On October 15, 2026, four AI labs shipped open-weight models inside a single news cycle. The models collected the headlines. The anomaly that deserved them was a single integer: 20,000,000. StepFun, Meta, Kimi, and Mistral โ four independent legal teams, four separate drafting processes, four distinct corporate cultures โ all planted the same dollar figure into their commercial-use thresholds. Tracing the gas trail back to the genesis block, that convergence is not a coincidence of taste. It is a distributed system reaching consensus on where "small commercial use" ends and "call our sales team" begins. I have spent enough hours inside assembly to distrust coincidence dressed as policy. When four isolated actors converge on one constant, the constant is doing work nobody is advertising. The work, in this case, is a conditional statement โ and conditionals are what I audit.
"Open" used to be a technical property. Somewhere between 2024 and this release cycle it became a distribution channel. The technical premise is thinner than the marketing: Meta's open line, Muse Spark Open, scores 39 on the Artificial Analysis Intelligence Index. Its closed sibling, Wonder, scores 56. A 17-point gap, roughly 30% relative. Read that as a limitation and you miss the design. Read it as a tier and the architecture snaps into focus โ the open line captures developers and ecosystem real estate, the closed line captures the inference budgets of anyone who needs frontier reasoning. This is freemium recompiled at the model layer. What the four labs actually shipped was not four products but three licensing architectures wearing product clothing: unrestricted (DeepSeek V4 and GLM-5.3 Flash under MIT and Apache 2.0), revenue-gated (StepFun, Meta, Kimi, Mistral), and hardware-aligned (NVIDIA's Nemotron Bolt under OpenMDW-1.1). None of this is described in a whitepaper. It is described in clause 2c-i, and clause 2c-i is where the real engineering lives.
Start with the integer, because the integer is a require() statement. Twenty million dollars is the boundary condition; cross it and execution reverts from self-service to enterprise sales. But here is the detail that separates a casual reader from an auditor: the clocks differ. StepFun meters monthly. Meta meters annually. Kimi and Mistral use a rolling twelve-month window. Same constant, different block times. Mistral's monthly snapshot is the aggressive configuration โ for a company with volatile revenue, a single strong month can trip the gate before the fiscal year closes, converting a free deployment into a negotiation overnight. That asymmetry is not sloppiness. It is a compliance-cost moat. Most deployers will never model the difference between a monthly and a rolling-annual trigger until they are already inside the funnel, and by then the switching cost is measured in retraining, not refunds.
Then there is StepFun's OEM gate, which the report flags as unique and which I read as the sharpest edge in the entire document. Clause 2c-i requires device manufacturers โ phones, cars, glasses, wearables, robots โ to negotiate a separate commercial license regardless of revenue. Not "above the threshold." Regardless. This is the first open-weight license I have seen that treats hardware embedding as a distinct execution path with its own gas cost. In audit terms, it is a modifier applied to the transfer function: you may move the weights, but you may not move them into silicon without a new signature. If this pattern propagates, model licenses stop being software licenses and become hardware-selection locks. An enterprise that embeds a gated model into a product loses the freedom to migrate that model to cheaper or more efficient silicon later, because the migration itself triggers renegotiation.
NVIDIA's move is the mirror image, and it is the one I find structurally honest. Nemotron Bolt ships fully open โ weights, training data, recipe โ under OpenMDW-1.1, and the RTX Spark hardware landed in the US market the following day. Every open model that runs efficiently on NVIDIA silicon pulls demand for NVIDIA chips; therefore a revenue gate on the model would be a landlord restricting the software that sells his own hardware. The subsidy is rational because the revenue was never in the model. It was in the die. This is the same vertical-integration logic I traced through EigenLayer's restaking thresholds in 2024 โ the security budget and the slashing condition have to be legible to each other or the whole structure is theater. Here the model, the chip, the inference stack, and the license are legible to each other by design. The catch is buried in the phrase "runs efficiently on NVIDIA hardware," which in practice means CUDA-optimized, which means a soft lock dressed as an open standard.
And then the clause almost nobody reads: license-follows-derivative. If that propagates, distillation inherits it. A model fine-tuned or distilled from a gated base may be forced to carry the original terms, producing a contagion effect closer to GPL than to anything the AI industry has shipped before. The downstream consequence is a distillation supply chain where provenance is legally load-bearing, and where a startup can wake up owning a model it did not know it had licensed. In the absence of trust, verify everything twice โ and here the thing to verify is not the weights, it is the inheritance graph.
My Uniswap V2 fork audit in 2020 taught me the cost of ignoring the boring layer. I spent 120 hours tracing the swap function's gas optimizations and found an arithmetic overflow in a custom fee-distribution path that the team's own reviewers had skipped because it looked like plumbing. The lesson was not that overflows are common. The lesson was that the value leaks through the mechanism everyone assumes is settled. The open-weight TCO conversation has the same shape. "Weights free" is the marketing line; inference compute, legal review, and ops are the fee-distribution path nobody traced. Long-context, high-batch inference can cost more than a closed API once you price the compliance team that has to read clause 2c-i for every product SKU.
Here is the blind spot the entire discourse is stepping over. Everyone is auditing the model. Nobody is auditing the license bytecode. As AI agents move on-chain โ and they are moving, I built a prototype in 2025 that let an LLM execute DeFi trades through a signing oracle โ the model's license becomes an off-chain invariant that gates an on-chain action. An agent running an OEM-gated model inside a DeFi vault is a latent compliance exploit: the contract executes, the license reverts, and the revert is not on-chain, so no one sees it until a counterparty's lawyer does. Code is law until the reentrancy attack, and the reentrancy here is a derivative model that inherits a license it never read. The 17-point capability gap, meanwhile, is not a flaw to be closed โ it is the tier boundary held deliberately open so that frontier demand stays behind the paywall. And the convergence of four labs on twenty million dollars deserves a question nobody is asking: independent decisions do not usually produce identical constants. That is either consensus or coordination, and one of those has an antitrust dimension.
Entropy increases, but the invariant holds. The invariant this cycle is that "open" no longer describes the weights โ it describes the gate, and the gate is the product. My forecast is unglamorous: within eighteen months, "permissionless" will migrate from a decentralization slogan to a security property, and the first serious AI-license incident will not be a model jailbreak but a deployment that inherited a clause it never audited. The labs are betting that developers read the README and skip the license. The question worth carrying forward is simple โ when your agent signs a transaction, do you actually know which contract it is running under, or only which weights you downloaded?