The Investor

The Divorce, Quarter by Quarter

Microsoft's Build 2026 keynote on Tuesday is being marketed as a product launch. It reads more like a strategic divorce announcement, or rather the first one Microsoft has been willing to put on a slide. Two months after restructuring its OpenAI relationship — keeping 27% of the new PBC, converting Azure exclusivity into a right of first refusal, and freeing OpenAI to contract directly with other clouds — Redmond is unveiling its own models for transcription, image generation, reasoning, and coding.

The framing is studiously humble: "good enough for simpler tasks." The economics underneath are not. AWS has its own stack, Google has its own stack, and Microsoft has now joined them, which means the business of reselling someone else's frontier — the channel that drove OpenAI and Anthropic enterprise growth through 2024 and 2025 — does not compound the way the bull case assumed.

Microsoft has free access to OpenAI's IP until 2032. That is six years of legal cover to replicate what OpenAI is still spending capital to build.

Three Scenarios Worth Modeling

The capex Microsoft is no longer obligated to spend on OpenAI's behalf can go three places, and it is worth being explicit about which one you are paying for:

1. Redirect to first-party models and Copilot inferencing, with the savings landing as margin inside four quarters. The upside case.
2. Spending continues at roughly the current pace while the customer mix broadens, and very little changes at the consolidated level. The base case.
3. OpenAI's new multi-sourcing freedom surfaces a capacity overhang somewhere in 2026 that nobody is currently modeling. The tail risk the sell-side will discover last.

Oracle already booked part of this shift in its $300B OpenAI commitment announced in September. CoreWeave and the merchant compute layer can now quote into a buyer that is no longer captive. The picks-and-shovels names — Nvidia, the power IPPs, the HVAC and switchgear suppliers — pick up another bidder at the same auction.

Foundation Model Lab Implications

For direct or SPV holders of OpenAI and Anthropic, the read-throughs are narrower than the headline suggests:

• OpenAI's compute is no longer a Microsoft annuity. Whatever lifetime value Azure was implicitly carrying on that single customer needs to come down.
• Anthropic's position with Amazon and Google now looks slightly less differentiated, because exclusivity was part of what made the comparison flattering in the first place.
• Pricing power on the "simpler tasks" tier is structurally capped the moment the hyperscalers' homegrown models ship.

This is probably wrong, but: Microsoft's own models could underperform and force a quiet retreat back into OpenAI dependence, in which case the "good enough for simpler tasks" framing was a hedge dressed as a victory lap. Even then, the IP clause running to 2032 keeps the failure scenario manageable. That is the deal.

The Co-Option Play

Nvidia putting $2 billion into Marvell alongside the NVLink Fusion announcement is not defensive. It is co-option, or rather the more interesting version of co-option: Nvidia is letting Google, Amazon, and Microsoft build their own ASICs on the explicit condition that the silicon plugs into Nvidia's networking fabric. The moat moved from the GPU to the interconnect while everyone was reading benchmark scores.

Compute scarcity is being solved in public while interconnect scarcity is being solved in filings. CPO and interconnect plays are where infra alpha migrates when the GPU trade gets crowded.

Why Now — The Workload Shape Is Changing

Two independent signals are converging. First, OpenAI multi-sourcing compute means more buyers at the networking auction, because every new data center relationship requires interconnect, not just more GPUs. Second, reinforcement learning has quietly become the dominant post-training methodology, which requires heavy inference loops rather than long pretraining runs, and that workload shape wants different cluster architectures with more fabric and fewer raw FLOPs.

The capex committed last year was committed against the pretraining shape. A meaningful slice of announced buildout is probably mispriced for the workload it will actually run. Not stranded. Mispriced. That distinction is the entire investment case for the networking layer, because the remediation itself creates demand.

The Private Landscape

Two distinct go-to-markets are emerging in optical interconnect and CPO:

Both already have strategic buyers, which is rare in deep tech infrastructure and worth pausing on. The counter-thesis is the obvious one: hyperscalers could coordinate on open interconnect, UALink or Ultra Ethernet, and undercut Nvidia's fabric lock-in entirely. They could. They have also historically taken 3-5 years to resolve coordination problems of this shape, and that window is the trade.

Jensen Huang's trip to Taiwan to check supply chains tells you the bottleneck has moved from GPU fab to component scarcity, meaning substrates, HBM, optics, advanced packaging, liquid cooling. Revenue visibility in the Tier-2 names is better than public market consensus is pricing.

The Reallocation

Reinforcement learning has become the post-training substrate of the frontier model industry. It is not a research curiosity or a system-card footnote, it is the step that turns a raw model into something a customer will actually pay for. OpenAI runs RLHF, Anthropic runs Constitutional AI, DeepSeek ships GRPO. The talent thesis written eighteen months ago assumed the scarce input was people who could scale pretraining, and that thesis is aging badly.

The scarce input now is people who can design reward models, run non-gameable evals, and ship RL pipelines that don't silently collapse in week three. There are perhaps a few hundred of them globally, they are not cheap, and the ones who are cheap are cheap for a reason.

RL is becoming what backprop was five years ago — table stakes. The multiples on the infrastructure layer will eventually price it that way.

The Investment Thesis

The RL-tooling opportunity has reached the point where the buyer set is legible, which is the only thing that separates a market from a research grant. Every frontier lab plus a growing roster of enterprise post-training teams adds up to a real TAM. Four subcategories are worth mapping:

• Reward modeling platforms — the quality of the reward signal determines everything downstream.
• Synthetic preference data — the RLHF pipeline's most expensive input is being automated.
• RLHF/GRPO tooling — purpose-built infrastructure for the post-training loop.
• RL-specific evaluation — the unsolved problem of detecting silent pipeline collapse.

These are still priced like speculative bets rather than infrastructure, which is probably wrong but is the version of wrong an early entrant gets paid for. The window for Seed/Series A entry at reasonable multiples looks like two to three quarters, based on the pace at which labs are buying rather than building.

The Risk Worth Pricing

Two failure modes matter. RL-as-post-training could collapse into a commodity inside the labs themselves, in which case the tooling layer gets squeezed from above and the talent premium decays into a normal engineering wage. Or, the more interesting version, the talent and the silicon both get re-sourced at the same time by the same buyers in the same quarter, which is where opportunity cost stops being abstract and starts hitting someone's income statement.

For portfolio companies doing custom post-training, model RL-engineer comp inflation into burn assumptions now rather than at the next board meeting. A managed post-training vendor already in the portfolio is both a hedge and, if the talent market tightens the way it looks like it will, a strategic asset the rest of the market will eventually quote a price on.