Waldo

AI data centers will eventually reach an equilibrium where those that can most efficiently turn power into compute will profit while the rest of the industry breaks even or loses money, just like Bitcoin mining. Early dedicated miners went obsolete and earned less than the power costs quickly.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

What I don't know, and where it seems like there is a lot of potential for nefariousness, is what fiance looks like in a partial operational situation. Some racks installed and producing while the construction is ongoing, the grey area that is "complete".

You could feasibly order the wrong long lead time item needed for project close out a few times and ride this completion grey area for years. I'm not sure how finance deals with this.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

All those big deals they were signing in 24 and 25, the majority that actually get built and make it on to balance sheets won't do so until 28-30. The big bubble inflator is that construction financing is off balance sheets and not counted. But they need it in hand before breaking ground.

AI sucking all available capital causing interest rates to rise and construction material and labor shortages are a sign of those big deals a couple years ago passing out of planning and design and into execution.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

They are going to need construction financing during design to order long lead time items and pay design contractors. Realistically it's more like 5 years from deal signed for compute to the construction loans for that compute to convert.

Numbers less than that are aggressive and meant to convince investors, or a repeat of those numbers by a press that doesn't understand them. Ed's and other articles about the structure of the bubble use compressed optimistic timelines for data center construction.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Anything under 2 years from shovels in the ground to energized and producing racks is an aggressive timeline. Testing and config could take months. The full build out to an operational big data center realistically is more like 3 years from shovels in the ground to producing.

Long story short, when you see a timeline of 2 years for a data center that is almost certainly way off. Financing isn't going to convert from short term construction financing to longer term bond debt until the very end of the project, when everything is done, all racks installed.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Construction likewise could be a multi year process and it certainly isn't going to be thrown up in a couple months. The steel frame goes fast, though it takes a while to get steel on site. The electrical, fire suppression, and HVAC systems are going to take a while.

At some point the building will be able to be occupied and the server racks installed. Chances are they will rapidly get something going to show investors, but the full electronic install is going to take a while. Again, it could take years.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Every project is late and over budget. Cost estimating and schedule estimating is hard. There is incentive to under count both. High costs and long timelines scare away investors. Reality doesn't have to hit until the middle of construction. This bias is built in to every public number.

This is a constant fact in the construction industry. Noone likes realistic cost estimates or schedules and even if they did, engineers are terrible at producing them. Any time a project schedule is reported it's always the best case scenario. Most projects under-budget funds for changes.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

The step I skipped between design and construction is not included in design-build contracts, but it still exists, its length depends on how efficient engineering is. This step includes hiring construction contractors and ordering materials, and permitting often bleeds to this step.

In the end, no matter how it's structured, from the point of creation of a project and early planning to shovels in the ground is going to be a years long process. Think 2-3. Then construction starts.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

This means that design likewise could take years to get to plans to submit for permitting. Timelines in the scale of weeks are wrong, it's going to take several months minimum. Permitting could take a while, these are complex plans for permitting departments.

To successfully compress timelines in the middle of design orders will have to be placed for the heavy HVAC and electrical equipment. The delivery timelines on this stuff is long. There is a process for approving vendor designs, it isn't off the shelf stuff, vendors have to build it.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Planning is basically always assumed to only take a few weeks when project timelines are in the public, but in reality it can take years. Lining up land, finance, utilities, preliminary permitting work, and design contractors. Opposition to data centers is only stretching the timeline further.

Data center design isn't some easy cookie cutter thing. The uninterruptible power needs are significant. The fire suppression and cooling systems are state of the art. These all need to be on drawings for permitting. These are complex designs with multiple vendors to coordinate.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Basically the financing structure of data center construction is very similar to the subprime mortgage bubble. While I don't know much about the details of finance, I know a whole lot about building large structures full of electronics equipment racks. How long does it take to build?

Every large project like this goes through 4 distinct phases. Planning, engineering, construction, and installation/testing. Some compression is possible (engineering and planning for example), and I purposely skipped a phase that occurs between engineering and construction, assuming compression.

Small thread about the AI bubble, specifically about a detail basically everyone gets wrong. Some get it wrong on purpose. Building data centers.

Professionally I plan construction of data center adjacent projects (working my way up through engineering), thus see the impacts of the bubble directly on construction. Im not going to link to any stories, but Ed Zitron has written about the gap between construction finance and operating debt.

Tesla stock is up because its confirmed that Elon is going to pursue merging Tesla with SpaceX. The Roadster is not a moneymaking project, it's advertisement to shareholders and regulators. Twitter shareholders did pretty good when Musk bought it.

Comfortably Boo@numb.comfortab.ly

Tesla stock is up on expectation people will pay $250,000 "reservation" cost for a non-flying model.

The new Tesla demo is the most "we need to merge with SpaceX" thing ever.

It's pretty much as if the design concept was use SpaceX tech in a car, make a fancy demo to excite people, then use it to facilitate a merge between his two big companies. It's just so blatant.

IMO - Eventually we are going to discover that the plethora of ancestor worshipping ancient religions actually have some signal there, related to where our instincts come from (multi time dimensions, time is only a vector to the observer). Instincts are a key ingredient to consciousness.

Einstein gave hints of the time multiverse with special relativity (if observer points of view can never merge...). It seems like a lot of physicists are there with various unified theories. Sci Fi though hasn't had much to say yet. Hard to wrap the mind around.

It's one of those things like heliocentrism, an alternate explanation of observed reality, that isn't easily accepted. That said, AI isn't conscious, it doesn't have its own relativistic point of view of the universe, unlike all living things.

My personal universal cosmology long predates the rise of "AI", as such I've been a skeptic of this industry from the start. Is AI conscious? Not even close. In the end it's still an algorithm. Consciousness needs the connection to the whole chain of life.

The need for the whole chain connection also means to train an AI with human-like instincts would take billions of years in vector time.

AI in its current form has bounding parameters but no instincts, and it can never have instincts because it is not connected to all other AI in the time multiverse like all life is to one another. Our feint ability for things like telephone telepathy comes from the same root. AI can't do this.

Green Bay stomping the Bears in the ground. Love it. Going deep when up by a ton.

In the days of yore, free range children used to know where the nasty kid hating old people lived and would mess with them for entertainment.

This post is unavailable.

Dr Earhart Sampson PhD 🌎 🇨🇦 🇺🇲 🇲🇽 🇩🇰 🇺🇦

They need to figure out how to compete with Mr Beast and his ilk; Hollywood is fast becoming entertainment for old people that kids these days largely ignore.

This post is unavailable.

Waldo reposted

@earhart.bsky.social

A Diver Volunteered to Clean Up Trash From a Wisconsin Lake. He Discovered a 1,000-Year-Old Dugout Canoe at the Bottom Instead share.google/wBaE10XbQaXt...

share.googleA Diver Volunteered to Clean Up Trash From a Wisconsin Lake. He Discovered a 1,000-Year-Old Dugout Canoe at the Bottom InsteadIndigenous communities near present-day Milwaukee likely hand-carved the white-oak vessel before shaping it with fire between roughly 992 and 1049 C.E.

Waldo reposted

Prof. Eliot Jacobson

@climatecasino.net

Breaking News! Code UFB!!! For the first time in recorded history, the Niño 3.4 sea-surface temperature (SST) is above 30.0°C, reaching a preliminary SST of 30.06°C on October 5th. The Niño 3.4 SST anomaly is now 3.41°C above the 1991-2020 baseline. Everything is not fine.

I took a course on the history of science and technology in the early modern period in engineering school 25 years ago. In hindsight it was one of the more important courses I took. NVM that critical thinking part, writing well is what gets you promoted as a practicing engineer, not doing calcs.

Mike Ross@eruptionchaser.bsky.social

Am I missing something here? In what universe would an engineering student ever study history?! Do many history students study engineering?! I *enjoyed* history at school, and I was *good* at it. But no way would I ever have studied it at university.

Waldo reposted

OGBobbyJoshson

@jawalker.bsky.social

Replying to @atrupar.com

It’s like when he renamed his dementia “the weave.”