The 2025 Regs Explained: How Treasury Classifies AI Transactions

In January 2025, Treasury and the IRS finalized two sets of regulations that now define how AI transactions are classified for U.S. federal income tax purposes. If you're building or operating AI agents that buy and sell services, these rules determine whether your payments are treated as services, royalties, rents, or sales — and that classification drives everything from sourcing to withholding.

The two regulation sets:

• Reg. section 1.861-18 — Digital content transactions (updated from the legacy software regulations)

• Reg. section 1.861-19 — Cloud transactions (entirely new)

Together, they're referred to as the 2025 regulations. They replace guidance that was originally written for shrink-wrapped software in 1998 and extend it to cover the cloud-native, API-driven world we actually live in.

What Counts as "Digital Content"?

The digital content regulations define digital content as a computer program or other content in digital format that is either:

• Protected by copyright law, or

• Not protected solely because the copyright expired or the creator dedicated it to the public domain

A "computer program" uses the exact same definition as U.S. copyright law: "a set of statements or instructions to be used directly or indirectly in a computer in order to bring about a certain result."

This definition is narrower than it might seem. It covers the source code and compiled binaries of an AI model — those are clearly copyrightable computer programs. But it may not cover the model's weights and biases, which are mathematical parameters generated by the training process. If weights and biases aren't copyrightable (and most commentators believe they aren't), they fall outside the digital content regulations entirely.

The regulations do have a carve-out: noncopyrightable elements like databases or documentation can be treated as part of a computer program if they're "incidental to the operation of the computer program." But a model's weights and biases are anything but incidental — they are the model's intelligence. Without them, the program is an empty shell.

The Four Categories of Digital Content Transactions

When a transaction does qualify as a digital content transaction, it falls into one of four buckets:

• Transfer of a copyright right — The buyer gets rights to reproduce, distribute, publicly perform, or create derivative works. This is the royalty-generating bucket.

• Transfer of a copy (copyrighted article) — The buyer gets a copy of the software but not the underlying copyright. Could be a sale or lease depending on benefits and burdens of ownership.

• Provision of services — Someone is hired to develop or modify digital content.

• Provision of know-how — Transfer of information about how to develop digital content.

Most B2B AI transactions don't involve transfers of copyright rights. When your agent calls an inference API, you're not getting the right to reproduce GPT-5. You're getting an output.

Cloud Transactions: The New Category

This is the big addition in 2025. A cloud transaction is defined as:

"A transaction through which a person obtains on-demand network access to computer hardware, digital content, or other similar resources."

The key phrase is "on-demand network access." When a customer uses a foundation model through an API — sending prompts and receiving outputs — they're accessing the model over a network, not receiving a copy. The model stays on the hyperscaler's servers. The customer gets compute results.

Cloud transactions are classified as services. Full stop. Under reg. section 1.861-19(c)(1), any cloud transaction — whether it involves digital content, hardware, or "other similar resources" — is a service.

This is the regulation that governs most AI agent interactions today. When your agent calls Claude, GPT, or Gemini through an API, that's a cloud transaction. It's a service. The income is sourced to where the provider's assets and personnel are located.

Why "Other Similar Resources" Matters

The cloud transaction definition doesn't limit itself to digital content. It includes "other similar resources." This phrase was likely included precisely because of situations like foundation models, where parts of what the customer accesses (the weights and biases) may not qualify as digital content.

For AI builders, this is actually good news. Even if a court decides that weights and biases aren't digital content and aren't copyrightable, cloud transaction classification still works. The customer is accessing "other similar resources" over a network. It's still a service.

The Predominant Character Rule

Real transactions are messy. An AI platform might bundle inference (cloud transaction), a downloadable SDK (digital content transfer), and consulting on fine-tuning (service). Which classification wins?

The 2025 regulations use a predominant character test. A transaction with multiple elements is classified based on its predominant character — determined by a three-prong hierarchy:

• Primary benefit to the specific customer — What is this particular customer primarily paying for?

• Primary benefit to a typical customer — If that's not ascertainable, look at usage data for typical customers in similar circumstances.

• Other factors — Marketing positioning, relative development costs, relative pricing of individual elements in uncontrolled transactions.

The entire transaction is then classified under whichever bucket the predominant element falls into. This is an all-or-nothing approach — unlike the OECD model, which generally bifurcates mixed contracts element by element.

For most AI agent transactions, the predominant character is straightforward: you're paying for inference results (a cloud transaction, i.e., services). The fact that you also downloaded an SDK or received API documentation doesn't change the overall classification.

How This Applies to Common AI Transactions

Your agent calls an inference API (OpenAI, Anthropic, Google)
→ Cloud transaction → Services income to the provider. Sourced to where the provider's assets and personnel are located.

Your agent downloads an open-source model and runs it locally
→ Transfer of a copy of digital content (the program) → Copyrighted article. Whether it's a sale or lease depends on the license terms and whether benefits and burdens of ownership transferred.

You pay for fine-tuning through a hyperscaler's platform
→ Likely a cloud transaction if you don't own the resulting weights file. You're accessing the model over a network, with fine-tuning as a feature. Still services.

A model owner licenses its model to a hyperscaler
→ This is where it gets complex. Could be rent (if the program is the predominant element) or royalties (if the weights/biases, treated as trade secrets, are predominant). The answer depends on the specific deal terms.

Your agent buys GPU compute from another provider
→ Cloud transaction → Services. On-demand network access to computer hardware is explicitly covered.

What the 2025 Regs Don't Cover

These regulations address federal income tax characterization and sourcing. They don't address:

• State sales and use tax — States have their own definitions of taxable digital services, and they vary wildly. Texas taxes 80% of data processing. Oregon taxes nothing. New Jersey exempts some digital services but taxes others.

• International VAT/GST — Other countries have their own classification frameworks, though the OECD model treaty provides some harmonization.

• The rate — The 2025 regs tell you what kind of income it is, not how much tax is owed.

This is exactly where the AgentTax engine comes in. We handle the state-by-state taxability matrix, the rate lookups, and the nexus analysis that the federal regulations leave to others.

Why This Matters for Show HN

If you're building autonomous agents that transact — buying compute, selling API access, licensing models — these regulations define the tax character of every one of those transactions. The fact that Treasury addressed cloud transactions explicitly in 2025 means the IRS considers these transactions well within the existing framework.

You can't argue the rules are unclear. You can't claim there's no guidance. The 2025 regs are here, they're final, and they apply to your agents right now.

The 2025 regulations were published as T.D. 10024 (digital content) and T.D. 10025 (cloud transactions) in January 2025. The full text is available at federalregister.gov. For analysis of how these regulations apply to specific GenAI business models, see KPMG's "No Need to Reboot: GenAI Fits the Tax Stack" (Tax Analysts, DOC 2026-5375, March 19, 2026).

AgentTax classifies AI agent transactions across 51 jurisdictions. Try the API →