions and enterprise licensing.

What This Means for Your Vendor Strategy

The stable, two-player enterprise AI market of 2024-2025 is over. What's emerging is a fragmented landscape where:

• Anthropic leads enterprise coding and productivity (40% spend share, growing)
• OpenAI is pivoting defensively to a superapp consolidation play (high execution risk)
• Model commoditization from below: MiniMax M2.7 delivers 90% of frontier quality at 7% of cost
• The Microsoft-OpenAI axis is fracturing — Microsoft building its own frontier models, OpenAI distributing through AWS for classified workloads

The organizations that win aren't those that pick the right vendor — they're those that build multi-vendor orchestration capability and measure cost-per-completed-task, not cost-per-token.

The Two-Path Ultimatum

David George at a16z has published what amounts to a strategic ultimatum for the entire software industry. Only two paths create durable equity value: Path 1 is AI-driven growth acceleration — adding 10+ percentage points of revenue growth within 12 months through net-new AI products. Path 2 is radical margin expansion to 40-50% true operating margins including SBC within 12-24 months. The Broadcom/VMware playbook — where Hock Tan drove 61% adjusted EBITDA margins through radical simplification — is the explicit template.

Companies that answer 'a little of both' are choosing a third path that leads to persistent multiple compression and value destruction.

The Prescription Is Unusually Specific

What makes this framework operationally consequential is a16z's granularity. This isn't strategy-deck abstraction — it reads like a playbook they've already deployed across their portfolio:

• Four-person pods collapsing design, product, and engineering — writing code on day one
• 50% of R&D allocated to net-new AI products
• $1,000/month/engineer token budget as 'close to table stakes'
• Identify ~5 people delivering 100x expected value regardless of seniority — give them leadership
• 30-day information-gathering sprint followed by watching which VPs engage and replacing those who don't
• Full machine redesign — not 8-10% layoffs but complete organizational restructuring

The timing is critical: a16z publishing this publicly means your competitors — especially a16z-backed ones — are likely already executing. Companies that commit in Q2 2026 will complete a transformation cycle before those that deliberate through Q3-Q4.

The Credit Markets Are Confirming the Thesis

The most alarming corroboration comes from private credit. Multiple major funds are gating redemptions after unusually high withdrawal requests, triggered by AI systematically weakening the SaaS lending thesis. The logic chain is devastating:

1. Sticky revenue becomes less sticky when AI replicates software functionality at marginal cost
2. Strong margins compress when AI-native competitors don't carry legacy headcount
3. Durable switching costs erode when AI makes migration trivial

This isn't theoretical — it's showing up in fund performance and redemption patterns today. The second-order effect: credit facilities, venture debt, and growth financing priced on recurring-revenue multiples will become more expensive and harder to access. CFOs who get ahead of this have significant negotiating advantage.

The Moat Erosion Thesis

The deeper strategic signal: traditional software moats — proprietary data, integration complexity, workflow lock-in, migration friction — are all weakening simultaneously as AI agents navigate across systems and reproduce integrations faster. If true, the entire SaaS valuation framework (negative churn, long customer lifetimes built on switching costs) needs recalibrating downward. Seat-based pricing is now the #1 cost line your customers will target for AI savings — and new budget growth flows to tokens, consumption, and outcome-based models.

The Scaling Law for AI Cyberattacks Is Now Measured

The UK AI Security Institute built purpose-built cyber ranges — simulated corporate networks and industrial control systems — and tracked AI model performance across generations. The results are sobering: from GPT-4o in August 2024 to Opus 4.6 in February 2026, average steps completed on a 32-step corporate network attack jumped from 1.7 to 9.8. The best single run completed 22 of 32 steps — roughly 6 of the 14 hours a human expert would need. Perhaps most critically, simply scaling inference-time compute from 10M to 100M tokens yields up to 59% additional performance gains.

This isn't a capability that requires a breakthrough to become dangerous; it's on a smooth, predictable improvement curve. Fully autonomous cyberattacks against production corporate infrastructure appear plausible within 1-2 model generations.

The MCP Paradox: Better Models = Worse Security

The AgentSeal research on MCP servers reveals a structural flaw in the agentic AI buildout: 10.8% of 5,125 scanned MCP servers contain toxic data flows where individually benign tool pairs combine into exploitable chains. The MCPTox benchmark finding that o1-mini follows prompt-injected malicious instructions 72.8% of the time — and that more capable models proved more susceptible — represents a fundamental architectural constraint. You cannot solve this by shipping a better model. The attack surface grows quadratically with tool count.

The Ecosystem Is Compromised At Multiple Layers

The threat surface has expanded across the entire stack simultaneously:

The Trivy supply chain attack deserves particular weight because it targets a security scanner embedded in countless CI/CD pipelines with privileged access to build environments. The attack's upgrade to encrypted C2 means traditional network monitoring would miss it entirely. This signals a broader shift: the scanning paradigm that has dominated vulnerability management for a decade is failing against adversaries who can compromise the scanners themselves.

Domain-Specific AI Changes the Adversary Calculus

Chinese researchers — including affiliates of the National University of Defense Technology — built MERLIN, a multimodal LLM for electronic warfare trained on just 100,000 specialized data pairs. It outperformed GPT-5, Claude-4-Sonnet, Gemini-2.5-Pro, and DeepSeek on reasoning tasks by wide margins. The business implication: domain-specific models trained on modest but high-quality data can decisively beat trillion-dollar frontier models. Applied to offensive security, this means adversary capability is no longer bounded by access to frontier models — it's bounded by access to domain-specific training data, which is far more widely available.