Is A.I. A Massive Lie? | Ed Zitron
Zitron questions whether the economics of large language models can support the more than $1 trillion invested in AI, pointing to subsidized usage, escalating compute costs, limited productivity gains, and persistent reliability problems. He argues that AI’s clearest realized profits are accruing to semiconductor sellers, while broad job displacement and a path to AGI remain unestablished. He forecasts that an AI investment unwind could trigger a market decline exceeding 50%, startup failures, historic hyperscaler impairments, private-credit losses, and a probable recession or worse.
Preview · 19 more points in the app
Read the Full SummaryKey Takeaways
Zitron argues that AI economics remain difficult to evaluate because users often access subsidized models, while the apparent customer base is concentrated in software companies with roughly $800 billion in global revenue. The mismatch is illustrated by compute costs doubling about every 45 days at Chamath Palihapitiya’s company while productivity rose only about 5%.
The episode describes an AI infrastructure cycle in which hyperscaler spending supports frontier labs such as Anthropic and OpenAI, while speculative data-center investment lifts Nvidia and other semiconductor stocks. Zitron estimates that $261 billion in capex has effectively been associated with one company and says additional frontier labs would be needed to absorb the resulting capacity.
Zitron says LLMs are better suited to bounded tasks such as summarizing, searching, and surfacing information than to complex financial or operational work. His examples include a Microsoft share-price error of $570—about $70 above its historical peak—and Klarna’s reversal of AI customer-service substitution after translation errors damaged support tickets.
According to Zitron, current evidence does not establish broad AI-driven labor displacement: the clearest losses have been in already-automated contract work such as translation, transcription, and some SEO tasks. He also disputes unsupported claims about software-engineer job losses and distinguishes corporate AI-investment claims from proof of actual displacement.
Zitron argues that the strongest realized profits are accruing to semiconductor sellers rather than AI adopters, citing Zillow’s millions in AI spending alongside approximately $40 million in net income. He places that spending against a SaaS sector facing slower growth and relying partly on price increases, with Adobe, Oracle, and Salesforce cited as examples.
TECHNOLOGY
“As far as LLM themselves, it's still not clear what the business model is because the majority of people that use LLM are using them in the subsidized fashion…”—
“It means they've spent $261 billion in capex on one company…”—
“he mentioned that he talked to his chief technology officer and said, "Look, give me a sense of how AI is doing for us right now."…”—
ECONOMY
“And the media's really crapped themselves on this one. The amount of stories it's like software engineers are having trouble finding work. Maybe it's AI…”—
MARKETS
“I think actually software as a service is it has been slowing for years. They've been raising prices. It's the only thing that's kept them ahead…”—
“On top of that the market isn't it's beyond irrational. The market is not to do with companies. It's to do with Twitter and it's to do with imaginary things…”—
FORECASTS
“As I said at the very beginning in the intro, the market cap of these AI related companies is eating the indices at this point in time and the spending…”—
“Revenue growth slows. And just to be clear, revenue growth will slow for Microsoft, Google, Amazon, and Nvidia. And it's not because the revenues will stop growing…”—
“Soft Bank, their liquidity, their future liquidity is contingent on OpenAI going public…”—
– Ed’s overall take on AI’s technological promises
Ed Zitron says AI's promised autonomous intelligence is unsupported by broad productivity gains, with current systems behaving more like unreliable automation than artificial general intelligence.
– Lack of proven productivity gains outside semiconductors
Zitron argues measurable AI-driven productivity or revenue growth is largely absent outside semiconductor companies and hyperscalers recycling spending among themselves.
– Unclear LLM business model and heavy subsidies
He says large language models lack a clear business model because users are heavily subsidized, making their true costs and economic value difficult to calculate.
– 70%+ of hyperscaler AI revenue tied to OpenAI & Anthropic
Zitron estimates that roughly 70% or more of hyperscaler AI revenue is connected to OpenAI and Anthropic spending, creating extreme dependence on two money-losing customers.
– Nvidia and semiconductors inflated by speculative data-center spending
He argues Nvidia and other semiconductor stocks are inflated by speculative data-center construction funded by expectations rather than demonstrated demand.
– Does AI actually deliver measurable ROI?
The discussion questions whether AI produces more than a dollar of value for every dollar spent, given escalating compute costs and uncertain practical benefits.
– Chimath Palihapitiya’s “compute costs doubling, 5% productivity” story
Zitron cites Chamath Palihapitiya's account of compute costs doubling while productivity rose only about 5%, illustrating AI's unfavorable cost-benefit ratio.
– Why productivity is almost impossible to measure outside sales
He explains that productivity is difficult to measure outside sales, while coding metrics such as commits or lines of code can increase without improving organizational outcomes.
– LLMs as unreliable research tools (Bloomberg Ask example)
Zitron describes Bloomberg Ask producing an obviously false Microsoft share price and argues that subtle LLM errors can invalidate financial research and models.
– Early-days bugs vs structural limitations of LLMs
He contends that LLMs are probabilistic systems unable to consistently detect their own mistakes, making reliability requirements fundamentally different from ordinary consumer use.
– Comparisons to AWS, Uber, the original iPhone, and the early internet
Zitron rejects the idea that current failures are merely early-stage bugs, arguing that LLMs struggle with contextual, experiential, and domain-specific research work.
– Neurosymbolic crutches and why 100% reliability is required
Comparisons with AWS, Uber, the early internet, and the original iPhone show that successful technologies solved clear problems or demonstrated obvious advantages rather than relying on hype.
– The “95% accurate + human QA” argument
He says LLM improvements depend on specialized training, tools, and external scaffolding, while the underlying systems remain unreliable and difficult to define as products.
– Why a modest “good-enough search” product could never justify the spend
Zitron argues that neurosymbolic systems and deterministic checks are necessary because LLMs cannot reliably perform tasks every time, especially where errors create cascading consequences.
– Will AI actually replace large numbers of jobs?
He rejects the claim that 95% accuracy plus human quality assurance solves the problem, saying review can erase the supposed efficiency gains and leave humans responsible for hidden errors.
– Why most job-displacement claims are overstated or false
Zitron says a modestly useful search or summarization tool might have a small market, but could never justify the trillion-dollar data-center investment currently being demanded.
– Is AI the new corporate DEI?
He argues AI's best current use cases are limited to searching, summarizing, and rough assistance, while subtle mistakes, bias, and missing context prevent dependable automation.
– Is the current spending just the necessary price of reaching AGI?
Zitron says AI is replacing some contract-based translation, transcription, and SEO work, but there is no evidence it will displace large numbers of core professional jobs.
– What the eventual reckoning looks like for markets and corporate America
He attributes many job-displacement claims to media speculation and executive decisions rather than proof that AI performs better, citing Klarna's reversal from AI customer service back to people.
– Extreme concentration of hyperscaler growth in OpenAI/Anthropic spend
Zitron compares corporate AI enthusiasm with DEI, but says executives embrace AI more eagerly because it flatters their intelligence and promises labor savings without requiring operational understanding.
– Why OpenAI and Anthropic cannot become the size of Microsoft or Google
He rejects the idea that current spending is simply the necessary cost of reaching AGI, arguing that researchers do not know what AGI is or possess a credible path toward it.
– Cascading impact on Nvidia, Taiwanese suppliers, SoftBank, and Oracle
Zitron says OpenAI, Anthropic, and hyperscalers have spent enormous sums producing mediocre LLMs, while the resulting systems remain far from conscious, reliable general intelligence.
– Potential 50%+ market decline and why a clean post-dot-com recovery is unlikely
He forecasts a reckoning in which slower AI spending undermines hyperscaler growth, Nvidia, Taiwanese suppliers, SoftBank, Oracle, venture capital, and private credit tied to data centers.
– Advice for retail investors: protect capital or look for opportunities?
Zitron predicts a market decline potentially exceeding 50%, historic corporate impairments, startup failures, recessionary damage, and a lasting collapse in trust toward financial and technology authorities.
Related Episodes
Important Disclaimers
Not Investment Advice: For education only — not financial, tax or legal advice.
AI-Generated: Made by AI from source transcripts and may contain errors.
No Guarantees: Past performance does not predict future results; investing carries risk.
Our Commentary · Request Removal