The Board Room

Three Hyperscalers, One Playbook

Microsoft's Build 2026 announcement closes a pattern that has been forming for two years. All three hyperscalers — Azure, AWS, GCP — are now positioned identically: homegrown models for margin capture, third-party models for optionality. Three independent actors arriving at the same strategy in the same window is not a coincidence to be debated. It is the shape of the market the next decade will be built on.

Microsoft's three-year arc is the clearest read on what happened. Distribute OpenAI models in 2024. Offer a marketplace of frontier models in 2025. Ship homegrown models covering 80% of enterprise workloads in 2026. The F10 family targets transcription, image generation, and routine coding. That is textbook disruption-from-below, run by Microsoft against its own partner.

The differentiation that justified single-provider commitments is thinning out faster than the procurement cycles built to evaluate it.

The 2032 Clause Nobody Else Has

Microsoft retains access to OpenAI's intellectual property through 2032. That clause lets them ship on OpenAI's frontier innovations while building model independence in parallel. No other hyperscaler has this hedge. It is the kind of optionality that lets Microsoft race toward independence without paying a reasoning-benchmark tax during the transition.

The Reinvestment Signal from RL

The technique that separates a competent base model from a frontier product is now well-understood. Reinforcement learning in post-training is the lever. OpenAI uses RLHF, Anthropic uses Constitutional AI plus RL, DeepSeek uses GRPO. The implication for enterprise buyers is that the quality gap between hyperscaler homegrown models and independent frontier labs will narrow as RL expertise diffuses, and it is already diffusing. Compensation for RL talent doubled in 12 months. That number is what aggressive investment looks like before the model rankings move.

What This Means for Your Platform Strategy

A reasonable skeptic would point out that frontier labs still hold a real lead on the hardest reasoning tasks. The skeptic is correct, for now. The skeptic does not explain why architectures built on a singular frontier provider with privileged hyperscaler access survive a market where every hyperscaler is shipping its own substitutes. The question is no longer which model to buy. It is where in the stack the differentiation lives. Above the model, in workflows and data and agents, or at the model layer the hyperscalers have already claimed.

The firms that treat this as a procurement exercise will discover in the next budget cycle that they solved the wrong problem. This is an architecture decision dressed as a vendor decision, and the two have very different shelf lives.

The Co-Option Strategy

Customer diversification into custom silicon was inevitable, and Nvidia appears to have stopped pretending otherwise. Computex was the reveal of a more patient play: co-opt the diversification through networking. The Marvell partnership, NVLink Fusion, and co-packaged optics together mean that whatever chip a hyperscaler builds, buys from Broadcom, or commissions from TSMC, it still talks to its neighbors through Nvidia's fabric. The chip decision moves. The interconnect decision does not.

Build whatever chip you want. The chips still have to talk to each other, and Nvidia's networking is how they will do it.

Why Scarcity Amplifies This

AI hardware supply constraints are now confirmed as a strategic constraint, not a temporary bottleneck. At thousands of coordinating chips per system, the interconnect layer carries more weight than any individual processor, and whoever owns the system architecture during scarcity also owns allocation priority. The repositioning from chip supplier to system architect is the move with stickier economics and broader capture, and it is the one being made now.

The Strategic Implication

A reasonable skeptic would say that Broadcom, Marvell, and hyperscaler-designed silicon already break Nvidia's grip. The reasonable skeptic is half right. Chip independence is not Nvidia independence. The networking fabric is a different form of lock-in, and probably a deeper one, because it sits underneath the chip choice rather than alongside it. The decision facing organizations evaluating custom silicon is not whether to use Nvidia GPUs. The decision is whether Nvidia's networking becomes the system-of-record for AI infrastructure topology, because that decision sets the next five years of vendor leverage.

This does not require immediate action but should reshape how any multi-vendor chip strategy or custom silicon program in planning is evaluated.

GitHub's Quiet Crisis

When GitHub's COO says publicly that the platform wants to attract "MacBook users," the tell is not addressable market. It is platform relevance. The threat to GitHub is not another code repository. It is the class of AI coding tools, Cursor and the autonomous coding agents behind it, that quietly make the repository itself less strategically central. If agents work around the clock without supervision, the open question is whether the repo remains the center of gravity for software development or becomes another data store that agents read from and write to.

Microsoft's response, unifying Copilot and cutting "bloat," concedes the product had become confused. Simplification is a defensive move, and a defensive move does not define a category.

Google's Counter-Play: Train the Next Million Agent Developers

Google's free AI Agents course is running at 1.5 million learners per cohort. That is developer relations at infrastructure-company scale. The curriculum tells you where Google thinks agent development is heading: a five-stage engineering discipline that moves from fundamentals to tool integration to context engineering to evals and security to production deployment.

Vendors only build curricula for capabilities they expect to be teaching for the next five years.

The "vibe coding" framing, natural language as the development interface, does two things at once. It expands the addressable developer population, and it deepens dependency on Google Cloud and Kaggle. The lock-in is not in the education. It is in the infrastructure assumptions baked into the notebooks, deployment templates, and evaluation frameworks that 1.5 million developers will internalize as default.

The Convergence Point

A reasonable skeptic would say these are two unrelated stories about two different vendors. The skeptic would be wrong. Both point to the same structural shift: the center of gravity for software creation is moving from repositories to agent orchestration layers. Organizations stuck between stages two and three of agent maturity, tool integration into context engineering, will find that the hard engineering work, the evaluation frameworks, the safety guardrails, the observability layer, all lives in stages four and five. The platforms chosen this year will be the ones those problems get solved on next year.