Open-weight models are reaching parity, so the moat moves from owning the model to orchestrating it. A model is a stateless function and commoditizes; a...

Infrastructure Is the New Moat

For the last few years the prevailing bet was that whoever owns the smartest model wins. Capital concentrates at the frontier, the best model attracts the most usage, usage funds the next training run, and the lead compounds into a natural monopoly. That story is starting to crack.

As open-weight models reach parity, the model stops being the moat and becomes a commodity input. The durable advantage moves upstream, from the model to the layer that orchestrates it.

The parity signal

Two reported developments push this from theory toward something concrete.

GLM-5, reported as a 744B mixture-of-experts model, ships under an MIT license with a low hallucination rate. Permissive licensing matters here as much as capability, because it removes the legal friction that kept open weights out of serious production.
Llama 4 extends usable context dramatically, with the Scout variant reported at roughly 10M tokens. Long context was one of the last clear advantages of closed frontier APIs.

If these reports hold, open weights are no longer chasing a moving target a year behind. They are landing close enough to closed frontier models that, for many tasks, model choice starts to look swappable rather than defining. That swap is rarely free. Prompts tuned for one model misbehave on another, and behavioral quirks differ enough that calling models fully interchangeable overstates the day-to-day friction. Parity is also task-dependent, and the strongest closed models still tend to lead on the hardest reasoning. The claim is narrower: open weights are good enough that the model is no longer the thing you build a company around.

Why commoditization moves the moat instead of destroying it

When a critical input commoditizes, value relocates to whatever layer is now scarce. Here that scarce layer is the execution environment: the framework that turns a raw model into reliable work.

That work is unglamorous and strategic. The environment swaps models underneath stable agent behavior, so a cheaper open-weight model can replace a closed one without rebuilding the system around it. It absorbs the differences between providers, the tool-calling formats, the context limits, and the idiosyncrasies in how each model follows instructions, so those differences stop leaking into the product. And it speaks shared protocols so agents built by different teams can interoperate. WebMCP, emerging as a W3C standard for agents, is the kind of substrate that could make the orchestration layer sticky the way HTTP made the browser sticky, if it matures.

A model that runs best on one framework loses to a framework that runs any model well.

Won't the framework commoditize too?

The obvious objection: if models are interchangeable, why won't orchestration frameworks become interchangeable glue code as well? It is the right question, and the answer depends on what each layer holds.

A model is close to a stateless function. You send tokens, you get tokens back, and nothing about your business persists inside it. That is exactly what makes it easy to commoditize, because there is nothing to leave behind when you switch.

A framework is the opposite. It accumulates state. Your agent definitions, tool integrations, memory, evaluation history, and the hard-won knowledge of which configurations actually survive production all live there. That context is specific to you and grows more valuable the longer it runs. Glue code commoditizes; the system of record around your agents does not, because the cost of recreating it climbs with every integration you add. Protocols reinforce the effect. Once interoperability standards settle, the framework that implements them well sits at the center of a network, and network position is far harder to copy than a feature.

A read on adoption

Developer attention is one early indicator, and it is moving toward the framework layer. The most starred agent frameworks on GitHub now sit in the hundreds of thousands of stars, which signals where builders are spending their time even though stars are interest rather than revenue or lock-in. Taken alone this would be thin evidence. It matters because it lines up with the structural argument: the network effects and switching costs that everyone assumed would accrue to the model provider look more likely to settle on the orchestrator. Once your agents, tools, memory, and integrations live inside a framework, the underlying model becomes a setting you change rather than a vendor you marry.

What this predicts

If the thesis holds, the second half of 2026 is where durable advantage forms at the orchestration and protocol layer rather than at the model layer. A few testable implications:

Spend on closed frontier APIs comes under pressure for routine agent work, since an open-weight model behind a good framework does the job at lower marginal cost.
Standards like WebMCP gain weight, because interoperability is worth more once no single model is irreplaceable.
The competitive question shifts from which model is smartest to which environment makes any competent model reliable, swappable, and integrated.

The timeline is the soft part of this. Frontier labs could reopen a capability gap and make the model the moat again. The structural claim is sturdier than any date: when an input commoditizes, value migrates to whatever stays scarce, and right now that is the layer holding your agents, your data, and your integrations.

The model is becoming a compute engine, and the framework is becoming the operating system that runs it.