JP Morgan izbacio izveštaj s prognozom vrtoglavih potrebnih investicija u AI do 2030.

https://x.com/x_delcourt/status/1988266379208986811

Neki tizeri:

Here is the killer lead quote from the JPMorgan report, written by a horde of analysts from the bank’s investment grade credit, high-yield bonds, industrials, semis, utilities, and asset-backed securities team:

The global data center and AI build-out will be an extraordinary and sustained capital markets event. Building out global data center and AI infrastructure and related power supplies could cost over $5 trillion in our view, with at least one consultant calling for $7 trillion of global AI-related capex.

Funding that extraordinary growth will likely require participation from every public capital market as well as private credit, alternative capital providers and even government involvement. The question is not “which market will finance the AI-boom”. Rather, the question is “how will financings be structured to access every capital market.”

The sheer scale of those investments has critical implications across the credit market landscape. Big picture, for various reasons public bond and syndicated loan market growth has slowed post-COVID recovery (shrunk in the case of High Yield Bonds). AI/Data Centers are likely to drive a re-acceleration of market growth. (...)

The scale of demand for compute remains astronomical, with actual growth somewhat constrained by physical limitations. Our base case estimates call for 122 GW of global data center infrastructure capacity installations from 2026-2030, at a rapidly accelerating rate. (...)

Power is the most important of those constraints. Current lead times for new natural gas turbines have ballooned to 3-4 years, and nuclear plants have historically taken 10+ years to build. Adding 150 GW of power in a timely manner is a remarkable challenge, particularly in light of grid upgrade requirements. (...)

Annual data center funding needs in 2026 are on the order of $700 billion, which could be entirely financed by hyperscaler cash flow and High Grade bond markets. However, 2030 funding needs are in excess of $1.4 trillion, which will likely require funding contributions from all capital providing markets. (...)

We think High Grade markets could absorb $300 billion of AI/data center related paper over the next year, and are assuming $1.5 trillion of funding from High Grade bond markets over the next five years. (...)

Mathematically, if anything like our forecast plays out, AI/Data Center related sectors could represent north of 20% of the market by 2030.

Big picture, to drive a 10% return on our modeled AI investments through 2030 would require ~$650 billion of annual revenue into perpetuity, which is an astonishingly large number. But for context, that equates to 58bp of global GDP, or $34.72/month from every current iPhone user, or $180/month from every Netflix subscriber. How that is apportioned between corporations, governments and consumers is, of course, a long-term debate. Regardless, even if everything works, there will be (continued) spectacular winners, and probably some equally spectacular losers as well given the amount of capital involved and winner takes all nature of portions of the AI ecosystem.

These are all extracts from JPMorgan’s synopsis. The full 52-page report is definitely worth reading if you have access. As the bank’s analysts note: “It is hard to imagine the world deploying $5 trillion of capital without at least some hiccups.”

bp = basis point (= 0.01%)

Ceo FT-ov članak (iza paywalla, još nema funkcionalne kopije na archive.is):

https://www.ft.com/content/59133fa3-3071-47b4-8761-c6922d07c34e

Edited November 12, 2025Nov 12 by vememah

AI bubble about to pop as returns on investment fall short?

Nik Martin

November 10, 2025

Billions have poured into AI, helping stock valuations soar. But the cracks are starting to show. Slowing adoption, surging costs and elusive profits are fueling warnings that the boom may be headed for a hard reset.

A computer security server room in an unknown location

The artificial intelligence (AI) party is still in full swing, with tens of billions globally pouring into infrastructure, startups and attracting the best talent.

Among the headline announcements this year: ChatGPT parent company Open AI, Softbank and Oracle pledged to invest $500 billion (€433 billion) in AI supercomputers, Open AI and chip giant Nvidia announced a $100 billion fund to maintain the United States' dominance in advanced chips, while Chinese tech giants Alibaba and Tencent hiked investments to help speed up China's ambition to lead AI by 2030.

Since ChatGPT’s debut in November 2022, AI-related stocks have added an estimated $17.5 trillion in market value, according to Bloomberg Intelligence, driving around 75% of the S&P 500’s gains and propelling companies like Nvidia and Microsoft to record-breaking valuations.

Corporations are hesitant over AI adoption

But signs of a hangover are getting harder to ignore. AI usage by corporations is slipping, spending is tightening and the machine learning hype has massively outpaced the profits.

Many economists think usage concerns, barely three years into AI going mainstream, dropkick the prevailing narrative that AI would revolutionize how businesses operate by streamlining repetitive tasks and improving forecasting.

"The vast bet on AI infrastructure assumes surging usage, yet multiple US surveys show adoption has actually declined since the summer," Carl-Benedikt Frey, professor of AI & work at the UK's University of Oxford, told DW. "Unless new, durable use cases emerge quickly, something will give — and the bubble could burst."

The US Census Bureau, which surveys 1.2 million US companies every fortnight, found that AI-tool usage at firms with more than 250 employees dropped from nearly 14% in June to under 12% in August.

AI’s biggest challenge remains its tendency to hallucinate — generating plausible but false information. Other weaknesses are inconsistent reliability and the poor performance of autonomous agents, which complete tasks successfully only about a third of the time.

"Unlike an intern who learns on the job, today’s pretrained [AI] systems don’t improve through experience. We need continual learning and models that adapt to changing circumstances," said Frey.

Unsustainable capital burn

As the gap widens between sky-high expectations and commercial reality, investor enthusiasm for AI is starting to fade.

In the third quarter of the year, venture-capital deals with private AI firms dropped by 22% quarter on quarter to 1,295, although funding levels remained above $45 billion for the fourth consecutive quarter, market intelligence firm CB Insights wrote last month.

"What perturbs me is the scale of the money being invested compared to the amount of revenue flowing from AI," economist Stuart Mills, a senior fellow at the London School of Economics, told DW.

Market leader OpenAI, which is backed by Microsoft, generated $3.7 billion in revenue last year, versus total operating expenses of $8-9 billion. The company says it is on course to make $13 billion this year but is still expected to burn through $129 billion before 2029, news site The Information calculated in September.

Mills thinks generative AI companies like Elon Musk's Grok and ChatGPT are "charging far less than they need to make a profit" and should raise subscription prices.

Few have quantified the AI bubble more starkly than Julien Garran, partner at UK-based research firm MacroStrategy Partnership. He argues that the sheer volume of capital flowing into AI — despite little evidence of sustainable returns — dwarfs previous speculative frenzies.

"We estimate a misallocation of capital equivalent to 65% of US GDP — four times bigger than the housing buildup before the 2008/9 financial crisis and 17 times bigger than the dot-com bust," Garran told DW.

Investors increasingly cautious

Recent earnings from Big Tech have sparked cautious optimism, but also fresh doubts about AI’s staying power. Data analytics and intelligence platform Palantir's Q3 revenue surged 63% year-over-year, but its stock price fell by up to 7% on the news. AMD and Meta also saw their strong AI-related earnings overshadowed by market concerns about sustainability.

That disconnect between soaring valuations and shaky fundamentals is exactly what worries Mills, who sees a widening gap between what AI promises and what it actually delivers to businesses.

"The data suggests that AI is not penetrating high enough up the value chain. Loads of people are using it, but it's not being used for tasks that directly contribute to value production," he told DW.

Nvidia's upcoming earnings on November 19 may prove a key test of whether the AI boom still has legs. In the second quarter, Nvidia's data center sales alone made up 88% of total revenue, which hit a record $46.7 billion. For Q3, the company has guided $54 billion, projecting 54% year-on-year growth, which would equate to a full-year total of more than $200 billion.

When will the bubble pop?

"With the exception of Nvidia, which is selling shovels in a gold rush, most generative AI companies are both wildly overvalued and wildly overhyped," Gary Marcus, Emeritus Professor of Psychology and Neural Science at New York University, told DW. "My guess is that it will all fall apart, possibly soon. The fundamentals, technical and economic, make no sense."

Garran, meanwhile, believes the era of rapid progress in large language models (LLMs) is drawing to a close, not because of technical limits, but because the economics no longer stack up.

"They [AI platforms] have already hit the wall," Garran said, adding that the cost of training new models is "skyrocketing, and the improvements aren’t much better."

Striking a more positive tone, Sarah Hoffman, director of AI Thought Leadership at the New York-based market intelligence firm AlphaSense, predicted a "market correction" in AI, rather than a "cataclysmic 'bubble bursting.'"

After an extended period of extraordinary hype, enterprise investment in AI will become far more discerning, Hoffmann told DW in an emailed statement, with the focus "shifting from big promises to clear proof of impact."

"More companies will begin formally tracking AI ROI [return on investment] to ensure projects deliver measurable returns," she added.

https://www.dw.com/en/ai-bubble-about-to-pop-as-returns-on-investment-fall-short-chatgpt-microsoft-nvidia-grok-elon-musk/a-74636881

AI’s biggest challenge remains its tendency to hallucinate — generating plausible but false information. Other weaknesses are inconsistent reliability and the poor performance of autonomous agents, which complete tasks successfully only about a third of the time.

1. Daje uverljive ali netačne informacije

2. Nedovoljno je pouzdan

3. Ima 66% šanse da zajebe

Bukvalno tri različita načina da se kaže ista stvar Samo mistifikuju činjenicu da je svima zinulo dupe na automatizaciju pa su uložili trilione na tehnologiju koja bukvalno ne postoji.

Može i ovde vest da je japanski SoftBank prodao sve svoje akcije NVIDIA za 5.8 milijardi dolara. Ali ne brinite ništa, planiraju da ulože još 22.5 milijardi u OpenAI

Kada prodje hajp, bude jasno neisplativo i nepouzdano - a bude tu, prisutno - bice to prava sajberpank distopija.

To je budalastina po definiciji.

Koliko uopste ljudi na ovom svetu uci da bi im to bilo isplativo? Proces ucenja je malo kompleksnija rabota od "hajde da naucimo ovo da bi milioneri zaradili vise para".

Treba da napravis masinu koja ce da rezonuje i razmislja na jos visem nivou nego sto mi to mozemo a treba da joj promaknu sve te brojke iznad i da ne bude u fazonu - jbt koliko se energije i para na dmevnom nivou sprca da ja radim, daj da se samougasim.

5 hours ago, vememah said:

Ceo FT-ov članak (iza paywalla, još nema funkcionalne kopije na archive.is):

https://www.ft.com/content/59133fa3-3071-47b4-8761-c6922d07c34e

U međuvremenu se pojavila:

https://archive.is/4eqHn

Edited November 12, 2025Nov 12 by vememah

1 hour ago, Downforce said:

To je budalastina po definiciji.

Koliko uopste ljudi na ovom svetu uci da bi im to bilo isplativo? Proces ucenja je malo kompleksnija rabota od "hajde da naucimo ovo da bi milioneri zaradili vise para".

Treba da napravis masinu koja ce da rezonuje i razmislja na jos visem nivou nego sto mi to mozemo a treba da joj promaknu sve te brojke iznad i da ne bude u fazonu - jbt koliko se energije i para na dmevnom nivou sprca da ja radim, daj da se samougasim.

Počeli su da kruže i skrinšotovi iz samog izveštaja:

https://x.com/pachos/status/1988324084921594062

Slike:

Edited November 12, 2025Nov 12 by vememah

3 hours ago, Downforce said:

Treba da napravis masinu koja ce da rezonuje i razmislja na jos visem nivou nego sto mi to mozemo a treba da joj promaknu sve te brojke iznad i da ne bude u fazonu - jbt koliko se energije i para na dmevnom nivou sprca da ja radim, daj da se samougasim.

Mašina koja rezonuje i razmišlja je Sveti Gral, tzv. Artificial General Intelligence. To će možda postojati a možda i ne, u nekoj dalekoj budućnosti.

Ovo što sad zovu AI tj. LLM nema veze s tim, ovo na osnovu sofisticiranih algoritama nabada otprilike šta mi želimo da čujemo pa sklapa rečenice, bez ikakvog razmišljanja. Ekstremno je neefikasno, svaki model troši sve više energije da bi se utrenirao a daje tek marginalno bolje rezultate.

Pare koje se sipaju u tu industriju su suštinski nepotrebne, svi ti data centri koji se otvaraju biće potrebni samo ako se optimistična predviđanja nekako ostvare pa "AI" bude implementiran svugde, u svakom aspektu ljudskih života.

Ali niko ne može ni da objasni kako bi to izgledalo, a kamoli kako će se do te tačke doći. Use cases su ekstremno limitirani, ROI nema ni od korova, ali pare se još uvek nemilice sipaju.

Sve 5 i meni je skroz sumanuta prica u ovom LLM konceptu. Nemam sta da dodam.

7 hours ago, Weenie Pooh said:

Ovo što sad zovu AI tj. LLM nema veze s tim, ovo na osnovu sofisticiranih algoritama nabada otprilike šta mi želimo da čujemo pa sklapa rečenice, bez ikakvog razmišljanja.

Iz dosadašnjeg korišćenja nisam imao takav utisak, da se "otprilike nabada što želim da čujem", plus kako bi uopšte LLM znao šta želim da čujem ako sam postavio neko uopšteno pitanje?

8 minutes ago, Time Crisis said:

Iz dosadašnjeg korišćenja nisam imao takav utisak, da se "otprilike nabada što želim da čujem", plus kako bi uopšte LLM znao šta želim da čujem ako sam postavio neko uopšteno pitanje?

Iz training data, naravno, plus iz RT pretraga ako te opcije koristiš.

LLM tako funkcionišu, tako formiraju rečenice i izjave - uprošćeno rečeno, pogledaju šta su drugi ljudi odgovarali na takva i slična pitanja, a na osnovu toga svakoj reči, rečenici, izjavi pripišu određenu verovatnoću - 77% šanse da treba odgovoriti "da", 23% da treba odgovoriti "ne", i vozi dalje, sledeći proračun.

Sve se to dešava izuzetno brzo, izuzetno uverljivo, ali i izuzetno neefikasno, potrošnja energije je enormna - ljudski mozak nikad ne bi mogao tako da funkcioniše, zato se mi i oslanjamo na heuristiku koju ne razumemo i ne umemo da simuliramo.

LLM zahteva konstantnu proveru toga sto tvrdi. Ume bas dobro da ide niz dlaku i ako mu ne protivrecis oladice te na keca nekom tvrdnjom. Neretko mi se desavalo da izmasta neke podatke i istrazivanja. Isto tako neretko zna da bude potpuno u pravu sto se potvrdi dubljom proverom.

Ali ozbiljan oprez i konstantna provera toga sto tvrdi je neophodna.

S tim u vezi, ne vidim kako investiranje u dalji razvoj ovoga ikako pomaze. Algoritam je algoritam. Njemu treba konstantna povratna sprega koja proverava stvari koje pronadje kao podatke i tvrdnja. To je ono sto mu nedostaje u algoritmu, reality feedback.

Takodje mu cesto nedostaje kontekst i ne moze kreirati ideje. Nema saznanje o tome sta je zapravo nesto sto bi trebalo preporuciti a sto je vrednovano na odredjeni nacin koji do sada najbolje opisuje odredjenu pojavu

Algoritam je takav da ce uvek odgovoriti na najbazicniji moguci nacin pa ako mu trazis ici dublje i dublje u problematiku...

Smatram ga jako dobrom alatkom ali ne vidim ga kao nesto sto opravdava tolika ulaganja i sto obecava ogromne profite. To su prosto izmastane price po meni.

Btw Wini, ne slazem se sa tobom da ne mozemo simulirati nacin na koji mi razmisljamo. Samo je problem sto bi takav model bio kao malo dete koje moras da ucis svemu i koje bi moralo razviti empatiju prema ljudima da bi nam na kraju koristilo. Potrebno je jako puno ljudskosti i truda da bi takva AI profunkcionisala. I pri tom sto je najvaznije - tu ne bi bilo nikakvog profita. Takav AI obesmisljava rad i novac kao takav.

8 hours ago, Weenie Pooh said:

Mašina koja rezonuje i razmišlja je Sveti Gral, tzv. Artificial General Intelligence. To će možda postojati a možda i ne, u nekoj dalekoj budućnosti.

to je vec prevazidjeno

Forget artificial intelligence. Heck, forget artificial general intelligence. Those are passé.

Now Silicon Valley is looking toward superintelligence. OpenAI (OPAI.PVT), Microsoft (MSFT), Anthropic (ANTH.PVT), Meta (META), and a slew of other AI companies are looking to a future where wildly intelligent AI can perform tasks better than most people.

Anthropic CEO Dario Amodei has written about AI that's smarter than a Nobel Prize winner across a number of fields and can "prove unsolved mathematical theorems, write extremely good novels, write difficult codebases from scratch, etc."

Meta CEO Mark Zuckerberg has spoken about developing personal superintelligence that "has the potential to begin a new era of personal empowerment where people will have greater agency to improve the world in the directions they choose."

And just last week, Microsoft AI CEO Mustafa Suleyman announced the formation of the group's MAI Superintelligence team.

"This year it feels like everyone in AI is talking about the dawn of superintelligence," Suleyman wrote. "Such a system will have an open-ended ability of 'learning to learn,' the ultimate meta skill."

Silicon Valley is going all in on 'superintelligent' AI, and there's plenty of hype

6 minutes ago, Downforce said:

LLM zahteva konstantnu proveru toga sto tvrdi.

...

Smatram ga jako dobrom alatkom ali ne vidim ga kao nesto sto opravdava tolika ulaganja i sto obecava ogromne profite. To su prosto izmastane price po meni.

Uopste nisu izmastane, samo to dosta ljudi gleda kroz prizmu pojedinacnog koristenja Chat GPT da bi se rijesio neki pojedinacni task. No, kad se dobar LLM primjeni na sistemke podatke u megafirmama. pa su jos ti podaci srednjeni po datotekam - tu itekako ima ogroman potencijal. Zato se uostalom i otpustaju silni ljudi.

Doduse, tu je problem je sto je to winner takes all stvar i sto rece Zuckenberg veci rizik za firme je necinjenje nego trosenje novca koji bi inace otisao na stock buyback. No, problem ce biti kad se firme koje to ne mogu da isprate iz vlastitog cash flowa pocnu zaduzivati, kao recimo Oracle.

6 minutes ago, Downforce said:

Btw Wini, ne slazem se sa tobom da ne mozemo simulirati nacin na koji mi razmisljamo.

Pogledaj ovo:

The Thinking Game (2024) - IMDb