The Dirty AI lie : How the GREATEST bet in human history started to crack in June 2026?
Executive Summary
Is AI the greatest business bet in human history or the greatest bubble ever inflated? The numbers in this case study: Big Tech will spend $725 billion on AI infrastructure in 2026 (8x 2020), consuming 94% of operating cash flows, while every major AI model company combined earns about $75 billion and loses at least $17 billion - JPMorgan says the spending only makes sense if AI earns $650 billion a year. Meanwhile the cost is landing on you: memory makers get 10x more per module from AI data centers, so DRAM is up 171% and Apple raised prices mid-year for the first time ever. The bull case cracked in June 2026 when enterprises like Lindy and Uber revolted over token costs and Palantir's CEO called the models 'irresponsibly oversold'. The historical mirror is telecom 1996-2001: the technology was real, the internet did change everything, and the bubble still burst - 97% of the fiber sat unused while $2 trillion vanished, then demand arrived and the survivors owned the backbone of the modern internet. The honest verdict: very high possibility of a bubble, not a certainty. FULL COVERAGE: captions appear complete. Sponsor: Odoo. Not investment advice.
Key Takeaways
The $650 billion question
JPMorgan's math: AI must generate $650 billion every year to give investors a bare 10% return on current spending. Reality: OpenAI (~$25B revenue, losing $14B), Anthropic ($26B target), Gemini (~$25B) add up to roughly $75 billion - while $725 billion goes out the door. Sequoia's David Cahn calls the gap the $600 billion question: a revenue deficit nobody knows who will fill.
94% of cash flow, one assumption
PIMCO reports Big Tech capex will consume 94% of operating cash flows over the next two years (up from 40% in 2023) - leaving $6 of every $100 for dividends, buybacks, salaries and everything else, all bet on one assumption: that in five years, today's data centers will look like a bargain.
You are already paying the AI tax
Samsung, SK Hynix and Micron - 90% of world memory supply - shifted 93% of production to high-bandwidth AI memory because data centers pay 10x more per module. Result: DRAM up 171% year-on-year, PC memory contracts up 105-110% in a quarter, 1GB of DRAM from $0.43 to $2.39 in six months, and Apple's first mid-year price hike in history.
The enterprise revolt of June 2026
The model assumed enterprises would pay ever more for tokens forever. Then Lindy (25 employees) moved all traffic to DeepSeek and cut costs 90%; Uber's CTO admitted the annual AI budget was blown in four months; Palantir CEO Alex Karp said enterprises feel they are paying for tokens that create no value - and McKinsey, BCG and MIT surveys show 73-95% of enterprise AI deployments failing to hit projected ROI.
Telecom 1996 is the mirror
After the Telecommunications Act, companies poured $500 billion into fiber on the story that traffic would double every three months. Traffic 'only' grew 100% a year - not enough. By the early 2000s as little as 2.7% of installed fiber carried data; bandwidth prices fell 90%; WorldCom and Global Crossing collapsed; $2 trillion of market value evaporated.
The twist: the wreckage became the backbone
The fiber did not vanish. When YouTube, streaming, cloud and smartphones arrived, demand finally caught up, and survivors bought the wreckage for pennies - the same overbuilt cable became the physical backbone of Google, Netflix and AWS. Technology real, companies dead, infrastructure triumphant.
Why it might not be 1999
The honest caveats: telecom bubble companies were debt-funded and profitless, while Nvidia earned $120 billion net income last year and Microsoft, Google and Amazon are the most profitable enterprises in history; the NASDAQ 100 forward P/E is around 26x, not the 60x of the dot-com peak. The host's verdict: very high possibility of a bubble, not a certainty.
The Conversation Breakdown
Apple's first-ever mid-year price hike
⏱ Untimed topicMacBook Air up 18%, iPad Pro up 20%, Apple TV up 54% - and the reason is a war over tiny memory chips thousands of kilometres away.
The capex explosion
⏱ Untimed topic$90 billion (2020) to $725 billion (2026) from just four companies, Nvidia's $5 trillion peak and the $320 billion single-week crack three days later.
What a data center actually is
⏱ Untimed topicA warehouse of 100,000 GPUs at $30-40,000 each, $3-4 billion of chips per building, $10-25 billion all-in - and 221 zettabytes of data by 2026 as the demand story.
The arithmetic that doesn't add up
⏱ Untimed topic$650 billion needed, $75 billion earned, $17 billion lost, $725 billion spent - plus the McKinsey/BCG/MIT ROI failure surveys and the Lindy/Uber/Palantir enterprise revolt.
How you pay the AI tax
⏱ Untimed topicThe memory-chip triopoly's shift to AI memory, DRAM up 171%, Dell's 5.5x price jump per GB, and Apple admitting it can no longer shield customers.
The capital cycle and the telecom mirror
⏱ Untimed topicFour steps of every bubble; Global Crossing's $47 billion profitless valuation, Corvis's zero-revenue IPO, 2.7% fiber utilization, WorldCom's record bankruptcy - and how the wreckage later powered Google, Netflix and AWS.
Two paths from here
⏱ Untimed topicPop: NASDAQ crashes and India's IT hiring freezes. No pop: token prices shoot up and cheap AI tools become luxuries. The miracle third path: token costs collapse - 'a teeny tiny possibility'.
Core Frameworks & Ideas
The Capital Cycle
The four-step pattern behind every infrastructure bubble: high returns attract capital; capital flows until overcapacity is built; overcapacity collapses returns; everyone dies except a few survivors who make a fortune when demand catches up. Railway 1846, fiber 2000, data centers 2026.
The 10% Justification Test
JPMorgan's gap analysis: any risky investment must promise at least a 10% return - so divide required revenue by actual revenue. For AI: $650B needed vs $75B actual, a 9-10x shortfall that no survey of enterprise ROI currently supports.
Technology Real, Price Wrong
The internet did change everything and the bubble still burst - demand was exploding, just not enough to justify overcapacity. The honest question is never 'is the technology real?' but 'does the price for it make sense?'
Actionable Next Steps
- ✓Before buying any AI-linked stock or fund, run the justification test yourself: what revenue does the current capex require, and who pays it? If the answer is 'someone, eventually', size the position accordingly.
- ✓If your company buys AI tokens, audit the bill quarterly against measurable ROI - Uber blew its annual budget in four months; Lindy cut costs 90% by switching models. The cheapest capable model wins.
- ✓Expect device inflation to persist while data centers outbid you for memory: if a laptop, phone or console purchase is coming, buying sooner may be cheaper.
- ✓Plan your career for both paths: if the bubble pops, IT services hiring freezes; if it doesn't, AI tool costs rise - either way, the durable skill is applying AI to a real business outcome, not the tools themselves.
Notable Quotes
“The problem with the AI capex boom is not only is it immense but a big chunk of it is funded with debt and that pain doesn't stay restricted. It spills over into the rest of society.”
“Every single enterprise I deal with, they're like, I am paying for tokens that create no value.”
“if anyone tells you for certain that this is a bubble, they're lying to you. And anyone tells you that it is definitely not a bubble is also lying to you”
The One Thing to Remember
Every infrastructure bubble follows the capital cycle: high returns attract capital, capital builds overcapacity, overcapacity collapses returns, and only the survivors get rich when demand finally catches up - the question is never whether the technology is real, but whether the price makes sense.