The Board Room

Three data points that should end the 'wait and see' posture

Ramp's analysis of spending patterns across thousands of companies reveals that top-quartile AI spenders have doubled revenue since 2023 while bottom-quartile companies flatlined. This isn't correlation — Anthropic's Economic Index confirms that early adopters develop compounding skills through learning-by-doing, meaning the gap widens with time, not narrows. And NBER research adds a dangerous wrinkle: executives report AI gains that are invisible to traditional metrics. Your dashboards may be telling you everything's fine while your competitors build advantages you can't even measure yet.

The AI adoption gap isn't closing — it's compounding. Every quarter of delay puts you further behind a curve that accelerates.

The METR curve changes workforce planning math

METR's tracking of AI agent autonomous task duration shows capability doubling every 4 months — accelerated from 7 months. The concrete progression: 50-minute tasks in early 2025, 5-hour tasks by late 2025. Extrapolate at the current rate and you reach full-day autonomous tasks by mid-2026, multi-day by early 2027. Every role in your organization that consists primarily of stringing together 4-8 hour cognitive tasks enters the displacement zone within 18 months. The token cost economics make this irreversible: a knowledge worker's entire annual cognitive output (~15 million tokens) costs $8 to $75 to process through a frontier model. The fully loaded human cost exceeds £150,000. That's not a 2x efficiency gain — it's a 2,000x cost collapse.

Anthropic's $20B ARR proves where value migrated

Anthropic going from $1B to $20B ARR in ~14 months — with 1.5-2x monthly growth in early 2026 — isn't a model quality story. The growth inflection came from a paradigm shift: Claude Code and Opus 4.6 moved the product from 'answer my questions' to 'do my work autonomously.' The market is telling us in dollar terms that value has permanently migrated from model quality to autonomous execution capability.

Why incumbents keep losing — and how to avoid their fate

Every AI coding paradigm shift (autocomplete → delegation → autonomous execution) was led by outsiders — Cursor by 'a bunch of kids,' Claude Code by engineers without Copilot experience. Microsoft had GitHub's codebase, the dominant IDE, and OpenAI partnership. Apple had the best silicon and 2B+ devices. Both lost because their organizational structures physically prevented crossing what one analyst calls the 'evolutionary valley' — the period where the next paradigm requires worse short-term metrics. The strategic imperative: if your AI exposure is primarily Microsoft Copilot because it's bundled with your stack, you're making a bet-the-company choice through the lens of operational convenience.

The Meta signal

Zuckerberg deploying one-agent-per-person to flatten Meta's management structure validates the thesis that traditional management hierarchies exist primarily to synthesize and relay information — something agents do faster with less distortion. Mature agentic implementations are expected to handle reliability and security by end of 2026. This isn't a 5-year vision; it's a 9-month infrastructure buildout.

What OpenAI actually did — and what it signals

OpenAI killed Sora less than six months after its viral launch, walked away from a $1 billion Disney partnership signed just three months ago, shelved consumer experiments including its erotic chatbot, and began consolidating ChatGPT, Codex, and Atlas into a single desktop superapp. All to free compute for 'Spud,' its next-generation frontier model that Altman claims will 'really accelerate the economy.' Simultaneously, Altman personally stepped away from safety oversight to focus on infrastructure. This isn't pruning — it's triage.

If the most well-capitalized AI company in the world kills a viral product for cost reasons, the era of 'ship capabilities, figure out economics later' is over.

The compute scarcity signal is industry-wide

The implication is stark: if OpenAI — valued at hundreds of billions, backed by Microsoft, with massive datacenter agreements — cannot simultaneously run a consumer video product and train its next model, the entire industry is compute-constrained. This thesis is reinforced by Microsoft leasing a 900MW datacenter site in Abilene, Texas originally developed for Oracle, and Google financing a multibillion-dollar Anthropic datacenter in the same state. Infrastructure has become a zero-sum land grab where power capacity and permitting are multi-year constraints. A site lost today isn't replaceable in six months.

Dual IPOs create a rare buyer leverage window

Both OpenAI and Anthropic are heading toward IPOs within 12 months, targeting a combined $135B+ in proceeds alongside SpaceX. For the next 6-9 months, both companies need to demonstrate enterprise revenue traction to justify public market valuations. This creates a rare window of buyer leverage: enterprise customers can extract pricing, commitment, and integration terms that will never be available again. The smart play is dual negotiations, using each company's competitive anxiety against the other.

Sources diverge on the strategic read

There's an important tension across today's intelligence. Some sources frame this as brilliant discipline — sacrificing near-term revenue for a model that could define the next era. Others frame it as a forced contraction driven by unsustainable unit economics and pre-IPO cost pressure. The truth matters: if Spud delivers a step-function improvement, the Sora sacrifice looks visionary. If it delivers incremental improvement, OpenAI will have burned a billion-dollar partnership, lost a product line, deprioritized safety, and handed Anthropic a competitive narrative — all for nothing. That execution risk should make any company with deep OpenAI dependencies uncomfortable.

Safety as competitive differentiator

Altman stepping away from safety oversight is a governance red flag and a competitive gift to Anthropic. In the pre-IPO period, Anthropic can now credibly argue it's the 'responsible' choice for enterprise AI, particularly in regulated sectors. Enterprise procurement decisions will increasingly be framed around this dichotomy: OpenAI optimizes for capability and speed; Anthropic optimizes for safety and reliability. For regulated industries, this isn't abstract — it's a vendor selection criterion.

Two simultaneous escalations

Today's intelligence reveals two governance crises converging. First: CanisterWorm, a self-propagating npm worm that steals developer credentials and then injects malware into the victim's own published packages — turning every compromised developer into an unwitting attack vector. For organizations with hundreds of internal npm packages, blast radius is exponential. This arrives alongside TeamPCP's campaign targeting Aqua Security's Trivy vulnerability scanner — the very tool organizations use to detect vulnerabilities — plus LiteLLM and Telnyx.

Second: Microsoft data shows 84% of security leaders are alarmed about unauthorized AI agents, and 62% of UK enterprises are already running them without authorization. This is shadow IT with agency — AI systems that don't just access information but take autonomous actions.

The 'trust by default' model for open-source package registries and the 'ignore by default' posture on AI agents are both simultaneously untenable as of this week.

The self-propagating worm is a step-function, not an increment

Previous supply chain attacks (which we covered last week including LiteLLM's compromised SOC2 certifications) were isolated — compromised package affects its direct users. CanisterWorm is fundamentally different: it weaponizes the victim's own publishing credentials to propagate through the dependency graph. Each compromised developer becomes a distribution node. The attack surface expands geometrically, not linearly. The TeamPCP campaign exploited a specific gap: Docker images published to Docker Hub without corresponding GitHub releases. This means any organization that validates packages only against source repositories — a common practice — is vulnerable.

Pinterest's MCP blueprint is the first real answer

Pinterest published what amounts to the first production-grade enterprise agent governance architecture: a registry-based MCP platform with centralized approval workflows, layered authentication, shared deployment paths, and IDE integration. This isn't a DevOps story — it's the first public blueprint for what enterprise agent infrastructure looks like when built for security and scale. If you're deploying AI agents without this kind of control plane, you're accumulating governance debt that compounds weekly.

The HubSpot calibration

HubSpot's data from their Prospecting Agent deployment provides a useful reality check: ~50% of users manually review outputs before sending. Human-in-the-loop isn't a transitional feature — it's the product architecture for enterprise AI for the foreseeable future. Google's transfer of Sashiko (an AI code reviewer that found 53% of bugs human reviewers missed in the Linux kernel) to the Linux Foundation sets a different timeline: within 12-18 months, AI-augmented code review will transition from innovative practice to expected baseline. Organizations without it will face questions from customers, auditors, and regulators.

The compounding risk

These two crises interact dangerously. Shadow AI agents pulling packages from compromised registries. Autonomous desktop agents (like Anthropic's Computer Use) executing on systems with poisoned dependencies. The attack surface isn't additive — it's multiplicative. And the governance infrastructure at most organizations was designed for neither autonomous agents nor self-propagating supply chain worms, let alone both simultaneously.