The Board Room

Deep dives

1. 01

   The 12-Month Platform Lock-In Window: Why Your AI Vendor Decision Just Became Irreversible

   Two Ecosystems, One Choice

   Fresh data from a16z's March 2026 Top 100 Gen AI Consumer Apps report quantifies what many suspected: the AI platform market has bifurcated into two distinct ecosystems with only 11% app overlap. ChatGPT's exclusive integrations skew consumer-transactional (Expedia, Instacart, Zillow, MyFitnessPal across 85+ transaction categories). Claude's exclusive integrations skew professional (PitchBook, FactSet, Snowflake, Databricks, Sentry, PubMed). This isn't head-to-head competition — it's iOS vs. Android forming in real time.

   OpenAI is executing a consumer internet platform strategy: 'Sign in with ChatGPT' identity layer, advertising tests, a 220-app marketplace, and 900M weekly active users. Sam Altman is building the next Google, not the next Microsoft. But there's a crucial contradiction: TD Cowen called OpenAI's retreat from its e-commerce checkout feature 'stunning' this same week. The super-app vision is announced; the execution is pulling back. OpenAI may be discovering that building commerce infrastructure requires organizational capabilities it doesn't have — six months before a planned $730B IPO.

   Anthropic's Enterprise Billing Trojan Horse

   Anthropic's Claude Marketplace is the more consequential platform move, despite less fanfare. By letting enterprises apply existing Anthropic committed spend toward third-party tools (GitLab, Snowflake, Replit, Harvey) with consolidated invoicing through Anthropic, they've replicated the exact mechanism that made AWS Marketplace and Salesforce AppExchange the gravitational centers of their ecosystems. Every dollar spent through the Marketplace deepens organizational dependency. The partner selection — GitLab and Snowflake — signals Anthropic is targeting the full enterprise development and data stack.

   Layer in the economics: Claude Code at $200/month against ~$5,000 in actual compute costs is a 25:1 loss ratio. This only makes sense as a land-and-expand play where the Marketplace captures margin from an installed base acquired below cost. It's the AWS playbook, executed at the foundation-model layer.

   The strategic question isn't 'which platform will win?' — it's 'which platform's user base is your customer, and are you building the right integrations before switching costs lock in?'

   The Policy Wildcard

   The White House's 'any lawful use' mandate — its direct response to the Anthropic-Pentagon standoff — threatens to remove ethical positioning as a platform-selection criterion entirely. If AI companies cannot restrict lawful government use, Anthropic's 'trusted alternative' brand may become a legal liability rather than a competitive advantage. Microsoft's hedge is instructive: Copilot Cowork built with Anthropic, Agent 365 featuring both Anthropic and OpenAI. They're treating model providers as interchangeable components — the real moat is M365's 400M+ seat distribution.

   The Bundling Compression Is Accelerating

   The a16z data provides the clearest evidence yet of platform bundling as an existential threat: Midjourney fell from Top 10 to #46 in three years as image generation was absorbed into ChatGPT and Gemini. Google's Nano Banana generated 200M images with 10M new users in its first week. Video, voice, and music tools are next. The survivors — Suno (#15, music), ElevenLabs (voice) — occupy modalities the platforms haven't prioritized yet. Notion's counter-example is instructive: 50%+ AI attach rate, roughly half of ARR from AI features, deeply embedded in workflow. Workflow lock-in beats model-quality differentiation.

   What to do

   1. →

      Conduct a 'bundling vulnerability audit' across your product portfolio — identify every capability ChatGPT or Gemini could absorb as a native feature within 12 months and map defensibility (workflow lock-in, proprietary data, enterprise integrations) for each

      This sprintMidjourney's fall from Top 10 to #46 happened in three years; video and voice tools are next in the compression cycle
   2. →

      Establish integration presence in both ChatGPT's app directory (220 apps) and Claude's MCP ecosystem (~210 connectors) within this quarter with dedicated integration resources

      This sprintOnly 11% app overlap means these are distinct distribution channels reaching different customer bases — missing either is leaving revenue on the table
   3. →

      Evaluate your Anthropic committed spend against the Marketplace lock-in dynamics — model whether consolidated billing works for or against your procurement flexibility

      This quarterMarketplace spending creates compounding switching costs; understand the trap before your finance team optimizes into it
   4. →

      Draft a formal 'government and defense' posture document for the board, articulating where your company stands on the compliance-resistance spectrum before the 'any lawful use' mandate forces a reactive decision

      This sprintThe White House mandate eliminates the luxury of ambiguity — your stance will be tested, and improvised answers carry brand and talent risk
2. 02

   The AI Productivity Illusion: Your Planning Assumptions Are Simultaneously Too Aggressive and Too Conservative

   The Hardest Data Point in AI Today

   METR's randomized controlled trial — the gold standard of evidence — with 16 experienced open-source developers found they were 19% slower when using AI assistance, while believing they were 20% faster. This isn't a survey; it's measured performance against a perception gap of nearly 40 percentage points. Combine this with an LLM-generated Rust rewrite of SQLite that ran 20,171x slower on primary key lookups (the query planner missed a single optimization flag), and a pattern emerges: AI coding tools optimize for code generation volume, not code quality or system-level correctness.

   If your planning assumptions include AI-driven headcount efficiency or accelerated delivery timelines, those assumptions need stress-testing against measured outcomes — not developer sentiment.

   But Capability Is Accelerating Faster Than Anyone Predicted

   Here's the paradox. While productivity gains disappoint at the individual level, capability timelines are compressing faster than top forecasters expected. Ajeya Cotra — one of the most rigorous AI forecasters in the field — publicly admitted her January 2026 predictions were 'much too conservative' by March, revising her year-end agent autonomy estimate from 24 hours to 100+ hours. Expert AI forecasts now have approximately a two-month half-life.

   A former Google engineer built a complete 4D reconstruction of a live military operation overnight using agent swarms and public data — work his previous team would have needed a quarter to complete. Claude Code hit $1B ARR in six months via a CLI tool invisible to traditional consumer metrics. Codex has 2M WAU growing 25% weekly. The capability explosion is real.

   The Reliability Wall Explains the Paradox

   The HKUST AgentVista benchmark reveals the binding constraint: the best multimodal agent (Gemini-3 Pro) achieves only 27% accuracy on real-world multi-step workflows. Open-source alternatives lag at 12%. Three out of four complex workflows fail. Karpathy's 'March of Nines' framework explains why: reaching 90% reliability is trivially easy (good enough for a demo), but each additional nine requires exponential engineering effort. For 10+ step processes, error compounding drives total success below 35%.

   ByteDance's CUDA Agent result offers a clue to where the value actually lies: their base model scored a mediocre 74% on KernelBench, but after finetuning on just 6,000 curated CUDA samples, it hit 100% on Level 1-2 and 92% on Level 3 — surpassing Claude and Gemini by ~40% on the hardest tasks. The moat is shifting from model scale to data curation and domain-specific optimization.

   What This Means For Your Organization

   The organizations seeing real AI productivity gains — AI-native startups running 40% leaner teams while raising larger rounds — have rebuilt processes around AI from scratch. They aren't layering AI tools onto existing workflows. This distinction is critical: the value capture requires fundamentally different organizational design, not better tooling on top of legacy processes.

   Metric	Perception	Reality
   Developer speed with AI	+20% faster	-19% slower (METR RCT)
   Complex task success	Demo-ready	27% in production (AgentVista)
   Agent autonomy by EOY	24 hours (Jan forecast)	100+ hours (Mar revision)
   Domain finetuning vs. frontier	Frontier wins	6K samples beat GPT-4 by 40%

   What to do

   1. →

      Commission a controlled internal benchmark of AI coding tool productivity using objective metrics (cycle time, defect rate, performance benchmarks) — not developer self-reporting — and present results to engineering leadership within 60 days

      This sprintThe 39-point gap between perceived and actual performance in the METR study means your current productivity estimates are almost certainly wrong; decisions based on self-reported data carry material planning risk
   2. →

      Launch a domain-specific finetuning initiative on your proprietary operational data this quarter — ByteDance proved this is higher-ROI than chasing the best foundation model partner

      This quarter6,000 curated samples beating frontier models by 40% on hard tasks means your unique data is more valuable than your model contract
   3. →

      Conduct an 'accelerated timeline' stress test on your 2027-2028 strategic plan — model what happens if AI agents can autonomously handle 100+ hour software projects by Q4 2026

      This quarterExpert forecasts now have a ~2-month half-life; any strategic plan built on 'conservative' AI timelines is almost certainly too slow
   4. →

      Ruthlessly prune your AI pilot portfolio to the 3-5 workflows where disciplined engineering can deliver 99%+ reliability — pause or kill everything else

      This sprintAt 27% complex-task success rates, most multi-step agent deployments will fail in production; concentrate resources where the March of Nines math actually works
3. 03

   Agentic Commerce: The $1.9T Infrastructure Race Nobody's Watching

   The Trust Layer Is Being Defined This Quarter

   Three infrastructure moves within days of each other signal that agentic commerce has crossed from concept to contested infrastructure category. Stripe launched Shared Payment Tokens enabling AI agents to transact on behalf of users. Mastercard launched Verifiable Intent — cryptographic authorization proofs backed by Google and IBM — positioning itself as the neutral trust infrastructure all agentic commerce flows through. Klarna partnered with Stripe to enable BNPL for AI shopping agents. These aren't product announcements — they're the first draft of the financial plumbing for the agent economy.

   Mastercard's play is strategically sophisticated: by building on open standards, they're not defending legacy card rails — they're repositioning cards as the authorization layer while stablecoins handle settlement. This directly counters the thesis that sent card network stocks down weeks ago. Cards aren't being disintermediated; they're being repositioned.

   The New Merchant Class That Can't Use Traditional Rails

   The demand side is equally significant. 36 million developers joined GitHub last year. 67% of Bolt.new's 5 million users are non-developers. 25% of Y Combinator's W25 cohort had 95%+ AI-generated codebases. These builders are launching API tools and data services at unprecedented velocity — but many lack the corporate entities, track records, and credit histories to qualify for traditional merchant accounts through Stripe, Square, or the card networks.

   This structural gap creates a natural adoption wedge for stablecoin-native payment rails. The x402 protocol, which embeds stablecoin payments directly into HTTP requests, targets exactly this use case. Meanwhile, Western Union's USDPT stablecoin — with 360,000 physical cash-out locations across 200+ countries on Solana — just solved the last-mile problem that kept stablecoins from mainstream adoption in emerging markets.

   The competitive window is 18-24 months before incumbents adapt their underwriting and onboarding to capture AI-born merchants. That's a meaningful first-mover opportunity for anyone positioned to capture it.

   Early Production Proof: The Balyasny Model

   Hedge fund Balyasny's deployment of GPT-5.4 across 95% of its 180 investment teams — cutting research cycles from days to hours with centralized platform governance, rigorous model evaluation, and embedded feedback loops — is the most operationally mature proof point for AI-mediated financial workflows in production. Morgan Stanley's 2,500 job cuts and projection of 200,000 European banking job losses by 2030 provides the macro frame. This isn't a pilot; it's full-scale production deployment of frontier AI in high-stakes financial decision-making.

   Stripe's Compounding Advantage

   At $159B valuation and $1.9T in processing volume, Stripe is becoming the operating system for AI-native businesses. Their LLM token cost billing feature — automatic margin markup on model costs — touches pricing, billing, revenue recognition, and margin management simultaneously. That's not a feature you switch away from. Combined with the Klarna partnership and agent payment tokens, Stripe is closing every gap between 'AI startup builds a product' and 'AI startup monetizes it.' John Collison's deliberate deprioritization of an IPO suggests they see the AI infrastructure opportunity as still in early innings.

   What to do

   1. →

      Audit your payment and commerce infrastructure against the emerging Stripe/Mastercard/Klarna agentic stack within 60 days — determine whether you're building on these rails, competing with them, or at risk of disintermediation

      This quarterThe trust layer for agent-mediated commerce is being defined now; the standards being set in this quarter will persist for years
   2. →

      Evaluate stablecoin payment rail integration (specifically x402 protocol) for any AI/developer-focused products — particularly if you serve the AI-born merchant class that can't qualify for traditional processing

      This quarter36M new GitHub developers and 67% non-developer builders on Bolt.new represent a structural underwriting gap that stablecoin rails fill natively
   3. →

      Study Balyasny's centralized AI deployment model as a reference architecture for your own production AI rollout — centralized platform, model evaluation gates, embedded feedback loops

      This quarter95% team adoption with measurable cycle-time compression is the most mature enterprise proof point available; most organizations are still running fragmented pilots that compound quarterly disadvantage