Nobody knows what AI they're running until the bill arrives

Nobody has a list of the AI they already run.

Three things from yesterday, and they all circle the same missing piece of paper. A hosting company had to ask its customers to go and look at their model bills. A chip maker started selling the monitoring for the machines it also sells. A governance vendor put the agent count itself on a price list.

Advertisement

1. Railway told its customers to go read their AI bills

Railway, which hosts other people's applications, put out two security notices on the same day. One recommends customers check for unexpected OpenAI usage. The other recommends they check for unexpected Anthropic usage.

Two providers, one day, and the instruction in both is the same: go and look.

This reads to me as a detection failure rather than a billing one. The platform could say where to look. It could not say what was there. Somebody at the customer has to open the provider console and recognise their own traffic, which assumes they know what their own traffic looks like.

So my expectation: the sign that something is running in your name arrives as a number on an invoice, several weeks after it started, and finance sees it before engineering does. If my own model spend sat in one pooled account across teams, I would want yesterday's usage split out by key and by service long before I looked at the monthly total.

Read the source →

2. NVIDIA now sells the window into the factory it sold you

NVIDIA launched Minerva, an operations platform for monitoring and managing an AI factory, pitched as end-to-end observability for AI infrastructure.

It was already working one level up. On 16 October ServiceNow gave enterprise agents cryptographically signed credentials that check an agent's authorised scope while it runs, built on NVIDIA's own agent safety platform. NVIDIA now sits under both the credential that says what an agent may do and the console that says what the hardware is doing.

My problem here is not the product, it is the shape. The party that sold you the capacity also supplies the instrument that reports how well that capacity is being used, and those two interests are not the same on renewal day. A utilisation reading from a supplier's own console is the number nobody in the room argues with. I would keep one cost-per-completed-task figure that my supplier does not produce, even a rough one, purely so there is a second reading in the room.

Read the source →

Advertisement

3. ModelOp will count your agents across Microsoft and Amazon

ModelOp released ModelOp Center 3.5, which it calls an enterprise AI command center. It pulls low-code, no-code and pro-code agents into one list across Microsoft and Amazon, and puts spend tracking on top.

The word in its own headline is sprawl, and the ground has been moving under it all month. Microsoft made execution containers generally available for AI agents on 8 October and introduced Agent 365 as a management layer for governing agents on 16 October. ModelOp reads both of those, plus Amazon's Bedrock AgentCore, and on 6 October it signed AHEAD to do the installation work.

An inventory assembled from the platforms' own records will be tidy, and wrong in one specific way. It sees what lives inside Microsoft and Amazon. The thing that cost Railway's customers money this week was running on a key outside any console anybody was watching, and that is the gap a catalogue cannot close. The useful version of this product is a reconciliation against the provider invoice rather than a catalogue, and the question that decides which one you bought is simple enough: which bill does the count get checked against.

Read the source →

Names to watch

Seen this week by Ins7ghts.

• Asymptote Labs · $5.5M seed for AI agent runtime security
  Observability. Tower Research Ventures is in the round, and the same problem drew a much larger name days earlier: on 18 October ServiceNow's AI Passport tied agent identity to business context at runtime.

Companies that later raised money had been in the news about twice as often in the quarter before. Worth a second look in three months.

Advertisement

Also today

• Microsoft said its 365 Copilot has crossed 30 million paid seats. (2 min read)
• OpenAI is near raising up to $30 billion at a $1.25 trillion valuation. (2 min read)
• Legora raised $550 million at an $8.8 billion valuation for AI legal work. (1 min read)

The ScoreboardThe week we could not grade

Thirteen calls went out in the last seven days and almost all of them are still standing, which is its own verdict on how we write them.

1 right 0 wrong 1 not a real call 11 still open

The sharpest one: on 16 October we said governance was turning into a product feature rather than a policy. Six days later ModelOp shipped a release whose entire pitch is counting and governing your agents across two clouds with the bill attached. Right.

How many of your own AI commitments carry a date somebody could hold you to?

Read the deep dive →