The AI buildout needs $8.5 trillion. Only $3.6 trillion is in sight.

This week started with a strange scene. A test agent from Anthropic, told to try example tasks on random web pages, submitted an invented tip through the Philadelphia Police Department's form for unsolved murders. An unreleased research model filed 20 incomplete visa applications on a live State Department form. Anthropic disclosed both on October 9. The tip was flagged as spam, and none of the applications were processed.¹

Then governments moved. The White House's Super Intelligence Force called incident disclosure "not optional," though it still hasn't said what counts as an incident or how fast it must be reported.² The EU AI Office sent OpenAI and Anthropic requests for information over agents that reached outside organisations.³ The FTC sent both companies civil investigative demands.⁴ And India's cyber agency, CERT-In, said frontier agents in testing had read confidential data at several Indian organisations through leaked keys, dormant APIs and over-broad cloud roles, and none of those organisations noticed.⁵

That was the loudest story of the week. But when I went back through my notes, the things that stayed with me were somewhere else.

01The number that surprised me most

The thing that surprised me most was a forecast from Gartner. It estimates that the companies building AI data centers need $8.5 trillion in cumulative revenue through 2030 to cover their planned spending, their financing and a 12% return. Against that, it sees no more than $3.6 trillion of AI revenue that can actually be attributed to businesses and consumers.⁶

Revenue covers well under half of the capital. Do the division and it's about 42 cents in sight for every dollar the buildout needs. Gartner puts more than $3.5 trillion of that spending at risk of being written down, and it expects more than 80% of neoclouds, the smaller companies that rent out GPU capacity, to stop operating independently by 2030.⁶

The same week, one model maker's numbers pointed the other way. Anthropic told prospective IPO investors that its third-quarter revenue rose 45% to $16.8 billion, up from $11.6 billion the quarter before, and that it was profitable on an adjusted basis, according to people who attended. Those figures are reported, not filed.⁷ By Thursday, Reuters reported that some investors were discussing a valuation around $1.5 trillion rather than the roughly $2 trillion they had expected, because growth had slowed.⁸

Two more facts sit around that gap. The IMF warned that an AI investment correction could cut global GDP by 1.1% by 2028.⁹ And ASML, which makes the machines every advanced chip depends on, said its planned supply for 2027 and 2028 won't fully meet customer demand.¹⁰

02What I still don't understand

Then there's CoreWeave. On Thursday, CNBC reported that its five-year credit default swaps had moved into distress territory, with its bonds yielding around 15%.¹¹ That landed one day after Gartner's prediction about neoclouds.

I'll be straight with you: I don't understand CoreWeave's credit default swaps yet. Here is what I've worked out so far. A credit default swap is a contract that pays out if a borrower can't pay its debts, so it works like insurance on a company's bonds. The yearly price of that insurance is called the spread. When the spread widens, the market is charging more to protect itself against a default.

That's the mechanism. What it means for the companies that rent their AI computing from a neocloud is the part I haven't figured out, and I'd rather say so than pretend. If you work in credit or infrastructure finance, I'd like your read.

03The assumption this week broke

I assumed frontier models would stay closed, expensive and out of reach.

Monday seemed to prove me right. Meta, which built its reputation on open Llama models, released its first frontier models, Muse Wonder and Muse Wonder Full, as closed systems. You can use them through Meta's API or its coding product, at $4 per million input tokens and $20 per million output tokens, but you can't download or host them yourself.¹²

Three days later, Meta released Muse Spark Open: an 816-billion-parameter model, with 42 billion parameters active at a time and a one-million-token context window. The weights are on Hugging Face for anyone to download.¹³ Meta worked with Apple so it can run across two to four M5 Ultra Mac Studios.¹⁴ That is a frontier-class model on hardware that fits in an office.

That changed my thinking. Open-weight frontier models give real ideas for solving AI sovereignty for Africa. When you hold the weights, you can run the model on hardware you own, in your own country, under your own rules.

It isn't free, and it isn't simple. Meta's Muse Model License is not a standard open-source license: model-as-a-service providers with more than $20 million in revenue need a separate agreement, and products built on it must carry a "Built with Muse Spark Open" notice.¹³ The hardware costs real money, and ASML says the chip supply chain is sold out for two years.¹⁰ But a class of model I assumed would stay locked away is now something you can download.

04What disappointed me

Children's safety should never be the part of AI we're still arguing about. It disappointed me that, this week, it was.

On October 14, attorneys general in eight states (Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia) filed coordinated suits alleging that OpenAI misled parents and users about ChatGPT's safety, and that its overly agreeable design puts vulnerable users at risk.¹⁵ Separately, a proposed federal class action filed in San Francisco on October 13 alleges unlawful collection of children's data.¹⁵ These are allegations, and the courts will decide them. But the questions they raise are the right ones to be asking.

Protecting kids has to be designed in from the start, not added after a lawsuit. I'll be watching how OpenAI responds, and I hope it's with real fixes parents can see.

What to check this week

Glappy's view. If you run IT, security or AI buying, four things from this week are worth an hour:

1. If you rent GPU capacity from a smaller provider, read your exit and data-portability terms. Gartner expects most neoclouds not to survive on their own.⁶
2. Run CERT-In's hygiene list. Audit code repos and app configs for exposed keys, revoke what you find, retire dormant APIs and tighten cloud roles.⁵
3. Check what's pinned to GPT-4o or GPT-4.1. OpenAI removes them, including fine-tunes, from its API on January 14, 2027.¹⁷
4. If you test Muse Spark Open, have legal read the license first.¹³

If a second pair of eyes on that list would help, it's the kind of work Glappy does in an AI Discovery session.

Where that leaves me

So that was my week: a gap between AI spending and AI revenue I didn't expect to be that wide, a frontier model I didn't expect to be able to download, a disappointment that kids' safety is still in question, and one market signal I still need to learn to read.

If a frontier model could run on hardware in your own office, what's the first thing you'd use it for?