This is part of WeGoAI's daily series on the things that actually shape our world. Figures from Western Digital and IDC, McKinsey, TechCrunch, Ars Technica, MIT Technology Review, Axios reporting and the Origina Software Tax report. All news dated 22 October 2026. Not investment advice.
“This is part of WeGoAI's daily series on the things that actually shape our world.”
One Day, Five Bills
Thursday 22 October 2026 handed us five AI stories. Only one of them was about tokens, and it was the least interesting one.
Uber told TechCrunch it is putting its entire global workforce - some 30,000 people - on ChatGPT Work, scaling from a 1,000-person pilot run earlier in the year. Uber did not disclose what it is spending. Password manager 1Password launched a product that lets an agent log into your accounts on your behalf without ever seeing your password, the same day Axios reported that US banks have started detecting agentic activity inside customer accounts and cannot tell whether the agent was asked to act or acted on its own. Security researchers at Claroc enumerated 6,800 Meta Muse conversation-share links with a single query and opened a few hundred of them, finding a candidate's confidential executive-search notes, a small-business loan application, a bank fraud alert with a partial account number, a real-estate investor's mortgage figures and a workplace's full staff vacation calendar. Meta suspended the sharing feature.
Then there was the cost nobody was talking about. MIT Technology Review published an analysis from Western Digital measuring what an inference request leaves behind: an 8-second video clip, delivered to the user as a 3.97MB 720p file, left more than 500MB of associated data on the back end - logs, cached context, immediate renders, backups and safety records. That is a 130x amplification of what the customer actually received, and none of it disappears when the job is done.
The theme is not "AI is expensive". It is more specific and more uncomfortable: the cost of AI is landing in budget lines that nobody modelled, and it is compounding.
The Part of the Bill Nobody Priced
Compute is the line item every AI budget has. Storage is the rounding error. That is precisely the wrong way round, and Western Digital's own measurements are the proof.
An 8-second clip is not an unusual unit of AI output. Product videos, catalogue images, ad creative, training data for the next model - every one of them writes a small artefact the user sees and a much larger residue the user never does. The delivered file is a rounding error next to the cached context, the safety records, the intermediate renders and the backups.
And the residue does not retire. A training run is a project: it starts, it burns compute, it ends, and you can budget and retire it. Inference runs continuously, and every single request leaves something behind that has to be kept, and increasingly re-used. As MIT Technology Review puts it, a model that idles while data loads costs about as much as one idling for lack of a GPU.
The direction of travel makes it worse, not better. McKinsey projects AI inference demand growing at a 35% CAGR through 2030, against 22% for training, which puts inference at more than 40% of total data-centre demand by the end of the decade.
So the workload that grows fastest is the one that never ends, and it grows fastest in the domain where every unit of output writes residue. Training is the campaign. Inference is the occupation.
The survey data says the same thing from the buyer's side. A Western Digital-sponsored IDC white paper surveyed 763 IT and business decision makers.
Nearly 95% report their data volume rose because of generative AI over the past 12 months. Roughly 61% report growth of 25% or more. And 96% anticipate needing faster archive retrieval to support inference or retrieval-augmented generation - which is to say, the data that accumulated quietly is now expected to become an input to the next workload. Most organisations cannot answer a single question about this with confidence, including who set their retention policy for AI-generated content, and whether their storage procurement is tied to training-time performance needs or to the actual life cycle of the data they are producing.
A CFO's word for this is a rounding error that has become a line item. As the analysis warns, a team that budgets AI as "compute plus a rounding error for storage" may find, in 12 to 18 months, that the rounding error has become a significant fraction of the programme's total cost - right around the time leadership is deciding whether to expand.
The Credential Bill, and the Leak That Bill Created
Now the other side of the same trade. An agent that can act needs authority, and authority used to mean a password. The industry has spent the last year pretending that this is a solved problem, and one day of news proved it is not.
Claroc's finding is worth stating precisely, because the technique is the story. Meta Muse, which Zuckerberg described last month as "built from the ground up for privacy and security", uses unguessable URL tokens for shared conversations - exactly like every other service. But a central design decision in constructing those URLs made them discoverable. The researchers enumerated 6,800 of them without effort.
"It took one query," they wrote - and went on to say that if asked beforehand, they would have guessed gathering this would take real effort.
What came out of the few hundred they opened was not embarrassing chat logs. It was a candidate's notes preparing for a confidential executive-search call with a frontier AI lab, naming the lab's internal contact and the meeting time; a national marketing company's internal revenue review, marked for its CEO; a small-business owner's loan application with the amount requested; a to-do list containing a bank fraud alert with a partial account number, a transaction amount and a time; a real-estate investor's internal performance report with two property addresses and exact return, income and mortgage figures; and a workplace's full-year staff vacation calendar, listing dozens of employees against the week each would be away.
Meta suspended the sharing feature while it reviews. Note what this is: not a model failure, not a jailbreak, not an adversarial user. A URL-construction choice, and a couple of hundred unlucky users whose lives happen to contain numbers worth stealing.
The industry's answer, arriving on the same day, was 1Password for Agents: the agent sends a unique digital key, 1Password confirms the agent's identity, the user gets an alert and personally approves, and the password manager fills in the credentials itself. The agent never holds the secret. That is a genuinely good design, and it is also an admission of how broken the current state is. As 1Password puts it, handing over a password to an agent "can leave people with sensitive information in places they can no longer control as a copy gets saved for the next time".
Meanwhile the banks are already seeing the traffic. Per Axios, US banks have begun to detect agentic activity inside user accounts. The bankers interviewed described the situation as "horrifying" - not because agents are stealing things, but because most banks cannot tell whether the agentic activity is tied to user requests at all. A Chase spokesperson told Axios that customers "should not give their usernames and passwords to anyone". An agent economy in which the counterparty cannot distinguish a customer's instruction from an agent's improvisation is not an authentication problem. It is an attribution problem, and it is arriving before the standards for it.
Meanwhile the vendors are shipping guardrails on their own initiative. Anthropic's new Claude Duet - built on Claude Opus 5.5, wired into over 6,000 connectors and plugins in the Claude Marketplace, able to buy things through Stripe's Link wallet or Shopify's Shop Pay - ships with a new classifier that checks every message, and by default the agent declines requests for credit card numbers and government identification. It will not trade securities or move money between accounts. Those are sensible boundaries. They are also a confession that the boundaries cannot be inferred from the model, and have to be bolted on.
Why This Week Made It Worse
Here is the part that should worry a CFO, and it is not in any of the announcements.
On the same day, OpenAI made its Decisions API model generally available - the direct answer to TypeSafe's Jev, the model that made that company a $7.6B business. OpenAI's product lead described seeing Jev and thinking "we have like a million use cases for this". The implementation is almost offensively simple: they took ChatGPT Luna and reconfigured it for speed and cost. The result is a model that turns a decision into an answer in 100 to 180 milliseconds instead of one to three seconds.
At Speak, which builds AI language tutors, the Decisions API cut the time between actions from 1.4 seconds to 0.13 seconds. Ten times faster.
This is unambiguously good, and it is also the mechanism that turns every other story above from a curiosity into a bill. Token prices have been falling for a year - the open-weight index has been falling steadily at the bottom of the book - and the industry keeps concluding that cost is therefore under control. But the token price was never the cost. The token price is the cheapest visible line on an invoice whose other lines were never raised.
When a decision costs 130 milliseconds instead of 1.4 seconds, an agent can make hundreds of decisions in the time it used to take one. Each of them writes logs, cached context, safety records and backups. Each of them may touch a credential. Each of them is another action on a bank account that the bank cannot attribute. Latency is the multiplier on residue. Making agents faster does not make the system cheaper; it makes the unbudgeted part of the system bigger, faster, and harder to catch up with.
The Tax That Was Already There
Meanwhile, enterprises were already bleeding money on a bill of the same shape, and nobody had connected the two.
Origina commissioned a survey, run by Censuswide, of 2,018 CIOs and senior IT decision-makers at companies with at least 10,000 employees and at least $1B in annual revenue - 1,011 of them in the US, the rest in the UK.
On average, 28% of these organisations' software budgets go on unwanted upgrades, migrations and licensing changes. 95% have paid for a vendor-driven upgrade, migration or licensing change in the past three years that they believed at the time would reduce performance, reliability or business value. 53% have done it more than once. 76% say they have very little real leverage in negotiations. 80% say moving away from a major vendor would cause significant disruption. Every single respondent runs at least one of IBM, HCL, VMware/Broadcom or SAP at meaningful scale.
The money does not vanish; it is reallocated from the things that actually matter. 36% of respondents named innovation as the area that loses funding, then human resources at 32%, security at 31% and customer-facing investment at 31%. Liberty Mutual's CIO, James McGlennon, put the human cost plainly: an upgrade cycle can occupy top technical workers for "the better part of a year", and every year it repeats is a year those people are not working on the initiatives that differentiate the business.
And the demand for protection is political now: 91% would support laws giving enterprises "protections comparable with those available to long-term residential tenants". Only 1% opposed.
Read the two halves together and the picture closes. Every AI agent connects to more vendors. Every agent connection is a new dependency inside a landscape where 95% of large enterprises already cannot refuse a vendor-driven change and 80% say leaving would cause significant disruption. Claude Duet alone connects to over 6,000 connectors and plugins. The agent economy multiplies the very lock-in that 2,018 CIOs say is already eating their budgets.
Marshall Van Alstyne, co-author of the Software Tax report, gives the remedy as a phrase worth stealing: "negotiate the exit before the entry." Set the price of migration assistance and cap renewal and expansion prices at the moment you sign. That advice was written for enterprise software. It is the correct advice for agent connectors too, and almost nobody is taking it.
The Number Uber Did Not Disclose
Back to Uber, because the missing number is the interesting one.
Uber piloted ChatGPT Work with 1,000 employees earlier this year, across finance, operations and product. 80% of early adopters were non-engineers. They used it to analyse operational data, localise training material for different markets and languages, and build personalised dashboards. Uber then paired 30 AI engineers with teams across the company for two weeks to automate parts of their jobs - "AI pods", they called them - and found, in Uber's own words, how much opportunity there is when you combine technical expertise with a deep understanding of a particular job.
Then it rolled out to 30,000 people, and did not disclose the spend.
That is the shape of the whole problem. The pilot was 1,000 users and 30 AI engineers. The rollout is 30,000 users and - as far as the reporting goes - a number of pods that is not stated. Uber's own framing is instructive: "our approach is to make powerful AI tools available across Uber while giving teams room to experiment... we'll look at how employees are using the technology, the kinds of workflows they develop."
That is a genuine strategy. It is also what you do when you do not yet know the unit economics. The cost of a seat is not the token bill. It is the token bill plus the residue, plus the credentials, plus the vendor connections, plus the security review, plus the storage tiering, plus - the one nobody has measured yet - the human time spent in workflows that the agent quietly changed.
What To Ask This Budget Cycle
Western Digital's analyst puts three questions to any organisation's AI infrastructure budget, and they are the right three:
What is our data retention policy for AI-generated content, and who set it? If the answer is "nobody, yet", the residue is unmanaged and will be discovered by an invoice rather than by a decision.
Is our storage procurement tied to training-time performance needs, or to the actual life cycle of the data we are producing? Storage strategy is still designed for hot-tier performance, while most AI data by volume accumulates in the warm, cool and cold tiers behind it - and those tiers are where retrieval latency and cost both bite.
If our AI usage doubled next year, would our storage cost double with it, or worse? Doubling the compute is a forecast you can make. Doubling the residue is a different and worse curve, because retained data does not shrink when compute demand falls.
Add two more from the agent side, and the list is complete:
For every agent connector we add, what is the exit? Price the migration assistance and cap the renewal before we sign, not after the dependency is load-bearing.
For every credential an agent can touch, can the counterparty tell it apart from the human? If a bank cannot, then neither can our fraud team, our auditors, or our regulators. That is the gap 1Password is selling into and Axios is reporting on, and it is the single fastest way for a useful agent programme to become a security incident.
The uncomfortable summary is this. Every enterprise agent programme is, at its core, a storage programme with a user interface. The token price is falling, the decision latency is collapsing, and the residue is growing at 130x per artefact - and none of that residue appears on the invoice that got approved.
The Numbers To Remember
---
The number on every invoice gets the budget meeting. The numbers on the invoices that arrive twelve months later get the exit interview. In 2026, the second set of numbers is the AI bill.
