The Tax Code Just Told You to Move AI Spend Where You Cannot See It

Gavin Newsom signed SB 122 on June 29, 2026. From January 1, 2027, California treats prewritten software, SaaS and seat-based AI application subscriptions as taxable tangible property. The rate is 7.25% at the state level, plus local district taxes, so the total runs from 7.25% to about 10.75% depending on the buyer’s address, and lands at 8% to 10% for most buyers in most real metros. Jason Lemkin’s reading of the bill puts the state’s own revenue projection at roughly $2B a year in combined state and local collections.

The exemption list matters more than the rate.

IaaS is exempt. PaaS is exempt. Custom software written for a single customer is exempt. Services that are mostly human effort, delivered electronically, are exempt. Taxable is the layer where you buy a named seat for a named person and the vendor can tell you who used it.

A procurement rule just became an observability rule, and it points the wrong way.

The exempt side is the unattributable side

Walk the two buckets from the perspective of somebody who has to answer “which team burned this budget.”

A seat-based AI subscription is the easiest AI spend to govern that exists. It has a per-user line. It has a login. When usage climbs in a quarter, you can name the people behind it and go ask them what changed. Attribution is a byproduct of the pricing model, and it comes free with the invoice.

Platform and infrastructure spend has none of that by default. It arrives as a compute bill against a project, an account, or a tag somebody set up long ago and never audited. Whether you can attribute it to a person, a team, or a workload depends entirely on tagging discipline your organization either built or did not. Nobody hands it to you.

SB 122 puts an 8% to 10% price on the first arrangement and zero on the second. It got there through a definitional line drawn between software and infrastructure decades before agents existed, by people whose subject was tangible property.

Nobody has to plan the migration for it to happen

A CFO deciding to trade visibility for savings would at least be making a visible decision. The mechanism at work here skips that step.

A team hits renewal on a seat-based AI tool at an effective price 8% to 10% higher. Somebody points out that the same model is available through a cloud provider’s platform offering at consumption pricing, with no per-seat tax exposure. The math favors the move on the spreadsheet in front of them. The spreadsheet does not have a column for “we lose the ability to say who ran this.”

Lemkin’s illustrative figure is a company spending $2M a year on tools facing $160K to $200K of new expense that buys nothing. That is an arithmetic example, not a survey result, but it is the right order of magnitude to change a renewal conversation. At that scale the restructuring pays for a full-time engineer to build the plumbing, and the plumbing that gets built is the plumbing that moves the workload, not the plumbing that preserves the audit trail.

We have argued that cost per completed task is the CFO metric and that variance, not a spending cap, is the control worth having. Both of those depend on knowing which unit of work produced which charge. A tax rule that rewards moving consumption to a layer where that mapping does not exist by default attacks the input to every cost-governance practice we have written about.

Consumption-based AI sits between the buckets

The reading gets genuinely uncertain at this point, and I would rather say so than paper over it.

The taxable definition Lemkin describes covers seat-based AI application subscriptions. The exempt definition covers IaaS and PaaS. AI spend that is metered by token, request or task, billed to an account rather than a person, is described by neither. It is not a seat. It is not obviously infrastructure either, since what you are buying is a model’s output.

I have not seen that classification resolved, and this is a commentary blog rather than the statute or a tax authority ruling. Treat the ambiguity as a live risk in both directions: budget for the possibility that consumption-priced AI gets pulled into the taxable bucket, and do not build a 2027 procurement plan on the assumption that it stays out.

What is safe to say is that uncertainty by itself creates pressure. When one arrangement is definitely taxed and another is arguably not, the arguable one wins the renewal.

The $5M threshold moves who does the accounting

Two more mechanics matter for anyone drawing up next year’s contracts.

Nexus did not change shape. A vendor collects if it has physical presence in California or $500,000 in California sales as a remote seller. SB 122 created no new threshold; it made software sales count toward the one that already existed. Vendors who were previously below the line because their revenue was “not tangible property” may now cross it on the same revenue.

Above the line, the obligation flips. When one vendor’s digital product sales to a single purchaser exceed $5M, the vendor is relieved of the collection duty and the purchaser self-assesses and remits use tax directly. Lemkin’s article states that threshold as “in a calendar year” in one place and as measured over twelve months in another, without resolving which controls. That distinction decides whether a January-to-December total or a rolling window triggers the flip, so confirm it with counsel rather than reading it off a blog post, including this one.

The governance consequence is the part nobody puts in the tax memo. Self-assessment means your finance team now has to itemize digital purchases by category, by jurisdiction and by vendor, for the largest AI relationships you have. That is an inventory obligation, and it is worth checking whether your current purchase record would survive an audit before the obligation arrives.

Do this now: run the classification pass before renewal season

Get finance and whoever owns your AI tooling budget in one room. Produce one table with a row per AI vendor and four columns.

Column one: the bucket. Seat-based subscription, platform or infrastructure, consumption-metered, or human-effort service. Mark the consumption rows as unresolved rather than guessing them into the exempt bucket.

Column two: attribution today. For each vendor, can you name the person or team behind last month’s charge? Answer for this afternoon, using the data you already hold. Anything you would first have to build belongs in column three.

Column three: attribution after a plausible move. If that vendor’s spend relocated to a platform arrangement to sit on the exempt side, what happens to column two. If the answer is that you would lose the per-user line, write down what would replace it and what it costs to build.

Column four: the $5M question. Which vendors put you within reach of self-assessment, and does your current purchase record support remitting use tax on them.

The output is a short list of relationships where the tax code is about to pay you 8% to 10% to stop being able to answer who did what. For each of those, decide deliberately: absorb the tax, or move and fund the attribution layer that the move destroys. Oversight is already a cost line in these systems, and the same discipline applies to an agent approving its own budget. A saving that arrives by deleting your own evidence trail is a loan against the next incident.

California now joins more than twenty states taxing SaaS in some form, and Colorado’s equivalent expansion takes effect the same day. The definitional line between software and infrastructure will be redrawn in more places than one, by people who were not thinking about AI governance when they drew it. Your job is to notice which side of it your visibility lives on before the renewal calendar decides for you.

---

This analysis synthesizes California Is Taxing SaaS and AI Tools: What SB 122 Does to Buyers and Vendors (SaaStr, Jason Lemkin, 2026), a commentary reading of SB 122 rather than the statute itself.

Victorino Group helps engineering and finance teams keep per-workload attribution intact when AI spend moves down the stack. Let’s talk.