Two openness headlines, pointing opposite ways
Two stories about openness ran through 2026, and they contradict each other on the surface. The first: the frontier models are closing. In April, Meta released Muse Spark as "proprietary only," its first model since Meta Superintelligence Labs was formed, available inside the Meta AI app and through a private API preview. The company that defined open weights with Llama now ships its flagship behind a door. The second story: the applications are opening, whether their makers like it or not. REA, an MIT-licensed reverse-engineering tool, lets a coding agent read a compiled binary and "build a version for your project."
Read them together and the contradiction disappears. In both cases the value moves to the same place, the firm that deploys the system. A closed frontier model sells capability by the token to deployers who keep the margin. An open clone hands a deployer the same capability for free, and the deployer keeps the margin again. The address of the margin never changes. Only the invoice does. Every AI tax proposal now in circulation keys on the invoice. That is the problem this piece is about.
The models are closing, and the numbers show it
Meta is the loudest example, not the only one. Menlo Ventures surveyed 495 U.S. enterprise AI decision-makers in late 2025 and found that Llama's stagnation "contributed to a decline in overall enterprise open-source share from 19% last year to 11% today." Enterprises, the report says, "remain cautious, preferring closed-source models." Chinese open-source models, for all their benchmark progress, accounted for 1% of LLM API usage in the same survey.
The capability gap tells the same story. A United Nations University summary of the Stanford AI Index 2026 notes that the gap between the top closed and top open models reversed direction, "widening back to 3.3% in 2025 after nearly closing in 2024." Those are secondary figures and small percentages. The direction is what matters. For two years the policy conversation assumed open weights would commoditize the frontier and keep the pricing power of labs in check. In 2026 the labs are keeping their best work, and enterprises are paying for it.
Meta's own claim about Muse Spark should worry anyone designing a tax on compute. The company says the model reaches its reasoning capabilities with roughly ten times less compute than Llama 4 Maverick. If that holds, a tax on inference compute would collect a tenth as much for the same output, from one model generation to the next. Capability goes up. The base goes down. That is not a base. It is a countdown.
The applications are opening, whether vendors agree or not
While models closed, the other half of the stack cracked open. REA connects coding agents to Hopper, Ghidra and IDA, the disassemblers professionals use, and lets them inspect native binaries, Electron apps and websites without source code. The repository was created in April and counted 149,178 GitHub stars on October 12. Around it came the clones. KDnuggets reports agent-built versions of Office and Adobe apps; WordCraft's developer estimates about 62% real-world feature parity with Word, and DeckCraft reports about 55% readiness as a replacement for PowerPoint.
The clones are rough. Testers found bugs and missing features, and the KDnuggets author is careful: "REA cannot automatically produce a perfect clone." But a software license is not defended by its last 40% of features. It is defended by the cost of rebuilding the first 60%, and that cost just collapsed. A proprietary application is now a proprietary binary with an open behavior. Anyone with an agent can read what it does and ask for the same.
So the stack inverts. The intelligence layer, once expected to be open, is closing. The application layer, once safely closed, is opening. A firm that buys closed intelligence and uses it to clone open applications keeps the margin from both. That firm is the taxpayer of the agentic age, and it is the one firm most AI tax designs do not name.
Why a tax on AI companies misses open weights entirely
Take the AI Tax and Work Protection Act, introduced in August by Representatives Casar, Foushee and Jacobs. It taxes large AI companies on the higher of token value sold or AI product revenue. Now put an open-weight model inside a bank. The bank downloads the weights, runs them on its own servers, and replaces a floor of back-office staff. No tokens are sold. No AI product revenue appears. The lab that trained the model gave it away. The bill collects nothing from the firm where the payroll left.
The same gap opens from the other side. When a firm clones its design software with an agent, the vendor's revenue falls, so any tax on the vendor's AI revenue falls with it. The clone itself is MIT-licensed and free. The margin, now sitting with the cloning firm, is invisible to a base that looks for AI sales. A tax built on the invoice vanishes exactly when the invoice does.
This is not an argument against the Casar bill's best feature, the unemployment trigger. It is an argument about where the base sits. A base on the seller moves with the vendor's business model, which is now changing every quarter. A base on the deployer moves with the work, which is the thing the tax is supposed to follow.
The margin lands on the deployer every time
Follow the money through all four routes. Buy a closed model by the token, and the margin from the work it does stays with the buyer, minus the token bill. Download open weights, and the whole margin stays with the deployer. Buy a SaaS license, and the vendor keeps a slice while the deployer keeps the rest. Clone the SaaS, and the deployer keeps everything. Four invoices, four different tax outcomes under a seller-based tax. One destination.
Tax The Agent proposes an agent contribution, a campaign idea and not law anywhere: a rate on the revenue a deploying firm's autonomous systems materially create or protect. It does not ask whether the model was open. It does not ask whether the app was bought or built. It asks what moved off payroll inside this firm and what revenue those systems now carry. The open-source developer who publishes a clone keeps no margin and owes nothing. The lab that sells closed tokens pays its ordinary corporate tax on that revenue, as it does today. The deployer pays on what it keeps.
Weights do not pay tax. Firms do. Any base that needs to know the license terms of the model before it can calculate what is owed will be rewritten by the next licensing decision at the next lab.
Three questions that sort every proposal
Here is the design test this year's openness shift makes necessary. First: does the proposal collect the same amount whether the model was open or closed? If an on-premise open-weight deployment pays nothing while a token purchase pays, the proposal is a tax on a vendor relationship, not on displacement. Second: does it collect the same whether the application was bought or cloned? If a cancelled license shrinks the base, the proposal rewards exactly the behavior that removes payroll fastest. Third: does it collect the same whether the vendor sits in Delaware or abroad? If not, the base migrates on the next contract.
Run the proposals through those gates. A tax on tokens fails the first and the third. A tax on compute fails the first and, after Muse Spark, erodes under its own feet. A tax on AI product revenue fails all three, because the clone wave removes the revenue. A royalty on attributable output at the deploying firm passes all three. That is not because it is clever. It is because it stands where the margin stops moving.
Audit your own stack for the openness gap
One task for this week. Take one organization you know and list its autonomous systems by how it got them: closed model by the token, open weights on its own servers, SaaS license, or agent-built clone. For each, estimate the revenue those systems create or protect, and note which ones would pay anything under a token tax, a compute tax or a vendor revenue tax. The rows that pay nothing under every proposal are the rows that are growing fastest.
Then run the total through the illustrative calculator below and set a rate you would defend in public. The calculator never asks whether your model was open. That is the point.
