The deployer inherits the bill

DEEP DIVE · iPROMPT #147 COMPANION

The deployer inherits the bill

In one week the price of running a capable model collapsed and the European Union switched on the penalty regime. The two are connected, and the connection runs one way: cheap inference pushes deployment outward, and responsibility travels with deployment — from a handful of well-lawyered labs towards thousands of companies with no AI governance function at all. What actually came into force on 2 August, who it primarily addresses, and the five-line registry that makes the rest tractable.

R. LAURITSEN · 5 AUGUST 2026 · 9 MIN READ

Last week this newsletter told several thousand people to download a model and keep it somewhere they control.

I still think that’s right. It is also, as of Sunday, a decision with a legal consequence attached — and the advice went out four days before the consequence commenced.

Between 27 July and 2 August, the cost of capable inference fell by roughly an order of magnitude. OpenAI cut GPT-5.6 Luna by 80%, to $0.20 per million input tokens. DeepSeek took V4 Flash 0731 to general availability at $0.14 in and $0.28 out, scoring 82.7 on Terminal-Bench 2.1 — better than its own larger flagship at agentic work. Moonshot’s 2.8-trillion-parameter Kimi K3 weights had gone live the Monday before, under a licence permissive enough for commercial use. Then on 2 August the European Union turned on Article 50, its enforcement powers over general-purpose AI, and the full penalty regime.

Most coverage treated those as four stories. They read better as one, and the connective tissue is who ends up holding the risk.

What actually commenced on 2 August

The consensus take on Sunday was that it had been defused. The Digital Omnibus, in force since 27 July, pushed full compliance for standalone Annex III high-risk systems from August 2026 out to December 2027, and Annex I embedded systems to August 2028. A great deal of relieved commentary followed.

That reading confuses the high-risk regime with the Act. What was deferred is the part requiring conformity assessments, quality management systems and notified bodies. What arrived is the part requiring you to tell people what you are doing — and the part with percentages of global turnover attached.

Why cheap inference moves the risk downhill

When a frontier model cost $15 per million output tokens and could only be reached through an API, the lab was your compliance partner whether you wanted one or not. Its acceptable use policy constrained you. Its filters caught things. Its terms allocated risk between you. You were renting a legal posture along with the compute, and most buyers never itemised it because it never appeared on the invoice.

When the weights are free and running on your own hardware, none of that arrives with them. You keep the capability and a much larger share of the obligation — and, more to the point, you keep it alone.

Simplified for orientation. Whether a given deployment attracts provider or deployer duties turns on facts specific to it.

That last row is the one people miss. Fine-tune an open-weight model and place it on the market under your own name and you may no longer be merely deploying someone else’s system. The Act distinguishes providers from deployers, and substantial modification is one of the routes across that line — where exactly the line falls for a given fine-tune is a question for counsel, not a newsletter, but the direction of travel is not in doubt.

The signature and the price cut, in the same week

The record: on 28 July, more than a thousand employees across OpenAI, Anthropic, Google DeepMind, Meta and Mistral signed a statement called Pacing the Frontier, asking the United States to build mechanisms that would make a coordinated slowdown of frontier development possible. Signatories included Dario Amodei, Jakub Pachocki, Shengjia Zhao and Anca Dragan. Anthropic and OpenAI endorsed it corporately within a day. Over the following seventy-two hours, those organisations and their competitors cut prices and expanded distribution.

My reading, which is inference and not reporting: the easy interpretation is hypocrisy, and I don’t think that’s it. Note what the statement asks for — the option to pace later, not a pause now. An industry that expects constraints to arrive has a standing incentive to build the installed base first, because adoption creates dependency and dependency creates constituency. A model with a hundred thousand production deployments is harder to pace than one with a hundred.

I can’t show you that any particular price cut was made for that reason, and I’m not claiming it. The incentive is structural and it doesn’t require anyone to have acted on it deliberately. What matters operationally is the effect, which holds either way: the installed base is growing quickly, and the compliance burden grows with it. The installed base is you.

Predictions, with timeframes

Everything above this line is the record. Everything below is forecast — dated, falsifiable, and mine. Mark it and check.

What operators do about it

Three moves, all available this week, none requiring a committee. Treat them as a floor rather than a compliance opinion — they cover the obligations most likely to catch a company that added a chat widget and forgot about it, and they are worth doing on their own merits regardless of what a lawyer later tells you about scope.

1. Say it first, in the first message. Every AI-facing surface states, unprompted and before the interaction begins, that the user is dealing with an AI system. Not in the footer. Not in the terms. In the opening line, where a person reads it. This is the cheapest obligation in the entire Act and the one most likely to be missed by a company that added a chat widget eighteen months ago and forgot.

2. Mark synthetic output at generation time. Generated images, audio, video and substantive text carry machine-readable provenance. C2PA is the practical standard. Do it at generation, because retrofitting provenance is impossible by definition — once the file exists without it, the claim is unverifiable.

3. Keep a five-line model registry. Per deployed system: which model, which version or weights hash, live from when to when, who approved the change, and where the human oversight record lives. Five lines. It is the difference between “we don’t know what was running” and an answer. Storage costs a fraction of a penny per record; a missing log costs the benefit of the doubt.

None of this needs a conformity assessment or a notified body. It needs a decision, an afternoon and the discipline to keep doing it in November. What it does not do is settle your position — three controls are a starting point, not the whole of anyone’s obligations, and if your deployment touches anything in Annex III the December 2027 work starts now rather than then.

Where I could be wrong

The obvious objection to the Q2 2027 call is enforcement capacity. National market surveillance authorities are still being designated in several member states, and the first year of any regime tends to produce guidance rather than fines. A regulator looking to make a point may also prefer a large, well-known model developer to an anonymous mid-market deployer — better headline, better deterrence, better-resourced opponent. If the first action lands on a lab, I lose the bet and the lesson is that enforcement follows attention rather than exposure.

But exposure is where the volume is. There are perhaps a dozen organisations that could be described as frontier model developers and rather more than a dozen thousand European companies that put a generative system in front of a customer this year. Most of them did it with a vendor’s name on the contract. An increasing number, from this week onward, will do it with weights they downloaded because they were free.

The mental model worth discarding is “AI capability is expensive, so we will be careful with it.” That was true and has stopped being true. Capability is nearly free. The scarce, expensive, differentiating asset is defensible accountability: knowing what you deployed, being able to prove what it did, and having told people it was AI before they thought to ask.

The organisations that get hurt over the next eighteen months won’t be the ones that adopted AI too slowly. They will be the ones that adopted it in the week it became free, and read the absence of an invoice as the absence of an obligation.

R. Lauritsen

EDITOR · iPROMPT

iPrompt

PUBLISHED BY FRONTWAVE MEDIA LTD · LIMASSOL, CYPRUS

iprompt.com