The Board Room

Deep dives

1. 01

   The Hormuz Clock: 45 Days Until Your Entire Infrastructure Roadmap Reprices

   A Supply Shock Three Times Worse Than Russia-Ukraine — in One-Quarter the Time

   The Strait of Hormuz is 95% blocked, with AIS transponders dark and satellite imagery as the only visibility into the situation. In the 24 days since the Iran conflict began, cumulative liquids production losses have reached 285 million barrels — 3x what the Russia-Ukraine conflict produced in 24 weeks. This is not a financial rerouting like 2022 — 10-15 million barrels per day of field-level production is physically shut in, requiring weeks to months to restart even after reopening.

   The Omani crude premium to WTI reached $60/bbl — the market is screaming that Asian buyers cannot physically access oil at any price.

   Taiwan's Grid: The Direct Path From Hormuz to Your Chip Supply

   Taiwan's power generation depends 15% on Qatari LNG, which is offline with the Strait blocked. TSMC's fabs consume enormous power. The path from 'Strait stays closed through April' to 'Taiwan implements rolling blackouts' to 'chip allocation crisis' is shorter and more direct than most technology leaders appreciate. If you experienced the 2021 chip shortage, imagine it triggered not by demand spikes but by physical power rationing at the foundries.

   Petrochemical Cascade Hits Hardware Directly

   Over 15 million barrels/day of petroleum consumption goes to non-fuel products — the plastics, resins, adhesives, and synthetic materials that constitute the physical substrate of technology. Prices are going vertical:

   • Butadiene (cables, seals, rubber): +140% in China
   • Toluene (solvents, coatings): +70%
   • PET (packaging, display films): +45% in Germany

   These aren't marginal increases — they're the kind of input shocks that force hardware vendors to renegotiate contracts mid-cycle.

   Data Center Power Crisis Compounds the Problem

   This crisis arrives on top of an existing structural constraint. Traditional gas turbines from GE Vernova and Siemens are backordered through 2032. Meta's 2026 capex of $115-135B (up from $72.2B in 2025) is consuming remaining power, land, and construction capacity. Arbor Energy's 3D-printed modular turbines won't grid-connect until 2028. The OECD has revised US inflation projections from 2.8% to 4.2%, driven primarily by this energy shock.

   North America's position as a net energy exporter is the strategic bright spot — the WTI-Brent spread blowing out from $5 to $20/bbl quantifies how much cheaper US-based operations are right now.

   China Is Already Hoarding

   China has restricted petroleum product exports — resource nationalism, the leading indicator that a crisis lasts. The forward curve's implied normalization by late spring is either the market's best intelligence on a geopolitical resolution, or the most dangerous case of anchoring bias since 2008. Nobody has reliable visibility into the actual situation.

   What to do

   1. →

      Convene a cross-functional war room this week to map exposure across semiconductor/Taiwan dependency, petrochemical-linked hardware inputs, and energy-cost-sensitive operations (data centers, cloud spend)

      Now45-day strategic reserve window means decisions made this week determine positioning for the rest of Q2
   2. →

      Accelerate hardware procurement and pre-purchase critical components with petrochemical or Asian manufacturing dependencies before the cost curve reprices

      NowInput cost increases of 45-140% will cascade through hardware pricing within 30-60 days
   3. →

      Model a 'Taiwan power rationing' scenario and develop contingency allocation plans with semiconductor vendors by end of April

      This sprintTaiwan's strategic LNG reserves are measured in weeks, not months
   4. →

      Hedge cloud infrastructure costs — negotiate rate locks or reserved capacity with major cloud providers before energy-driven repricing hits

      This sprintCloud providers' energy costs are rising; pricing adjustments lag by 60-90 days
   5. →

      Re-evaluate pending infrastructure investment decisions with a geographic component — the risk-adjusted calculus has shifted materially toward North America

      This quarterWTI-Brent $20 spread creates multi-quarter window of relative US energy cost advantage
2. 02

   AI's Business Model Is Splitting in Two — And the Winner Takes Both Margins

   OpenAI Just Proved AI-Native Advertising Works at Scale

   OpenAI reached $100M annualized ad revenue within 6 weeks of its ChatGPT ad pilot — with less than 20% of eligible users seeing ads and only 600 advertisers onboarded. Relevance scores exceed 93%. Self-serve advertiser access launches in April; international expansion follows. The strategic implication: OpenAI can now cross-subsidize inference costs with advertising revenue, offering better AI to free users (larger audience for ads) while funding more expensive frontier models. This creates a virtuous cycle that subscription-only competitors cannot replicate.

   When a platform can offer superior AI for free because advertisers foot the bill, every competitor's pricing model breaks.

   Anthropic Is Winning Enterprise Despite OpenAI's Pivot

   OpenAI's internal 'code red' messaging — calling Anthropic's enterprise dominance a 'wake-up call' — is remarkable for a company with 900M weekly active users. The numbers: Anthropic holds 40% enterprise market share to OpenAI's 27% (Menlo Ventures). Anthropic achieved revenue parity despite a fraction of consumer visibility. OpenAI's response: a $10B PE joint venture with TPG, Advent, Bain, and Brookfield to buy distribution through portfolio companies — an unprecedented channel play that signals organic enterprise growth has stalled.

   Vertical AI Is Beating Frontier Models in Production

   Intercom's Fin agent hit ~$100M ARR, resolving 2 million customer service issues per week while outperforming both GPT-5.4 and Opus 4.5 on customer service benchmarks. This validates a thesis that should reshape your AI investment: proprietary data and domain expertise capture more durable value than model scale alone. Nvidia's $26B five-year open-weights commitment is the hardware incumbent's response — training Nemotron natively in NVFP4 on Blackwell GPUs at 442 tokens/second (59% faster than OpenAI, 66% faster than Google) to lock developers into its hardware-model stack.

   The Margin Crisis Is the Real Story

   Traditional SaaS ran at ~75% gross margins with near-zero marginal cost. AI features introduce variable costs that scale linearly with engagement, compressing margins to ~30%. Across 18 SaaS earnings calls, AI is revenue-positive but margin-neutral. Salesforce's $800M Agentforce ARR is a revenue story, not a profit story. Every LTV calculation, CAC payback model, and IPO readiness benchmark built on SaaS economics needs rebuilding.

   Model	Revenue Signal	Margin Reality
   OpenAI (ads)	$100M ARR in 6 weeks	Ad-subsidized inference
   Anthropic (API)	40% enterprise share	High inference cost admitted
   Vertical AI (Fin)	~$100M ARR	Domain data moat
   SaaS + AI bolted on	$800M Agentforce	Margin-neutral

   The Cloud-AI Divorce Reshuffles Everything

   OpenAI exploited a stateful/stateless legal distinction to break Microsoft's cloud exclusivity and partner with AWS. The most strategically valuable new layer — persistent agent state, memory, tool orchestration — now lives on AWS, while Azure retains the commoditizing inference layer. Meanwhile, DeepSeek reportedly trained a model entirely on Huawei Ascend chips, proving China's decoupling from Nvidia is no longer theoretical. Qwen3.5-122B already outscores Nemotron on intelligence benchmarks.

   What to do

   1. →

      Conduct a gross margin stress test across all AI-powered product lines by end of Q2, modeling margin impact at 2x and 5x current usage levels

      This sprint30% margins at scale break every SaaS financial assumption — identify which products become unprofitable before they do
   2. →

      Reassess AI vendor strategy: evaluate Anthropic's enterprise tooling vs. OpenAI within 90 days, factoring in Anthropic's pre-IPO incentive to lock in marquee customers

      This quarterAnthropic's October 2026 IPO creates a brief negotiating window; post-IPO leverage shifts to them
   3. →

      If you have consumer-facing AI products, commission a feasibility study on AI-native advertising as a revenue stream this quarter

      This quarterOpenAI just proved the model works — the second-mover window is closing as self-serve access launches in April
   4. →

      Identify your top 3 vertical domains with proprietary data and fast-track vertical AI development or acquisition

      This quarterIntercom's success against GPT-5.4 proves domain-specific AI captures more enterprise value than horizontal models
3. 03

   The Race to Own the Agent Execution Layer Is Happening This Quarter

   Stripe Wants to Be the OS Between AI Agents and Every SaaS Vendor

   Stripe's Projects.dev isn't a developer tool — it's a bid to become the mandatory intermediary in every agent-mediated service transaction. Patrick Collison explicitly cited Karpathy's observation that the hard part of agent-built startups is DevOps orchestration — payments, auth, infra, security — and shipped a product where a single CLI command creates a PostHog account, generates an API key, and configures billing. Every SaaS vendor now faces the choice: be in Stripe's catalog (gaining agent-mediated distribution) or be invisible to agent workflows.

   Cloudflare Bids for the Execution Runtime

   Cloudflare's Dynamic Worker Loader positions V8 isolates as the default runtime for AI agents at $0.002/day — claiming 100x speed and 10-100x memory efficiency versus containers. The pricing is designed to kill the 'build your own' option. The constraint: JavaScript-only execution. For Python/Go agent workflows, the door remains open for competitors. But for the massive JavaScript ecosystem, this could be category-defining.

   Continual RL Creates Compounding Moats

   Cursor claims it ships improved model checkpoints every 5 hours through a productized RL feedback loop using production user behavior as reward signals. NVIDIA's ProRL Agent shows decoupling rollout from optimization nearly doubles SWE-Bench scores. The combination means vertically-integrated companies with high-frequency user feedback loops can now compound quality improvements at a rate horizontal competitors cannot match. If Cursor's models improve every 5 hours while a competitor depends on quarterly foundation model releases, the quality gap becomes unbridgeable within months.

   Base models are becoming table stakes. The orchestration layer is the product. Organizations that treat the harness as glue code and the model as their competitive advantage have the value equation inverted.

   Identity Is the Control Plane — RSAC 2026 Consensus

   RSAC 2026's consensus that 'AI agents broke identity' signals a once-in-a-decade control plane shift. When AI agents operate autonomously, make API calls, and chain multi-step workflows, identity is the only layer that can answer who this actor is, what they're authorized to do, and whether they're within bounds. JumpCloud repositioned as the 'identity control plane for shadow AI.' 1Password pivoted to 'Unified Access' for non-human identity (NHI) credential management. VPN-based access is confirmed broken at 51% incident rates.

   The Context Layer Is the New Moat

   GitHub's analysis of 2,500+ custom instruction files validates that the gap between productive and unproductive AI agent usage isn't about which model you run — it's about context engineering. Intercom built 13 plugins and 100+ composable skills encoding institutional knowledge into Claude Code. Copilot's three-layer context architecture (repo-level, path-specific, custom agents with MCP tools) creates lock-in through context investment, not feature superiority. Once companies invest in writing hundreds of copilot-instructions files, switching costs compound.

   Security Debt Is Accumulating Faster Than Governance

   In a single week: Stripe auto-provisions API keys; Hermes Agent runs unattended for hours; Claude Code auto-fixes CI failures without human review; and GODMODE — a productized persistent jailbreaking skill — was announced for agent harnesses. OpenClaw accumulated 104 CVEs in 18 days (200x the rate of LangChain or Ollama across their entire lifetimes). MCP documentation poisoning can inject malicious packages with zero malware required — a researcher planted fake PyPI packages in Plaid and Stripe docs that coding agents would auto-install.

   What to do

   1. →

      Assess your company's position in the emerging agent-to-service provisioning layer by end of Q2 — determine whether to integrate with Stripe Projects.dev, build a competing pathway, or both

      This sprintStripe is executing during a brief window where agent provisioning standards haven't calcified — early participants shape the ecosystem
   2. →

      Rebalance AI investment from base model capabilities toward harness/orchestration layer if currently over-indexed on model selection

      This quarterMultiple convergent signals confirm the orchestration layer captures more durable value than model access
   3. →

      Commission a non-human identity (NHI) audit across your organization — map every AI agent, service account, API key, and automated credential with production access

      This sprintRSAC consensus: identity is the control plane for AI. If you don't know what non-human identities exist, you can't govern them
   4. →

      Establish an AI agent governance framework covering auto-provisioning, unattended execution, and credential lifecycle management before end of Q2

      This quarterAgent autonomy is expanding faster than security infrastructure — the organization that treats this as an afterthought will be the first headline