[Eufrat]

Are these AI-driven data centers actually getting built? I was reading about Kevin O’Leary’s initial proposal for his Utah data center and it’s just preposterous, it’s inconceivable anyone needs that much compute at any realistic demand level. The ask is so massively large that even the scaled down version is just absurd.

We also have the Stargate UK data center that the Starmer government trumpeted before it was exposed as a glorified PR stunt without any substance. [1]

[1] https://www.theguardian.com/technology/2026/jul/04/openai-ap...

[copper4eva]

I happen to know some people who work for a company that helps build data centers. Business is booming for them, lots of projects all over the country (USA). So just to answer the straight up question of if they're actually getting built, yes they certainly are.

[crawfordcomeaux]

Have you considered the compute demands for nationwide AI video surveillance?

[serial_dev]

AFAIK, video surveillance is less computationally expensive than LLMs, and much of the work can be done at the edge devices.

So don’t worry, there is already probably a nationwide video surveillance system out there, even without Kevin

[shit_game]

Regardless of the computational needs to facilitate mass surveilance, there are inescapable storage needs precipitated by surveilance. Governments want to retain data because its value in the future is unknown in the present. This is the entire premise behind the Harvest Now Decrypt Later[0] principle that fuels much of the previous surveilance actions taken by governments. But HNDL is speculative; it either hedges on the bet that computational breakthroughs will weaken encryption methods to the point where brute force attempts at decryption are not financially feasible for most actors but still physically possible, or rely on a backdoor in an encryption algorithm to decrypt data that is valuable. There has been much speculation about this feasibility and value RE: quantum computation and also the design of encryption algorithms from the ground up (see things like the Clipper Chip and RSA's relationship with the NSA).

With things like video, photo, audio, and liturgical surveilance, the data is largely out in the open (or simply purchased from their hosting platforms, leading to social media and adtech companies like Facebook and Google being defense contractors by nature) and does not need to be decrypted (in most cases) - rather, it needs to be correlated. This is a great task for machine learning because that's what the field has largely been engineered to do for the last 60 years, but it relies so, so, so heavily on having data to both train on and use in its practice. Storage of all data is the perfect end-goal of a surveilance state. Storage requires data centers. Data centers of this calibre are effectively black sites in that they are nigh impenetrable in both practical and legal senses, and I'd wager many data centers host actual digital black sites simply due to their ubiquity as means of doing so coupled with the necessity of the hardware and facilities they provide that are needed for doing so.

Theres really nothing that will ever be publicly available that would say "this data center has a hundred thousand SOTA GPUs in it that are doing gait recognition on every camera in every airport in the world in real time" versus "this data center is a bunch of tape machines that store everything ever recorded from every facsimilie of Room 641A[1]".

0: https://en.wikipedia.org/wiki/Harvest_now%2C_decrypt_later

1: https://en.wikipedia.org/wiki/Room_641A

[leonidasrup]

Harvest Now Decrypt Later principle holds very well for intelligence work, the harvested information could be useful for very long time.

You can see this in the long time intervals specified for Declassification of classified documents.

"Executive Order 13526 establishes the mechanisms for most declassifications, within the laws passed by Congress.[1] The originating agency assigns a declassification date, by default 25 years. After 25 years, declassification review is automatic with nine narrow exceptions that allow information to remain as classified. At 50 years, there are two exceptions, and classifications beyond 75 years require special permission."

https://en.wikipedia.org/wiki/Declassification#United_States

"Decrypt Later" should generalized to "Use Later", many recordings are useful without decryption, just from the metadata you can establish who, when communicated. Identification of relations between persons, building of social graphs, is the basic of intelligence work. Much of this relationship analysis can be done just from metadata.

One example of such analysis is the NSA Co-Travel Analytics

"CO-TRAVELER does not simply collect location information. It creates a portrait of travel times and people who crossed paths, revealing our physical interactions and relationships. The cell site information goes beyond email and phone calls and ordinary telephony data, allowing the U.S. government to know who we are with in-person and where. This is information that would be impossible to collect using traditional law enforcement methods. "

https://www.eff.org/deeplinks/2013/12/meet-co-traveler-nsas-...

https://www.eff.org/files/2013/12/11/20131210-wapo-cotravele...

[shit_game]

Well put; the scope of usefulness of surveilance data lies well beyond what is in that data itself, but how it relates to other data. You don't necessarily need to decrypt data to distill value from it, as its metadata (location, datetime, medium, etc.) is also (and potentially more) useful when related with other metadata. Storing this data is also a monumental task, given the fidelity of records that many consumer devices create, store, and upload.

[michaelt]

Video-processing AI can be very flexible in its demands.

Even if a camera supports 4k 60fps video, the 'edge AI' can downscale the input to 224x224 2fps and use a small network.

Often the 'edge AI' just needs to detect humans/cars/pets and trigger saving some footage to an SD card/turning on a light. It's not the end of the world if it misses some things or raises false alarms sometimes.

And if you buy a Reolink on Amazon expecting it to be able to recognise a face from 20 yards away, you'll be disappointed.

[DeepSeaTortoise]

IMO you highly underestimate the amount of processing and storage video surveillance requires. There is just no good way to determine if you took that pen or made the newly illegal handgesture towards the wrong person 12 years ago unless you have enough storage to keep all the footage around and the processing power to search through it or prepare it for new types of queries in reasonable time.

Also, have you considered AI being necessary to fill in for missing surveillance footage if the cameras are failing (the ruling class)? E.g. the whole Epstein (PR) disaster could have been avoided if the cameras could have been kept running and the footage post-processed in time.

[protocolture]

I think like 30 - 40% of all the announced projects turned out to be financially viable. Possibly less.

Conversely I see heaps of already in progress data center builds and upgrades tacking AI on as the purpose as just PR to sell colo.

[rolph]

if you can actually simulate a stock market with fidelity, i believe that would be a distinct advantage.

probably not short trading, but investigating leveraging positions may be worth it.

[peer2pay]

I’m not too deep into the matter but don’t they all have to wait in the interconnection queue first anyway? My understanding is that that alone will take some years

[bpodgursky]

Most serious projects are building onsite power generation using natural gas to avoid this.

[hylaride]

Kevin O’Leary is a grifter and he should be completely ignored. He literally got rich conning HP (though they were stupid for going ahead with it).

That being said, there is a lot of actual buildouts happening right now. The question is how much of the planned future capacity will actually get developed.

[quickthrowman]

I thought he got rich conning Mattel to buy The Learning Company in 1999? They paid $3.5B for it and gave it away a year later. Then there was a lawsuit with a $100M+ settlement.

Kevin O’Leary also published pictures of him shedding tears while buying an Audemars Piguet watch, which should be brought up and mocked every time he’s mentioned.

[hylaride]

You're right, I misremembered. HP was conned by a British guy named Mike Lynch.