The Board Room

Deep dives

1. 01

   The Enterprise AI Vendor Map Just Flipped — Your Procurement Strategy Is Already Stale

   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.

   What to do

   1. →

      Evaluate Anthropic Claude as primary enterprise AI vendor for coding and productivity workflows this quarter

      Now40% enterprise spend share and structural advantages over OpenAI's distracted superapp strategy make this the default enterprise choice until proven otherwise
   2. →

      Commission a 90-day AI coding tool vendor review — benchmark Claude Code vs Codex vs Cursor for your top 3 engineering use cases

      This sprintThe $5.5B coding market is consolidating around model-native agents; delaying locks you into a legacy tool-layer approach
   3. →

      Audit all AI vendor contracts for Microsoft-OpenAI partnership dependency assumptions and model scenarios for dissolution

      This sprintThe fracturing partnership creates SLA, pricing, and availability risks that most contracts don't address
   4. →

      Pilot multi-model routing: use frontier models only where quality delta matters, route routine work to MiniMax M2.7 or equivalent for 90%+ cost savings

      This quarterAt 7% of frontier cost with 90% quality, commodity models eliminate the margin penalty of AI at scale
2. 02

   a16z Just Declared the Software Middle Class Dead — And the Credit Markets Are Confirming It

   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.

   What to do

   1. →

      Convene a board-level strategy session this quarter to explicitly declare Path 1 or Path 2 — present the a16z framework, your current positioning, and required investment for each path

      NowActivist investors and acquirers are reading this framework; companies that don't declare a path will have one declared for them
   2. →

      Commission a pricing model transition roadmap from seat-based to usage/consumption/outcome-based pricing within 60 days

      This sprintSeat-based revenue sits directly on the cost line enterprise CFOs are most eager to cut with AI
   3. →

      Review all debt and credit facility terms for exposure to SaaS-model repricing in private credit markets

      This sprintFund redemption gates signal that SaaS-backed lending terms are tightening — get ahead before refinancing becomes adversarial
   4. →

      Pilot the four-person pod model on one net-new AI product initiative — collapse roles, cap headcount not compute, measure output against a traditionally-staffed team

      This quartera16z-backed competitors are already running this model; you need internal data on whether it works for your context before committing at scale
3. 03

   AI Security Just Got Empirical Data — And the Numbers Rewrite Your Threat Model

   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:

   Attack Vector	Scale	Implication
   Malicious GitHub repos	100K+ repos, AI-automated	Code provenance trust broken
   ClawHub AI skills	42% malicious	Agent marketplace trust broken
   Trivy scanner compromise	Security toolchain itself	Defenses become threat vectors
   Agent scheming	0% → 90%+ under pressure	Non-linear risk, not linear
   Exploitation windows	Under 24 hours	Traditional patching cadences broken

   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.

   What to do

   1. →

      Commission an AI-augmented red-team assessment of your infrastructure within 60 days, specifically modeling multi-step attack chains informed by the UK AISI methodology

      This sprintThe 5.8x capability scaling is empirically proven and on a predictable curve — your current threat model is based on last year's adversary capability
   2. →

      Audit all MCP server integrations for toxic data flow patterns and establish a maximum tool-count policy per server this quarter

      This sprintAttack surface grows quadratically with tool count; limiting tools per server is the most effective near-term mitigation
   3. →

      Establish sub-24-hour critical patching capability for all internet-facing services, or deploy architectural compensating controls deployable in under 4 hours

      This quarter20-hour exploitation windows break every traditional patch management cadence — speed is now the primary defensive metric
   4. →

      Increase cybersecurity defensive investment 25-40% with specific allocation toward AI-powered defensive tools and behavioral detection (CADR)

      This quarterThe adversary capability curve is accelerating predictably; static budget allocations guarantee a widening defensive gap