The Board Room

The Convergence That Changes Everything

Three data points landed this week that, individually, are impressive. Together, they describe a structural break in how software gets built. Anthropic confirmed Claude writes over 90% of its own code. GitHub's CPO disclosed 17 million agent-generated pull requests in March 2026 alone — with platform growth running at 3x internal forecasts. And Bain reported that human oversight is now the primary friction slowing AI-driven cost savings. Read together: AI code is production-ready, it survives human review at scale, and the humans reviewing it are the bottleneck.

The question was never whether agents could generate code. The question was whether they could generate code that survives review at scale. Seventeen million is not a number you reach by failing review.

The December 2025 Capability Jump

GitHub's CPO Mario Rodriguez pinpointed December 2025 as the inflection — the shift from "micro-delegation" (autocomplete) to "macro-delegation" (agent completes defined work units, human reviews rather than corrects). This single capability shift produced the 17M PR number three months later and has driven infrastructure demand that surprised even GitHub, requiring physical capacity expansion.

The Cost Structure Decoupling

The less-discussed but operationally devastating development: GitHub moves to usage-based billing on June 1, 2026. Your Copilot spend is no longer a predictable seat-based line item — it now scales with agent activity, which is growing at multiples. GitHub simultaneously released Chronicle (session analytics) and cheaper routing models, acknowledging this is the friction point that stalls enterprise rollout. Organizations that don't establish AI FinOps governance before the switch will face surprise bills in Q3.

The Workforce Planning Implication

The Kauffman Foundation data provides the macro frame: startup job creation has fallen 33% since 1997 (7.9 to 5.3 per 1,000 people) — and that decline predates the current AI cycle. A company founded in 2026 will compete for your market with 15 people and a stack of agentic systems replacing two or three departments. The revenue-per-employee gap between lean entrants and incumbents is widening in the same direction.

The headcount model is not just inefficient. It is a signal to the market about how the company believes leverage is created, and the market is repricing that belief.

Where Sources Diverge

There is a tension worth surfacing. Multiple sources confirm the autonomous code production reality, but they diverge on timeline. One source frames this as a 12-month window to restructure; another suggests 18 months; a third implies the repricing happens in 2-3 quarters as the usage-based billing hits and agent PRs reach 30-50% of total volume. The most conservative reading still demands action within the current planning cycle. The aggressive reading says you're already late.

What This Means for the Engineering Org

The human role shifts from builder to architect and judge. The org design, leveling ladders, and hiring profiles that match this shift are a 12-18 month project, not a quarterly one. Companies that figure out human-to-agent structure first will run at 2-3x feature velocity at the same headcount. That window is measured in quarters, not years.

The Vendor Landscape Just Changed

SpaceX is now booking $2.17 billion per month in committed compute revenue from Google and Anthropic alone. Annualized, that clears twenty-six billion dollars, which puts an aerospace company in the revenue band of the named cloud incumbents. The reasonable read is not that this is a side experiment or a pilot dressed up for press. It is a hyperscale cloud operation that arrived outside the procurement vendor matrix every enterprise was using last quarter. The roster of firms operating at hyperscaler cadence grew by one, and nobody's RFP template has a row for it.

A reasonable skeptic would call SpaceX a niche provider. The skeptic would have to explain how a niche provider books two billion dollars a month.

The Physical Constraint Is Real

Meta is putting GPUs inside 125,000 square-foot temporary tent structures with off-grid power because conventional data center construction takes two to three years and the workload schedule does not. A company that can write a ten-figure check chose fabric over an eighteen-month building. That is not a flex. It is a concession, and it tells you where the compute supply curve actually sits. The gap between what is needed and what can be delivered through normal construction is structural, not cyclical.

Contract Volatility

The detail in Google's thirty-billion-dollar SpaceX deal that matters for procurement is the 90-day cancellation clause. Both sides are treating current compute pricing as provisional. Either GPU supply catches up and prices fall, or contracts get renegotiated higher. Any cloud commitment signed today for 2027 workloads should be modeled the same way. Long-dated capacity is the asset. Month-to-month flexibility is the exposure.

The Infrastructure Buildout Is Not Slowing

Several signals point to expansion rather than optimization:

• SoftBank has committed seventy-five billion euros to French data centers.
• Google rented 110,000 Nvidia chips from SpaceX, which is what sourcing from an aerospace company looks like when conventional supply is tapped out.
• Buffett placed a ten-billion-dollar bet on Alphabet's AI infrastructure thesis, and Buffett is not a man known for chasing momentum.
• SpaceX IPO is targeting June 28, which will further capitalize the compute ambitions already in motion.

The implicit market view is that agent-era workloads will be orders of magnitude more compute-intensive than chatbot workloads. The firms betting against that are betting against more than a hundred billion dollars of already-committed capital.

What This Changes

GPU allocation, power contracts, and long-dated capacity are now being competed for by a wider pool of buyers than any procurement model from last year accounts for. The firms able to sign the largest commitment fastest hold pricing power for the next cycle. Anyone modeling vendor leverage on the previous distribution of named buyers is modeling a market that no longer exists. The contracts signed in the next four quarters will look decisive in two years.

Codex disappears into ChatGPT

The decision to merge Codex into ChatGPT is being read as product simplification. It is bundling. Microsoft ran this play on browsers in the nineties. Salesforce did a version of it to point CRM tools a decade later, and AWS did a more thorough version to standalone infrastructure providers. Folding the coding product into a general-purpose assistant with more than 200 million users puts a clock on every standalone AI coding tool in the market. Cursor, Replit, and the rest have raised enough capital to outlast a bundle on an eighteen-month view. The five-year view is harder to argue with a straight face. The category now has an expiration date on it.

Cognition picks a different fight

Cognition's simultaneous repositioning as the "Switzerland of AI Agents" is the more interesting move of the week. It is what a company does when it has decided that competing on per-agent capability against OpenAI, Anthropic, and Google is not a fight it can win. The structural bet is that enterprises will demand interoperability at the agent layer even while they accept vendor lock-in at the model layer. The multi-cloud management arc took roughly a decade to play out in cloud. The agent version looks compressed to two or three years.

The AI market is crossing from the land-grab phase, where point solutions proliferate and everyone gets funded, into the consolidation phase, where platforms absorb features and the remaining independents must find a structural position they can actually defend.

Open weights close the gap

Consolidation is being accelerated by a development we flagged a few weeks ago and which has only firmed up since: open-weight models have reached near-parity with frontier closed models. Moonshot's Kimi K2.5, Zhipu's GLM-5, MiniMax M3 with its million-token context, and Google's Gemma 4 running laptop-class multimodal are close enough on the benchmarks that matter that pricing power for closed-model providers erodes from here. Any product whose moat had quietly reduced to "we have access to the best model" has watched that moat drain over two quarters.

The three positions still worth defending

The middle is being squeezed from both ends at the same time. Undifferentiated AI products without platform scale, neutrality positioning, or deep vertical expertise are not in one price war. They are in two. The bundle compresses pricing from above. Open-weight commoditization compresses it from below.

Inference gets cheaper, again

Efficient reasoning techniques (RLVR auto-verification, Qwen's sparse MoE, Gemini's adaptive thinking) are driving 3-5x inference cost reductions inside twelve months. A reasonable skeptic would point out that we said something similar last year and the savings mostly got absorbed into longer reasoning traces. The reasonable skeptic is correct. What is different this cycle is that features previously shelved on margin grounds are being re-priced and redeployed in production rather than in demos. The organizations that ship at the new cost point first capture user habits before slower competitors finish updating their business cases.