The Board Room

Deep dives

1. 01

   Middleware Is the Kill Zone — Own a Vertical or Own the Control Plane

   When the model itself performs distributed systems engineering — writing programs to coordinate tools, holding state across multi-hour sessions, choosing between scripting and direct interface manipulation — the customer's question is brutal: why pay for your orchestration when the runtime orchestrates natively? At monthly release cadence, 'the labs won't ship our feature' is no strategy.

   Follow where a billion-dollar week did not go. Norm Ai ($1.2B, agentic law) and Lovable ($13.2B, code generation) got funded for vertical depth; SambaNova, Positron, and Prime Intellect for infrastructure. Horizontal tooling — wrappers, thin orchestration, general-purpose middleware — is squeezed from both sides, and investors are pricing that before operators are.

   Two quieter moves show layer ownership. NVIDIA open-sourced 110+ agent skills portable across Claude Code, Codex, Cursor, and Kiro — quietly teaching every agent to depend on CUDA-X libraries. Open spec, closed gravity: the Android playbook for AI infrastructure. And the routing proxy is emerging as the API gateway of the AI era — open-source proxies like Plano claim 2x cost reduction with zero code changes. Prompt caching is model-specific, so a cache-hot expensive model beats a cache-cold cheap one: per-call routing is broken; session-level pinning is the pattern. Whoever owns that layer owns cost observability and vendor optionality.

   ---

   Next 90 days:

   1. Map every product against what GPT-5.6, ChatGPT Work, and Muse Spark do natively.
   2. Where overlap exists, differentiate on proprietary data and vertical expertise — or reposition as infrastructure these runtimes consume.
   3. Own the control plane: adopt routing as configurable infrastructure now, not single-provider volume commitments that trade away optionality.

   In an agentic-runtime market, the only defensible positions are below the model (infrastructure) or above it (vertical data) — everything in between is a feature awaiting absorption.

   What to do

   1. →

      Map every product line against GPT-5.6, ChatGPT Work, and Muse Spark 1.1 within 30 days; flag features the runtimes now ship natively and assign a differentiate-or-reposition call to a named owner

      NowLabs ship monthly; every quarter of tolerated overlap compresses pricing power and exit options simultaneously
   2. →

      Stand up an open-source routing proxy (Plano or equivalent) as the control plane between all agent services and model providers this quarter, with session-level model pinning as default policy

      This quarterRouting is the chokepoint for cost observability and vendor independence — the proprietary-advantage window is closing, but the optionality window is wide open
2. 02

   Your Engineers' Keystrokes Trained a Competitor's Model

   Assume worst case as base case: if your engineers use Cursor, your proprietary code patterns, API designs, and architecture are plausibly in Grok 4.5 — rentable by any competitor at $2 per million input tokens. xAI's disclosure that it trained on the 'entire data from Cursor' is a structural trust failure across the developer-tool ecosystem, and it resets what procurement must demand from every AI vendor.

   The second-order effect is a new moat: vendors who can credibly commit to training-data isolation will win enterprise deals on compliance, not capability. Expect data-usage guarantees as table stakes in renewals within two quarters — and expect your customers to ask you the same question about the AI in your products.

   The sovereignty response is already in production. JPMorgan runs SambaNova SN40/SN50 on-prem for inference. Ollama hit 8.9M monthly developers across 85% of the Fortune 500 — doubling in six months with 14 employees. Prime Intellect raised $130M to let enterprises train their own models instead of renting APIs. This is risk management for regulated workloads, not anti-cloud ideology, and it's creating a parallel infrastructure market. Companies whose tooling works identically on-prem and in-cloud will own the next enterprise buying cycle.

   The geopolitical layer compounds it: China classifying Claude Code's telemetry as a 'backdoor' while Anthropic calls it abuse prevention previews regulatory incompatibility forcing parallel AI stacks for global operators. One-world AI architecture is becoming untenable as a planning assumption.

   Data provenance just became a procurement gate: the coding AI vendor who can prove what they did NOT train on wins the enterprise.

   What to do

   1. →

      Audit data exposure across every developer tool (Cursor, Copilot, all coding assistants) within two weeks; demand written training-data isolation guarantees as a condition of every renewal

      NowThe Grok-Cursor precedent makes interaction data a training asset by default — contract silence is consent
   2. →

      Pilot a private-instance or on-prem coding AI deployment for your most sensitive repositories this quarter, using the JPMorgan/SambaNova and Ollama patterns as reference architectures

      This quarterSovereign inference is in production at the largest regulated enterprises — the tooling maturity risk is gone
3. 03

   Rising Inputs, Subsidized Outputs: The Squeeze Your 2026 Budget Missed

   The contradiction is the useful signal. API prices are falling while underlying compute repriced upward. H100 contract rates bottomed at $1.70/hr in October and rebounded 38% to $2.35/hr. Spot is up 10% year-to-date. The GPU glut thesis, which this column entertained a quarter ago, is done. Meta and xAI are losing money per token to buy share. Both dynamics cannot hold, and the honest posture assumes 3-5x API price normalization within 18 months. A P&L built on today's subsidized rates is a bet on the subsidy, not the technology.

   The memory market is running the same cycle from the other side. Memory stocks up 200-700% YTD, with Micron margins above the 2018 peak, is textbook commodity cyclicality. 2023 losses stopped factory investment. AI demand met constrained supply. Prices exploded. The new capacity that pricing is now attracting ends, as it always does, in oversupply. Every multi-year contract at today's memory pricing will look expensive in 24 months. The asymmetric position is short-duration commitments now, scaling as the cycle turns, so the cost basis falls while competitors sit locked in at the peak.

   The procurement tell is in the paper. SpaceX's three-year infrastructure deal carries 90-day cancellation provisions. When the most committed buyer still demands quarterly optionality, hedging against custom silicon making current fleets obsolete, the clause is worth more than the headline discount. It survives only until a seller's market removes it.

   ---

   Board framing: compute is a structurally repriced input, not a declining one. The posture is a barbell. It consumes subsidized API pricing near-term while writing contract structures that assume both the subsidy's end and the memory cycle's reversal.

   The arbitrage is real for a limited window, and the infrastructure contracts that survive are the ones that assume the subsidy ends in 18 months, because it will.

   What to do

   1. →

      Re-baseline 2026-27 AI infrastructure budgets against 25-40% higher compute costs before annual planning locks, and stress-test embedded-AI product P&Ls against 3-5x API cost normalization

      This quarterMost 2026 plans assumed declining compute — October's price floor already falsified that
   2. →

      Insert 90-day exit clauses into every infrastructure contract signed this quarter and refuse multi-year memory or compute lock-ins at current pricing

      NowMemory margins are at levels that historically precede collapse; peak-price lock-ins convert a temporary condition into a permanent cost disadvantage
4. 04

   Four Cash-Burning Giants Just Put Themselves on the Clock

   Unprofitable companies that IPO during speculative phases face a reckoning 4-8 quarters later, when public investors demand a path to profitability. OpenAI, Anthropic, SpaceX, and Stripe just accepted that clock. For profitable incumbents competing against them, plan two phases: aggressive spending through mid-2027, then forced rationalization — with the first earnings-pressure window opening Q3-Q4 2026. Pre-position pricing, hiring, and retention plays for phase two while competitors are still in phase one.

   Netflix is the cautionary mechanism. Revenue grew 16.2% in Q1 — yet market cap fell 44% in a year while the S&P gained 22%, because projected growth decelerated from 15.9% to 12-14%. A 2-3 point deceleration triggered a valuation regime change no execution could offset, and management's failed bid for Warner Bros. Discovery (outbid by Paramount Skydance) shows they saw it coming and tried to buy a new narrative. Every growth-multiple company should identify its own WBD-equivalent move — and the trigger date.

   The rotation signal is the actionable part. Quality began outperforming in June-July, Berkshire is beating the S&P, and quality assets trade at a 15% discount to the index while compounding intrinsic value at ~20%. While momentum capital chases IPO paper, genuine moats — pricing power, recurring revenue, switching costs — are available at rational prices. That window closes when the rotation re-rates them.

   The market is about to start paying for profitability again — the disciplined acquirer's window is open precisely because everyone else is watching the IPO tape.

   What to do

   1. →

      Build an acquisition shortlist of quality assets trading at the current discount this quarter, prioritizing durable-moat targets ignored by momentum capital

      This quarterThe quality rotation is early; the discount compresses as it accelerates, and the best assets go first
   2. →

      War-game newly public AI competitors' first earnings cycles (Q3-Q4 2026): pre-position pricing, talent, and customer-win plays for their rationalization phase

      This quarterQuarterly discipline forces cash-burning rivals to retreat on a predictable schedule — advantage goes to whoever planned two quarters early