Zitron Warns $1.1T AI Compute Debt Threatens Oracle, Pension Funds

The Hater's Guide To The AI Bubble With EZPR — Ai4

Ai4

Unpayable foundation model infrastructure commitments and debt-financed data center buildouts will fail to generate sufficient compute revenue, shifting structural losses onto public pensions and corporate balance sheets.

AI-assisted summary of The Hater's Guide To The AI Bubble With EZPR — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by the channel Ai4 on August 18, 2026, this presentation features Ed Zitron, writer at EZPR, host of the Better Offline podcast, and author of Where's Your Ed At. The talk outlines how the tech industry has diverted over a trillion dollars into generative AI, breaking down the flawed unit economics, questionable return on investment, and impending financial fallout of the AI bubble.

## TL;DR Ed Zitron delivers a scathing financial indictment of the generative AI boom, arguing that tech hyperscalers and foundation model labs are sustained by distorted annualized metrics and massive off-balance-sheet debt. He demonstrates that planned data center buildouts require trillions in annual compute revenue—vastly exceeding the entire global software market—while providers face negative unit economics on probabilistic, hallucination-prone models. Characterizing OpenAI as the systemic 'Lehman Brothers of AI' and highlighting debt-fueled neoclouds, Zitron warns that foundation labs will never achieve the scale needed to cover their infrastructure liabilities. As these unpayable obligations mount, the financial damage is poised to cascade onto public pension systems and private credit markets.

## Atmosphere & style The presentation is an intense, uncompromising broadside delivered with razor-sharp financial forensics, mordant humor, and palpable urgency. Zitron cuts through Silicon Valley tech optimism with aggressive clarity, drawing nervous chuckles and stunned silence as he systematically dismantles industry claims. It feels like an urgent whistleblower briefing rather than a conventional tech keynote, vibrating with righteous indignation and challenging the audience to confront an impending economic reckoning.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) The generative AI boom is built on unsustainable financial engineering, where hyperscalers and startups deploy annualized run rates and off-balance-sheet liabilities to obscure weak broad-market demand. - [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) Just two companies, OpenAI and Anthropic, account for roughly 70% of all hyperscaler AI revenue, demonstrating extreme market concentration rather than healthy ecosystem-wide adoption. - [01:50](https://www.youtube.com/watch?v=CldB-5PuFH4&t=110s) The transition from heavily subsidized flat-rate pricing to raw token billing has exposed the broken ROI of large language models, triggering strong enterprise pushback against paying for probabilistic, hallucination-prone tools. - [04:40](https://www.youtube.com/watch?v=CldB-5PuFH4&t=280s) Historical comparisons to the capital burn of AWS ($29.7B over 12 years) or Uber ($33B) fail, as single AI fundraising rounds and annual commitments now routinely exceed tech's largest historical infrastructure investments. - [06:36](https://www.youtube.com/watch?v=CldB-5PuFH4&t=396s) The probabilistic nature of LLMs causes unpredictable execution costs and recursive agent loops, making flat-rate SaaS subscription models structurally unprofitable even at high gross margins. - [09:15](https://www.youtube.com/watch?v=CldB-5PuFH4&t=555s) Planned data center buildouts requiring $2.7 trillion in annual compute revenue are macroeconomically unviable compared to the entire global software market, which will reach only ~$800 billion by 2026. - [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) Specialized 'neocloud' GPU hosters operate primarily as debt-fueled special-purpose vehicles to acquire Nvidia chips, relying heavily on artificial demand and subsidized foundation lab spending. - [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) OpenAI's audited financials reveal severe structural losses, projecting a $20.9 billion net operating loss on $13.07 billion in revenue for 2025. - [14:17](https://www.youtube.com/watch?v=CldB-5PuFH4&t=857s) Core industry arguments for reaching profitability—including inference margins, pricing power, search ad monetization, and halting model training—collapse under scrutiny. - [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) OpenAI represents a systemic 'Lehman Brothers' risk to the tech sector, having chained itself to $750 billion in compute obligations that threaten hyperscaler partners like Oracle. - [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) The ultimate financial fallout of the AI infrastructure bubble risks being socialized, as speculative data center private credit is increasingly underwritten by public pension funds and insurance capital.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) The AI Industry Con and Hyperscaler Accounting Swindles

Ed Zitron introduces himself as an independent AI critic determined to voice the uncomfortable truths that the broader tech industry refuses to acknowledge. He opens with an immediate warning: anyone pressuring you to back artificial intelligence or nudging you to say things like "well, of course I love AI" or "of course I think AI is the future" is pulling you into a con. As the host of the *Better Offline* podcast and the author of the *Where’s Your Ed At* newsletter—which surged from 50,000 subscribers at the start of 2025 to over 110,000—Zitron highlights his access to OpenAI's leaked financials as proof that he understands the actual numbers behind the hype.

Four years into the generative AI boom, the underlying business fundamentals remain bleak. Citing analyst estimates from UBS, Wells Fargo, and Barclays, Zitron notes that just two companies—OpenAI and Anthropic—account for roughly 70% of all hyperscaler AI revenue. There is nowhere near enough broad market demand to justify the colossal buildout of data centers. To conceal this reality, startups and hyperscalers lean heavily on annualized run rates (multiplying a single month by 12, or four weeks by 13). Tech giants avoid disclosing true AI revenue because the media uncritically credits them with winning bets, even as hyperscalers take on $1.1 trillion in commitments from OpenAI and Anthropic alongside an estimated $1.3 to $1.6 trillion in off-balance-sheet debt and liabilities.

- **Newsletter reach**: *Where’s Your Ed At* grew from 50,000 subscribers at the beginning of 2025 to 110,000. - **Concentrated revenue**: OpenAI and Anthropic generate ~70% of all hyperscaler AI revenue per UBS, Wells Fargo, and Barclays. - **Hidden liabilities**: Hyperscalers have committed over $1.1 trillion to OpenAI and Anthropic, backed by $1.3 to $1.6 trillion in off-balance-sheet liabilities and debt. - **Run rate metrics**: Annualized run rates are routinely deployed to obfuscate actual, recurring top-line figures.

[01:50](https://www.youtube.com/watch?v=CldB-5PuFH4&t=110s) The Broken ROI of Large Language Models and Token Pricing

The return on investment (ROI) conversation reached a turning point when OpenAI and Anthropic abruptly altered their pricing structures. Previously, enterprise subscribers on flat monthly tiers could burn astronomical sums of compute—up to $8,000 in tokens on Anthropic's $200-a-month tier, or up to $14,000 on OpenAI's enterprise tier. In March and April, however, the providers effectively told customers they would have to pay the true per-million token costs. Microsoft made a similar shift with GitHub Copilot, where users paying $39 a month had previously been consuming $1,500 to $5,000 worth of compute. When forced onto raw token rates, customers balked at the sheer expense of paying for probabilistic models that charge them regardless of whether the output works, fails, or hallucinates.

```mermaid flowchart TD accTitle: Subsidized vs Actual Token Economics accDescription: Flowchart illustrating the transition from heavily subsidized flat-rate AI access to raw token billing and enterprise budget pushback.

subgraph Subsidized ["Heavily Subsidized Tier (Pre-2025)"] A["Flat Monthly Fee ($39 - $200/mo)"] --> B["Massive Token Consumption ($1.5k - $14k Value)"] B --> C["Hyperscaler & VC Cash Burn"] end

subgraph RealCosts ["Raw Token Metering (Post-Shift)"] D["Per-Million Token Billing"] --> E["Cost Incurred Regardless of Success/Hallucination"] E --> F["Enterprise Pushback & ROI Scrutiny"] end

Subsidized --> RealCosts ```

Enterprises are waking up to the fact that large language model ROI cannot be reliably measured. The industry is only now confronting this dilemma because AI labs realize they "can't keep giving away $40 for a dollar." High-profile enterprise leaders are pushing back; Uber COO Andrew McDonald recently voiced difficulty in justifying AI investments because the technology fails to connect cleanly to useful product features or faster shipping cadences. Meanwhile, OpenAI CEO Sam Altman offered little reassurance to enterprise buyers facing exploding bills, effectively leaving the industry without answers while peddling probabilistic models that still hallucinate.

- **Subsidized consumption**: Anthropic allowed up to $8,000 in token burn on a $200/month plan; OpenAI allowed up to $14,000 on its flat enterprise plan. - **GitHub Copilot disparity**: Microsoft subsidized heavy users consuming $1,500 to $5,000 in compute against a $39/month fee. - **Uber executive skepticism**: Uber COO Andrew McDonald openly questioned the ability to link AI spending to tangible engineering outcomes and speed. - **The hallucination reality**: Claims that hallucinations are solved are demonstrably false across all frontier models and search integrations.

[04:40](https://www.youtube.com/watch?v=CldB-5PuFH4&t=280s) Debunking the AWS Myth and the Absurdity of AI Burn Rates

Proponents of the AI boom frequently justify staggering losses by drawing historical parallels to Amazon Web Services (AWS) or Uber. Zitron systematically dismantles the AWS comparison: AWS launched in 2003 and reached profitability in 2015, burning an inflation-adjusted cumulative total of $29.7 billion. By contrast, Anthropic raised $30 billion in a single month and $65 billion within a year, while OpenAI raised up to $122 billion across its funding tranches. Similarly, Uber burned roughly $33 billion before achieving an uneven profitability. The entire historical capital burn of tech's biggest infrastructure success stories can now fit inside a single AI venture funding round.

The defense that generative AI is in its "early days" rings completely hollow four years and over a trillion dollars into the cycle. Zitron demands to know when the supposed early phase ends and how many hundreds of billions must be vaporized before investors expect sustainability. The massive capital burn across foundation model developers is not building enduring software margins; it is feeding an unsustainable mythology that ignores fundamental unit economics.

- **AWS historical burn**: Amazon Web Services cost $29.7 billion (inflation-adjusted) over 12 years (2003–2015) before achieving profitability. - **AI fundraising scale**: Anthropic raised $65 billion in a single year; OpenAI secured commitments up to $122 billion. - **Uber comparison**: Uber's entire cumulative burn of $33 billion is eclipsed by individual AI financing events. - **Time and capital expended**: Four years and more than $1 trillion have been committed since the generative AI rush began.

[06:36](https://www.youtube.com/watch?v=CldB-5PuFH4&t=396s) Unpredictable Token Economics and the Unsustainable Subscription Model

The fundamental economics of token-based billing make it incompatible with standard software purchasing. Because LLMs are probabilistic, developers and enterprises cannot anticipate what a specific workflow will cost. An autonomous agent or coding platform like Replit can easily get caught in a recursive loop and burn $50 in tokens in minutes without completing the task—or worse, corrupt the existing software stack. Because no consumer or enterprise will willingly pay unbounded fees for runaway loops, AI wrappers are forced to hide actual token costs beneath subsidized subscription flat fees.

```mermaid flowchart LR accTitle: The Runaway Token Feedback Loop accDescription: Flow showing how non-deterministic models cause unpredictable API expenses and budget collapses.

A["User Prompt / Harness"] --> B["Probabilistic LLM Execution"] B --> C{"Loop Success or Failure?"} C -- "Runaway Loop" --> D["Unplanned Token Burn ($50+ wasted)"] C -- "Code Corruption" --> E["Damaged Software Stack"] D & E --> F["50% Token Budget Slashes"] ```

Enterprises attempting to rein in runaway spending are slashing token budgets by 50%, yet managers cannot explain how to cut spend without metrics for what good output looks like. Zitron highlights that reported ARR surges at companies like Anthropic and OpenAI were not stable software revenue, but rather spikes in non-recurring API token maxing driven by corporate mandates. Analysis from SemiAnalysis revealed that even if foundation model providers operated at an unrealistic 70% gross margin, a subscriber utilizing just 25% of their plan capacity would plunge the account deep into negative margins. These subscriptions exist exclusively because venture capital cushions the structural deficit.

- **Replit and agent runaway costs**: Unmonitored loops frequently consume tens of dollars in tokens per session while breaking codebases. - **Artificial ARR**: Surges in annualized revenue figures reflect transient, non-recurring API token maxing rather than predictable recurring software subscriptions. - **SemiAnalysis findings**: Even under an assumed 70% gross margin, just 25% capacity utilization by a flat-rate subscriber pushes the provider into the red. - **Growth obsession**: Corporate mandates forcing employees to adopt AI with no metrics for success stem from growth-at-all-costs venture logic.

[09:15](https://www.youtube.com/watch?v=CldB-5PuFH4&t=555s) The Data Center Fantasy vs. Global Software Reality

The narrative of infinite, insatiable demand for AI compute collapses when evaluated against the macroeconomic capacity of the software industry. Approximately 322 gigawatts of data center capacity are currently in planning, translating to roughly 225 gigawatts of critical IT infrastructure (GPUs, server racks, and power delivery). At an industry-standard benchmark of $12 million per megawatt, operators would require $2.7 trillion in annual compute revenue to justify the capital outlay.

The required $2.7 trillion annual compute spend dwarfs the entire global software market, which will barely reach $800 billion in 2026. Furthermore, with Nvidia CEO Jensen Huang citing infrastructure costs ranging up to $100 billion per gigawatt, hyperscalers face insurmountable capital requirements to build out capacity for demand that does not exist outside of subsidized venture experiments.

- **Power projections**: 322 GW of planned capacity yields 225 GW of critical IT capacity. - **Capital mismatch**: Supporting planned buildouts requires $2.7 trillion annually, compared to an entire global software industry generating ~$800 billion by 2026. - **Capex benchmarks**: Infrastructure costs range from $50 billion to $70 billion per gigawatt, with Jensen Huang quoting up to $100 billion per gigawatt.

[10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) The Impossible Math of Gigawatt Compute and the Neocloud House of Cards

Ed Zitron challenges the fundamental premise of the AI infrastructure boom: there simply isn’t enough actual demand for AI compute to justify even a quarter of what is currently being constructed. Rejecting the label of being merely a "doomer" or a "hater," Zitron emphasizes that this situation is genuinely terrifying rather than satisfying. Building these mega-facilities takes vastly longer than advertised, and the underlying economics are completely detached from commercial reality. To make a single gigawatt of compute capacity revenue-positive, a company needs roughly $30 billion in annual revenue. Yet Oracle is charging just over $10 billion for the planned 1.2-gigawatt Stargate facility in Abilene, Texas for OpenAI—only half of which is even built. As Zitron puts it, "You ain't getting $30 billion, mate."

Meanwhile, the entire generative AI software market remains pathetically small: roughly $100 billion in annual revenue, or perhaps $130 billion if adopting the inflated models of outlets like Exponential View ("who should be ashamed of themselves; they can't count"). That total generously bundles together every single dollar of compute spending from OpenAI and Anthropic, who are already burning tens of billions just to serve inference to existing customers at a loss.

When factoring in specialized "neocloud" GPU hosters like CoreWeave, IREN (Iris Energy), Nebius, Cipher Mining, and TeraWulf alongside the traditional hyperscalers, Zitron argues that 80% to 90% of entire demand profiles are artificial. These neoclouds are essentially debt-fueled special-purpose vehicles that exist solely to borrow money and purchase Nvidia chips. Even Jensen Huang candidly acknowledged that without Nvidia, CoreWeave and Nebius would not exist. Zitron sees all the hallmarks of a "quick visit to Enronville," while bank analysts, tech CEOs, and journalists blithely brush off structural insolvency with a collective shrug of "they'll work it out."

```mermaid flowchart TD accTitle: The Neocloud and Foundation Model Debt Spiral subgraph DebtFunding["Capital Inflow & Infrastructure Debt"] VC["Venture Capital & Private Credit"] Pensions["Public Pensions & Insurance Funds"] VC --> Neoclouds["Neoclouds: CoreWeave, Nebius, IREN"] Pensions --> Neoclouds end

subgraph GPUHoarding["Compute Commitments"] Neoclouds -->|"Debt raised to buy GPUs"| Nvidia["Nvidia Hardware"] Hyperscalers["Hyperscalers & Hosts: Oracle, Microsoft, AWS"] -->|"Gigawatt Data Centers"| Labs["AI Labs: OpenAI & Anthropic"] end

subgraph BrokenEconomics["Revenue Deficit & Insolvency"] Labs -->|"Unprofitable Inference & Huge Capex"| MassiveLosses["Massive Net Operating Losses"] MassiveLosses -.->|"Inability to cover $1T+ liabilities"| SystemicRisk["Socialized Losses: Bailouts & Pension Impairment"] end ```

- **Gigawatt economics**: Achieving positive cash flow on 1 GW of data center capacity requires ~$30 billion in revenue, contrasted against Oracle’s ~$10 billion pricing for OpenAI's Abilene site. - **Industry scale**: Total generative AI industry revenues hover around $100B–$130B, propped up largely by foundation labs subsidizing inference. - **Neocloud exposure**: Specialized GPU infrastructure players (CoreWeave, Nebius, IREN, TeraWulf, Cipher Mining) rely almost entirely on debt financing tied directly to Nvidia allocations. - **OpenAI 2025 audited burn**: Financial documents reviewed by Zitron and the *Financial Times* show OpenAI projecting a $20.9 billion net loss for 2025, against $13.07 billion in revenue (over $800 million of which originated from SoftBank for an obscure project dubbed "Crystal Intelligence").

---

[14:17](https://www.youtube.com/watch?v=CldB-5PuFH4&t=857s) Mythbusting the Path to AI Profitability

The prevailing hope in Silicon Valley is that foundation model builders will eventually grow into their capital liabilities through scale, higher pricing, or ancillary business lines like advertising. Zitron tears down each of these arguments as mathematically impossible. OpenAI plans to spend an astronomical $750 billion on compute by the end of 2030, while Anthropic follows an almost identical financial trajectory—generating around $5 billion in lifetime revenue as of March while burning over $10 billion on compute alone. When reports surfaced claiming Anthropic had achieved an operating profit for a single quarter, the reality behind the curtain was that Elon Musk temporarily slashed the costs of their data center capacity at SpaceX. Zitron dryly quips: "I would be profitable too if it wasn't for all the costs."

Between Oracle, Microsoft, Amazon, and Google, commitments from AI labs stand at $1.1 trillion—climbing toward $1.3 trillion once CoreWeave, IREN, Nebius, and Cerebras are counted. Addressing the popular defenses made by venture capitalists and corporate PR, Zitron systematically dismantles four central industry myths:

| Myth | Industry Claim | Economic Reality | |:--- |:--- |:--- | | **Inference Profitability** | Labs earn strong positive margins on every API call and user query. | There is zero audited proof. Unverified claims stem largely from isolated analyses of DeepSeek research papers. | | **Pricing Power** | Model providers can simply hike subscription and token prices at will. | When prices were raised in March, customers "screamed like they were being pecked to death by birds," severely stalling uptake. | | **Ad Monetization** | Search ads embedded inside chatbot responses will unlock search-engine-tier margins. | eMarketer estimates total industry generative search ad revenues at only $1 billion today, reaching a mere $5 billion by 2030—a rounding error against data center capex. | | **Halting Training** | Labs can curb spending once frontier models are finalized. | Impossible due to competitive pressure (OpenAI, Anthropic, China) and the relentless technical need for pre-training, post-training, synthetic data generation, and countering model drift. |

---

[16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) OpenAI as Lehman Brothers, Oracle's Gamble, and Socialized Losses

Zitron explicitly identifies OpenAI as the systemic risk at the center of the bubble—the "Lehman Brothers of AI." OpenAI commands over 50% of generative AI revenues and 60% of total compute demand, yet it has chained itself to roughly $750 billion in compute obligations that neither it nor its partners can ever hope to service. That total exceeds the combined infrastructure commitments of Nvidia, Microsoft, and Amazon put together.

This exposure reaches critical mass with Oracle. While the audience chuckles at the mention of Larry Ellison, Zitron warns that Oracle’s staggering 7.1-gigawatt buildout for OpenAI threatens to bankrupt the legacy software titan. Oracle cannot genuinely afford to build that much infrastructure, and OpenAI cannot afford to pay for it once it exists. Across the aisle, Anthropic mirrors this structure under CEO Dario Amodei—whom Zitron characterizes as running a carbon copy of OpenAI while using an "Elizabeth Holmes voice" to tell the press that society is on the cusp of an exponential curve, all while sitting atop $300 billion in commitments and burning tens of billions annually.

The most catastrophic element, however, is that ordinary citizens are unwittingly underwriting the entire edifice:

- **Private credit exposure**: Hundreds of billions of dollars from private credit funds are pouring into data center construction for speculative, non-existent end-user demand. - **Tapping pension and insurance capital**: Private credit managers have sourced this liquidity directly by acquiring insurance companies and securing capital allocations from municipal and state public pension funds (readily verifiable via PitchBook). - **Socialized downside**: When these unpayable infrastructure loans default and data centers are left empty, the losses will cascade onto retirement funds and local state budgets. "We are going to subsidize this whether we like it or not."

Zitron concludes with a fierce condemnation of the financial engineers and tech cheerleaders who inflated the bubble: "It is time to call bullshit on the AI industry." He calls for a software economy rooted once again in genuine utility and measurable return on investment, rather than "unproductive, ineffective, unprofitable, unsustainable pieces of software" pushed by labs obsessed solely with vanity metrics and growth at all costs. Inviting critics to confront him directly, he closes his talk by offering his personal email and Bloomberg Terminal IB handle to debate any comers.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) well, of course I love AI — Ed Zitron

> [01:50](https://www.youtube.com/watch?v=CldB-5PuFH4&t=110s) can't keep giving away $40 for a dollar. — Ed Zitron

> [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) You ain't getting $30 billion, mate. — Ed Zitron

> [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) who should be ashamed of themselves; they can't count — Ed Zitron

> [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) quick visit to Enronville — Ed Zitron

> [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) they'll work it out. — Ed Zitron

> [14:17](https://www.youtube.com/watch?v=CldB-5PuFH4&t=857s) I would be profitable too if it wasn't for all the costs. — Ed Zitron

> [14:17](https://www.youtube.com/watch?v=CldB-5PuFH4&t=857s) screamed like they were being pecked to death by birds — Ed Zitron

> [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) Lehman Brothers of AI — Ed Zitron

> [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) Elizabeth Holmes voice — Ed Zitron

> [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) We are going to subsidize this whether we like it or not. — Ed Zitron

> [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) It is time to call bullshit on the AI industry. — Ed Zitron

> [16:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1008s) unproductive, ineffective, unprofitable, unsustainable pieces of software — Ed Zitron

## Mentioned Ed Zitron, EZPR, Better Offline, Where’s Your Ed At, OpenAI, Anthropic, UBS, Wells Fargo, Barclays, Microsoft, GitHub Copilot, Uber, Andrew McDonald, Sam Altman, Amazon Web Services (AWS), Amazon, Replit, SemiAnalysis, Nvidia, Jensen Huang, Oracle, Stargate, Exponential View, CoreWeave, IREN (Iris Energy), Nebius, Cipher Mining, TeraWulf, Financial Times, SoftBank, Crystal Intelligence, SpaceX, Elon Musk, Google, Cerebras, DeepSeek, eMarketer, Larry Ellison, Dario Amodei, PitchBook, Bloomberg Terminal

## Video description Ed Zitron - Writer, EZPR

What if everybody was wrong about AI?

Ed Zitron walks you through how the entire tech industry got diverted by generative AI, burning over a trillion dollars on a dead-end technology, the broken economics of LLMs, the questionable ROI of AI, and what you're going to want to do to avoid getting burned when the bubble pops.

[Source](https://www.youtube.com/watch?v=CldB-5PuFH4)

中文简报

Zitron警告：1.1万亿美元AI算力债务威胁Oracle与养老基金无法偿还的基础模型基础设施承诺与债务融资的数据中心扩建将无法产生足够的算力收入，从而将结构性损失转嫁给公共养老金和企业资产负债表。 Ed Zitron 对生成式 AI 热潮提出了尖锐的财务控诉，认为科技超大规模企业与基础模型实验室依靠失真的年化指标和巨额表外债务来维持。他指出，规划中的数据中心扩建每年需要数万亿美元的算力收入——这一数字远超整个全球软件市场的规模——而服务商在易产生幻觉的概率性模型上面临着负单位经济效益。Zitron 将 OpenAI 定性为具有系统性风险的“AI 领域的雷曼兄弟”，并特别指出了由债务推动的新兴云服务商（neoclouds），警告基础模型实验室永远无法达到覆盖其基础设施负债所需的规模。随着这些无法偿还的债务不断累积，金融破坏势必将蔓延至公共养老金体系和私人信贷市场。生成式 AI 的繁荣建立在不可持续的财务工程之上；OpenAI 和 Anthropic 两家公司就占据了超大规模云厂商约 70% 的 AI 收入，而年化运行率（ARR）被大量用于掩饰疲软的市场真实需求。 (0:04) 从高额补贴的固定包月制转向原生 Token 计费后，企业对不可预测的推理成本与模型幻觉强烈反弹，大语言模型破碎的 ROI 彻底暴露。 (1:50) 将 AI 资本消耗类比 AWS（12年累计烧钱297亿美元）或 Uber（330亿美元）完全站不住脚，如今仅单轮融资或年度算力承诺就已超越以往科技史上最大的基础设施投资。 (4:40) 全球规划中的数据中心建设每年需支撑 2.7 万亿美元的算力收入，这在宏观上根本无法实现，因为到 2026 年全球整个软件市场规模预计仅约 8000 亿美元。 (9:15) OpenAI 凭借超 7500 亿美元无法履约的算力承诺成为整个泡沫中的“雷曼兄弟”，而私人信贷对数据中心的盲目投资已将公立养老金与保险基金置于巨大系统性违约风险中。 (16:48)

Reading summary

Technology critic Ed Zitron argues that the generative artificial intelligence sector is an unsustainable financial bubble driven by structurally negative unit economics, artificial demand, and massive infrastructure overbuild. He contends that over a trillion dollars in compute commitments and off-balance-sheet liabilities pose severe systemic risks that threaten to cascade into public pensions and broader debt markets. Subsidized Pricing and Broken Enterprise ROI (1:50) Foundation model providers initially masked the high costs of probabilistic AI through subsidized flat-rate tiers. Anthropic permitted up to $8,000 in compute burn on a $200 monthly plan, OpenAI allowed $14,000 on enterprise accounts, and Microsoft subsidized GitHub Copilot users consuming up to $5,000 in compute for a $39 fee. When labs transitioned to raw per-million token metering, corporate buyers faced severe budget inflation for tools that charge regardless of whether outputs succeed, fail, or hallucinate. Because models are non-deterministic, autonomous workflows can burn tens of dollars in recursive loops without completing tasks or while damaging existing codebases. Corporate leaders have begun resisting these costs, with Uber COO Andrew McDonald openly questioning AI's inability to demonstrate clear engineering returns or accelerate shipping velocity. Analysis from SemiAnalysis indicates that even with an assumed 70% gross margin, an enterprise subscriber utilizing just 25% of plan capacity plunges the provider into negative margins. Zitron argues that surges in annualized run rates do not represent durable software subscriptions, but rather transient spikes in token consumption driven by top-down corporate mandates lacking clear success metrics. Flawed Historical Parallels and Accelerating Burn (4:40) Zitron dismantles industry attempts to justify unprecedented losses by citing the growth trajectories of Amazon Web Services or Uber. AWS required an inflation-adjusted cumulative total of $29.7 billion across twelve years before achieving profitability, while Uber consumed roughly $33 billion prior to reaching uneven profits. In contrast, foundation model developers now burn through comparable sums within months: Anthropic raised $65 billion in a single year, while OpenAI has secured commitments of up to $122 billion. Financial documents reviewed by Zitron and the Financial Times project OpenAI to generate $13.07 billion in revenue against a $20.9 billion net operating loss in 2025. Similarly, Anthropic has burned over $10 billion on compute against approximately $5 billion in lifetime revenue. Zitron rejects the notion that advertising or pricing power will close this deficit, noting that enterprise customers balked when prices rose in March, while eMarketer projects total generative search advertising will reach only $5 billion by 2030—a negligible fraction of ongoing capital expenditure. Data Center Expansion Exceeds Global Software Revenue (9:15) The scale of planned data center construction is fundamentally detached from global software spending. Developers have proposed approximately 322 gigawatts of capacity, representing 225 gigawatts of critical IT infrastructure. At an industry benchmark of $12 million per megawatt, operators require roughly $2.7 trillion in annual compute revenue to justify the capital layout. This required spend dwarfs the entire worldwide software market, which is projected to reach approximately $800 billion in 2026. Achieving cash-flow-positive operations on a single gigawatt requires approximately $30 billion in annual revenue, yet Oracle charges roughly $10 billion for OpenAI's planned 1.2-gigawatt facility in Abilene, Texas. With the entire generative AI market generating only $100 billion to $130 billion annually—roughly 70% of which is concentrated within OpenAI and Anthropic—hyperscalers and specialized GPU hosting providers are building infrastructure for commercial demand that does not exist. Neocloud Leverage and Socialized Financial Risk (16:48) Hyperscalers and specialized neoclouds—including CoreWeave, Nebius, and IREN—have accumulated an estimated $1.3 trillion to $1.6 trillion in off-balance-sheet debt and liabilities to purchase Nvidia chips. Zitron characterizes OpenAI as the systemic "Lehman Brothers of AI," citing its $750 billion in compute commitments through 2030. These commitments place intense structural pressure on partners such as Oracle, whose 7.1-gigawatt buildout for OpenAI creates financial obligations neither company can reliably service. Much of this expansion is financed through private credit funds that have raised capital by acquiring insurance companies and securing commitments from municipal and state public pension systems. Zitron warns that because these data center assets are tied to debt-fueled artificial demand, eventual defaults will not remain confined to venture investors, but will transfer financial losses onto retirees and municipal budgets. Zitron concludes that the technology industry must abandon vanity metrics and unviable capital burn to re-anchor itself in software that demonstrates measurable utility and sustainable margins. The central unresolved issue is how long private credit and hyperscaler balance sheets can absorb tens of billions in foundation lab losses before mounting infrastructure debt forces widespread capital write-downs.

精炼阅读摘要

科技评论人 Ed Zitron 指出，当前的生成式人工智能热潮是由不可持续的巨额补贴、失真的财务指标以及狂热基建构筑的经济泡沫；以 OpenAI 和 Anthropic 为代表的核心企业缺乏真实的商业回报能力，其累积的万亿美元级未公开负债与过度扩张，正在将系统性金融风险转嫁至公共养老金与传统科技巨头。虚幻的营收规模与财报掩护 (0:04) Zitron 指出，在生成式 AI 繁荣四年后，该行业的基本面依然极其严峻。援引瑞银（UBS）、富国银行（Wells Fargo）和巴克莱银行（Barclays）的分析估计，仅 OpenAI 和 Anthropic 两家公司就占据了云厂商（Hyperscalers）全部 AI 收入的约 70%，整个市场远未形成能够支撑庞大数据中心扩张的广泛真实需求。为了掩盖这一困境，初创企业和科技巨头普遍依靠年化运行率（Run Rate，即将单月数据乘以 12 或四周数据乘以 13）来人为夸大营收表现。与此同时，科技巨头背负着惊人的隐性债务。云厂商已向 OpenAI 和 Anthropic 提供了超过 1.1 万亿美元的承诺，同时承担着估计达 1.3 万亿至 1.6 万亿美元的表外债务与负债。据泄露的财务数据与《金融时报》（Financial Times）披露，OpenAI 预计在 2025 年录得高达 209 亿美元的净亏损，而在其 130.7 亿美元的预期营收中，有超过 8 亿美元来自软银为“Crystal Intelligence”项目提供的单笔注资。 Token补贴破裂与企业端ROI幻灭 (1:50) 大语言模型（LLM）的商业模式正经历剧烈震荡。此前，基础模型开发商依靠风投资金以极低固定订阅费提供高额算力——例如 Anthropic 在每月 200 美元的套餐中曾允许消耗高达 8,000 美元的 Token，OpenAI 企业套餐允许消耗高达 14,000 美元，微软 GitHub Copilot 每月 39 美元的账户则实际消耗 1,500 至 5,000 美元的算力。然而，随着服务商在 3 月和 4 月转向按每百万 Token 的真实成本计费，企业客户面临成本失控，纷纷对不论输出成功、失败还是产生幻觉都强制收费的概率模型产生抵触。 SemiAnalysis 的分析证实，即便假设模型供应商享有不切实际的 70% 毛利率，只要固定付费用户使用其套餐额度的 25%，账号就会陷入严重负毛利。由于自主智能体（Agent）和代码平台极易陷入递归死循环并在数分钟内白白消耗数十美元，企业对不可预测的账单愈发难以承受。例如 Uber 首席运营官 Andrew McDonald 公开表示难以将 AI 支出与实际工程成果挂钩；许多企业已开始将 Token 预算削减 50%。算力基建脱节与“新云”债务链条 (9:15) AI 算力需求的叙事与全球软件业的宏观支付能力存在不可逾越的鸿沟。目前全球规划中的数据中心供电容量达到 322 吉瓦（GW），对应约 225 吉瓦的关键 IT 负载容量（GPU、服务器与配电）。按行业基准每兆瓦 1,200 万美元造价计算，运营商每年需获得 2.7 万亿美元的算力收入才能支撑资本开支。然而，到 2026 年全球整个软件市场的总规模预计仅为 8,000 亿美元左右；英伟达（Nvidia）CEO 黄仁勋更直言每吉瓦基础设施的建设成本高达 500 亿至 1,000 亿美元。在微观运营层面，维持 1 吉瓦算力实现现金流转正需要每年约 300 亿美元收入，但甲骨文（Oracle）为 OpenAI 在得克萨斯州阿比林规划的 1.2 吉瓦 Stargate 设施仅收费略高于 100 亿美元。此外，CoreWeave、Nebius、IREN、TeraWulf 以及 Cipher Mining 等被称为“新云”（Neocloud）的专门化 GPU 托管商，本质上是依靠举债维系的特殊目的载体，其唯一业务就是借贷购买英伟达硬件。Zitron 警告，此类缺乏终端需求的重度杠杆操作，呈现出安然（Enron）式的财务脆弱性。盈利神话破灭与系统性损失社会化 (14:17) 针对业界常将当前亏损类比为 AWS 或 Uber 早期阶段的说法，Zitron 进行了数据反驳：AWS 在 2003 年推出至 2015 年盈利的 12 年间，通胀调整后的累计亏损总额仅为 297 亿美元；Uber 累计亏损约为 330 亿美元。而 Anthropic 一年内就筹集了 650 亿美元，OpenAI 融资承诺累计达 1,220 亿美元，以往科技基建的全部建设亏损已被单轮 AI 融资轻松吞没。不仅如此，业界预期的变现路径接连落空：推理毛利缺乏审计证实，涨价尝试遭到客户强烈抗拒，而 eMarketer 预计生成式搜索广告收入目前仅为 10 亿美元、到 2030 年仅增至 50 亿美元，对于抵消数千亿美元的算力开销微不足道。 Zitron 直指 OpenAI 是整个 AI 泡沫中的“雷曼兄弟”，其背负的 7,500 亿美元算力承诺超过了英伟达、微软和亚马逊基建承诺的总和；甲骨文为其推进的 7.1 吉瓦基建极可能导致自身陷入财务危机。更为危险的是，数千亿美元的私人信贷资金正涌入数据中心建设，而这些信贷机构通过收购保险公司和吸纳各州及市政公共养老金来获取资金。一旦发生债务违约与数据中心空置，其损失最终将被迫由普通退休人员和公共财政承担。 Zitron 总结认为，生成式 AI 领域正在兜售缺乏实际效用与持续盈利能力的软件，技术狂热与风险投资逻辑脱离了基本的单位经济学。他呼吁行业认清泡沫破裂的不可避免性，回归关注切实的投资回报率与商业价值，而不是继续依赖虚荣指标和无法兑现的增长幻想。

Detailed notes — zh

## Background 本内容由频道“Ai4”于 2026 年 8 月 18 日发布在 YouTube 平台。页面简介指出，EZPR 创始人兼科技评论作者埃德·齐特隆（Ed Zitron）在演讲中深度剖析了整个人工智能行业的虚假繁荣，揭露其如何耗费上万亿美元追逐死胡同技术、大语言模型破损的经济学逻辑、虚无缥缈的 ROI，并为公众提供了在泡沫破裂时不被波及的避坑指南。 ## TL;DR 埃德·齐特隆通过详实的数据与财务审计报告，系统性地戳破了生成式 AI 行业的虚假叙事。他指出，超大规模云厂商的大部分 AI 收入其实仅来自 OpenAI 和 Anthropic 的自我算力消耗，且全行业通过年化运行率（ARR）等财技严重掩盖了巨额亏损与缺乏真实端到端需求的事实。从底层逻辑无法闭环的按 Token 计费与补贴订阅，到动辄需要数万亿美元年收入才能回本的数据中心基建狂潮，AI 领域正演变为充斥着投机性债务的“安然式”空中楼阁。随着私人信贷大肆借道退休与养老基金为无客户数据中心输血，这场狂欢破灭后的代价最终恐将被转嫁给普通纳税人，实现恶劣的“损失社会化”。 ## Atmosphere & style 齐特隆的演讲语调犀利、直击要害且充满煽动性的批判力量，宛如在行业盛典上当头浇下一盆冰水。台风毫不留情地撕碎科技巨头的公关滤镜，以排山倒海般的财务数据和尖锐比喻步步紧逼，将听众从对技术乌托邦的盲信中猛然拽回严峻的商业现实。整场分享节奏紧凑、火药味十足，充斥着冷峻的末日警示感与对资本盲目挥霍的愤怒，带给人一种目睹雪崩前夕既震颤又清醒的强烈冲击。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) 超大规模云厂商 70% 的 AI 相关收入仅来自 OpenAI 和 Anthropic 两家公司的自身算力采购，整个行业依赖“年化运行率”（ARR）财技掩盖缺乏真实终端需求的困境。 - [01:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=108s) 过去极度扭曲的订阅补贴难以为继，大模型按 Token 计费的底层机制让企业面对无法预测的成本与持续存在的幻觉死局，ROI 根本无法量化证明。 - [04:39](https://www.youtube.com/watch?v=CldB-5PuFH4&t=279s) 拿 AWS 和 Uber 烧钱史来洗白 AI 属于逻辑偷换：Anthropic 和 OpenAI 仅单年或单轮的融资规模，就已数倍超过 AWS 经历 12 年盈利周期或 Uber 历史烧钱的总和。 - [05:39](https://www.youtube.com/watch?v=CldB-5PuFH4&t=339s) 按 Token 计费的黑盒属性使单次调用成本无法预期，智能体死循环不仅烧尽预算更可能破坏代码库，早期的 ARR 暴涨多是盲目行政指令下的“刷 Token”虚假繁荣。 - [08:13](https://www.youtube.com/watch?v=CldB-5PuFH4&t=493s) 现有 AI 订阅模式在底层算术上严重倒挂，而全球规划的 225 GW 算力基建年均需要 2.7 万亿美元收入才能回本，远超全球软件行业仅 8000 亿美元的总盘子。 - [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) 生成式 AI 真实市场年营收仅约 1000 亿至 1300 亿美元且多为内部自耗，1 GW 算力需 300 亿美元年收入回本，基础设施扩张与商业回报之间存在无法跨越的鸿沟。 - [12:02](https://www.youtube.com/watch?v=CldB-5PuFH4&t=722s) 全行业 80% 至 90% 的资本支出具有高度投机性，CoreWeave 等新型云厂商（Neoclouds）沦为举债采购 NVIDIA 算力卡的工具，重演安然式的债务绑定泡沫。 - [13:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=805s) OpenAI 与 Anthropic 深陷财务深渊，二者背负着高达 1.1 万亿至 1.3 万亿美元的算力采购长期承诺，任何风投或云巨头都无力填补其惊人亏损漏洞。 - [15:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=925s) 业内关于“推理盈利、大幅提价、AI 广告救场、停训止血”的四大商业自救神话全部破产，身负 7500 亿美元承诺的 OpenAI 已成为可能拖垮 Oracle 的“AI 雷曼兄弟”。 - [17:57](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1077s) 私人信贷将大量退休养老金与公众保险基金注入缺乏真实客户的算力基建，泡沫破裂后坏账将迫使财政兜底，最终导致全社会普通纳税人承担“损失社会化”。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) 破除 AI 泡沫神话：被“年化运行率”掩盖的虚假繁荣作为独立 AI 批评家、播客《Better Offline》主理人以及科技时事通讯《Where's Your Ed At》的作者（其订阅人数从 2025 年初的 5 万快速攀升至 11 万），埃德·齐特隆（Ed Zitron）直接拿到过 OpenAI 泄露的内部财务数据。面对满场的行业参会者，他开门见山地泼下一盆冷水：任何强迫你为 AI 行业站台、施压让你在大会期间违心说出“我当然热爱 AI”、“AI 就是未来”的人，都根本不值得信任，因为你正身处一场彻头彻尾的骗局与行业高压之中。这场 AI 泡沫已经浑浑噩噩、可悲地持续了整整四年。根据瑞银（UBS）、富国银行（Wells Fargo）和巴克莱（Barclays）的分析师估计，超大规模云厂商（Hyperscalers）70% 的 AI 相关收入竟然仅来自两家公司——OpenAI 和 Anthropic，而且主要来自于这两家初创公司自身采购的算力支出。市场上根本没有足够的真实端到端需求来支撑如此庞大的数据中心建设。为了掩盖真实营收的窘境，初创公司和云厂商普遍采用名为“年化运行率”（Annualized Run Rate, ARR）的金融财技——把单月收入乘以 12，或者把四周收入乘以 13 来拼凑出一个看似庞大的数字。云厂商之所以对具体的 AI 实际创收讳莫如深，是因为媒体总会盲目替它们贴金，鼓吹其“AI 豪赌正在获得回报”，却对它们已经烧掉上万亿美元、背负着来自 OpenAI 与 Anthropic 约 1.1 万亿美元采购承诺，以及高达 1.3 万亿至 1.6 万亿美元表外负债与债务的严峻现实视而不见。 - OpenAI 与 Anthropic 两家公司贡献了超大规模云厂商 70% 的 AI 业务收入。 - 云计算巨头背负着 1.1 万亿美元的采购承诺，以及 1.3 万亿至 1.6 万亿美元的表外隐性负债与债务。 - 行业普遍使用“年化运行率”（ARR）包装财务指标，以此规避披露真实的月度与年度实际营收。 ```mermaid flowchart TD accTitle: AI 泡沫的虚假营收循环 subgraph CapitalFlow["资本与虚假繁荣循环"] VC["风险投资 / 融资注入"] --> AIStartups["AI 头部巨头 (OpenAI / Anthropic)"] AIStartups -->|"采购巨额算力 (贡献 70% 收入)"| Hyperscalers["超大规模云厂商 (Hyperscalers)"] Hyperscalers -->|"巨额承诺与表外债务 ($1.1T - $1.6T)"| DataCenters["庞大数据中心建设"] end subgraph Narrative["公关与资本包装"] Hyperscalers -.->|"公布虚高'年化运行率' (ARR)"| Media["科技媒体与资本市场"] Media -.->|"鼓吹'AI 豪赌回报丰厚'"| VC end ``` --- ### [01:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=108s) 真实成本浮出水面：无法衡量的 ROI 与幻觉死局整个行业一直刻意回避的 ROI（投资回报率）问题在今年三四月份终于被迫摆上台面。此前，Anthropic 和 OpenAI 对企业订阅用户推行了极度扭曲的月度补贴政策：用户每月仅需支付 200 美元订阅费，就允许在 Anthropic 上烧掉高达 8,000 美元的 Token，在 OpenAI 上甚至能烧掉 14,000 美元。直到今年 3 月，两家巨头终于意识到“不可能永远用 1 美元去倒贴 40 美元”，转而要求企业客户承担真实的“每百万 Token 计费成本”，这一举措立刻引发了客户的强烈反弹。随后微软的 GitHub Copilot 也做出了类似调整——此前它以每月 39 美元的超低价允许用户消耗 1,500 到 5,000 美元的 Token 额度，提价后许多使用了一年多的老订阅用户纷纷怒斥新成本简直是“不可理喻的疯狂”。用户之所以对按 Token 真实计费感到愤怒，是因为大语言模型（LLM）从底层机制上就无法精准衡量投入产出比：无论模型最终是否输出了预期的可用结果、无论使用者是否理解其运作逻辑、甚至无论你究竟有没有切实用到 AI，只要发起请求，系统就会照单全收地扣费。Uber 首席运营官安德鲁·麦克唐纳（Andrew McDonald）近期公开坦言，公司极难证明在 AI 上的支出是合理的，因为高昂的成本根本无法与“上线实用功能”或“加快交付速度”产生直接关联。面对用户对真实成本的质疑，行业掌舵者、OpenAI 首席执行官萨姆·奥特曼（Sam Altman）除了轻描淡写地抛出一句“突然之间这变成了一个大问题”之外，根本拿不出任何解决方案。大型语言模型本质上是概率模型，幻觉问题从未被解决；如果未来几天有任何人宣称“幻觉已被彻底攻克”，那他绝对是在公然撒谎。 - 企业此前每月花 200 美元，可烧掉 Anthropic 高达 8,000 美元、OpenAI 高达 14,000 美元的 Token 额度。 - GitHub Copilot 过去以每月 39 美元的价格补贴用户消耗 1,500 至 5,000 美元的 Token。 - LLM 的概率性特征决定了它每次调用的不确定性与持续存在的幻觉，使企业无法将按量付费的 Token 成本转化为确定性的业务产出。 --- ### [04:39](https://www.youtube.com/watch?v=CldB-5PuFH4&t=279s) 拿 AWS 与 Uber 洗白的破产逻辑面对 AI 泡沫破裂的质疑，支持者最常用的经典辩护辞令就是拿历史做挡箭牌：“当年的亚马逊云计算（AWS）在盈利前也烧了海量的钱”。但如果把两者的真实财务数据摆在同一张天平上，这种历史类比就会瞬间崩塌。AWS 于 2003 年启动，历经 12 年在 2015 年实现盈利，将通胀因素计算在内，AWS 在这期间消耗的全部成本为 297 亿美元——这个数字甚至比 Anthropic 仅在今年 2 月单轮融到的资金还要少 3 亿美元。如今 AI 领域的烧钱速度早已脱离了正常的商业逻辑。Anthropic 仅在今年就募资了 650 亿美元，OpenAI 的融资规模更是高达 1,220 亿美元。即便拿曾经疯狂烧掉 330 亿美元才勉强实现惨淡盈利的 Uber 来对比，当下头部 AI 初创企业单单一轮融资的吸金体量，就能把过去科技巨头整个生命周期的总烧钱额度轻而易举地“吞噬”数次。用过去的科技神话来为如今毫无节制的资金浪费洗白，完全是一种颠倒黑白的叙事陷阱。 | 对比项目 | 历史案例：AWS（2003–2015 实现盈利） | 历史案例：Uber（实现微弱盈利前） | 现状：Anthropic（2026年单年） | 现状：OpenAI（单轮/累计融资） | |:--- |:--- |:--- |:--- |:--- | | **总资金消耗 / 募资规模** | 约 297 亿美元（已调整通胀） | 约 330 亿美元 | 650 亿美元 | 约 1,220 亿美元 | | **资本效率评价** | 建立起全球最大的云计算基础设施与稳定盈利体系 | 经历漫长补贴战后才形成微利 | 仅单年融资就超过 AWS 12 年总成本的两倍以上 | 资金规模相当于数个历史级独角兽全生命周期烧钱总和 | --- ### [05:39](https://www.youtube.com/watch?v=CldB-5PuFH4&t=339s) 混乱的 Token 经济学与“刷 Token 狂热” 基于 Token 的收费模式对于实际业务而言极其昂贵且充满不可预测性。无论是 Perplexity 还是 Replit 等建立在底层大模型之上的“套壳”（wrapper）类初创公司，之所以不敢直接向普通消费者按真实的百万 Token 费率收取全额费用，就是因为他们深知消费者绝不会为模型的失控买单——比如 Replit 用户经常遭遇模型陷入死循环、莫名其妙烧掉 50 美元 Token 却一无所获的情况。当用户抱怨时，行业往往以“现在还处于早期阶段”来搪塞。但齐特隆尖锐地反问：行业已经投入了四年时间、烧掉了超过一万亿美元，“早期”到底要持续到什么时候？究竟要烧掉多少钱才算够？按 Token 计费模式（包括开源模型自建部署在内）的致命缺陷，在于用户永远无法预知单次操作的真实开销。受不同的调用框架（harness）、提示词写法、用户使用方式甚至是模型当天的响应状态影响，输出成本完全是个黑盒；一旦智能体循环发生崩溃，不仅会在几小时内把预算在无意义的死循环中烧光，甚至可能反向污染或损坏企业原有的软件代码库。现在不少企业高管盲目要求“将 Token 预算削减 50%”，但在根本不知道如何衡量 AI 产出价值的情况下，这种削减完全是盲人摸象。过去两家巨头吹嘘的 ARR 暴涨，本质上并非来自于健康的客户留存与订阅复购，而是源于企业在“不惜一切代价实现增长”的畸形压力下盲目指示员工去“刷 Token”（token maxing）所催生的虚假繁荣。 - 按 Token 计费让单次操作的成本与质量彻底脱节，死循环不仅烧毁巨额预算，更存在破坏软件技术栈的风险。 - 企业在没有明确评估体系的前提下盲目削减 50% 预算，印证了 AI 部署缺乏量化标准的混乱现状。 - 头部模型的所谓业务增长建立在非持续性的 API 消耗浪潮与高管盲目的行政指令之上。 --- ### [08:13](https://www.youtube.com/watch?v=CldB-5PuFH4&t=493s) 补贴订阅的算术崩溃与 2.7 万亿美元的数据中心天堑目前的消费级与企业级订阅模式从底层算术上就无法闭环。以 SemiAnalysis 的测算为例，用户每月支付 200 美元订阅费，可以在 ChatGPT 上消耗价值 14,000 美元的 Token，在 Claude 上消耗价值 8,000 美元的 Token。退一步讲，即便按照 70% 的超高毛利率做出极其慷慨的假定（现实中远低于此），只要订阅用户用掉了其总额度上限的 25%，这笔订阅业务就会立刻转为亏损。市面上几乎所有现存的 AI 订阅产品在经济模型上都是完全畸形的，它们之所以能苟延残喘，完全依赖于风险投资源源不断的输血；一旦资本泡沫破裂、热钱退场，整个订阅体系将瞬间崩塌。更严峻的危机潜伏在数据中心的疯狂扩张中。所谓“全球对 AI 算力有着无休止的饥渴需求”纯属虚假炒作。目前全行业处于规划阶段的电力容量高达 322 吉瓦（GW），即便折算为实际可用于承载 GPU 和服务器的关键 IT 负载，也有约 225 吉瓦。按照行业每兆瓦（MW）1,200 万美元的建设成本计算，要收回这些硬件基础设施的投资，每年需要产生惊人的 2.7 万亿美元算力收入。然而，整个全球软件行业到 2026 年的总盘子预计也仅能勉强触及 8,000 亿美元。 ```mermaid flowchart LR accTitle: 全球软件总规模与 AI 算力所需收入的悬殊鸿沟 subgraph Gap["年度资金量对比（美元）"] A["全球整体软件行业总盘子 (2026 预期)\n$8,000 亿 ($0.8T)"] B["支撑 225 GW 算力基建所需的年收入\n$27,000 亿 ($2.7T)"] end A -.->|"存在 3.3 倍以上的巨大资金缺口"| B ``` 面对这一高达 2.7 万亿美元的荒谬数字，齐特隆直言正常人的理智都会感到震惊。雪上加霜的是，建设每一吉瓦的数据中心本身的建造成本就高达 500 亿至 700 亿美元，英伟达（NVIDIA）CEO 黄仁勋甚至公开宣称每吉瓦的造价需要 1,000 亿美元。即便未来真能凭空出现足以支撑这些基建的营收需求，单是筹措资金去完成这些数据中心的天价建设，都将把整个科技行业拖入万劫不复的深渊。 - 现有订阅机制存在严重倒挂：在假定 70% 毛利的前提下，用户仅需消耗额度上限的 25%，平台即陷入实质性亏损。 - 全球规划中的 322 吉瓦电力容量对应约 225 吉瓦的关键 IT 算力负荷。 - 维系现有规模的数据中心建设每年需要 2.7 万亿美元的算力收入，而 2026 年全球软件行业的全部市场总额仅为约 8,000 亿美元。 - 数据中心单吉瓦（GW）建设成本高达 500 亿至 700 亿美元，黄仁勋的最新估算更是达到了每吉瓦 1,000 亿美元。 ### [10:41](https://www.youtube.com/watch?v=CldB-5PuFH4&t=641s) 算力需求与商业回报的巨大鸿沟当前生成式 AI 产业正面临着极其严峻的危机：市场对 AI 算力的真实需求，根本无法满足甚至支撑哪怕四分之一的产能扩张。这绝非单纯的“末日论（Doomerism）”或纯粹的唱衰，而是一个令人不寒而栗的现实——AI 数据中心等基础设施的建设周期远比预期漫长，且经济模型根本无法自洽。要让 1 吉瓦（GW）的算力容量实现正向营收，大约需要 300 亿美元的年收入作为支撑；然而以 Oracle 为例，其为 OpenAI 在德克萨斯州阿比林（Abilene）打造的“星际之门（Stargate）”数据中心规划了 1.2 GW 算力，收费仅略高于 100 亿美元（且目前只建成了一半）。这中间存在巨大的收支缺口，整个行业根本无法产生 300 亿美元的回报。整个生成式 AI 市场的年收入规模其实小得可怜，仅在 1000 亿美元左右（即便按照极其乐观但缺乏计算常识的机构估算，也至多 1300 亿美元）。更致命的是，这笔营收中还打包了 OpenAI 和 Anthropic 等公司自身消耗的所有算力支出——这两家公司单是为了给现有客户提供推理服务，就不得不消耗数百亿美元，自身依然深陷严重亏损。 ```mermaid flowchart LR accTitle: AI 基础设施投入与产出失衡图 subgraph Cost["资本投入与成本"] A["1 GW 算力中心建设与运营"] B["数百亿美元推理与云服务成本"] end subgraph Revenue["微薄的商业回报"] C["收支平衡门槛: 300 亿美元"] D["实际收费: 约 100 亿美元"] E["全行业规模: 1000-1300 亿美元"] end A --> C A -.-> D B --> E ``` - 1 GW 算力中心要实现营收正循环，需达到约 300 亿美元的收入规模。 - Oracle 为 OpenAI 规划的 Stargate Abilene 1.2 GW 容量收费仅约 100 多亿美元。 - 整个生成式 AI 行业的年营收规模仅约 1000 亿至 1300 亿美元，且严重依赖内部计算开销支撑。 --- ### [12:02](https://www.youtube.com/watch?v=CldB-5PuFH4&t=722s) 新型云服务商与“安然式”债务泡沫整个 AI 产业目前完全是建立在风险投资、炒作与巨额债务之上的空中楼阁，其所谓的“需求故事”本质上充斥着谎言。如果算上 CoreWeave、IREN、Nebius、Cipher Mining 和 TeraWulf 等新型云服务商（Neoclouds），整个市场中 80% 至 90% 的资本支出都带有极高的投机性。这类新型云厂商的核心运作模式就是通过借债来购买 NVIDIA 的 GPU。就连黄仁勋（Jensen Huang）自己都曾直言：“没有 NVIDIA，CoreWeave 就不会存在；没有 NVIDIA，Nebius 也不会存在。”这种彼此绑定的运作机制，充斥着当年安然（Enron）丑闻崩塌前的所有典型特征。面对如此巨大的隐患，行业内部却弥漫着一种盲目的乐观主义。不论是分析师、科技巨头 CEO，还是大型银行的高管，在谈及这些风险时往往轻描淡写地回应一句“船到桥头自然直（They'll work it out）”。这让人匪夷所思，就像普通人在面临付不起房贷的危机时妄想“一切总会好的”一样荒谬。这也是为什么公众对 AI 的抵触情绪愈发强烈——除了数据盗用、污染少数族裔社区环境以及极其过分的资本开销外，更因为这些科技公司完全能在没有任何实质业务支撑的情况下随心所欲地大肆圈钱。 - 计入 CoreWeave、IREN、Nebius 等 Neoclouds 后，投机性资本支出占比高达 80% 到 90%。 - 新型云厂商普遍沦为“举债买卡”的空壳工具，极具系统性金融风险。 --- ### [13:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=805s) OpenAI 与 Anthropic 的财务深渊无论舆论如何粉饰，财务数字不会说谎：AI 巨头们正在以惊人的速度烧毁现金。《金融时报》披露并经过审计的财务文件显示，OpenAI 在 2025 年的亏损高达 209 亿美元。在其账面录得的 130.7 亿美元相关资金中，有超过 8 亿美元来自软银（SoftBank）针对一项名为“Crystal Intelligence”的模糊项目的注资，甚至无人能解释该项目究竟代表什么。更为惊人的是，OpenAI 计划到 2030 年底在算力上投入整整 7500 亿美元。当被问及资金来源时，业内只会含糊其辞地声称“继续融资”，但没有任何风投或超大规模云厂商（Hyperscalers）能填补 7500 亿美元的无底洞。 Anthropic 的财务结构也如出一辙。截至 2025 年 3 月，Anthropic 的累计终身收入仅约 50 亿美元左右，而在算力上的投入却已高达 100 亿美元，这根本不是正常企业的运作逻辑。坊间传闻 Anthropic 曾实现“单季度盈利”，其背后的真相只是埃隆·马斯克（Elon Musk）在特定季度为他们削减了 SpaceX 数据中心的算力成本——“如果不算所有成本，那我其实也能盈利”。两家公司在 Oracle、微软、亚马逊和 Google 等巨头处的长期采购承诺已达 1.1 万亿至 1.3 万亿美元。 ```mermaid flowchart TD accTitle: OpenAI 与 Anthropic 现金消耗与长期承诺 subgraph OpenAI["OpenAI 财务状况"] OA1["2025 年亏损: 209 亿美元"] OA2["2030 年算力支出计划: 7500 亿美元"] end subgraph Anthropic["Anthropic 财务状况"] AN1["终身营收: 约 50 亿美元"] AN2["算力开销: 100 亿美元"] end subgraph Commitments["全行业云与算力采购承诺"] COM["累计承诺: 1.1 万亿 - 1.3 万亿美元<br/>(涵盖 Oracle / MSFT / AWS / GCP / CoreWeave 等)"] end OpenAI --> COM Anthropic --> COM ``` - OpenAI 2025 年审计亏损达 209 亿美元，规划到 2030 年算力投入 7500 亿美元。 - Anthropic 终身营收约 50 亿美元，算力支出已达 100 亿美元。 - 头部 AI 厂商累计绑定的长期算力承诺高达 1.1 万亿至 1.3 万亿美元（包含 Cerebras 约 200 亿美元协议等）。 --- ### [15:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=925s) 击碎 AI 行业的四大商业神话业内普遍流传着关于大模型实现盈利的几大“救命稻草”神话，但逐一剖析后会发现它们根本站不住脚： 1. **“推理成本已经实现盈利”**：这完全是无稽之谈。没有任何公开确凿证据支持这一结论，仅有部分人拿着 DeepSeek 的单篇研究报告通过纸面推导自欺欺人。 2. **“可以通过提价实现盈利”**：各大厂商在 3 月份尝试提价时，客户的抗议声“简直像被群鸟活活啄死一样惨烈”，提价空间极其有限。 3. **“搜索与生成式广告能拯救营收”**：市场调研机构 eMarketer 预估，整个生成式 AI 搜索广告市场的年收入规模仅有 10 亿美元左右，到 2030 年全行业也仅能达到 50 亿美元。这笔钱连 OpenAI 在佐治亚州建设的数据中心所需 GPU 成本的十分之一都覆盖不了，在庞大的资本消耗面前只是微不足道的四舍五入误差。 4. **“可以停止模型训练来止血”**：厂商根本无法停下训练步伐。为了应对模型漂移（Model Drift），以及面对来自中国和竞品厂商的军备竞赛，预训练、后训练以及采购海量训练数据的开销永远无法避免。 OpenAI 占据了全行业 50% 的营收和 60% 的算力需求，本质上就是“AI 领域的雷曼兄弟（Lehman Brothers）”。其高达 7500 亿美元的算力承诺，甚至超过了 NVIDIA、微软和亚马逊三家巨头承诺的总和，没有任何机构能兜得住这种规模的风险。与此同时，Oracle 正在为 OpenAI 兴建高达 7.1 GW 的算力中心，但 OpenAI 既付不起账单，Oracle 自身也无力承担这笔建设成本，这甚至可能将 Larry Ellison 和 Oracle 拖入破产深渊。而 Anthropic 同样背负着 3000 亿美元级别的采购承诺，每年亏损数百亿美元，本质与 OpenAI 毫无区别。 --- ### [17:57](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1077s) 泡沫破裂与损失的社会化最令人心惊的现实在于，全社会的大众正在不知不觉中为这场 AI 泡沫买单。如今私人信贷（Private Credit）正将数千亿美元资金砸向那些根本没有真实客户的 AI 数据中心，而这些私人信贷资金的源头正是各大退休养老基金、商业保险基金以及公共养老金。 ```mermaid flowchart TD accTitle: 私人信贷与养老金风险传导链条 A["公众养老金 / 退休基金 / 保险基金"] --> B["私人信贷机构 (Private Credit)"] B --> C["缺乏真实客户的 AI 数据中心建设"] C -. 泡沫破裂坏账.-> D["地方与州政府财政救助"] D --> E["纳税人与全社会承担损失 (损失社会化)"] ``` 一旦坏账爆发、贷款无法偿还，养老和退休基金将面临巨额亏空，最终只能迫使地方政府动用财政兜底。这意味着 AI 行业的投资损失最终将被转嫁给普通纳税人，实现“损失社会化”。这一切正是任由资本野蛮扩张、盲从商业决策的结果，所有推波助澜制造这场泡沫的人在崩溃到来时都应当感到羞愧。软件行业亟需回归理性，建立在切实的投资回报率（ROI）和真正有用的工具之上；而不是像现在这样，由一群只追求眼球效应与不计代价增长的 AI 实验室，制造出大量低效、不可持续且无法盈利的代码废料。面对这个注定走向覆灭的行业，是时候戳穿谎言、直面真相了。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=CldB-5PuFH4&t=4s) 任何强迫你为 AI 行业站台、施压让你在大会期间违心说出“我当然热爱 AI”、“AI 就是未来”的人，都根本不值得信任，因为你正身处一场彻头彻尾的骗局与行业高压之中。 — Ed Zitron > [01:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=108s) 不可能永远用 1 美元去倒贴 40 美元 — Ed Zitron > [01:48](https://www.youtube.com/watch?v=CldB-5PuFH4&t=108s) 大型语言模型本质上是概率模型，幻觉问题从未被解决；如果未来几天有任何人宣称“幻觉已被彻底攻克”，那他绝对是在公然撒谎。 — Ed Zitron > [05:39](https://www.youtube.com/watch?v=CldB-5PuFH4&t=339s) 行业已经投入了四年时间、烧掉了超过一万亿美元，“早期”到底要持续到什么时候？究竟要烧掉多少钱才算够？ — Ed Zitron > [12:02](https://www.youtube.com/watch?v=CldB-5PuFH4&t=722s) 没有 NVIDIA，CoreWeave 就不会存在；没有 NVIDIA，Nebius 也不会存在。 — 黄仁勋 (Jensen Huang) > [13:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=805s) 如果不算所有成本，那我其实也能盈利 — Ed Zitron > [15:25](https://www.youtube.com/watch?v=CldB-5PuFH4&t=925s) 这笔钱连 OpenAI 在佐治亚州建设的数据中心所需 GPU 成本的十分之一都覆盖不了，在庞大的资本消耗面前只是微不足道的四舍五入误差。 — Ed Zitron > [17:57](https://www.youtube.com/watch?v=CldB-5PuFH4&t=1077s) 软件行业亟需回归理性，建立在切实的投资回报率（ROI）和真正有用的工具之上；而不是像现在这样，由一群只追求眼球效应与不计代价增长的 AI 实验室，制造出大量低效、不可持续且无法盈利的代码废料。 — Ed Zitron ## Mentioned Ed Zitron, Better Offline, Where's Your Ed At, EZPR, OpenAI, Anthropic, 瑞银 (UBS), 富国银行 (Wells Fargo), 巴克莱 (Barclays), GitHub Copilot, Microsoft, Uber, Andrew McDonald, Sam Altman, AWS, Perplexity, Replit, SemiAnalysis, NVIDIA, 黄仁勋 (Jensen Huang), Oracle, Stargate, CoreWeave, IREN, Nebius, Cipher Mining, TeraWulf, 安然 (Enron), 金融时报 (Financial Times), SoftBank, Crystal Intelligence, SpaceX, Elon Musk, Google, Cerebras, DeepSeek, eMarketer, 雷曼兄弟 (Lehman Brothers), Larry Ellison ## Video description Ed Zitron - Writer, EZPR What if everybody was wrong about AI? Ed Zitron walks you through how the entire tech industry got diverted by generative AI, burning over a trillion dollars on a dead-end technology, the broken economics of LLMs, the questionable ROI of AI, and what you're going to want to do to avoid getting burned when the bubble pops. [Source](https://www.youtube.com/watch?v=CldB-5PuFH4)

Source quotes

[0:04] Ed Zitron: 任何强迫你为 AI 行业站台、施压让你在大会期间违心说出“我当然热爱 AI”、“AI 就是未来”的人，都根本不值得信任，因为你正身处一场彻头彻尾的骗局与行业高压之中。 [1:48] Ed Zitron: 不可能永远用 1 美元去倒贴 40 美元 [1:48] Ed Zitron: 大型语言模型本质上是概率模型，幻觉问题从未被解决；如果未来几天有任何人宣称“幻觉已被彻底攻克”，那他绝对是在公然撒谎。 [5:39] Ed Zitron: 行业已经投入了四年时间、烧掉了超过一万亿美元，“早期”到底要持续到什么时候？究竟要烧掉多少钱才算够？ [12:02] 黄仁勋 (Jensen Huang): 没有 NVIDIA，CoreWeave 就不会存在；没有 NVIDIA，Nebius 也不会存在。 [13:25] Ed Zitron: 如果不算所有成本，那我其实也能盈利 [15:25] Ed Zitron: 这笔钱连 OpenAI 在佐治亚州建设的数据中心所需 GPU 成本的十分之一都覆盖不了，在庞大的资本消耗面前只是微不足道的四舍五入误差。 [17:57] Ed Zitron: 软件行业亟需回归理性，建立在切实的投资回报率（ROI）和真正有用的工具之上；而不是像现在这样，由一群只追求眼球效应与不计代价增长的 AI 实验室，制造出大量低效、不可持续且无法盈利的代码废料。

Source quotes

[0:04] Ed Zitron: well, of course I love AI [1:50] Ed Zitron: can't keep giving away $40 for a dollar. [10:41] Ed Zitron: You ain't getting $30 billion, mate. [10:41] Ed Zitron: who should be ashamed of themselves; they can't count [10:41] Ed Zitron: quick visit to Enronville [10:41] Ed Zitron: they'll work it out. [14:17] Ed Zitron: I would be profitable too if it wasn't for all the costs. [14:17] Ed Zitron: screamed like they were being pecked to death by birds [16:48] Ed Zitron: Lehman Brothers of AI [16:48] Ed Zitron: Elizabeth Holmes voice [16:48] Ed Zitron: We are going to subsidize this whether we like it or not. [16:48] Ed Zitron: It is time to call bullshit on the AI industry. [16:48] Ed Zitron: unproductive, ineffective, unprofitable, unsustainable pieces of software

Original transcript

[0:04] What's up everyone? This is me. I'm an independent AI critic and I'm going to [0:08] say a bunch of stuff that no one else is saying. I want to open by saying something. [0:12] Anyone pressuring you to back the AI industry is not to be trusted. And when I say this, I mean [0:18] if you feel the pressure to say at any point in this conference, well, of course I love AI. [0:23] Well, of course I think AI is the future. know that you are in the presence or the pressure of [0:28] some kind of con. I'm the host of Better Offline and I'm the author of the Where's Your Addict [0:32] newsletter. 110,000 subscribers was 50,000 at the beginning of 2025. It's growing fast. Oh yeah, [0:39] also I got OpenAI's leaked financials, so I really know what I'm talking about. [0:43] Now, let's talk about the AI bubble four boring sad years in, shall we? [0:48] Open AI and anthropic based on analyst estimates from UBS, Wellelss Fargo and Barclays make up 70% [0:55] of all hyperscala AI revenues. We do not have the demand to substantiate the data centers [1:01] and startups and hyperscalers are all engaging in a kind of financial swindle that some of you might [1:06] know about called annualized run rates. That's month x* 12, 4 weeks time 13 or whatever you want [1:12] it to be. And anytime you hear someone mention run rate, it's because they don't want you to [1:16] know their real revenues. Funnily enough, along with that, hyperscalers don't want to tell you [1:21] how much they're making from AI because, well, the media gives them credit. They say, "Wow, [1:25] their AI bets are paying off even as they refused to share why they've spent over a trillion dollars [1:31] or taken on $1.1 trillion of commitments from OpenAI and Anthropic and basically nobody else. [1:38] And there's of course that 1.3 1.6 $6 trillion of offbalance sheet, well liabilities and debt [1:44] you could call it. So let's go on to the most important thing, the ROI conversation. You [1:50] probably heard a little bit of this recently and that's because back in March, April of this year, [1:55] Anthropic and OpenAI after having their enterprise subscribers on a kind of monthly subscription plan [2:01] where you can burn as many tokens as you like up to eight $8,000 on the 200 buck a month for [2:06] Anthropic and 14 grand on Open AIS. Well, they allowed the enterprises to do that all [2:11] they want up until March when they said, "Okay, you're going to pay the real cost of AI." Well, [2:17] I don't think people were very happy. And when Microsoft followed example with GitHub Copilot, [2:23] which allowed people for 39 bucks a month, kind of a bargain, to burn 1,500 to $5,000 a month in [2:29] tokens. Well, check out what this guy thinks about the real costs. And just to be clear, he's saying [2:35] here he's been a C-pilot Pro subscriber for a year and he thinks the new costs are insane. [2:41] The new costs he's referring to are the actual costs, the per million token costs of AI. [2:47] And people hate paying these costs because they pay whether the model does it does what they want [2:51] or not. And they pay regardless of whether they actually understand what they're doing. [2:56] And they pay whether or not they actually really know whether they're using AI at all. [2:59] And that's because there's societal and industry pressure to use AI regardless of its outcomes. [3:04] And you cannot measure the ROI of large language models. And for the first time, we're having this [3:10] conversation. We should have had it years ago, but we're having it now because Anthropic and OpenAI [3:15] realized that they can't keep giving away $40 for a dollar. And Andrew McDonald, the COO of Uber, [3:22] recently said that he's having trouble justifying the cost of AI because it's hard to connect to, [3:28] I don't know, useful features, shipping things faster. And I want to be clear, if you're in here [3:33] and you're hearing this and this is frustrating or upsetting you, I'm not here to be mean to you. I'm [3:38] here to illuminate you and to start a conversation that's necessary because we are in a bad place [3:44] and the tech industry is too. And right now, this conversation is going to get worse and [3:49] worse and worse. And if you're an AI company, you likely have subsidized subscriptions, too, [3:54] and know that this is already a problem. Now, thankfully, of course, our industry at the [3:59] helm of the industry is Clammy Sam Alman, the CEO of OpenAI. Thank God for Sam Alman. Sammy, [4:04] what do you have to say? Well, all of a sudden, it's a huge issue. Brilliant. Thanks, mate. That's [4:10] exactly what we needed. Here's another hundred billion dollars, Matt. This is what Sam Wolman had [4:15] to say to sell people on paying the actual cost of AI. And that's because no one has an answer. [4:21] Large language models are probabilistic. They still hallucinate. If anyone says to [4:25] you in the next few days that hallucinations are a solved problem, they are lying to you. [4:29] It's not even an opinion. They're still there. Go use Google. Even use the f the frontier models. [4:35] They still hallucinate. It's an absolute joke. Now, there are lots of counterarguments to the [4:40] AI bubble narrative. You're going to hear them all over the place. You've heard them, I'm sure. [4:44] AWS cost a lot of money, right? Amazon Web Services. The classic justifier for all of [4:51] this waste is that Amazon Web Services cost a lot of money before it became profitable. [4:56] In 2003, Amazon Web Services launched and in 2015 it became profitable. And adjusted for inflation, [5:02] it cost $29.7 billion, also known as $300 million less than Anthropic raised in February. [5:10] Anthropic has raised $65 billion this year and OpenAI has raised up to $122 billion. Who knows [5:15] when the trenches come in, but nevertheless, even compared to Uber, which burned $33 billion [5:20] before becoming a messy kind of profitable again, you can just drop the entire burn rate [5:26] of the examples that people use to justify AI into a single funding round, sometimes multiple times. [5:32] And it's frustrating because so much money has been burned as a result of this mythology. Now, [5:39] we need to be clear about how bad the economics are, too. Tokenbased billing is very expensive. [5:45] It's very expensive. And I think anyone doing a rapper of any kind of a for any kind of AI [5:49] startup knows this. The reason that no one is charging the actual token rates for things like [5:54] Perplexity or Replet or what have you is because they know regular consumers are not going to pay. [5:59] You know that they're not actually going to pay when a I don't know, Replet users probably [6:04] know this one. when it goes in an insane loop and decides to burn $50 worth tokens [6:09] and then you're told, "Well, that's just that's just how AI works." Well, it's early days, right? [6:14] It's early days. It isn't early days. We're four years and over a trillion dollars into this. We're [6:19] anything but early days. And if you're saying early days, my only question is when do those end? [6:25] How much is enough? Or how much is too much? These are questions we should have been having [6:30] years ago. Sorry, conversations we should have been having years ago. We need to have them now. [6:36] But the big problem with large language models served at per million token rates, the same [6:41] with open source, by the way, is you cannot actually anticipate how much an action costs. [6:47] You don't know based on the harness, based on the prompt, based on the user, based on what the model [6:52] feels like that day, what you're actually going to get out the end. You don't know if you're running [6:57] loops whether one of those loops is going to break and it's going to burn a few hours worth of tokens [7:01] on nothing. or to make matters worse, actually damaging the software stack you're working on. [7:06] And thus, it's hard to make a budget. You've probably heard now that people are cutting [7:10] their token budgets by 50%. The problem with that is that how do you cut something by 50%. [7:17] If you don't know how to measure what good looks like, the answer is you don't. And we haven't [7:23] heard much from Anthropics ARR in a while. Sorry, I mean run rate, annualized recurring revenue, [7:28] even though they're basing that off of API spend, which isn't recurring. And I think that's because [7:33] most of their revenue growth, same with OpenAI, came from this rush of token maxing. From people [7:38] spending tokens arbitrarily without thinking about it or being told by business idiot bosses [7:44] that you need to spend thousands of dollars on AI a month. That's the only way we'll go forward. [7:49] There is an industrialsized pressure on each and every person in this room and [7:53] most of the people in the economy to adopt AI services with no regard for what they can do, [7:58] what their outcomes are, or the actual costs. It is a con. It is a grift. And it is a result of a [8:04] tech industry obsessed with growth. Growth at all cost thinking is driving us into a godamn void. [8:13] So let's talk about the subscriptions. Now I imagine quite a lot of you have clawed or chat [8:18] GPT subscriptions. know that those are probably going to go away one day because as I mentioned [8:23] earlier semi analysis found that on the 14 you can burn 14 gram worth of tokens on chat GPT 200 bucks [8:30] 8,000 on clawed now I'll get to the mythology about that soon but just let's be clear they [8:35] also found that even if the gross margins were 70% they are not just really clear they're not [8:42] all it took was a user using 25% of their utilization to make it in the red [8:47] and this is making an agree egregious assumption about the economics. This does not make sense. No [8:53] AI subscription makes sense. Yet, it's most of the ways that people interact with and buy software. [9:00] And it's deeply worrying that we still have these subscriptions going. The only reason [9:04] we do is endless venture capital. Venture capital that will run out when the air leaves the bubble. [9:10] And just let's be just clear, I've got far worse news. Let's talk about data centers. [9:16] So, you've probably heard phrases like there's insatiable demand for AI compute, all useful [9:21] compute will be taken. That just isn't true. As I mentioned earlier, UBS, Wells Fargo, Barclays, [9:28] 70% of hyperscaler AI revenue comes from OpenAI and anthropic and predominantly from their compute [9:34] spend. That's only at best in a few years going to be a few hundred billion. We have 322 gawatt of [9:42] capacity under planning right now. Don't know how much that is under construction, but nevertheless [9:47] translates into about 225 gawatt of useful critical IT. So the GPUs, the servers they're [9:53] in and that all that and the other. At $12 million a megawatt, that means we need $2.7 trillion of [9:59] annual annual compute revenue. To give you some context for that, the entire software industry, [10:06] the entire global software industry will barely scratch $800 billion in 2026. [10:13] Now I ain't no mathematician, right? But that's a lot smaller than 2.7 billion uh trillion. See, [10:20] even trying to say 2.7 trillion, my mind's like, "Nah, mate. It couldn't be that stupid. [10:25] Couldn't possibly be that. We want to build that." And each gigawatt of this costs 50 billion to 70 [10:30] billion. And Mr. Jensen Hong recently said, "$und00 billion a gigawatt." So even if this [10:36] revenue existed, they needed to build the capacity and it's going to cost all this money to build it. [10:41] And then hopefully the companies are left over at the end. But as I've discussed, [10:44] there actually isn't that much demand for AI compute. At least not enough to even remotely [10:49] satiate a quarter of this. This is a crisis. It's a genuine crisis. And people frame me as a doomer [10:56] and a hater. And perhaps they're right on the last part. But this isn't me being a doomer. [11:00] I don't love this. This is actually quite terrifying to me because when you actually [11:06] look at how difficult these things are to build, they're taking so much longer than you'd expect. [11:12] And even when they're built, you need about $30 billion of revenue to make a gigawatt of compute [11:17] capacity actually revenue positive, which is a real problem cuz Google, sorry, Oracle [11:22] is charging, I think, a little over 10 billion for the 1.2 2 GW of capacity that's in Stargate [11:29] Abene for Open AI of which only half is built but nevertheless you ain't getting 30 billion mate [11:35] and the overall industry is kind of pathetically small 100 billion of annual revenue maybe 130 if [11:41] you're feeling real sexy with it like exponential view who should be ashamed of themselves they [11:45] can't count and that also bundles in every ounce of compute spend from open AI and anthropic [11:51] we're talking tens of billions of dollars just to serve inference for the customers they have [11:56] which is still unprofitable doesn't really make sense. So we have this entire industry [12:02] that's propped up by venture capital hype and debt and the demand story is fundamentally align [12:10] and I say 80% here the 70% number I referred to earlier referred specifically to hyperscalers [12:16] when you bring in coreweave iron nebus cipher mining terowolf it's 80 if not 90%. [12:23] All of these Neoclouds are going to zero. They are vehicles that exist to raise debt, [12:28] to buy Nvidia GPUs. Jensen Huang himself even said, "Without Nvidia, Coreweee wouldn't exist. [12:33] Without Nvidia, Nebius wouldn't exist." These are all the hallmarks of a quick visit to Enronville. [12:40] And we're all sitting here kind of just going, "It'll be all right. What will happen?" I talk to [12:44] analysts. I talk to CEOs. I talk to people every day. Talk to people at big banks and they say, [12:49] "It'll work out. Christ, I wish I could live my life like that. Can I make my mortgage this week? [12:55] You know, it'll be all right, mate. And I think that this is why so many people are anti-AII. [13:00] That and the stealing and the poisoning black neighborhoods and the egregious data center [13:05] capex is the fact that these people can raise as much money as they want based on nothing. [13:11] And the crazy thing is is that when you bring this up to an alarming amount of reporters, they go, [13:17] "They'll work it out." In fact, that's what everyone will tell you. They'll work it out. [13:21] And perhaps they're right, you know, perhaps I'm just completely wrong. [13:25] But these companies lose so much money. And I'm definitely not wrong. Look at how much [13:29] cash they lose. It's so insane. $20.9 billion loss for OpenAI in 2025. Financial documents [13:39] seen by me in the Financial Times. That is not a that's not doomerism. That's just what it looks [13:45] like. And those were audited, by the way. And here's the other sad thing. $13.07 07 billion, [13:51] but over 800 million of that came from SoftBank for something called Crystal Intelligence, which [13:57] no one can tell you what it means, and I've asked. And this company plans to spend $750 billion [14:04] in compute by the end of 2030. Again, when you ask people where's the money going to come from? Oh, [14:10] they'll just raise it. Okay, but but but where from? Yay, venture hyperscalers, I guess. [14:17] $750 billion is what they need. And the thing is when you look at Anthropic, they're much the same [14:23] company. As of March, Anthropic made, I think it was around $5 billion in lifetime revenue, [14:29] spent 10 billion on compute. These are not real businesses. And if you've read, by the way, [14:34] that Anthropic had a profitable quarter. Bad news, folks. The only reason that they had [14:38] a profitable quarter is Mr. Elon Musk reduced the cost of their SpaceX data center capacity. [14:44] And uh yeah, just for the magical quarter where they happen to be profitable. You know, [14:48] it's kind of like if I I would be profitable too if it wasn't for all the costs. [14:57] And talking of costs, um they have $1.1 trillion of commitments across Oracle, [15:02] Microsoft, Amazon, and Google. These actually go up to much like I think it's like 1.3 trillion [15:07] when you include corewave, iron, nebas by proxy of course, uh cerus, $20 billion. [15:13] just throwing out numbers at this point. I could just be saying name, random names, [15:17] but these are real. This is from the information. Again, you talk to people about this and they go, [15:21] "Ah, you know, they'll work it out, will they?" Now, but let's start for some myths [15:25] cuz I love myth busting. Open pardon me. I know how to say the word. I promise. Open AI [15:32] and anthropic are profitable and inference. This is a myth. No one has proof of this. [15:36] Someone has a single document from the deepseek. I think it's the deepseek research. They say, "Well, [15:41] based on this and based on my calculations, I've done their problem." There's no proof of this. [15:45] Oh, they'll they'll just raise prices to reach profitability, right? Yeah, they did that. They [15:50] did that in March, and their customers scream like they were being pecked to death by birds. [15:55] Oh, well, ads will save them. Bad news, folks. Bad news. Earketer said that the estimate for [16:00] the entire industries ad revenue from generative AI search ads was a billion dollars. That's not [16:08] even going to cover what, like a quarter, no, like a tenth of the GPUs for the Georgia data center [16:14] that OpenAI is working on. And they're expecting $5 billion again for the entire industry by 2030. [16:21] It's a rounding error. It's a footnote. Open AI will be quite dead by then. But nevertheless, [16:25] they're not going to do it. And of course, the argument they can just stop training models. [16:29] No, they cannot. Just they can't. China will do it. Anthropic will do it. Open AI will do it. [16:35] Someone will do it. They have to keep training to keep up with model drift as well. And they [16:39] they have to pre-train. They have post- train. There's so much training. Then there's the [16:42] cost of the training data too. There's no way out of this. And this cannot continue. [16:48] Open AI is the layman brothers of AI. They account for more than half of all AI compute demand and [16:53] 60% of sorry 50% of revenues, 60% of compute demand. And it those $750 billion in compute [17:00] commitments they cannot afford. No one can. Microsoft, if Microsoft had those, they could not. [17:07] In fact, I'm pretty sure $750 billion is larger than the over overall commitments of Nvidia, [17:14] Microsoft, and Amazon combined. And then there's this Oracle problem. Now, [17:19] I know I hear people laughing. Don't laugh at Larry Ellison. He's such a lovely guy. We [17:25] Larry Ellison may actually bankrupt Oracle. I know this sounds extreme, but they're building [17:30] 7.1 gawatt of capacity for open AAI that open AAI cannot afford and Oracle can't afford to build. [17:36] It's a bad situation for everyone. And by the way, Anthropic is much the same company. [17:41] Other than the fact we have this weirdo Wario Amade over there. Other than the fact we have to [17:45] hear a guy do Elizabeth Holmes voice and say we're nearing the exponential. They're the same company. [17:51] $300 billion worth of commitments losing tens of billions of dollars of a year. They're the same. [17:57] And then there's the sad part. As if this was happy, I realize we are funding the AI bubble. [18:05] Private credit is sinking hundreds of billions of dollars into AI data centers for customers that [18:10] do not exist. And they're funding it by buying retirement and insurance funds and working with [18:17] public pension funds. Go and look at pitchbook. Go and just do any basic research here. [18:23] So we are going to see the losses socialized because when or if these retirement funds or [18:28] pension funds or insurance funds are left kind of with nothing at the end of this from these loans, [18:35] they're going to go to the local state funds. We are going to subsidize this whether we like it or [18:39] not. And this is terrifying to me. And I think everyone who inflated the AI bubble when this [18:45] explodes should feel nothing but shame because this is what happens when you allow capitalism [18:50] to explode at this rate. And this is what happens when you base everything based on what business [18:56] idiots do. It is time to call [ __ ] on the AI industry. I am I mean you no harm, [19:03] but I'm a trutht teller and I want the world to be a better place. And it starts [19:07] with building a software industry on meaningful returns on investment and actual useful tools. [19:14] Instead, we have these unproductive, ineffective, unprofitable, unsustainable pieces of software [19:20] run by AI labs that have no interest in anything other than attention and growth at all costs. [19:26] And in the end, I think this industry is doomed. If you disagree, that is my email [19:31] address. Shoot me an email. If you've got a Bloomberg terminal, hit me up on IB. I'll be [19:36] taking all comers for 10 minutes afterwards. Thank you so much for your time, everyone.