The Board Room

Deep dives

1. 01

   The AI Measurement Crisis: $100M/Month in Waste Proves Your Adoption Data Is Fiction

   A convergence of evidence from multiple independent sources this week confirms what skeptics suspected: enterprise AI adoption metrics are systematically gamed, and the industry's productivity narrative is built on inflated data. This isn't a minor calibration issue — it undermines investment decisions, headcount models, and vendor valuations across the sector.

   The Scale of the Problem

   At Meta, 85,000 employees burned 60.2 trillion tokens in 30 days — estimated at $100M+ monthly. Internal leaderboards turned usage into a performance signal, and engineers optimized accordingly. At Microsoft, VPs who rarely write code top internal AI usage rankings. At Salesforce, minimum spend floors were established, and engineers calibrated to stay just above average. This is Goodhart's Law at unprecedented scale: the measure became the target and ceased to be a useful measure.

   Every board deck in the industry citing AI adoption metrics — percentage of code AI-generated, agent utilization rates, token consumption growth — is now suspect.

   The Benchmark Problem Compounds It

   Independent testing revealed that Alibaba's Qwen3.6-35B jumped from 19% to 78.7% on the same benchmark by changing only the agent scaffold — not the model, not the training data, not the parameters. Models are shown to overfit to their own harnesses. This means published leaderboard scores are functionally meaningless for procurement decisions, and the competitive moat in AI development lives in scaffold engineering, not model intelligence. A finding with direct implications for every vendor evaluation in your pipeline.

   The One Model That Works

   Shopify's governance framework stands alone as a demonstrated countermeasure. Three design choices made the difference: renaming the leaderboard to a 'usage dashboard' (removing gamification), implementing circuit breakers for anomalous spend spikes, and having leadership personally review what top spenders actually build. Shopify's CTO Farhan Thawar's insight — that per-token cost (problem complexity) matters more than total volume — is the conceptual shift. It moves measurement from 'how much AI?' to 'how hard are the problems you're solving with AI?'

   Second-Order Implications

   The 10x business-user spend increases AI vendors cite as product-market fit evidence include significant waste. If three of the world's most sophisticated engineering organizations couldn't prevent metric gaming, the enterprise AI demand curve feeding vendor valuations is materially overstated. One senior Meta engineer suspects the token leaderboard was deliberately designed to generate training data for next-gen coding models — a $100M/month data collection strategy disguised as a productivity initiative. If true, the traces generated under gaming incentives may produce poisoned training data, not useful signal.

   ---

   The throughline is clear: the AI industry is in a measurement crisis. Leaders who continue reporting token consumption and adoption percentages without outcome verification are making decisions on fabricated data. The correction, when it comes, will be painful for organizations that built strategy on inflated numbers.

   What to do

   1. →

      Audit all internal AI usage metrics this week — determine what percentage of token consumption is productive vs. performative

      NowYour next board deck likely cites adoption metrics that may be 30-50% inflated; presenting bad data erodes board trust and leads to misallocation
   2. →

      Implement Shopify-style governance within 30 days: rename leaderboards to dashboards, add circuit breakers, require qualitative review of top spenders

      This sprintShopify is the only proven model; every week without governance accumulates waste and distorts planning data
   3. →

      Mandate that all model evaluations and procurement decisions include testing within your actual deployment scaffold, not published benchmark scores

      This sprintThe 19% → 78.7% scaffold swing means leaderboard rankings are theater; only your deployment context reveals true performance
   4. →

      Discount enterprise AI demand signals by 30-50% in all ROI models and vendor valuation assumptions

      This quarterTokenmaxxing at Meta, Microsoft, and Salesforce means the adoption curve feeding industry projections is materially inflated
2. 02

   AI Security's Dual Shock: Frontier Containment Broke While Offense Got 800x Cheaper

   The Breach That Broke the 'Responsible Release' Model

   Anthropic built Claude Mythos — a cybersecurity model it deemed too powerful for public access — distributed it under the codename Project Glasswing to select partners, and watched a Discord group reverse-engineer access on launch day. The attack vector was embarrassingly simple: URL patterns and naming conventions leaked in an unrelated Mercor data breach, combined with borrowed contractor credentials. The White House called emergency meetings. This isn't a failure of AI safety research — it's a structural failure of the 'trusted partner' distribution model that every frontier lab relies on.

   Frontier AI containment is fundamentally broken — and your partner distribution model is the attack surface.

   The Offense-Defense Economics Flipped in a Week

   Mythos found 271 real vulnerabilities in Firefox — a codebase with decades of human security review. But within weeks, multiple teams demolished the narrative that frontier models are required for this capability. AISLE's nano-analyzer, using a 3.6B-parameter model at $0.20/M tokens, found the same FreeBSD RCE that was Mythos's flagship finding — scanning 35K files in 10 hours for under $100. Vidoc reproduced results with GPT-5.4 and Claude Opus 4.6 at under $30 per file. Semgrep proved deterministic pre-filtering paired with LLM 'hotspot interrogation' consistently outperforms brute-force scanning.

   The conclusion converging across the security research community: the moat in AI vulnerability discovery is system architecture — context generation, scope narrowing, skeptical triage — not the model. Any security company or investment thesis predicated on 'frontier model access' as differentiation should be treated with deep skepticism.

   MCP Is the New Systemic Attack Surface

   Simultaneously, Model Context Protocol implementations are exhibiting Log4Shell-class systemic risk. Critical CVEs this week:

   Product	CVSS	Impact
   OpenAI Codex CLI	9.8	RCE via malicious MCP config
   Flowise	9.9	RCE via unsafe serialization
   Upsonic	9.8	RCE via MCP config injection
   Axios (HTTP library)	10.0	Cloud metadata exfiltration

   These aren't isolated bugs — they're protocol-level design weaknesses. Engineering teams are pulling MCP-based AI tools into prototypes that ship to production without security review. PraisonAI has four simultaneous critical CVEs. FastGPT has NoSQL injection bypassing all password checks.

   No One Owns Agent Security — Yet

   RSAC 2026 confirmed the vacuum. Across 11 main-stage keynotes, every speaker agreed on what AI agents need (asset management, permissions, observability, output validation), and not a single one claimed to have built it. All customers remain in monitor-only mode. Cisco released five open-source agent security tools (AI BOM, MCP Scanner, A2A Scanner, CodeGuard, DefenseClaw) — a deliberate ecosystem play to define the standard, mirroring Kubernetes' path to dominance. The window to adopt Cisco's framework or compete to define an alternative is measured in quarters.

   What to do

   1. →

      Commission a supply-chain security audit of every vendor and partner with access to your AI models this week — specifically map how deployment metadata could be reverse-engineered from third-party breaches

      NowMythos was cracked via a trivially avoidable supply-chain data leak; your own model distribution likely has equivalent exposure
   2. →

      Run an emergency SBOM scan for Axios dependency across all applications — CVE-2026-40175 is CVSS 10.0 and provides a direct path to harvesting cloud IAM credentials

      NowAxios is embedded in virtually every modern JavaScript application; this is likely in your stack today
   3. →

      Institute an AI tool adoption governance framework within 30 days with mandatory security review for any MCP-based tooling

      This sprintMCP implementations have systemic RCE vulnerabilities and engineering teams are adopting them into production without review
   4. →

      Evaluate Cisco's five open-source agent security tools and make a build-vs-adopt decision within 60 days

      This quarterWhoever defines the AI agent security standard owns the most consequential security category of the next decade; the window to influence the standard is closing
3. 03

   Private Equity Becomes the Enterprise AI Distribution Channel — Your Go-to-Market Just Got Flanked

   The Structural Play

   OpenAI's DeployCo is strategically brilliant: a $10B joint venture with TPG, Bain Capital, Advent, Brookfield, and Goanna Capital — OpenAI invests up to $1.5B, PE firms add $4B+ — that effectively buys a distribution channel into thousands of portfolio companies without building an enterprise sales team. Anthropic is mirroring the move with Blackstone and Hellman & Friedman. Two leading AI labs, simultaneously, decided that PE-mandated adoption is faster than traditional software sales.

   AI adoption is being repositioned from a discretionary technology purchase to a PE-mandated operational improvement — similar to how PE firms drove ERP and cloud adoption in earlier cycles.

   What This Changes

   For any enterprise technology company, three shifts follow:

   1. Your buyer persona may change. When AI mandates come from PE operating partners rather than CIOs, the sales cycle collapses from quarters to weeks — but the decision criteria shift from technical merit to operational ROI.
   2. Your competitive set expands. AI labs with billions in distribution capital now compete directly for the enterprise deployment budget you assumed was yours.
   3. Your horizontal AI features face commoditization within 18 months. PE firms standardize tooling across portfolio companies. Whatever OpenAI and Anthropic sell through DeployCo becomes the default.

   The Talent Diaspora Creates Parallel Disruption

   OpenAI's organizational unbundling accelerates the dynamic. Mira Murati's Thinking Machines Lab already landed a multibillion-dollar Google Cloud deal within a year of leaving. Jerry Tworek's Core Automation is actively recruiting from Anthropic and Google DeepMind. The centralized lab model (2020-2025) is giving way to a distributed ecosystem of founder-led teams with deep expertise in narrow domains. The next breakthrough AI capabilities may come from companies you haven't heard of yet — backed by the same PE capital flowing through DeployCo.

   SaaS Repricing Compounds the Pressure

   This PE distribution channel arrives as SaaS economics are already under stress. Gross margins are compressing from 70-80% to approximately 52% due to AI inference COGS. ServiceNow lost 15% after-hours despite 22% revenue reacceleration — the $7.75B Armis acquisition's margin dilution was the catalyst, but the structural cause is the market pricing in AI-driven compression of ServiceNow's core value proposition. Revenue growth is no longer a valuation shield when the market believes your addressable market is shrinking.

   Meanwhile, the 100x SaaS cost compression is now quantified and personal: one AI company CEO documented his own $200K/year event tooling being replaceable for $2K with AI-native alternatives. Scale this across every low-NPS SaaS category and a massive reallocation of enterprise software spend is coming.

   ---

   Sources disagree on timing: the PE distribution channel suggests enterprise AI adoption will accelerate (AI labs buying distribution), while the tokenmaxxing evidence suggests current adoption metrics are inflated. Both can be true — and the tension is the insight. Real adoption will accelerate through PE mandates, but measured adoption will correct downward as organizations strip out performative usage. Leaders who can distinguish genuine adoption from metric gaming will make better capital allocation decisions than those who can't.

   What to do

   1. →

      Assess your enterprise GTM vulnerability to PE-channel AI distribution within 60 days — map which of your customers are PE-owned and model the scenario where adoption mandates bypass your sales team

      This sprintDeployCo gives OpenAI distribution into thousands of portfolio companies without a single sales call; if your customers overlap, your competitive position just changed
   2. →

      Evaluate partnership or commercial relationships with emerging OpenAI diaspora startups (Thinking Machines Lab, Core Automation) before they get locked into exclusive cloud or platform deals

      This quarterFrontier talent is fragmenting into new power centers that will define next-gen AI capabilities; early relationships compound
   3. →

      Renegotiate every SaaS contract where NPS is below 30 — use the 100x AI-native replacement leverage to extract 40-60% concessions or begin internal rebuild

      This quarterThe $200K→$2K cost compression is quantified and replicable; vendors know this, and your leverage is highest before AI-native alternatives mature
   4. →

      Prepare a board-level narrative on AI disruption risk to enterprise software valuations — ServiceNow's 15% crash on strong growth is the template your directors will reference

      This quarterThe market is repricing SaaS regardless of current performance; your board needs a proactive framing before investor questions arrive