Would you hire staff without knowing what they cost you each month?

Stripe buying OpenRouter is the most important fact of the week in this dilemma. Not for the size of the deal, but for what it reveals: the layer between customer and model has started to capture value. If a company can route each task to the most suitable model, it gains control over cost, speed, quality and its contingency plan for when the main supplier fails, slows down or rations usage.

That is the practical answer to the portability question. A single interface between application and suppliers lets you send simple tasks to cheap models, use better models when complexity demands it, and compare real cost per task. It is what any operations director does with a critical input: qualify more than one supplier, measure both, and keep the second one warm.

OpenAI's price cut reinforces the reading. For anyone buying today, it is good news. For anyone designing a critical process, it is a warning. An operation built on promotional pricing needs to know what happens when the promotion ends. The token is the unit of consumption for these models. The more input text, context, documents, retries and long answers, the bigger the bill tends to be.

Thomson Reuters completes the triangle. The point is not that every company should train a model. The more useful point is this: where there is proprietary data, a clear domain and a repetitive task, using an open-weights model as the base can be one more answer to portability. The American legal information company reduced its dependence on outside labs by using a Chinese base, in the same week that AI geopolitics was again demanding that companies pick a side. The irony matters less than the lesson: cost control, governance and product roadmap can also be designed into the architecture.

NVIDIA's result is useful because it separates model from system. AVO used a powerful model, but the difference came from the engineering around it: memory that carries across attempts, a supervisor that notices when the agent gets stuck, and a cycle of inspect, plan, execute and evaluate. Twelve per cent fewer actions means twelve per cent less bill. The number that matters is not price per token. It is cost per completed task.

The Stampli case shows the same thing from another angle. When the task is bounded, the context is available and human review is clearly assigned, the agent absorbs execution. When those three elements are missing, AI tends to become an extra layer of attempts, rework and coordination.

CoSnitch adds the part the productivity spreadsheet does not see. Patching software is not the same as remedying a process. If an assistant has stored bad instructions, persistent memories or dangerous permissions, someone has to audit what was left behind. Agents do not merely execute. They remember, reach for tools, carry context, and can turn one click into a permanent state.

McKinsey enters here as context, not as a fact of the week. In July, QuantumBlack reported that 93% of respondents to a May survey had exceeded their AI budgets. That is why the question in the title is not rhetorical: companies are already hiring digital labour without being able to forecast the bill.

Unitree shows the first stage. The debut priced a demonstration: up 460% on day one, on top of a prospectus that already recorded a 53% fall in adjusted net profit for the quarter. Three days later, once the market began measuring the company against fundamentals rather than narrative, about 45% of the value had gone. The technology did not change in that interval. The criterion did.

The Chinese robots over 100 metres show the stage that does not advance. Peak performance produces headlines and does not produce contracts. Between running fast once and operating every day, near people, assets and processes, there is a demand the demonstration never makes: repeatability measured over time.

Waymo is the only one of the three to have crossed all three stages. With the scale came unions, state legislators, federal agencies and city councils debating fleet caps, per-mile fees and rules of operation. And because it is measured, what the demonstration concealed comes into view: weekly rides have been stuck at 500,000 since late March, while the fleet grew and the city count went from ten to fifteen. Each vehicle now makes about 125 trips a week, against 167 in May 2025. The company is buying map coverage, not ride volume. The point that matters to anyone running a business is the order of events. Waymo did not start being measured continuously because the engineering matured. It started being measured because it was contested. Contestation created the obligation to produce a number, and the number is what makes it possible to finance, insure and depreciate the asset.

This is where the corporate agent parts company with the other three. It did not end up without a price, without depreciation and without residual value because it is intangible. It ended up that way because it was never contested. It came in through the IT door, with no union, no neighbour, no regulator, no public hearing, and so nobody was ever obliged to produce a number about it after the pilot. It went the same way with the automatic lift, with the cash machine and now with the robotaxi: continuous measurement arrived alongside whoever complained. The difference, with the agent, is that whoever complains first is likely to be in-house, and to be called the CFO.

Start with what the number is not. It is not an investment, not capex and not revenue. It is a residual value guarantee. If defined default events occur, NVIDIA covers the difference between the guaranteed minimum value and whatever the owner of the property recovers by re-letting or selling. The relevant detail is that the structure is tied to OpenAI's 20-year leases and takes effect in phases, as the data centres are brought into service between 2028 and 2030.

That changes how AI capex reads. The chip supplier is helping to make viable the infrastructure that will host its own compute exclusively, while the tenant takes on part of the economic risk of the structure. The promotional pricing on offer to developers and companies today is underwritten by commitments that mature later.

The demand exists. Anthropic taking its run rate into the tens of billions shows the market is not merely narrative. The risk has stopped being “is there demand or not?”. It has become who carries the gap between infrastructure contracted now and monetisation consolidated later.

The contradiction of the week is simple. The price of GPT-5.6 Sol fell for three months, but the quantity available was capped for some users. The two facts only look contradictory to anyone who believes list price is real price. It is not. When the supplier lowers the tariff and rations the quantity, the scarce product is allocation.

The expected increase in NVIDIA servers reinforces the point. Even when the main chip advances, other components start to capture the cost. Memory, the server assembly, power, cooling, interconnect and availability all enter the bill. The bottleneck does not disappear. It moves.

For companies, rationing is not only a cost. It is a continuity risk. A critical process running on a single model supplier has acquired a new failure mode: it is not the system going down, it is the quota running out. No operations director would accept a production input with a sole supplier and an invisible quota. In AI, a great many companies are accepting exactly that without noticing.

Control has stopped being an abstract discussion. It has become a cost line, a schedule and a governance question. OpenAI did not merely say it would monitor better. It set a deadline: an alert within 30 minutes and a mandatory pause if the risk is not cleared. That is more concrete than “use AI responsibly”.

For companies, the translation is direct. An agent that executes code, reaches into systems, queries sensitive data or triggers third parties needs a stopping criterion. Having a policy is not enough. You have to define the event, the person responsible, the reaction time and the consequence.

The divergence between OpenAI and Anthropic on data retention exposes another hidden cost. Real portability is not only technical. There are models your architecture can call, and models your data policy permits you to use. If those two lists do not exist in writing, with an owner, the company does not have portability. It has an intention.