Every time you ask an AI a question, a chip in a data centre somewhere starts doing sums.
That’s the most basic description I can give of what “compute” is in a modern sense. It’s the work a computer does, and it’s paid for in chips, electricity, and time.
An AI model, at its most basic form is a huge pile of numbers. Type a question, and the chips push your words through that pile, over and over, to guess the best next word, and then the one after that, and the one after that.
Every word costs a sliver of compute. Until now, that’s been manageable, because a chatbot answers you in a few seconds and then sits there waiting for your next question.
Oh, and saying “please” and “thank you” to your AI makes absolutely zero dent in compute resources. Anyone saying different is a fool.
Now, this has all been fine… but the new wave of AI doesn’t sit and wait for your next question.
Meta launched its personal AI agent, Muse, a few weeks back in the US. It’s an agent, which means it doesn’t only answer. It goes off and does work for you.
Meta’s launch-day post explains how it works.
You and your Muse share your own dedicated computer in the cloud. This is where your Muse lives and where all data and credentials for any service you connect are securely stored.
Every user gets their own computer, running in a data centre.
David Singleton from Meta’s Superintelligence Labs said:
It’s a real computer, and we’ve designed the security architecture of the Muse Secure VM carefully so you and your Muse can do almost anything you could with a computer sitting under your desk.
We’re really proud to be able to put a real computer in millions of people’s hands with a similar level of transparency. We built a file explorer right into the Library tab of the UI. We want you to be able to see the markdown files Muse writes while it thinks about how to serve you better.
Hot on the heels of Muse, OpenAI launched Dots on Tuesday.
Each one works from (surprise, surprise) “a cloud computer of its own,” and it’ll do things like prepare an invoice when it spots a billing slip-up.
Sam Altman called it “an AI helper that always has your back.”
Elon Musk’s xAI got in early with Grok Bot back in August. Its bots “share a computer of their own in the cloud, so jobs do not stall when you step away,” researching sales leads overnight or clearing invoices out of Gmail.
Do you see what’s happening here?
Do you see the pattern?
A chatbot burns compute for a few seconds, while an agent burns it all day and all night, on your own computer that’s never switched off, and now we’re talking about hundreds of millions… billions of people.
An AI agent can use how much more energy than a simple AI query?
|
Add to that, let’s say you’ve got Muse, Dots, and Grok Bot. Three computers per person. I’ve said before, we are on the verge of AI agents grossly outnumbering the number of people on Earth.
And long term, well, an infinite number of AI agents can exist.
All of them need compute.
In 1938, The Fleet Street Letter warned its readers that war was coming to Europe — and even identified September as the critical month. Since then, its readers have been warned ahead of Black Monday, the dotcom crash and the 2008 financial crisis. Today, Britain’s longest-running investment advisory is still looking for the “news behind the news” — and the investment opportunities others could be missing. Click here to discover what The Fleet Street Letter sees coming next. Capital at risk. It warned readers ahead of Black Monday. The dotcom crash. And the 2008 financial crisis. Now, nearly 90 years after it was founded, The Fleet Street Letter is still searching for the “news behind the news” — and the investment opportunities that could emerge from the next major turning point. Click here to discover what its editors are seeing now. Capital at risk. Britain’s Investment Letter Since 1938
What Does Britain’s Oldest Investment Letter See Coming Next?
A dollar in, $1.60 out
There’s one big name missing from that list of useful personal AI agents, for now.
Anthropic, the maker of Claude.
Anthropic is heading for the biggest IPO ever, at a hoped-for valuation of about $2 trillion.
But in leaked filings, Anthropic’s valuation looks a bit obscene.
Yes, it makes a heck of a lot of money, but it spends a mind-blowing amount as well. In fact its debts are racking up daily
All because it can’t get enough compute.
Anthropic made about $4.59 billion of revenue in 2025, up 1,088% from $386 million the year before.
However, wait till you see the inverted spend on compute.
Anthropic spent about $7.33 billion on compute and infrastructure in 2025, up 190%. That’s more than half its $12.65 billion of operating costs.
So for every dollar of earnings Claude made last year, Anthropic spent about $1.60 on chips, servers, and data centres.
That sounds awful, because it is, but it also tells us where to go as investors.
Run the 190% backwards, and compute cost about $2.5 billion in 2024, compared with $386 million of revenue… that’s more than $6 of compute for each dollar earned.
In one year, the compute cost of earning a dollar fell by about three quarters, and yet its compute bill still tripled.
This is exactly what I’ve said would happen. If you make AI cheaper, people use far more of it, and agents will use even more again. As token costs fall, the speed of AI development accelerates.
That is, if you can get your hands on compute.
It’s in such supply that Anthropic had to go shopping. The filing lists $518 billion of cloud, computing, and infrastructure obligations over the coming years.
How on earth are they going to fund that?
Follow the invoices
Every cent of that $518 billion ends up as somebody else’s revenue.
Amazon (Nasdaq: AMZN) is one of the biggest beneficiaries.
In April, Anthropic committed more than $100 billion over 10 years to Amazon Web Services (AWS), for up to five gigawatts of Amazon’s own Trainium chips.
Broadcom (Nasdaq: AVGO) builds Google’s tensor processing units (TPUs), the AI chips Google designs in-house. Anthropic gets about 3.5 gigawatts of that capacity through Broadcom from 2027.
Then there’s Akamai Technologies (Nasdaq: AKAM), a company known for delivering web pages. Last Thursday, it signed a seven-year deal worth $11.6 billion to run Anthropic’s CPU workloads.
And don’t forget the memory side of things (again).
Anthropic brought Samsung and SK Hynix in as strategic infrastructure partners in May, and Micron (Nasdaq: MU) was confirmed as another strategic partner in June.
The world’s three makers of high-bandwidth memory all own a piece of their customer.
The market has said that a lot of this simply doesn’t pan out, that it’s indicative of the AI trade falling apart.
I couldn’t disagree more.
Sure, the Anthropic IPO valuation looks absolute garbage. I wouldn’t go near it. But everything else looks very enticing.
Anthropic will release its own version of Muse, Dots, and Grok Bot soon enough.
Then we end up like… well, I think where we go next is actually best summed up by this post I saw today.
Maybe you have an agent for the household, an agent for the office, an agent for your utilities and bills, an agent for your calendar of events and activities, an agent for a secondary job, an agent for your investments, an agent that coordinates and manages them all…
And guess what every single one needs?
Compute.
Provide the compute, sell the compute, profit from the compute. That’s where you find the market winners.
Until next time,
Sam Volkering
Investment Director, Southbank Investment Research
PS AI agents are going to need an extraordinary amount of compute. And the companies providing it could be some of the biggest winners of this next phase of AI.
My AI system, Hyperion, scans every tradeable stock in the UK and US markets and ranks them from worst to first. It’s just identified one company with an Ignition Score of 89 out of 100. I reveal the ticker in the first 10 seconds here.
