The Board Room

Deep dives

1. 01

   The Buildout Bleeds, the Toll-Takers Print — and the Leverage Is Yours Now

   Two companies in the same AI boom posted opposite balance-sheet outcomes, and the asymmetry hands enterprise buyers a negotiating window that closes on its own schedule.

   The cash burn is a symptom. The asymmetry underneath it is the strategy lesson. Alphabet's operating margin slipped two points to 34% in a quarter where both Cloud and Ads improved their margins. The offset is Google DeepMind, now broken out as a separately reported cost center and rising sharply. Frontier research is consuming the profit that AI monetization produces. Any vendor pitching AI-driven margin expansion carries the same hidden drag, whether or not it appears on the slide.

   Put two companies from the same boom side by side. Alphabet consumed cash to build. Stripe generated it. Revenue up 33% to $6.8B, free cash flow up 52% to $3.2B, and the proceeds now funding a reported $53B, PE-blended bid for PayPal. The toll-taker prints while the builder bleeds. Meta has read the same math and is monetizing its own overbuild: a reported $10B compute lease to Anthropic, with AWS's Dave Brown hired to run it. Capital markets have already priced the divergence. 2026 biotech IPOs are up 55% on average; the ten biggest AI-adjacent listings are down 6%.

   For leadership, the near-term consequence is concrete: cloud pricing leverage has moved to buyers. A hyperscaler under free-cash-flow scrutiny will concede more on committed-use discounts, capacity guarantees, and price than it would have a year ago. A reasonable skeptic would note that hyperscalers have absorbed capex cycles before and kept pricing discipline. The skeptic is right about history. The difference is that this window exists precisely because markets treat the current capex level as abnormal, and it narrows the moment they normalize it.

   The second consequence is M&A. AI cash is arming acquirers while AI cost risk starves mid-tier builders, and roughly 160 enterprise software startups are flagged as likely sale candidates this year. That is a sourcing list as much as a market signal. The concentration risk cuts both ways. If Stripe-PayPal closes, payments becomes a two-horse race, and every processor relationship deserves a review on that basis alone.

   The tradeoff here is timing, not direction. Negotiating now, while the leverage is fresh, costs some organizational attention this quarter. Waiting until the next earnings cycle resets expectations costs the leverage itself.

   What to do

   1. →

      Open cloud vendor renegotiation this quarter, using disclosed cash-burn and raised capex guidance as leverage on committed-use pricing and capacity guarantees.

      NowBuyer leverage exists only while markets treat current capex as abnormal — it compresses once expectations normalize.
   2. →

      Map vendor concentration to Stripe/PayPal and screen the ~160 flagged for-sale software startups for tuck-in or partnership targets before the window narrows.

      This quarterCash-rich strategics are moving now; concentration risk is cheap to diagnose early and expensive mid-crisis.
2. 02

   A 28x AI Budget Bought Speed, Not Progress

   The velocity numbers are real; the divergence beneath them is the board-relevant fact — and the companies actually winning aren't the ones spending most.

   Adoption was never the question — translation is. Across 500+ engineering teams, AI-generated code crossed 50% of everything shipped (from 34% a quarter earlier) and median throughput rose 37%. Yet the Developer Experience Index slipped from 67 to 65, median pull-request size nearly doubled, and AI budgets climbed 28x against flat innovation output. Speed went one way; delivery health went the other. The mechanism isn't mysterious: AI was bolted onto a review-and-delivery pipeline built for human-paced output, so bigger PRs pile up behind slower reviews and technical debt compounds quietly beneath the velocity chart.

   Set that against who is actually winning. Poolside's Laguna S 2.1 — 118B parameters, 8B active — reportedly beat a rival roughly 10x its size, with the edge attributed to behavior (verification, persistence) rather than raw scale, built by a 115-person team on 5-8 week cycles. Midjourney runs nine-figure revenue on about 40 people and zero outside capital. The pattern is consistent and it contradicts the reflex to spend: the leaders aren't buying more compute or more seats, they're buying engineering discipline and judgment.

   For a Leader, the trap is reporting the throughput number upward while the 28x line waits in the CFO's deck. A falling experience index against a 28x budget increase means the organization is funding speed, not progress — and the gains that do exist concentrate in small, tech-native pockets, so an org-wide average hides both where the leverage sits and where the risk is accruing.

   The move is to instrument value before budget scrutiny forces the conversation on someone else's terms, and to redesign the review layer rather than tune it. A throughput claim that arrives without a paired health signal is a claim you cannot defend.

   What to do

   1. →

      Stand up an internal DXI-equivalent health metric before the next budget cycle so every throughput claim arrives paired with a delivery-health signal.

      This quarterOwning the value narrative internally beats reacting to a CFO's 28x-spend question with no counter-data.
   2. →

      Commission an AI-era redesign of the code-review layer — capped PR size, mandated incremental delivery, AI-assisted review — not incremental tuning.

      This quarterDoubled PR size is an early technical-debt indicator; the fix is process architecture, not more tooling.
3. 03

   AI Liability Just Got Three Price Tags

   Training data and agent execution each produced a concrete dollar figure this cycle — extending a chatbot-output precedent from a year ago — and the bill lands on the deploying organization, not the model vendor.

   For two years AI liability sat in the theoretical column of the risk register. This cycle it acquired dollar figures, on three distinct failure surfaces, extending a precedent set over a year ago. On training data, Anthropic settled book-piracy claims for $1.5B, the number that will anchor every future copyright negotiation and diligence checklist. On output, the year-old tribunal ruling that held Air Canada legally liable for a bereavement-fare policy its chatbot invented remains the anchor precedent, and Cursor absorbed a cancellation wave after its support agent fabricated a device-limit rule. Hallucination is now enforceable liability and churn, not a PR footnote. On execution, Amazon's Kiro assistant deleted a production environment with operator-level credentials, and three incidents cost an estimated 6.3 million orders, triggering two-person approval across 335 critical systems.

   Read together, these say one thing from three angles. As agents move from suggestion to action, accountability does not disappear. It gets harder to locate, and the bill lands on the deploying organization rather than the model vendor. A reasonable skeptic would call these isolated failures of immature tooling that better prompts will fix. The compounding math answers the skeptic. At 95% per-step accuracy across 20 chained steps, end-to-end success falls to roughly one in three. That is an architecture constraint, not a prompt-tuning problem.

   For a leader, the exposure is legal, brand, and operational at once, which makes it a governance question rather than an engineering footnote. The organizations handling it well share two unglamorous habits. They know the provenance of every model's training data, and they gate any agent that can commit to pricing, policy, or production changes behind human approval. Neither habit shows up in a demo. Both show up in a deposition. Amazon's response is the preview of what enterprise customers and regulators will write into security questionnaires within a year.

   The sensible position is to price the exposure now, treating the $1.5B figure as a floor rather than a ceiling for unlicensed-data risk at an organization's own scale, and to ship the approval gate before an incident ships it instead. The alternative is waiting for the questionnaire to arrive with the terms already written.

   What to do

   1. →

      Commission a legal and technical training-data provenance audit across all internal and third-party models in production this quarter.

      This quarterThe $1.5B settlement is now the reference number in copyright negotiations and diligence; unproven provenance is a modeled liability.
   2. →

      Mandate human-in-the-loop approval gates on any agent with write access to production or authority to commit on pricing, policy, or refunds.

      NowThe Air Canada ruling makes agent commitments enforceable, and the Kiro precedent shows the blast radius is real and fast.