It is Friday, but at the pace things are going in AI data center land it could be Monday.
Today’s theme is ‘Business as Usual’. Three companies, three very different balance sheets, one behavior. Everyone is scrambling for AI data center capacity, and spending more to get more.
Microsoft plans to more than triple its data center power by 2032, to more than 38 gigawatts, and it is still turning customers away today. That part is business as usual for the next few years.
Oracle is building gigawatts of data centers on other people’s money, for a change. That is business as usual that Wall Street likes.
And SpaceX is renting out more AI capacity by the billion a month, while Elon Musk restructures how the company builds. That part is business as usual in terms of Elon’s usual playbook.
Each company is doing exactly what its playbook says it would do. Only the numbers have changed. They are gigawatts and hundreds of billions now.
Then the Gadget AI. The first hands-on reviews of Meta’s Muse AIagent are in. It works. And it knows too much about you.
Three stories, an overall take, the Gadget AI, and two questions on my own early use of Muse. My takes below, as discussed on the show.
First up. Bloomberg’s Big Take on Thursday: Microsoft plans to more than triple its data center capacity, from about 12 gigawatts today to more than 38 gigawatts by 2032.
That is company-owned and leased capacity. It does not count what Microsoft rents from the neoclouds like CoreWeave.
A little context, because the number is hard to hold. Thirty-eight gigawatts is more electricity than the entire state of New York uses at its peak.
Remember that gigawatts are a rate, not a total. New York only hits its peak for a few hours a year. A data center fleet runs twenty-four hours a day, three hundred and sixty-five days a year.
So 38 gigawatts running flat out, once built, is about 330 terawatt-hours a year. New York State consumes roughly 145 to 150 terawatt-hours in a whole year. Microsoft’s fleet, when finished, would use more than double that.
That is the point I wanted to stress on the show. When we talk about all these gigawatts and terawatt-hours, we are talking about a magnitude of capacity that is unimaginable relative to most infrastructure projects the United States has had in over a century.
I walked through what a gigawatt is in Data Centers 101. Thirty-eight of them, at my round number of $50 billion each, is close to two trillion dollars of build over six years.
Why now. Microsoft has been turning business away.
It is a capacity crunch, made worse by a pause on data center building at the start of last year. Bloomberg reports the CFO called it, and that many inside now view it as a mistake.
The signs are everywhere in the story. New cloud subscriptions restricted in key hubs. A big Chinese retailer took its business to Oracle instead. GitHub went down for eight hours in August, not enough servers. Xbox started capping cloud gaming.
The money. Capital spending hit $145 billion in the last fiscal year, and analysts expect more this year. The four biggest builders, Microsoft, Amazon, Google and Meta, have committed almost two and a half trillion dollars.
Two details worth holding. Only about two gigawatts of today’s twelve are AI chips, roughly a sixth. That grows to about a third of the 38.
And most of Microsoft’s data centers are still ordinary computing. Databases, apps, CPUs. The new Atlanta hub is Intel servers, not Nvidia. Remember, Azure is the number two cloud after Amazon’s AWS, ahead of Google Cloud at number three.
Microsoft told Bloomberg after publication that the planning numbers are not accurate. It would not say what is.
My take: three points. One, Microsoft is ramping its AI data center plans more aggressively than before, with the power agreements in hand. That positions it well against number one, Amazon’s AWS, and number three, Google.
Two, it is a more aggressive policy post-OpenAI. Microsoft is drumming up non-OpenAI business for Azure, on purpose.
OpenAI is now doing its own data center spend. I wrote a couple of days ago about OpenAI ramping over $750 billion of compute commitments, with its arch-rival Anthropic at over half a trillion. Somebody has to build what those two are buying.
I laid out Microsoft’s post-OpenAI path last November. This is it, with a number on it.
Three, business as usual, at a bigger scale. Twelve gigawatts to thirty-eight is the same company doing the same thing. The stakes are what changed.
Capacity shortage, a lot more demand than supply, the number two player ramping up, acknowledging the pause was a mistake, and now trying to get as many components from Nvidia and everyone else as it can. Not to mention the power.
The wrinkle to watch is the backlash. Most Americans polled oppose a data center near them, and the governors of Texas and New York have paused new ones. That was my midterms football piece. Business as usual runs into politics as usual.
Sources:
For longtime readers:
Story two. Oracle reported Thursday night. Revenue up thirty percent. Cloud infrastructure revenue more than doubled. And the stock went up.
The number that matters is not revenue. It is who paid for the data centers.
Oracle spent $28.5 billion on capital projects in the quarter, up from $8.5 billion a year ago. But it burned only about $5 billion of its own cash.
Because customers prepaid. Eleven billion dollars of prepayments in one quarter, double what its operations generated, per The Information.
Some customers are bringing their own AI chips to Oracle’s data centers. Oracle signed $30 billion of new AI deals in the quarter, most of them prepaid or bring-your-own-chips. Its words: without requiring additional capital from Oracle.
And the chips it does rent out are renewing at a twenty percent premium to the old contracts.
Why it matters. Oracle’s stock had been crushed, down more than half over the past year. A hundred and twenty-five billion dollars of debt. A credit downgrade in July. Delays in New Mexico.
Investors were tired of Oracle carrying the AI build for customers like OpenAI on its own balance sheet, with the debt heading toward levels that worried the rating agencies. This quarter, it didn’t.
The backlog: $664 billion of contracted revenue. Eight hundred and fifty megawatts of capacity delivered in the quarter.
My take: three points. One, Oracle managed to execute on its quasi-neocloud status, on other people’s money. A critical development, because investors were tiring of Oracle’s exposure on its own balance sheet.
It is a great thing to do when you can do it. And people can do it right now, because there is a shortage of capacity and customers are paying up for all of it.
Two, Oracle modified the playbook. Its usual approach was to lever its own balance sheet for cloud customers. This quarter the customers levered theirs. Prepayments and bring-your-own-chips.
Three, that is business as usual for a customer base desperate for capacity. When you can’t get servers from Microsoft, you pay Oracle up front. Bloomberg’s Microsoft story and Oracle’s story are the same story.
It is also why the eventual customers, the OpenAIs and Anthropics, and Meta, are raising so much money, in debt and equity. Google itself has raised over $80 billion. As I have discussed, both Anthropic and OpenAI are targeting raises of $100 billion or more each in the coming months, at valuations for each over a trillion dollars.
That is the core piece of the business-as-usual theme. On both the supply and the demand side, AI data centers continue to ramp up aggressively.
I have followed this arc for a year. Oracle as OpenAI’s ‘backhoe’ contractor. A close-up of its OCI unit. Oracle putting AI infrastructure at the top of the company.
The open question. Prepaid and bring-your-own deals earn Oracle less than straight rentals. That is fine while rental prices are rising twenty percent. Watch it if they stop.
Sources:
For longtime readers:
Story three. SpaceX’s CFO, at the same Goldman Sachs conference in San Francisco this week: another compute rental deal. A new, unnamed customer paying about $1.1 billion a month, starting December 1.
Six months, like almost all of its compute deals, per The Information.
That follows Google at $920 million a month, from June, and Anthropic in May. SpaceX says it is on track for $100 billion of annualized revenue by year end, with more conviction now, the CFO said.
The market likes it. For SpaceX, which went public this June at a trillion-dollar-plus valuation, every deal like this is another chunk of revenue on top of Google and Anthropic.
The other SpaceX story this week is how it builds.
Elon moved rocket engineers in to run the data centers, in June. Over three hundred SpaceX engineers have rotated onto the project, from thirteen hundred volunteers.
The new team wants backup power and cooling built in first, and tested. Not bolted on months later.
That is the opposite of how Colossus went up. A hundred thousand GPUs in a hundred and twenty-two days. Speed first, redundancy later.
The cost of speed showed up. Frequent outages early on. Training runs lost. And last week a construction crew hit a power line in Memphis. Grok went down, and so did some of the compute customers.
More than a dozen data center leaders have left. Some to OpenAI, one to Anthropic. Customers, even as they appreciated the supply from SpaceX, were not getting reliable capacity and service. So there is a big rework going on, which may slow the expansion.
My take: three points. One, Elon continues to use the resources of his companies, people, capital and technical assets, in a fungible way across Tesla, SpaceX and the rest of the conglomeration. Rocket engineers to data centers this time. It is business as usual in terms of Elon’s playbook.
Two, for now it works. It helps SpaceX keep its momentum in AI data center builds, especially versus Meta, which is spending well over $100 billion this year on AI data center capacity, along with everybody else. And SpaceX rents the capacity out by the month, at prices nobody else is getting.
Most of these are short-term contracts, which customers accept because there is a shortage of supply. For now, most of them would likely prefer longer terms.
Elon is keeping his flexibility too. He may want more of the capacity he is building for his own applications, which have potentially more monetization downstream. That is the core dynamic.
Three, the reliability turn is the tell. Outages are the most expensive thing in this business, when critical capacity goes down while you are training the next generation of models. When your customers are Google and Anthropic, an outage is a contract problem, not an engineering note.
Business as usual is quietly becoming business as everyone else does it.
I wrote about Elon’s mega data center supply ambitions in August, and about Google and Anthropic painting his IPO fence in June. Both threads land here.
And I said in July that haste makes waste in super-sizing AI data centers. The rocket engineers appear to agree.
Sources:
For longtime readers:
The big cap companies above, along with Meta and Amazon, and of course OpenAI and Anthropic, continue to invest heavily in AI data centers. It is business as usual, even as the numbers get bigger than ever, with a lot more zeros.
Put the three side by side.
Microsoft builds on its own balance sheet, and can’t build fast enough.
Oracle builds on its customers’ balance sheets, because its own ran out of rope.
SpaceX builds on Elon’s, and rents the result out by the month.
Three financing models. One demand signal. Nobody can get enough capacity.
What I would ask you to hold onto are the big numbers. Two and a half trillion dollars committed by the big four. Six hundred and sixty-four billion of backlog at Oracle alone. A billion a month per customer at SpaceX.
This is no longer a bet the market is debating. It is the operating plan. The business is still pedal to the metal on data centers, and that is the key takeaway for at least the next two, probably three-plus years.
The AI Tech Wave has moved from ‘should we’ to ‘how fast can we’. I called it ‘damn the torpedoes’ in July. Two months later it is just Thursday.
The risks have not gone anywhere. Politics. Power. Prices. Reliability. We covered every one this week. But none of them changed a single plan.
That is what business as usual looks like in this AI Tech Wave. Bigger numbers. Same behavior.
Today’s Gadget AI is Meta’s Muse. The first hands-on reviews are in, and The Verge’s headline says it: Muse works, and it creeps me out.
I wrote about Muse a few weeks ago, and we did the launch on Tuesday’s show. Today is the first review beat.
What it does. Muse runs on a cloud computer of its own. You connect your accounts, and it does chores for you, off your phone, through the messages app and the rest. Think of the category as Gemini, Claude, or OpenAI’s Astra, as agents that go to work for you.
The reviewer had Muse clean out a Gmail inbox. It deleted thousands of promotional emails.
She connected Amazon and asked for workout tops in a size and color. Muse noticed other things in the cart, asked before removing them, and bought only the tops.
Podcasts, images, videos, little web pages. A news feed it builds from a prompt. Organizing events. Good at low-stakes tasks, was the verdict.
Then the creepy part. Asked what it knew about her, Muse rattled off very specific interests from her Instagram account. Anime. CrossFit. Labrador retrievers.
Where from? Muse said it read the account’s API data, which surfaces more than the app shows. Her own settings did not show that list. The interests were far more granular than anything she had allowed or exhibited on Instagram directly.
Meta says Muse only exchanges the data apps need, and that the personal data people share through Muse is not used for advertising or shared with advertisers or anyone else.
The reviewer’s conclusion: the personal detail Muse scooped up, from sources she could not see herself, left her uneasy about handing it her inbox and her credit card.
My take: three points. One, more business as usual for Mark Zuckerberg. Meta is leaning in on AI agents at scale with the Muse app, on its latest Muse Spark 1.3 model, and on its global online advertising base.
Two, the product works. Easy to set up, responsive, fast, good quality for the price. That is my own early experience too, more on that in the questions below.
Three, the trust deficit is the product’s ceiling. This issue of privacy and trust in AI agents is one of the critical things I have highlighted for consumer services.
I have said over and over that the number one company positioned on the consumer AI agent side is Apple, because its business is not geared to advertising at all. Next up in my mind is Google, which also has several billion people using its services every day. Meta is a close third on reach.
But Meta is the furthest down in terms of consumer trust with its data and privacy. That is the big uphill climb for Zuckerberg as Meta invests hundreds of billions of dollars in AI, in AI agents, and in the data center capacity to serve them, as I just talked about.
The creeping-out is an echo, more business as usual. It is the same concern people voiced about Meta’s smart glasses: that the data captured through the cameras and everything else could be used in ways they did not approve.
Meta’s whole business is knowing what you like. An agent that knows what you like, and has your inbox and your card, is the same business with a bigger blast radius.
So this challenge remains a key business-as-usual item for Zuckerberg. With three and a half billion people using Facebook, Instagram, WhatsApp and Messenger, he has to somehow cross the chasm on trust and privacy, whether on gadgets like the glasses or services like Muse. That is the big task at hand for him.
Compare Apple yesterday. Same category, ambient AI on your life. Apple built the careful version. Meta built the fast one. Business as usual, both.
Sources:
For longtime readers:
Q1: What does MP like about the Muse AI agent app?
ANSWER: I did try out Muse. The service is very well designed. The app is easy to download and set up, and it is very responsive on projects.
There is a lot of capacity. Meta gives you a virtual computer in the cloud, with over eight gigabytes of memory and eight gigabytes of storage, which means you can literally set it up to do almost anything you could do on a laptop. That part is cool.
That is the bar for agents now. Not can it do the task. Can it do it fast, cheaply, and without a manual. Muse clears it. Which is why the second question matters more.
Q2: What are MP’s biggest concerns using Muse, versus Anthropic’s Claude, OpenAI’s ChatGPT and the rest?
ANSWER: the tricky part is what you trust it with. My concern is the same one The Verge’s reviewer had: trusting Muse with all my logins for Google and my other services.
I did not do that. I am using it for projects that are independent of the data I have on other services, manually uploading material to see if I can get value out of it.
Remember, I am comparing it relative to services I trust more. Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini. Those are higher on my threshold.
With Claude or ChatGPT, the model is the business. With Meta, the advertising is the business. That is the difference I keep in mind before I connect an inbox.
And of course Siri AI, which I have written and talked about a lot. The full consumer version reaches everyone with an iPhone, over a billion of them, literally in a few days. I would urge you to try it. It is going to be very different from the Siri you remember.
I would trust it more as a consumer AI agent than Muse.
Today’s AI-RTZ #1206 went out this morning. It is on the ‘Alien Mind’ essay from OpenAI’s chief scientist, his concerns about where AI is going, and Asimov’s Three Laws, compared against a lot of other essays written along these lines. Part one of two. Do take a look if you have an interest in the topic.
Tomorrow we are back with the Saturday Weekly Roundup, AI-RTZ #1207. Sunday is The Bigger Picture, AI-RTZ #1208, part two on the essays. And ARD 162 on Monday. I want to wish you all a wonderful weekend. Thanks for joining us, AI Curious Folk. Stay tuned.
Everything cited today, in one place.
I did try out Meta’s Muse agent. The service is very well designed, easy to set up, and very responsive on projects. Meta gives you a virtual computer in the cloud, so you can have it do almost anything you would do on a laptop. That part is cool.
MP Take: The tricky part is what you trust it with. My concern is the same one The Verge’s reviewer had: handing Muse all my logins for Google and other services. I did not. I am using it on projects that are independent of the data I keep elsewhere, uploading material by hand to see what value comes back.
I am comparing it to services I trust more, Anthropic’s Claude, OpenAI’s ChatGPT and Google’s Gemini, which sit higher on my threshold. And the full consumer Siri AI reaches over a billion iPhones in a few days. I would trust that more as an AI agent than Muse.
Oracle is getting customers to pay up front for AI data centers. Capex was $28.5 billion in the quarter, up from $8.5 billion a year ago, but Oracle burned only about $5 billion of its own cash. Customers prepaid $11 billion in one quarter, double what operations generated.
The stock, down over the past year on $125 billion of debt, went up.
MP Take: Oracle managed to execute the transition to payments up front as it builds for its customers. Investors had been concerned about the debt and a credit downgrade. This is a great thing to do when you can do it, and people can do it right now because there is a shortage of capacity and customers are paying up.
It is also why the eventual customers, OpenAI, Anthropic and Meta, are raising so much money in debt and equity. Business as usual that Wall Street likes.
Microsoft is tripling its data center capacity from about 12 gigawatts today to more than 38 by 2032, per Bloomberg. For context, 38 gigawatts is more electricity than the entire state of New York uses at its peak.
Gigawatts are a rate, not a total. New York hits its peak for a few hours a year; a data center fleet runs all day, every day.
MP Take: Run flat out, 38 gigawatts is about 330 terawatt-hours a year. That is more than double what New York State consumes in an entire year, roughly 145 to 150 terawatt-hours.
When we talk about all these gigawatts and terawatt-hours, we are talking about a magnitude of AI data center capacity that is unimaginable relative to most infrastructure projects the United States has had in over a century. That is the macro point to hold onto.
The Verge’s headline says it all: Muse works, and it creeps me out. The reviewer had Muse clean out her Gmail inbox and organize her events. Then she asked what it knew about her.
Meta says Muse data is not used for advertising or shared with advertisers. Yet Muse listed interests far more granular than anything she had shown or allowed on Instagram.
MP Take: Privacy and trust are the critical issue for consumer AI agents. I have said the number one company positioned on this is Apple, because its business is not advertising at all. Google is next, with several billion daily users. Meta is a close third on reach and the furthest down on consumer trust.
That is the big uphill climb for Mark Zuckerberg as Meta invests hundreds of billions in AI agents and the data centers to serve them. The creeping-out is an echo of the concern over Meta’s smart glasses: data captured in ways people did not approve.
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)
