Over the past three weeks, I published a detailed long-form article in three installments, analyzing the potential triggers for an AI bubble burst, the triggering events, and the subsequent chain reactions.
Last week, the industry winds shifted.
Just as Uber COO Andrew Macdonald stated that AI investments are 'increasingly hard to justify,' making it difficult to reconcile the spending with the input-output logic of practical consumer-facing features (the company's CTO previously revealed that Uber exhausted its annual token budget in just four months), Axios reporter Madison Mills broke the story that one enterprise, having failed to set spending limits, accidentally burned through $500 million on Anthropic models in a single month.
Days later, Mills reported again that multiple enterprises have begun scrambling to cut AI-related expenditures.
The root cause, as I previously argued, lies in the fact that large language models are inherently prone to hallucinations, and the proliferation of various invocation frameworks and agentic interfaces means there is currently no method to accurately calculate the ROI of AI investments, nor even a unified standard to measure the cost of a single task.
Different prompts, projects, and interaction scenarios can all produce unpredictable errors, forcing users to stay constantly on guard against catastrophic blunders from these so-called 'intelligent systems'—after all, LLMs possess no autonomous thought or consciousness, and beyond pre-training and fine-tuning, they have no capacity for self-directed learning.
If a product cannot quantify its actual effects, account for usage costs, or measure return on investment, people have every reason to question why they paid for it in the first place.
The industry's current debate over AI returns falling short of expectations is understandable, but in my view, the inability to quantify costs is the far more fatal problem.
Yesterday, Microsoft's GitHub Copilot overhauled its billing rules, switching from a subscription plan based on usage quotas to token-based metered billing. Previously, users paid just $39 per month for unlimited token consumption.
Users were furious: one reported a single prompt consuming 50% of their monthly quota; another used 60% in just a few hours; a single prompt cost one user 31% of their allowance; another estimated five hours of continuous use would exhaust their entire monthly allocation; eight commands ate nearly half the quota for one user; two prompts consumed 14% for another.
Others complained that 33% of their quota vanished into thin air within hours, turning GitHub Copilot overnight from a favorite subscription product into a nagging headache.
It should be noted that the platform is still in a promotional period, during which users can claim $11 or $21 in free credits each month.
Like the vast majority of users on subsidized AI subscription plans, these users never knew the actual cost of each operation from the start. Microsoft deliberately obscured the true pricing of prompts, allowing users to consume resources heavily to drive GitHub Copilot's user growth.
Such chaos pervades the entire industry.
Every AI subscription product on the market uses subsidized pricing to hide true usage costs. Consequently, all the hype pieces praising Claude Code and the AI sector fall into a cognitive illusion: beautifying the product's actual value and overestimating the industry's current state.
A quick explainer
Let me break down the AI subscription profit model: users pay a monthly fee to access services from Anthropic, OpenAI, etc., and can use the product freely within daily or weekly rate limits. However, vendors never clearly label what those limits correspond to in actual resource usage—just a vague percentage progress bar—leaving cost estimation entirely up to the user's guesswork.
When invoking an AI model, input content is converted into input tokens (one token ≈ 0.75 English words), and the model's output is converted into output tokens. Vendors bill per million tokens.
Even with caching mechanisms to reuse existing data and reduce redundant reads/writes, every interaction incurs a cost, regardless of output quality or whether the result is usable in production.
This is also the core reason why the vast majority of AI startups are structurally unprofitable: VC funding flows almost entirely to Anthropic and OpenAI to prop up their own loss-making services. While top AI labs can build their own compute infrastructure to trim some costs, there is zero empirical evidence that this achieves profitability.
For example, every $1 a user spends on an Anthropic subscription can correspond to $8–13.50 in token costs. Even if AI boosters claim Anthropic's inference business is profitable, that claim rests solely on CEO Dario Amodei's theoretical reasoning, with no audited financials to back it up.
Put simply: under a subscription model that only shows call limits and hides actual unit prices, model failures like infinite loops or erroneous outputs can always be dismissed as 'early-stage tech immaturity'—after all, the user only paid $20, $100, or $200 for the whole month.
Anthropic, OpenAI, and the entire AI industry deliberately hide true billing standards because once users have to pay for the model's basic blunders, mass complaints and rights-protection actions would inevitably follow.
Broken Promise: AI Usage Costs Rising, Not Falling
In recent months, the cost chaos has fully surfaced: in Q1 2026, Anthropic and OpenAI quietly switched all enterprise customers to token-based metered billing. Enterprise decision-makers are typically removed from the front lines; until recently, management forcibly pushed company-wide AI adoption, with some even tying AI usage rates to performance reviews and job retention.
After long-term use of subscription products that hide real costs, employees developed a habit of treating AI as near-zero-cost. Combined with middle managers pressuring for large-scale deployment, no one knew the actual AI cost of any single business function throughout the process.
This business model was doomed from the start: the vast majority of AI users neither understand nor are allowed to see the true token cost, leading to unrestrained, extensive usage that continuously generates uneconomic resource consumption.
Every viral article hyping AI is written by authors blinded by industry rhetoric, ignoring real costs, and vigorously promoting a technology that is fundamentally unreliable, unstable in effect, persistently high-cost, and rising in price.
Incidentally: even if the per-million-token price drops, the total tokens required to complete the same task have surged, ultimately driving up overall inference costs. An analogy: gasoline gets cheaper, but the commute keeps getting longer, so total travel expenses rise instead of fall.
OpenAI, Anthropic, and other AI firms colluded to deliberately conceal AI's true costs, and this marketing playbook worked seamlessly until the switch to metered billing.
Less than a quarter into the new billing cycle, major enterprise clients plunged into cost panic: Walmart set token consumption caps for its internal AI coding tool Code Puppy, with a spokesperson saying they would 'guide employees to use AI only in scenarios that create real value'; just days earlier, Amazon Senior VP Dave Treadwell warned staff 'don't use AI blindly just for the sake of using AI.'
The AI boom of the past several years was built on layers of lies. The entire industry colluded to create the illusion that AI is affordable, controllable, commercially sustainable, soon-to-be-profitable, and that hallucination defects are curable—portraying current problems as mere growing pains.
The reality: the AI industry has absorbed over a trillion dollars in capital, massive top-tier talent, the vast majority of startup funding and media attention, and the output of tens of millions of creators—yet it has failed to cure the visible foundational flaws.
Whenever an insider points out logical flaws, they are met with false analogies to the Uber bubble (which is incomparable) or to Amazon Web Services' early massive investment (AWS invested $52 billion cumulatively over 14 years and achieved positive cash flow in year nine)—and fed the empty talking point that 'costs will eventually fall, just wait for the inflection point.'
After four years and a trillion dollars in capital, overall AI costs have risen, not fallen; leading firms burn cash ever more aggressively; product reliability hasn't improved; and boosters double down on fabricating lies to cover reality, all so a handful of billionaires can harvest the gains.
OpenAI and Anthropic's products perfectly cater to managers unwilling to deeply engage with the business, producing plausible-looking half-finished products that fool lay executives and middle managers; the core reason this model has survived is the vendors' deliberate concealment of true usage costs.
I'll say it again: today's AI deployment costs are far higher than three years ago, and there is no downward pricing trend. Sam Altman's vision of 'intelligence too cheap to meter' is pure fantasy; NVIDIA's Blackwell GPUs failed to lower compute costs, and the upcoming Vera Rubin GPUs won't either; Google TPUs, Amazon Trainium and Inferentia chips, Vera Rubin CPUs, OpenAI's custom silicon, and DeepSeek hardware—none can achieve AI compute price reductions.
The public's conviction that AI costs will keep falling rests on the historical trend of electronics getting cheaper, but that logic doesn't apply to AI: model training and inference costs include not just hardware procurement but also ongoing high electricity bills. LLMs rely on high-power parallel computing, requiring expensive GPUs; as model parameters grow and architectures become more complex, the number of GPUs needed for training and inference skyrockets in lockstep.
After three generations of product iteration, NVIDIA GPUs have not brought down compute costs—proof enough that the native business model of generative AI has a fatal flaw.
Don't Compare the AI Bubble to the Dot-Com Bubble: The AI Crash Will Be Far Worse, with No Dot-Com-Style Recovery
The industry loves benchmarking the AI boom against the dot-com bubble, but AI's risks are far greater. People prefer to misuse history to rationalize the industry's chaos, avoiding the fact of the largest capital misallocation in global history.
The dot-com bubble had two main segments: the e-commerce/startup bubble and the telecom infrastructure bubble.
According to Justin Kollar's analysis, the telecom bubble stemmed from a severe misjudgment of market demand.
Behind the massive nationwide fiber rollout was a widely circulated fallacy—that internet traffic doubled every 90 days. This claim filled brokerage reports, earnings calls, and investor decks; if true, exponential demand growth would forever outpace infrastructure supply, making every inch of fiber perpetually profitable.
But data shattered the lie: AT&T researcher Andrew Odlyzko's analysis of actual traffic showed U.S. backbone traffic doubling annually—rapid growth, but nowhere near the absurd 90-day doubling rate.
Meanwhile, wavelength-division multiplexing allowed a single fiber to carry multiple light signals in parallel, exponentially increasing per-strand capacity.
In the end, infrastructure deployment far exceeded actual demand: the global internet user base was small, and dial-up speeds were painfully slow.
Everyone cites 'the internet was also doubted back then' to justify AI's prospects, but let's clarify history: in 2000, only 52% of U.S. adults used the internet, rising to 61% by 2003; World Bank data shows global internet penetration was 16% in 2005, reaching 71% only in 2024.
The key difference lies in broadband adoption: early 56K dial-up was metered by the minute and sluggish; from 2000–2002, average U.S. speeds topped out at 400 kbps (~50 KB/s); today the U.S. average exceeds 200 Mbps (~25 MB/s).
Back then, page load times are unimaginable today; later, web design optimization and mobile adoption further revitalized the internet ecosystem. Early 2000s e-commerce was nascent, and the massive fiber infrastructure left by the dot-com bubble became a long-term industry dividend—one of the bubble's few positive legacies.
The core divide between AI and the dot-com bubble: once idle fiber was lit, the resulting internet services offered better experience at lower cost than dial-up. Failures like TheGlobe, WebVan, and Pets.com lost money due to their own broken business models, not because internet access was too expensive.
Their successors—Facebook, Instacart, Chewy—achieved profitability without overturning the underlying science of goods delivery or network access; the first-wave companies collapsed due to reckless expansion and exorbitant customer acquisition costs, up to $400 per user.
Dell and CoreWeave have deployed the first Vera Rubin GPUs, yet no vendor mentions profitability or sustainable operations—because NVIDIA's R&D focus is on raising hardware prices, not improving energy efficiency.
Per founder Jensen Huang's public statements: current AI data centers cost ~$50 billion per GW, with future costs projected to climb to $80–100 billion per GW.
Pricing trends show zero sign of compute costs declining; even as single-device theoretical performance rises, end users see no real price drops or efficiency gains, and the industry still cannot quantify the deployed value of hardware upgrades.
In summary: the dot-com bubble was driven by capital overheating, and its burst left usable communications infrastructure that enabled ongoing cost reductions and efficiency gains. AI data centers and the AI startup track lack any such recovery logic.
The AI bubble will leave no reusable quality infrastructure—no dot-com-style redemption rally.
People often cite the internet or railway bubbles, claiming the industry won't vanish after the burst, but such analogies fundamentally misunderstand the nature of the AI industry.
GPU-packed AI data centers are highly specialized for AI compute; GPUs can handle big-data parallel statistics and scientific simulations, but their architecture is inherently ill-suited for fragmented, miscellaneous tasks, making them a poor fit for the vast majority of general-purpose compute needs.
The dot-com redemption relied on massive fiber surplus—adding a little more build-out brought cheap broadband to everyone. AI has no reusable infrastructure dividend.
AI data centers require billions upfront, operate at a loss for five to six years, and may never recoup construction costs. Five years from now, the ops costs for Vera Rubin and Blackwell racks—electricity, space, on-site generators—will be the same as today; the costs to complete and grid-connect abandoned data centers, with their dedicated power, won't decline over time either.
After the dot-com crash, failed firms sold servers and office gear cheap, letting founders build startups on a shoestring: a Sun Microsystems Ultra Enterprise 3000 server sold for $43,000 then (~$89,000 today), drawing 1,200–1,500 watts to run a full-stack business. A single B200 Blackwell GPU draws 1,200 watts alone, and a single user needs 4–12 GPUs for complex AI coding tasks. A handful of GPUs can't support a profitable project—monetization and scaling are non-starters.
Idle fiber plus transceivers lights up broadband; an AI data center is a purpose-built facility with custom cooling for specific chips—converting it to a general-purpose data center requires a total teardown and rebuild, writing off the massive upfront capex.
Even buying bankrupt compute firms' idle Blackwell GPUs cheap doesn't help individuals: the dedicated facility and custom cooling carry high fixed overhead; even if the chips are free, colocation costs remain prohibitive.
Internet and railroads survived bubble cycles because their upfront infrastructure was a one-time investment with controllable marginal costs thereafter.
Taking over a GPU-packed idle AI facility on the cheap leaves you with a bottomless pit of electricity and labor costs; idle hardware still incurs fixed costs, and profitability demands 24/7 full utilization—but the industry has no proven profitable business model, and even massive funding hasn't closed the loop.
Model training costs are purely ongoing operational expenses: training a large model from scratch can require tens of thousands of H100/H200 GPUs, and the electricity bill is paid in full regardless of outcome; a single failed training run can lose tens or hundreds of millions, and post-burst, no capital will absorb such wasteful investment.
I mentioned a theoretical scenario in my previous paid column: once AI reverts to true pricing, most enterprises won't afford the usage costs; but today's fake demand is fueled by subsidized pricing, and the moment firms face real costs, they slash budgets and halt investment.
Post-burst, industry AI investment enthusiasm will plummet: startup funding dries up, enterprise AI budgets are cut, data center financing stalls.
The entire AI bubble chain is propped up by industry hype: hyping 'AI factories' to issue debt for NVIDIA hardware and build facilities; touting AI software deployment value to keep enterprises buying OpenAI and Anthropic services; painting AI deployment potential to secure startup funding; relying on nonstop media hype to hide sky-high costs. Once this lie system collapses, the bubble instantly disintegrates.
The industry sustains heat through lies, false advertising, and capital-captured financial media, using vague messaging to inflate AI valuations. Mainstream media habitually use vague lines like 'AI can develop software,' misleading the public into thinking a single prompt yields production-ready commercial software; in reality, model output is riddled with vulnerabilities, only fooling lay managers and lazy journalists.
This creates public-pressure FOMO—'refusing to embrace AI means missing the next internet dividend'—forcing blind corporate investment. Without that pressure, AI products would have to stand on genuine merit; yet even with market-wide hype, top vendors still rely on subsidies to sustain pricing, proving the product's intrinsic value is weak.
The only way to inflate the bubble on both software and hardware fronts is to hide AI's true costs and deployment results from the public and investors.
Now enterprises face real bills and are scrutinizing deployment efficacy, throwing the industry into collective cost panic.
ROI Controversy Erupts: Incalculable ROI = No Real AI Return
Last week, SemiAnalysis released a highly misleading report, 'AI's Hidden Output: Visible Costs for Invisible Outputs,' arguing that 'AI deployment gains land first, get quantified later,' and claiming 'AI may bring an Industrial Revolution-scale transformation, with enterprises doubling down on AI investment while vast new economic output remains in an unmeasurable hidden state.'
SemiAnalysis specializes in semiconductor research and has a vested interest in propping up the AI bubble; their whitewash—'the returns exist but can't be measured'—only underscores the industry's desperation.
They define 'hidden output' as AI-created economic increments that don't show up in GDP, prices, employment, or industry financial stats, split into two categories:
1. Substitution-type hidden output: work originally done by humans handed to AI; estimates suggest AI can currently optimize or automate ~$1.5 trillion worth of labor.
2. New-output-type hidden output: brand-new work that wasn't done before AI because costs were too high; long-term incremental potential supposedly far exceeds the substitution segment.
Their illustrative argument is full of holes: 'A legal document, whether drafted by a lawyer or generated by AI, has the same inflation-adjusted economic value to the user, corresponding to the same GDP increment.'
Hiring a lawyer buys professional experience, case-law research, and litigation risk control—AI merely stitches text together and hallucinates frequently, incapable of replicating legal risk management; the economic value is worlds apart.
The report adds: 'After AI takes over document work, human service fees disappear, costs become token expenses; official stats show legal service averages rising, simple docs all handled by AI, creating a GDP data gap with only scattered token consumption counted in other industries.'
Four years into AI commercialization, the industry still relies on hypothetical cases to prove value.
In reality, simple documents aren't fully outsourced to AI; law firms' core value is using practitioner experience to mitigate legal risk, from junior drafting to partner sign-off—a full risk-control chain. This hollow logic is the foundation inflating the AI bubble.
So-called new AI work is narrowed by the authors to literature reviews and email summaries, backed by anecdotal claims that most token spend goes to brand-new incremental business rather than replacing existing labor.
If AI Had Positive ROI, Layoff Waves Would Be an Undeniable Fact
All news about AI replacing jobs is filled with ambiguous, vague language.
Earlier, multiple outlets cited Oxford Economics claiming entry-level roles were being widely replaced by AI; the source only noted employment declines in some sectors post-2022 and potential signs of AI substitution—no hard deployment data.
CNBC hyped an MIT study claiming AI could replace 11.7% of U.S. labor—the data came from a labor simulation model, not actual AI deployment measurements.
The vast majority of companies laying off in AI's name do so to appease stock prices and grab headlines, not because tech deployment forces cuts. If AI truly replaced labor at scale, soaring global unemployment and social unrest would be unmistakable economic signals.
If AI deployment matched the hype, every industry would face disruptive transformation:
1. Software industry collapse: anyone could generate commercial software with a sentence, no one would buy packaged software, cloud vendors' models would crumble. The SaaSpocalypse is not a valuation correction from sector overheating (PE firms inflated software valuations 2018–2022; Apollo's John Zito called asset pricing completely detached from fundamentals, unrelated to AI).
2. Accounting industry extinction: tax filing fully automated, professional accountants lose livelihood.
3. Legal industry chain collapse: firms stop hiring juniors en masse, law schools dry up, legal salaries plummet.
4. Cross-disciplinary research restructuring: deep reports in every field generated with one click, human research loses value.
The precondition for these scenarios is zero hallucination, autonomous reasoning, and originality—current products meet none of these standards.
The inability to measure AI layoff data stems from the fact that AI cannot independently perform a full job—it can only make shoddy substitutions in a few scattered outsourced roles, lacking end-to-end duty fulfillment. Models have no workplace experience, industry knowledge, or subjective judgment; all outputs derive from training data.
Sam Altman and Dario Amodei raised hundreds of billions by stoking unemployment fears, then both walked back the layoff narrative; industry practitioners who blindly amplified the hype share the blame.
Products with Real ROI Don't Need Constant Potential-Pitching
ChatGPT, Claude, Gemini are globally available for anyone to try. If AI truly enabled efficient deployment, we'd see a flood of AI-profitable startups, industry leaders pulling ahead via tech, and service prices plummeting.
Four years later, no benchmark enterprise has disrupted its industry with AI.
The industry wouldn't need vendor fabrications—like Anthropic claiming Mythos was too powerful to release, then launching it months later; executives wouldn't need to repeatedly verbally attest to AI's value; deployment results would be visibly obvious, no potential-pitching required.
Any Quantified ROI Calculation Points to Missing Returns
Bain recently surveyed executives at 951 companies with >$100M revenue; data cannot confirm stable positive AI returns: only 37% achieved 10–20% cost reduction, 40% saw <10% reduction, and globally only 4% achieved >30% AI-driven cost cuts.
The cost-reduction base is vague: 10% on tens of millions vs. thousands is worlds apart; statistical definitions are murky.
Even more damning: 44% of firms plan new AI investments based on the previous round's vague savings data, yet some firms' previously promised savings have yet to materialize.
Bain's punchline exposes the farce: 'The tech works, but the value hasn't landed. Rolling forward new investments based on past fictitious gains looks like risk control, but it's actually a self-referential gambling loop with a built-in flaw.'
In plain terms: the product runs, but creates no economic value—equivalent to deployment failure.
Bain advises firms to verify actual automation gains before deploying AI, not expected gains, to avoid blindly amplifying risk.
A top consultancy with billions in revenue having to explicitly remind clients to calculate ROI underscores the absence of AI deployment value.
Facing runaway compute costs, Sam Altman told an interviewer: companies widely report soaring AI spend and murky returns, and the industry will supposedly self-correct the input-output balance.
Reporter David Faber missed the follow-up, pivoting mid-interview to the irrelevant topic of space-based compute—exemplifying how the AI bubble inflates via media acquiescence: key questions are dodged, diverted to ethereal future visions.
Worth billions and heading a $852B-valued OpenAI, Altman shifts industry-reform responsibility onto the whole industry while dodging the core problem himself.
LLM Users Reduced to Scam Victims
If AI had real ROI, the industry could cite concrete revenue-generating deployments. Praising Spotify engineers' reduced workload is meaningless: platform iteration, system stability, and feature delivery haven't materially improved; massive AI-generated code has instead worsened software bugs.
In actual cost terms, AI coding tools have raised R&D spend across the board, with firms paying huge bills to Anthropic and Cursor while code quality broadly declines. Some engineers enjoy the tools, but no one can produce proof of corresponding deployment gains.
I'll say it straight: practitioners attempting full business-process deployment via LLMs are all victims harvested by marketing rhetoric.
Light use—daily scripts, simple office assistance, voice transcription, search aid—is fine (though search results still need source verification); binding the entire business chain to an LLM equals falling for a marketing scam.
Deploying an AI automation pipeline requires stacking massive error-handling scripts to compensate for model hallucinations; users spend far more hours debugging the tool than the original manual work would take; their sense of achievement comes from product-marketing brainwashing.
A few use models to simplify Python data scripts—the benefit stems from Python itself, not an LLM breakthrough.
A trillion dollars in capital yielded a product that merely lowers the barrier to writing starter scripts—the epitome of capital waste.
All AI automation projects are essentially scams packaged with massive subsidies and false advertising. When pressed on deployment returns, proponents can only recycle irrelevant analogies like Uber or the dot-com bubble.
The fervent defenders' zeal comes from inability to answer basic questions about deployment gains and cost accounting; all current LLMs are sold on paper potential, with actual deployment far below marketed expectations.
The Ultimate Answer to Unquantifiable ROI: AI Has No Positive Return
LLMs only marginally speed up simple tasks; as complexity rises, accuracy and cost-effectiveness both drop. Expanding compute clusters only widens the range of tasks models can attempt—it doesn't deliver on profit promises.
The few deployment cases either deliberately ignore inherent model flaws or throw massive human effort at error correction, yielding extremely low-value half-finished products. LLMs packaged as automation magic rely entirely on humans to fix errors and absorb defects; the labor meant to be replaced is instead spent maintaining the AI.
Capital and executives brainwash media and markets with lies, harvesting industry resources; media practitioners and lay managers lack technical and industry awareness, swept up by hype. LLMs precisely cater to the herd's vanity, making people think the model creates value, when in reality humans pour more resources into patching the product's holes.
Trillions spent and industry-wide resources yield low-quality copy, shoddy spreadsheets, and copyright-collaged images—no disruptive deployment outcomes whatsoever.
As paying users, our job is to judge product quality, not to make excuses for defects or shill for vendor hyperbole; the constant need to defend 'potential' proves the AI industry's business model is fundamentally flawed.