Three weeks ago, four independent AI labs shipped flagship models on the same day. The weights were free. The licenses were not. StepFun, Meta, Kimi, and Mistral โ separate teams, separate jurisdictions, separate balance sheets โ all landed on the same number: $20 million in revenue before the "open" model turns commercial. Nobody coordinated. That's the part that should worry you.
I've spent 24 years watching protocols promise openness and then quietly meter it. I audited a Mumbai DEX in 2017 that marketed "permissionless liquidity" while its admin keys could freeze any pool. The code was open. The power wasn't. The AI labs just did the same thing at industrial scale โ except this time they wrote the metering into the license itself, in plain text, with clause numbers.
Here's the landscape. Open-weight models now fall into three camps, and the camp matters more than the benchmark.
First, the unrestricted camp. DeepSeek V4 and GLM-5.3 Flash ship under MIT and Apache 2.0. No revenue triggers, no hardware clauses, no negotiation. You can build, sell, embed, and walk away.
Second, the revenue-gated camp. StepFun, Meta's Muse Spark Open, Kimi K3, and Mistral Large 4 all set a $20 million threshold โ but they measure it differently. Some count monthly. Some annual. Mistral uses a rolling twelve-month snapshot. Same number, different clocks.
Third, the hardware-aligned camp. NVIDIA's Nemotron Bolt ships under OpenMDW-1.1 with weights, training data, and recipe fully open. Zero restrictions. NVIDIA doesn't sell model access. It sells the silicon the model runs on.
Four labs, one day, three philosophies. That's not a product cycle. That's an industry splitting along the fault line of how it intends to get paid โ and the split matters more than any benchmark score, because it decides who can build on what, and at what price.
For anyone who has lived inside token launches, this should feel familiar. You've seen "fair launch" become "fair launch with a 4-year vest." You've seen "community-owned" become "community-owned pending multisig approval." The words stay. The constraints migrate.
Let me get into the actual mechanics, because the details are where the strategy lives.
The $20 million convergence is not a coincidence, and it's not collusion. It's calibration. Four labs independently drew the line between "small commercial use" and "we want a sales conversation." Below the threshold, you self-serve. Above it, you don't get a price sheet โ you get a negotiation. The threshold isn't a revenue share. It's a lead-qualification filter. The moment you cross it, you stop being a user and start being a pipeline entry. And the number travels. Once one lab sets it, the next lab's lawyers have a precedent.
That's the same logic behind tiered RPC access and "enterprise" API plans. The free tier isn't generosity. It's a funnel.
Mistral's choice to measure monthly is the sharpest detail in the whole set. A monthly snapshot punishes volatile revenue. A seasonal business โ retail, ad-driven, anything cyclical โ can trip the threshold in a good quarter and wake up owing a commercial license in a bad one. A rolling twelve-month window smooths that. Mistral didn't smooth it. Choosing the monthly clock is a deliberate decision to convert faster. It's the licensing equivalent of a liquidation threshold set one tick below where you're trading.
Then there's StepFun's OEM clause. Section 2c-i requires device makers โ phones, cars, glasses, wearables, robots โ to negotiate a separate commercial license regardless of revenue. Not "above $20M." Regardless. A two-person startup shipping a smart speaker has to enter a negotiation it has no leverage to win.
I ran a version of this math during the 2022 bear. I was auditing Layer 2 state root calculations across Optimism and Arbitrum, over 100,000 transactions, and what I found wasn't a throughput problem โ it was a dependency problem. Sequencers that looked decentralized depended on a handful of keys. The architecture said "open." The operational reality said "gated." The AI labs have simply moved the gate from the runtime to the contract. It's cleaner. It's more honest, in a way. At least you can read it before you deploy.
Now consider NVIDIA, and the case gets sharper. Nemotron Bolt dropped with a fully open license. The next day, RTX Spark hardware hit the US market. That's not two events. That's one product launch in two halves. Every open model that runs "efficiently" on NVIDIA hardware pulls demand toward NVIDIA chips. When you don't sell the model, giving the model away is not altruism โ it's demand generation. The landlord opens the apartment because the rent comes from the building, not the door.
And "efficiently" is doing quiet work in that sentence. Efficiency, in practice, means CUDA-optimized. A fully open license plus a proprietary inference stack equals a soft lock that never appears in the license text. Speed is a feature, not a bug, until it breaks โ and this one breaks toward the incumbent.
NVIDIA can afford total openness because its moat sits a layer down, in silicon. That's the strongest position in the stack and the hardest to attack. It's also the one most likely to draw a self-preferencing question from regulators โ a company that owns the model, the inference stack, and the chip is one the antitrust crowd will eventually notice. They always do. Just late.
The Meta split tells the same story from the other direction. Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index. Meta's closed Wonder scores 56. Seventeen points, roughly a 30% relative gap. Meta calls this a "structural moat." Read it plainly: the open model is good enough to capture developers, and deliberately not good enough to threaten the paid tier. Open source is no longer where a lab puts its best work. It's where a lab puts its best distribution. The capability gap is the product segmentation.
One more truth the marketing buries: the weights are free, but the inference isn't. A fully open Nemotron Bolt still needs compute to run, and at long context and high batch, that bill can dwarf a closed API. The license removes the fee and hides the cost. That's not a discount. That's a migration.
This is the part the crypto-native crowd keeps missing. We spent a decade arguing that open beats closed on principle. These labs agreed โ and then engineered the openness so that it feeds the closed tier. Free weights, paid capability. Free tokens, paid liquidity. Free testnet, paid mainnet. The metaphor isn't strained โ it's the same curve, drawn by different hands.
The crypto parallel isn't decorative; it's structural. Revenue gating is vesting with a different name. Unrestricted licensing is a fair launch. Hardware alignment is a chain subsidizing its own validators. The AI labs reconstructed, in eighteen months, the exact playbook DeFi spent five years refining. Curation is the new consensus mechanism โ and the license is where the curators write the rules.
Here's where I'll push back on my own instinct.
The reflexive take is that this is betrayal โ that "open" was supposed to mean free, and these labs broke the promise. I don't buy it. The promise was never real. Every open-source project in crypto history has drawn a line somewhere: the token that gates governance, the fee that funds the treasury, the multisig that can pause the contract. The protocol is neutral; the user is the variable โ and so is the license. A license is just a protocol for human coordination, and every protocol has parameters.
The real contrarian read is colder. The $20M threshold, the OEM clause, the hardware alignment โ these aren't restrictions on openness. They're admissions that openness was never the business model. It was the go-to-market. And the fact that four labs converged on the same number without coordinating tells you the market found the equilibrium on its own. No cartel needed. Just four teams reading the same demand curve.
What actually worries me isn't the gating. It's the licensing contagion. If "license follows derivative works" becomes standard, then a model fine-tuned on Nemotron Bolt inherits its terms. Distill it, and the terms follow the distillate. That's GPL-style propagation arriving in machine learning โ and it could freeze the fine-tuning economy before it matures. We watched this movie with forks in crypto. The license war is downstream of the code war, and it always arrives late.
Yields are transient; infrastructure is permanent. The weights are the yield. The license is the infrastructure. And the infrastructure is being rebuilt, right now, by people who learned from crypto's mistakes โ and crypto's playbook.
So the question isn't whether open source is dying. It's whether the next generation of builders reads the clause numbers before they ship, or after. In 2017 I caught an integer overflow because I read the code and not the whitepaper. In 2026, the exploit isn't in the contract. It's in the license. Read it first.