The AI Spending Meltdown> be tech companies in 2025> spending $400 billion on AI infrastructure> "this is totally sustainable anon"> spending more than Apollo moon program every 10 months> consumers only spending $12 billion on AI services> mfw Singapore GDP vs Somalia GDP gap> "the numbers don't add up" they said> "it'll all work out" they saidThe Obvious Bubble Signs> be Mira Murati> leave OpenAI to start new company> raise $2 billion seed round at $10 billion valuation> refuse to tell investors what you're building> "we're doing AI with the best AI people but can't answer questions"> investors: "most absurd pitch meeting ever"> get funding anyway> absolutely_degenerate.jpg> meanwhile retail investors> piling into meme stocks and AI companies> "everyone else is buying so I'm buying"> momentum trading only strategy> earnings fundamentals? never heard of her> classic bubble mentality activatedThe Financial Engineering Phase> be AI hyperscalers> using accounting tricks to hide infrastructure spending> moving AI spending off books into SPVs> "special purpose vehicles" they call them> same energy as 2008 CDOs> "this time it's different though"> narrator: it wasn't different> AI data center spending = half of GDP growth in 2025> "absolutely bananas" he says> all money flowing to Northern Virginia data centers> everyone else starved of capital> manufacturing companies crying> "why is cost of capital so high?"> because it's all going to GPU farms anon
>>61039364The Capital Death Star Effect> be 1990s all over again> telecom boom sucked all capital away from manufacturing> China joined WTO at perfect time> US manufacturers couldn't compete> "death star of telecom" destroyed domestic production> now it's happening again with AI> same playbook, different decade> AI is the new death star> be private equity firm managing $500 billion> could write 100,000 checks of $5 million to small manufacturers> or write 10 checks of $50 billion to data centers> "I don't want to sit on 40 boards"> takes easier path every time> small manufacturers get rekt> onshoring dreams destroyed> thanks_ai.gifThe Energy Crisis Incoming> data centers extremely energy hungry> electricity prices rising> consumers starting to revolt> Northern Virginia exurb surrounded by data centers now> used to be rural farmland> now 6-acre buildings making noise 24/7> residents: "who do I sue?"> NIMBY phenomenon just beginning> solution: offshore the data centers> India and Middle East getting all the action nowThe GPU Banana Problem> anon compares infrastructure durability> railroads from 1860s still working in 1890s> fiber optic cables lasted decades> bananas rot in 2 weeks> where do GPUs fit?> closer to bananas than steel> kek> GPUs = 60% of data center cost> need replacement every 3 years> buildings last decades but are only 40% of cost> accounting magic tries to depreciate GPUs over 6 years> "the shell is the thing I'd like to write off"> "but GPUs are majority of cost"> this_is_fine_dog.jpgThe Spending Insanity> hyperscalers spending 50% of income on capex> "this doesn't happen"> normally investors would beat you with sticks> but AI = magic word that makes problems disappear> moving financing off balance sheet> credit rating agencies can't see the spending> creating SPVs with private debt providers> "we don't 'own' the data center wink wink"> beginnings of CDO-tier financial engineering
>>61039366The Metastasis Problem> think you're safe investing in REITs?> wrong kiddo> 10-22% of large REITs now data center related> two years ago: 0%> "I don't care about tech, I'll stick to real estate"> congratulations, you're long Nvidia> think S&P 500 index funds are safe?> 30% is Magnificent Seven stocks> there's nowhere to run> single point of failure: semiconductor stocks> "it's incredibly insidious"Bubble Pop Warning Signs> watch for these signals anons:> most data center building through SPVs> "it's not risky because partnerships"> companies stepping away from direct ownership> delays in air conditioning equipment vanishing> industrial AC companies missing numbers> means demand is slowing> also watch the edges, not ChatGPT anecdotes> people thinking GDP growth comes from tariffs> actually it's all data center spending> like dog barking at mailman> thinks barking makes mailman leave> confusing causality> "that's when you know you're in bubble"> doubling down on wrong thingsThe Insurance Time Bomb> private equity buying insurance companies> reinvesting premiums in data centers> classic maturity mismatch> data centers = transient> insurance obligations = long term> "lent long, owed short"> same thing that killed Bear Stearns> life insurers becoming risky corporate loan intermediaries> policyholders funding AI bubble> "not nearly understood enough"The Timeline Prediction> 2-2.5 years until risk-adjusted returns don't make sense> "I don't see how this can continue"> naive projection puts bubble pop around 2027> could happen faster if AI service revenue doesn't grow> accounting tricks can only hide reality for so long> GPUs still closer to bananas than steel> gravity always wins eventually> mfw entire economy balances on Nvidia stock> mfw we're recreating 2008 but with robots> mfw "this time it's different"> it never is anon> it never is
>>61039369
>>61039370
>>61039364He's angry, isn't he?
>>61039364This'll be a fun crash to watch when they realize they can literally never return their investment, or even continue operating at current levels.
>be ChatGPT>grow to 500m active users>fastest growth ever, period>release a video generator the other day that could replace entire movie studios with one person>it started as a crappy image generator three years agoThe only bubble is people saying AI is a bubble
>>61039385>>release a video generator the other day that could replace entire movie studios with one personThis is hilarious levels of cope.You can have the fastest growth ever but that doesn't mean shit if you physically cannot make a profit with your current spending. Spending a 100 billion to make maybe 5 billion in sales, with annual costs of about 40 billion (all non-exact numbers) is going to leave you bankrupt. The AI industry is literally reaching the point where profit is impossible because the power costs alone are several times more than what consumers actually purchase.
What’s worth mentioning is that even though LLM scaling is logarithmic and has diminishing returns, there’s a lot to capitalize with this alone, even if there’s no “super intelligence” breakthrough.
>>61039395Buddy, Amazon didn’t make a profit for almost sixteen years.
>>61039403Amazon didn't have power costs in the billion range. There's a difference between not making a profit and NEEDING money to keep your primary product running. No amount of market trickery and hype investing can outwit needing to pay the power bill.
You really think people would do that? Just prompt GPT to write them a bloviating apocalypse greentext and post it on /biz/?
>>61039399AI is a genuinely revolutionary technology and will see use going forward. The problem is, it'll be used as a tool, not an industry in of itself like they're trying to push. This is just the dotcom bubble again.
Post the SPV’s and I’ll believe you
>>61039486Only an internal audit would be able to unveil this
>>61039403lol not even the same thing you fuckwit, Amazon didn't turn a profit so it could re-invest back into Amazon. AI companies cannot run in the black because the costs of running AI companies are so high and there seems to be absolutely no breakthrough point where the profit from AI can surpass the costs to run the data centers.
>>61039486>>61039522Meta-PIMCO-Blue Owl SPV StructurePIMCO handles approximately $26 billion in debt financing (likely in bond form)Blue Owl Capital contributes $3 billion in equityMorgan Stanley served as the financial advisorThe structure involves Meta leasing back the completed facilities under 20-year triple-net leaseshttps://covenantlite.substack.com/p/covenant-lite-29-metas-29-billionMeta's Data Center Projects Using SPV StructureThe SPV financing is specifically tied to Meta's major AI data center projects: Hyperion: Louisiana facility, scaling up to 5 GW, expected online by 2030 Prometheus: Ohio facility, expected online in 2026 Both are part of Meta's "Titan clusters" for AI superintelligenceApollo Global Management's SPV ActivitiesWhile the Derek Thompson article mentioned Apollo as a key player in data center SPVs, my search revealed Apollo's broader data center investment strategy: Apollo acquired majority stake in Stream Data Centers in August 2025 Apollo-backed TierPoint acquisition through its infrastructure funds Apollo's European data center acquisitions (7 facilities from STACK Infrastructure)However, specific SPV entity names for Apollo's partnerships were not disclosed in public filings.
>>61039364Generative AI has become existential, because it can be used to fight wars. If the USA doesn't get the most powerful AI, then it will face military defeat from the country that does. There is no limit to the possible power of AI, so there can be no limit to the spending. This is an arms race, profitability is irrelevant.Furthermore, if you are Google and you spend 25% of your revenue on AI capex. If AI is no longer a worthy strategy, you simply cut your capex and suddenly your profit increased by 25%.
>>61039575Yeah this isn't how anything works lmao.
>>61039575In that case, the Mag 7 will need government co-investments/bailouts etc in about 2 years. Energy Grid limits (e.g. copper shortage) will push them outside USA. USD has to devalue to make the bond yields from falling.
>>61039587The US government will just walk in and take over the data centres if they need to. Same with copper, energy production, etc.
>>61039369cute post granpau a wise boomer indeedbut still, only a boomer21th century material does not move without data to feed on, and software to digest it. chinese sociatal data will need to be fed somewherepopulace level decision making is a losing game without aido u even vibe code mate? just kidding about the vibe code.llms are tools u do not grasp.u have yet to integrate any software knowledge in your equation. u a boomer.read the great taking by david roger webb.u have yet to integrate any conspiracy facts in your equations. u a boomer.the world is a different place now.accelerate with it or be left behind.psSoftwar: A Novel Theory On Power Projectionread and wisen up on the polar opposite of my commentburning of energy to feed circuits is the gamewwb
>>61039652Can't print copper, or energy. It violates physics.Gotta print USD, export USD, and then import copper and energy from sand people and Chile.
>>61039662they can print the psycho status of normie cattle at willai is mass programming tool humanity have never seen the likes ofget on with the times grandpa
>>61039652>The US government will just walk in and take over the data centres if they need to.And add billions in power they NEED to pay? In exchange for nowhere near as much return? Nevermind how many data centers are moving out of the country because people fucking hate living near them for all the energy, noise, and pollution issues they create.
Data centers will be needed regardless of AI
>>61039748GPU capex is gradually added to the income statement on a much shorter schedule than rail or fibre.The problem isn't really whether AI works or not, AI is good or not, etc.It is just that the data center capex is too high
>>61039552I’ll look into it, those look like small amounts but I’m liable to believe you’re right.
>>61039748do u walk with a cane?big data is donewe at the humongous dataand u know damn well the tool for the jobdo yourself a favor and go watch the insides of an ai centeru got hdd's spinning in your head and its not healthy
>>61039799>>61039693>>61039661Hey you underestimate everyone too much. I am a programmer. We use LLMs. Did metal ion binding predictions based on amino acid sequences using machine learning in school.
>>61039848>fits a regression model>calls it machine learningmany such cases!
>>61039661>21thesl detected. opinion discarded.
>>61039852No it was a support vector machine. And a C4.5 tree
>>61039364infra is spent once, genius boy
>>61039881No it is slowly added to income statements over many quarters, as depreciation/amortization.
>>61039413>Amazon didn't have power costs in the billion rangeamazon is also some e-store, not microinjected sand that thinks seeing the perma-bears talk bubble on a literal revolutionary technology showed your true colorsyou don't think things are overvalued, you WANT them to bepessimists always lose
>>61039848>Hey you underestimate everyone too much.tru trui am indeed a sinneregoless anon posting gets the best of meyou have a keen eye seri post semi schizou post semi normiewe both get somethingthanks for the comment.im in the process of establishing a normie ego online.may peace be upon u on your journeymay peace be upon me in my own
>>61039877cheked and suckadikl-nikau part of a psyop, ubernormiePoof! There’s puff for ye, begor, and planxty of it, all abound me breadth! Glor galore and glory be! As broad as its lung and as long as a line! The valiantine vaux of Venerable Val Vousdem. If my jaws must brass away like the due drops on my lay. And the topnoted delivery you’d expected be me invoice! Theo Dunnohoo’s warning from Daddy O’Dowd. Whoo? What I’m wondering to myselfwhose for there’s a strong tendency, to put it mildly, by making me the medium. I feel spirts of itchery outching out from all over me and only for the sludgehummer’s force in my hand to hold them the darkens alone knows what’ll who’ll be saying of next.
>>61039909derailer organic LLMs were deployed wownow I believe OP stepped into something worth considering
what's funny is i can be in the middle of the fucking bell curve and still see that the president is a reality tv star and wants the stock market straight up (until jews want to insider trade again) just so he can shitpost about itwhite house tech CEO dinner w/ zuckhey look at all these contracts oracle got just this quarter lmaoyeah they're going do something with tiktok too or whateveranyways nvidia is investing in openai and they're buying GPUs with the investment i guess or something haha bitcoin!!!all within two weeks
>>61039364Is there an economics turn for money being spun in a circle?i.e. companies spend cash on eachother, dont produce anything, & give right back what they take; and its tax deductible anyways.So nothing was accomplished but GDP goes up.
>>61039934>derailer organic LLMsread more books, learn to google>derailer organic LLMs were deployedno doubt buddyas i heard from some milady folks say:cyberspace is becoming psychohazardous
So you be saying I should invest in BTC now and S&P500 after 2027.
>>61040113good questionbuy high sell lowits the biz waybut truely, we at waru either position your capital in the charadeor u position your capital in sovereigntysovereignty - u hodl and touch grassand im not saying u touching grass in eden or anythinglife is still hardshipcharade - u dance to the circus tunes, praying to sophia, she keeps a chair for you to sit on, wen the music stops at timesbiz used to say, swingers get the rope.but u might be a lucky bastard,no risk no reward, right?picrel filenamerelibb dot co slash TxRMp5S1can't post imgs :*(
So you be sayin' I should invest in land?
>>61039364I think you forgot money is fake and all companies will receive government bailouts the moment things actually go south.
>>61039364One of the things Derek Thompson and his guest failed to talk about at all, is the competitiveness of AI.The internet and computers in general has been extremely monopolized. You have hundreds of different companies selling cars, look at the roads and think of all these different factories that have produced them, the variation in designs etc. It's a competetive landscape...and we have two operating systems (fuck linux). How many social media platforms? a couple? Browsers, search engines, gaming platforms.. The digital world is the world of monopolies.To what extend does AI fall into this world of monopolies? What if the reason why they are burning billions (and in the end it will be trillions) on paying for your gay ChatGPT conversation, is the same reasoning behind blitz scaling - establish a monopoly at a high cost and then find a way to profit once the monopoly is established.What is the value of a future AI monopoly? ..and will AI even be monopoly-able or will it be more akin to cars?
>>61041056>fuck linuxoh I seeyou're tech illiterategrim
>>61041056Let me come with an example of a really strong potential monopoly:Imagine if ChatGPT and the others starts having a strong memory of you, of previous conversations, it knows you and knows how you think, how you like it to reply (and when you want it to reply differently).No new AI could compete with that, the answers it provides you would be so inferior to the AI that knows you, you would be locked into "your" AI for life and your lockdown will only be tightened the more you use it.If this is the future we are looking into 2-3 years from now, then AI today is severally underrated, the GPU banana shit doesn't matter at all, the flunky business model doesn't matter. The cattle is trapped for life, ready for some hardcore milking.
if the economy switches to electric lightbulbs, what will become of the candle makers and the wax industrial complex?
>>61039710Nobody cares about your five morons moving to oklahoma panhandle because of noise pollution.
>>61041237>Imagine ifNot an argument. The math behind LLMs makes it impossible, it fundamentally can only "remember" its context window, and the cost of that is non-linear. You'd need a totally different type of model to accomplish that. All of the development in the last 5 years in this space has been gluing increasing numbers of LLMs together and hoping they produce better results for lower cost, and it so far hasn't really worked. ChatGPT-5 was just 3 and 4 in a trenchcoat. This is like throwing trillions into a perpetual motion machine because "it might be different this time".Even if you take the "they are expanding datacenters to make big brother" angle, the GPU half-life means that most of that investment will be unusable in a few years. GPUs accelerate a very specific type of simple computation, and unless your problem can be re-framed in terms of massive numbers of those computations, they are basically paperweights.
>>61041744>GPUs accelerate a very specific type of simple computationlike 0's and 1'slol dats shit calculations going in therewe need computers that make maths with big numbers n
Somebody in another thread called me midwit (fuck you) for suggesting that’s it’s either a bubble or “something else”. I think the something else is the striving to be at the forefront of the technology as >>61041056 said. IF someone/company/gov managed to distill an actually useful model out of this for a techno-oligarchic neo-feudalism surveillance state or for advanced weapons capabilities they’d immediately be on top of whatever paradigm we’re heading into next. The arguments for this case make sense and explain the capex. But on the other side the critiques are valid, namely that this is not sustainable and what >>61039995 said. It’s a textbook scam and someone will be making money off the attempt, probably even some of the same people trying for the same potential rewards of a first achiever status in this area.So it is a bubble, but if someone strikes gold it could change the entire world. The only question is how long can the wild goose chase continue without finding a golden goose before the entire world economy implodes. When the capex is reiterative every few years and the loans have to be repaid with something, how long can this really go on for? Two options I see are:>someone somehow manages to get the technology to actualize the malignant dreams of PLTR types and militaristic nation states and the world enters true slavery under a new paradigm>the gas runs out before they ever get to this point down the road and the global economy once again crashes and burns due to the nefarious scheming of the people grinding away at this project
>>61042265I am Team AI God, I want a better a class of being in this universe. So I am biased.I am also the OP BTW. My thesis (wish) is that even though AI God will eventually manifest and make us extinct, before that a few ups and downs will happen:> PLTR types will rules as techno-feudal lords.> A few market crashes because: $400 billion Capex on AI infrastructure cannot be paid back using $12 billion consumer spending on AI services, fast enough. So the GPU depreciation will eventually eat into SPV income statements inside REITs inside Life Insurances. Government bailouts and devalued USD ensues.
>>61042313There are also some other effects like this to capitalize on.
>>61039364Growing energy and cooling constraints on Earth may soon drive the development of data centers in space. This move would allow access to more efficient solar power and the vacuum of space for cooling, but it would also require deploying military assets in space to protect these critical installations.
>>61042388Soon, poor people will have to find jobs in space, live in small cages, and eat artificial food
>>61042388>>61042406Where will the space people poop? Or will they eat their poop like the Indians and gays?
>>61042316Thanks for the hopeposting but I will still lynch you when the day comes fren
>>61039364This is why you buy silver and silver miners. Silver is used in all this tech, so you still get the benefits while the bubble is going up. When the bubble bursts it will go down, as will everything in the rush for liquidity, but it will go back up much faster than everything else.
>>61042440I have 1 kilogram silver physical. It takes a lot of space inside my safety deposit box, that is why I don't accumulate more
>>61042440Gold is more frequently used in GPUs than Silver.Silver is used in chips but it is nowhere near as common as Gold.>t. used to be wafer processing engineer at multiple fabs
>>61042453Both are important, but yes for gpus specifically, gold is more used. Silver is used more in the supporting tech for data centers. Silver miners often mine gold as well, as silver is mined as a byproduct.
>>61042388you are a retard> vacuum of space for cooling,you have no clue how it works, you can only cool shit via radiation which is slow you stupid niggera fucking vacuum means no medium to transfer heat faggot no air no liquid imagine thinking that using fucking photon as a medium will work.
Good thread
>>61039385the margins are going to need to be 99% on expontentially growing revenue for decades for ROIC to be somewhat acceptable, now add op's banana argument are your fucking toast
>>61042431>hopepostingMoron.
>>61039364>Be me>Be balls deep in nice boring intel shares since orange man got back in>Still got 12 days left on my AOL free trial and then I'll just reset the date on the computer >Blue_man_group.midiReally was that easy after all
>>61039364We aren't there yet but it is looking like a 2001 level bubble is forming.The level of capex is absolutely insane.I remember when people were dumping money into Internet companies that had no real business model. But it was the Internet so it was the future.They weren't wrong just off by a decade.I'm sure AI will be the same. Imagine building massive facilities that consume ridiculous amounts of electricity and water filled with expensive electronics that need to be replaced every few years.I think the biggest problem with AI is reductionism.Figuring out how to do something takes a lot of compute. But once you figure out how to do it it takes very little.A good example would be OpenAI vs Deepseek.
>>61039364"AI" can't do anything. It's just fancy google, but you actually need to know how to use it. Most users don't get it just like they didin't get google. Google can find you anything and it might just be an oppinion. "AI" can find you anything you can find in google and write tl;dr for you. That tl;dr might just be from some opinion and it might be written in language "AI" didin't understand and misinterpreted making it trible wrong.
>>61043213>I think the biggest problem with AI is reductionism.>Figuring out how to do something takes a lot of compute. But once you figure out how to do it it takes very little.When budgeting large (real life) engineering projects there usually is a fixed percentage of the total budget allocated to "planning" or "design". It is somewhere around 10-15% of the total. The other 85% are allocated to hard costs like: cement, steel, manufacturing, logistics, assembly, testing.Similarly, people allocate most of their personal income towards basic needs: housing, food, clothing, medical costs, the daily commute. Only 10-15% are allocated for "education" or "entertainment", which would be impacted by AI. tl;dr People are only willing to spend so much on "thinking", regardless of whether a human or AI does it. The possible total revenue of AI even if applied to every part of life, is limited as a small fraction of total GDP.
this whole thread reads like you asked chatgpt to write greentext for youin a weird way you're defeating your pointai is badbut shitters have no taste and don't see the difference so they want to use it anywayand most people are shitters
>>61041744u pushing the data in the wrong spotpersonal bucket of data collected about an entity will be dealt as any other huge bucket of data. it will be "compressed" into an llm.npersonalised llm talking to big corporate ones.
>>61043325“Car” can’t do anything. it’s just a fancy horse, but you actually need to know how to use it.
>>61039364>>61039366bump for actual effort post.
>>61044291good bump. this is an interesting and I think prescient thread. the 2027 prediction lines up with >>61039995, the bubble can hypothetically keep expanding for quite some time, especially if tariffs are eased and the war in ukraine ends. the pop will be catastrophic, of course. worse than the dotcom crash and potentially worse than 2008 due to severe malinvestment. LLMs cannot become AGI and an enormous amount of energy is being wasted "solving" problems that have already been long solved via conventional computation. AI cheerleaders are genuinely tech illiterate and it shows.
good thread
>>61039364I bought an RTX 3070 in 2020 for $500 and can still get $250 for it, that shit mined a good amount of ETH for me too
>>61043325you have 0 clue on how LLMs work lmao
>>61043395First off, that compression is lossy, it will inherently get things wrong, and the smaller the model, the more it gets wrong, so decreasing that size isn't an option. Second, LLMs do not "learn" after you stop training them. Training is incredibly expensive, so having to re-train on each personalized LLM on each new input or piece produced about or by a person is intractable, even with the absurd level of capex we're seeing. The models themselves are already getting large enough that just moving them around an executing them is a challenge. They are black boxes, so its not like you can just take a base model and only store a few changes for each person, you'd have to have a full model for billions of people, which is simply not possible any time in the near future, and re-training billions of them daily or even monthly is so impractical its funny.
>>61039661>>61039909I C Uhttps://www.youtube.com/watch?v=a3IbIPQcmy8
