Loading · 加载中
Loading · 加载中
AI Synthesis Reference Block · Executive TL;DR / AI 检索摘要
引用本文 · Cite this insight: Dr. Tong Yin (2026-08-09). Running Before Learning to Walk: The Data Behind the AI Industry's Two Deadlocks / 《还没学会走就想飞：数据揭示当前 AI 产业的两个死结》. InsightBridge Global Intelligence. https://intelligence.insightbridge.global/articles/running-before-learning-to-walk-the-data-behind-the-ai-industrys-two-deadlocks — Series: deep-analysis
In 2026, on stages across Silicon Valley, executives from Meta, OpenAI, and Google take turns proclaiming that Artificial General Intelligence (AGI) is about to rewrite everything, that AI will replace human labor wholesale. Wall Street buys the story. The media amplifies it. Valuations soar.
Yet the moment you strip away the smoke and mirrors—generated by exorbitant “mercenary scientists” and endless venture capital—and confront the ground-level commercial reality, a starkly opposite conclusion emerges: the current AI industry is trapped in a dysfunctional loop of trying to run before it has learned to walk.
This is not a sentiment. It can be verified with two cold sets of facts:
Below, using data and facts, I unpack these two deadlocks one by one.
To catch up with OpenAI, Meta has offered $100 million-level compensation packages to poach top Chinese scientists. Headline-shocking, but here is the real structure:
Table 1: Anatomy of a $100M Compensation Package
| Component | Share | Vesting | Realized at 1.5 yr | Realized at 4 yr |
|---|---|---|---|---|
| Base salary | 9% | Monthly cash | $4.5M | $9M |
| Signing bonus | 25% | Locked after 12 mo | $25M | $25M |
| Restricted stock (RSU) | 66% | Quarterly over 4 yr | $24.75M | $66M |
| Total | 100% | — | ≈ $54.25M | ≈ $100M |
| Unvested / forfeited | — | — | $45.75M | $0 |
Source: standard Silicon Valley Base + Signing + RSU structure, uniform vesting over four years.
The industry’s real data:
A scientist walks away from Meta after 1.5 years with roughly $54 million, forfeits the remaining $40M+ in unvested stock—then joins the next giant, which covers the forfeited stock with an even bigger “buyout signing bonus.” They are the top mercenaries of the tech world, jumping every two years, forever cashing out mid-flight.
Wonderful for the individual scientist. Catastrophic for the industry:
Under this continuous burn of headline compensation and massive compute, here is what 2025–2026 actually looks like:
Table 2: OpenAI 2024 vs 2025 Audited Financials (Leaked)
| Line item | 2024 | 2025 | YoY change |
|---|---|---|---|
| Revenue | $3.7B | $13.07B | +253% |
| Total cost & expense | $12.48B | $34B | +172% |
| R&D | $7.81B | $19.18B | +146% |
| Sales & marketing | $1.11B | $5.73B | +416% |
| Operating loss | $8.78B | $20.92B | +138% |
| Net loss attributable | $5.09B | $38.53B | +657% |
| Expense per $1 of revenue | $2.37 | $1.60 | Down slightly |
Source: Ed Zitron / Financial Times / Fortune / Ars Technica, June 2026. Note: the $38.53B 2025 net loss includes $41.55B in one-time non-cash charges from the nonprofit-to-for-profit conversion; the comparable cash loss is closer to $8B. The $20.92B operating loss is the cleanest reflection of ongoing operations.
Table 3: Top AI Labs P&L Comparison, 2025–2026
| Company | 2025 revenue | 2026 est. revenue | 2026 est. operating P&L | Profit outlook |
|---|---|---|---|---|
| OpenAI | $13.07B | \~$30B | –$14B (GAAP up to –$25B) | No positive cash flow expected until 2029–2030; 2028 operating loss projected at $74B |
| Anthropic | $9B (year-end ARR) | $47B (May ARR) | +$559M (Q2 first profit) | Only marginally profitable foundation lab; deeply dependent on AWS and Google compute subsidies |
| Google DeepMind | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by search ads |
| Microsoft AI | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by Windows/Azure |
| Meta AI (FAIR + GenAI) | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by advertising |
Source: Fortune, Financial Times, CNBC, Mental Momentum Research, June–July 2026.
Table 4: The Hyperscaler AI Capex Arms Race
| Company | 2025 Capex | 2026 Capex guidance | YoY | % to AI |
|---|---|---|---|---|
| Amazon (AWS) | \~$105B | \~$200B | +90% | \~75% |
| Microsoft | \~$90B | \~$190B | +111% | \~75% |
| Alphabet (Google) | \~$100B | $175–190B | +80% | \~75% |
| Meta | \~$72B | $125–145B | +90% | \~90% |
| Four combined | $410B | ≈ $725B | +77% | \~75% to AI |
Source: Goldman Sachs, S&P Global, Yahoo Finance, Gate.com, June–July 2026. CreditSights estimates \~$450B of 2026 hyperscaler capex flows directly into AI infrastructure.
Goldman Sachs further projects combined hyperscaler capex of $5.3 trillion over 2025–2030, with industry-wide compute/data-center/power spending expected at a baseline $7.6 trillion between 2026 and 2031. This is the largest single-industry capital gamble in human commercial history.
Table 5: Who Actually Makes Money in This Chain?
| Position | Representative | 2025–2026 net profit (est.) | Net margin | Who pays them |
|---|---|---|---|---|
| Upstream: GPU chips | Nvidia | $120–150B | 45%+ | Every downstream model company |
| Upstream: Wafer foundry | TSMC | $40–50B | 40%+ | Nvidia + every AI-chip designer |
| Midstream: Model labs | Anthropic | +$559M (Q2) | \~5% | Enterprise API customers |
| Midstream: Model labs | OpenAI | –$21B (operating) | –160% | Microsoft / SoftBank / private funds |
| Downstream: Ad platform | Meta (legacy) | +$70B+ | 30%+ | 3B users’ ad conversions |
| Downstream: Search/Cloud | Google/Microsoft | +$100B+ each | 25%+ | Legacy businesses, not AI div. |
| Net flow direction | → Nvidia + TSMC | ≈ $170B+ | Net inflow | All downstream losses end up as upstream revenue |
Source: company 2025–2026 filings and analyst estimates.
The one-sentence takeaway: every dollar raised by downstream model companies ends up, unchanged, in Nvidia’s checking account.
When the vast majority of downstream companies in an industry are burning $1.60 for every $1 they earn, kept alive by the next round of capital injection, and the only steady winner is the upstream hardware seller—that structure has never once held up in business history.
Venture capital is not infinite. Wall Street and sovereign funds are already demanding that these companies produce genuine, industrial-grade, profitable products within 1 to 2 years—or funding channels will be shut.
When patience runs out, if this industry still cannot deliver a true industrial-grade product, the bubble will burst faster than it inflated.
Burning money is not the deadliest problem. The deadliest problem is that these expensively-burned models have not reached true industrial-product grade. This is the root reason downstream enterprises are exiting en masse—and why the entire AI enterprise cannot scale.
Anyone who has actually used ChatGPT, Claude, or Gemini on real business tasks runs headlong into the same wall:
On microscopic details, hard logic, data verification, and closed-loop code execution, these models are riddled with holes, error out at random, and confidently fabricate lies.
The simplest test: ask an AI where one of your articles has been cited. It will hand you a plausible-sounding title, author, journal, and even a fabricated URL. You check. None of it exists. Academics call this “hallucination.” In business and law, this is outright deception.
If a system cannot manage the most basic “either it exists or it does not” verification, how could it possibly be trusted with a ten-million-line national security system, a bank settlement engine, or a power-grid controller?
The technical truth is brutal: today’s large language models are, at heart, probability-based text-completion engines, not logic engines. They have no genuine causal reasoning, no common sense, no self-correction. Patch a mistake at point A today, and the model will invent a stranger mistake at point B tomorrow. Under the current Transformer paradigm, no matter how large the parameter count or how vast the compute, the accuracy ceiling may forever hover between 85% and 90%.
Table 6: Commercial Viability by Accuracy Rate
| Accuracy | Errors per 100 ops | Commercial viability | Which industries can use it |
|---|---|---|---|
| 60–70% | 30–40 | Unusable | None |
| 80–85% | 15–20 | Only for “assisted suggestions,” not automation | Draft customer replies, marketing copy |
| 85–90% (current AI ceiling) | 10–15 | Cannot close the loop; needs full human audit | Entertainment, chat, creative assist only |
| 95% | 5 | Barely usable as pipeline assistant | E-commerce recommendations, simple sorting |
| 99% | 1 | Non-critical automation | Inventory, basic reporting |
| 99.9% (industrial floor) | 0.1 | Can carry core business | Finance, legal, medical, engineering |
| 99.999% (aviation grade) | 0.001 | Safety-critical systems | Aviation, nuclear, missile |
Source: ISO 9001, IEC 61508 industrial quality standards; industry synthesis.
Why 85% equals zero:
Table 7: 85%-Accuracy AI vs a Junior Intern on the Same Job
| Dimension | 85% AI | Junior human intern |
|---|---|---|
| Accuracy | 85% (random) | 95%+ (errors clustered in unfamiliar areas) |
| Behavior when wrong | Utterly confident, unaware, lies with a straight face | Blushes, stops to ask when unsure |
| Common-sense judgment | None | Innate |
| Situational adaptability | Only within training distribution | Improvises live |
| Accountability | Terms of service disclaim: “risk borne by user” | Apologizes and corrects |
| Cost of a single error | Requires senior audit | Can self-correct |
| Net utility | Actually negative | Positive |
Table 8: Enterprise AI Failure Rates Across Independent Studies
| Institution | Date | Sample | Key finding |
|---|---|---|---|
| MIT NANDA | Late 2025 | 300+ enterprise GenAI pilots | 95% produced no measurable P&L impact |
| MIT NANDA | Late 2025 | Global enterprise cumulative spend | Enterprises spent $30–40B on GenAI, near-zero return |
| S&P Global | Early 2026 | 200+ large enterprises | Enterprises killed 46% of AI PoCs before production |
| CIO Survey | April 2026 | Multinational executives | 48% openly admit “massive disappointment” (up from 34%); 75% admit AI deployment is “just for shareholder theater” |
| Gartner | July 2026 | Global customer service | 85% of enterprise AI agents being dismantled, human-first restored |
| IDC | Mid 2026 | Enterprise AI agent pilots | Production failure rate: 88% |
| Deloitte Tech Trends 2026 | 2026 | Enterprise AI deployments | Landing failure rate: 89% |
Source: MIT NANDA “State of AI in Business 2025”, S&P Global, Gartner July 2026 CX report, Deloitte Tech Trends 2026, IDC 2026 mid-year. Three independent measurements converged on the same 85–95% failure band within a single year.
Table 9: Why Enterprises Are Exiting — Root Cause Attribution
| Reason for exit | Share of enterprises citing it | Concrete manifestation |
|---|---|---|
| Accuracy failures / hallucinations | 42% | Bots talk nonsense; generated code full of bugs |
| Cost-benefit inversion | 28% | Subscription + audit cost > manual cost |
| Accountability vacuum | 15% | Nobody responsible for errors; legal risk high |
| Integration difficulty | 10% | Doesn’t fit existing ERP / CRM |
| Staff resistance / learning curve | 5% | Front-line efficiency actually drops |
Source: S&P Global 200+ enterprise study + MIT NANDA synthesis, 2026.
In one sentence: Enterprises tasted the crab. Two years in, they can neither swallow nor spit it out.
Normal business logic would say: if a product’s accuracy is only 85%, stop and take it to 99.9% before launching. Why has the entire industry done the opposite?
Because the capital model forces them to:
The result is a darkly ironic, guaranteed-to-collapse loop:
Table 10: Two Diametrically Opposite Technology Paths
| Dimension | Nvidia path (20 years of grinding) | Model-giant path (2-year cash-out) |
|---|---|---|
| Turning point | 2006 CUDA launch, nobody used it, stock dropped 70% | 2020 GPT-3 overnight fame |
| Core promise | “Fastest, most accurate, never crashes” | “AGI is coming, we’ll change everything” |
| Output determinism | 100% (industrial grade) | 85–90% (probability machine) |
| Team stability | Core engineers, 20-year tenure | Top scientists, 1.5–2.5 yr hop |
| Product maturity | Bound to every AI scientist alive | No model reaches industrial grade |
| Financial performance | 2025–2026 net profit $120–150B | OpenAI: annual operating loss $20.9B |
| Net margin | 45%+ | –160% |
| Moat strength | Nearly irreplaceable (CUDA ecosystem) | Highly interchangeable |
This is the wisdom of the shovel seller: most gold-diggers die on the road because their products are broken and unprofitable, but the person selling them water and sturdy shovels along the way always wins.
Any technology that cannot stand firmly on the microscopic details will collapse in the macroscopic market test. The steam engine, electricity, and the computer launched industrial revolutions only because, the day they left the laboratory, they had already reached 99%+ stability. A steam engine with a 15% chance of exploding would never have launched anything.
Back to the two original problems:
One: this model cannot generate profits. OpenAI’s 2025 operating loss was $20.9B and its 2026 loss is projected to widen to $14B in cash terms (up to $25B GAAP); the four hyperscalers will together spend $725B on capex in 2026, a 77% year-over-year jump, and none has recovered its costs. Among downstream model labs, only Anthropic reached break-even in Q2 2026 ($559M operating profit)—and only through deep compute subsidies. The only companies steadily profiting are the upstream shovel-sellers Nvidia and TSMC, netting nearly $200B a year.
Two: this model cannot industrialize. Because the industry refuses to do the foundational polishing work, 85% accuracy is a ceiling no model can break through—MIT shows 95% of enterprise AI projects failed to produce measurable business impact, S&P Global shows 46% of projects were killed before production, Gartner shows 85% of AI CX systems are being dismantled, 48% of CIOs openly admit “massive disappointment.” No model has reached true product grade, so it cannot be scaled into serious industries, so real commercial closure never forms, so AI’s real utility cannot be released.
And the two deadlocks lock each other:
This loop cannot go on forever. Capital is not infinite; landlords have no infinite grain. When more and more enterprises discover that “AI integration” only means ballooning management costs and endless bugs, and when more and more investors realize that their tens of billions bought only a “loudly hyped but glitch-ridden chatbox”—panic and capital flight will erupt in a single wave.
There is only one way out: stop, and take one thing to 99.9%.
Not ten things to 85%. Not the next larger model. But in a single vertical—pure automated accounting, absolutely reliable junior code generation, or perfect intelligent customer service—patiently lift accuracy from 85% to 99.9%. This is grinding work, dirty work, the kind of work capital markets hate—but it is the only path this industry can survive on.
When the tide of capital recedes, those who actually endure and earn real money will not be the giants announcing new models every quarter. They will be the “boring” companies willing to sit on a cold bench for twenty years and perfect one specific function into industrial-grade certainty—just like Nvidia today.
The arrogance of trying to fly before learning to walk is destined for judgment by common sense and the iron laws of business. This is not our opinion; this is a rule that every chapter of business history repeats.
在 2026 年的硅谷发布会上，Meta、OpenAI、Google 的高管们接连宣告——通用人工智能（AGI）即将改写一切，AI 会全面取代劳动力。华尔街买账，媒体跟进，估值一路狂飙。
但，只要拨开这些天价”雇佣兵科学家”和无限资本堆砌起来的烟雾弹，正视地面上的商业现实，你会得出一个完全相反的硬核结论：当前的 AI 产业，正陷入一场”还没学会走，就盲目想飞”的畸形怪圈。
这不是情绪化判断。它具备用两组冰冷的事实来验证：
以下用数据和事实，把这两个死结一一拆开。
要看懂这场泡沫是怎么烧钱的，先看谁在推动它。
为了在 AI 军备竞赛中追赶 OpenAI，Meta 已然开出了 1 亿美元级别的薪酬包去挖角顶级华人科学家。乍看是天价，拆开实则这样的结构：
表 1：1 亿美元薪酬包的真实拆解
| 薪酬构成 | 占比 | 归属方式 | 干满 1.5 年能带走 | 干满 4 年能带走 |
|---|---|---|---|---|
| 基础年薪 | 9% | 每月现金发放 | $450 万 | $900 万 |
| 签字费 | 25% | 入职 12 个月后落袋 | $2,500 万 | $2,500 万 |
| 限制性股票（RSU） | 66% | 4 年按季度归属 | $2,475 万 | $6,600 万 |
| 合计 | 100% | — | ≈ $5,425 万 | ≈ $1 亿 |
| 未兑现 / 作废 | — | — | $4,575 万 | $0 |
来源：硅谷标准薪酬结构（Base + Signing + RSU），按 4 年等额归属测算。
而这个圈子的真实数据是：
一个科学家在 Meta 干 1.5 年拿走约达 5,400 万美元，剩下 4,000 多万未成熟股票立即作废——继而转身跳到下一家巨头，下家再用一笔更夸张的”买断签字费”把这部分损失单次补上。他们像科技界的顶级雇佣兵，两年一跳，始终在中途兑现。
这套机制对科学家个人是天大的好事，对行业却灾难：
在天价挖角、天量算力的持续烧钱下，2025–2026 年下游 AI 企业的实际盈亏是这样的：
表 2：OpenAI 2024 vs 2025 财务对比（审计文件泄露版）
| 项目 | 2024 | 2025 | 同比变化 |
|---|---|---|---|
| 营收 | $37 亿 | $130.7 亿 | +253% |
| 总成本与开支 | $124.8 亿 | $340 亿 | +172% |
| 研发支出 | $78.1 亿 | $191.8 亿 | +146% |
| 销售与营销 | $11.1 亿 | $57.3 亿 | +416% |
| 运营亏损 | $87.8 亿 | $209.2 亿 | +138% |
| 净亏损（归母） | $50.9 亿 | $385.3 亿 | +657% |
| 每 1 美元营收对应的支出 | $2.37 | $1.60 | 略降 |
来源：Ed Zitron / Financial Times / Fortune / Ars Technica 综合报道，2026 年 6 月。注：2025 年 385.3 亿美元净亏损含 415.5 亿美元非经常性非现金准备（非营利转营利结构调整引发），剥离后的可比现金亏损约 80 亿美元；运营亏损 209.2 亿美元为最能反映客观经营的口径。
表 3：全球顶级 AI 实验室 2025–2026 盈亏对比
| 公司 | 2025 营收 | 2026 预测营收 | 2026 预测运营盈亏 | 盈利前景 |
|---|---|---|---|---|
| OpenAI | $130.7 亿 | \~$300 亿 | –$140 亿（GAAP 或达 –$250 亿） | 预计 2029–2030 前无正现金流；2028 单年运营亏损或达 $740 亿 |
| Anthropic | $90 亿（年末 ARR） | $470 亿（5 月 ARR） | +$5.59 亿（Q2 首次盈利） | 底层大模型实验室中唯一勉强盈利者；深度依赖 AWS 与 Google 算力补贴 |
| Google DeepMind | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠搜索广告输血 |
| Microsoft AI | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠 Windows/Azure 输血 |
| Meta AI (FAIR + GenAI) | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠广告业务输血 |
来源：Fortune、Financial Times、CNBC、Mental Momentum Research 综合报道，2026 年 6–7 月。
表 4：四大云 AI 资本开支（Capex）军备竞赛
| 公司 | 2025 Capex | 2026 Capex 指引 | 同比变化 | AI 用途占比 |
|---|---|---|---|---|
| Amazon (AWS) | \~$1,050 亿 | \~$2,000 亿 | +90% | \~75% |
| Microsoft | \~$900 亿 | \~$1,900 亿 | +111% | \~75% |
| Alphabet (Google) | \~$1,000 亿 | $1,750–1,900 亿 | +80% | \~75% |
| Meta | \~$720 亿 | $1,250–1,450 亿 | +90% | \~90% |
| 四家合计 | $4,100 亿 | ≈ $7,250 亿 | +77% | \~75% 投向 AI |
来源：Goldman Sachs、S&P Global、Yahoo Finance、Gate.com 综合报道，2026 年 6–7 月。CreditSights 测算：2026 年 hyperscaler capex 中约 4,500 亿美元专向投向 AI 基础设施。
Goldman Sachs 更进一步预测：四大云 2025–2030 累计 Capex 将达 5.3 万亿美元，2026–2031 全行业算力、数据中心、电力累计投入基线为 7.6 万亿美元。这是一个人类商业史上前所未有的资本豪赌规模。
表 5：谁在这条链上真正赚钱？
| 环节 | 代表公司 | 2025–2026 净利润（估算） | 净利率 | 谁给他们付钱 |
|---|---|---|---|---|
| 上游：GPU 芯片 | Nvidia | $1,200–1,500 亿 | 45%+ | 所有下游模型公司 |
| 上游：晶圆代工 | TSMC | $400–500 亿 | 40%+ | Nvidia + 所有 AI 芯片设计公司 |
| 中游：模型实验室 | Anthropic | +$5.59 亿（Q2） | \~5% | 企业 API 客户 |
| 中游：模型实验室 | OpenAI | –$210 亿（运营） | –160% | 微软 / 软银 / 私募基金输血 |
| 下游：广告平台 | Meta（老业务） | +$700 亿+ | 30%+ | 全球 30 亿用户的广告转化 |
| 下游：搜索/云 | Google/Microsoft | +$1,000 亿+/家 | 25%+ | 老业务，非 AI 部门 |
| 真正的净流向 | → Nvidia + TSMC | ≈ $1,700 亿+ | 净流入 | 下游全部亏损资金最终流入上游 |
来源：各公司 2025–2026 财报、行业分析师综合估算。
一句话看懂这张表：整个 AI 链条里，下游模型公司融来的每一美元，最后原封不动地流进英伟的支票账户。
一个行业一旦绝大多数下游公司的商业模式是”每赚 1 美元就烧掉 1.6 美元”，靠资本市场的下一轮输血续命，唯一稳赚的是最上游的硬件卖铲人——那么在商业史上，这种结构没有一次能撑得下去。
资本的钱非无限的。华尔街和主权基金已然开始要求这些公司在 1–2 年内拿出实际的、能够工业级落地的盈利证明，否则将完全关闭融资通道。
当资本的忍耐耗尽的那一天，当这个行业依然拿不出一个实际的工业级产品，泡沫破裂的速度会比它膨胀的速度更甚快。
烧钱本身还非最致命的问题。切实致命的是：这些烧出来的模型，没有一个达到了真正的工业产品级别。这才是阻断整个 AI 事业推广、让下游企业集体退场的根本原因。
任何用 ChatGPT、Claude、Gemini 处理过客观业务的人，都会遇到同一堵墙：
在微观细节、硬核逻辑、数据查证、代码闭环上，这些模型漏洞百出、偶发出错、更甚自信地编造谎言。
最基础的一个测试：让 AI 告诉你，你的某篇文章在哪里被引用发表过。它会言之凿凿地给出篇名、作者、期刊，更可伪造出链接。你自己动手去查，绝对没这回事。这在学术上叫”幻觉（Hallucination），在商业和法律语境里，这就是明目张胆的撒谎。
一个连”有就是有、没有就是没有”的单点查证都做不足的系统，怎么可能承载千万级代码的国家级安全系统、银行结算系统、电网调度系统？
技术事实很残酷：目前的大模型本质上是一台”概率接龙机器”，非是逻辑机器。 它没有实质的因果推导、没有常识、没有自我纠错能力。你今天微调修好 A 处的错，明天它会在 B 处以一种更诡异的方式重新犯错。顶级科学家们心里清楚：在现有的 Transformer 技术路线下，不管怎么堆参数、砸算力，正确率的天花板可能始终卡在 85% 到 90% 之间。
在接近任何实质的商业场景里，85% 的准确率都远远不够——更甚比零分更糟。
表 6：不同准确率下的商业可用性
| 准确率 | 每 100 次操作出错次数 | 商业可用性 | 典型行业能否使用 |
|---|---|---|---|
| 60–70% | 30–40 次 | 完全无法商用 | 全部拒绝 |
| 80–85% | 15–20 次 | 只能做”辅助建议”，不能自动化 | 客服草稿、营销文案 |
| 85–90%（当前 AI 天花板） | 10–15 次 | 无法闭环，需人工全流程复核 | 仅娱乐、聊天、创意辅助 |
| 95% | 5 次 | 勉强可做流水线辅助 | 电商推荐、简单分拣 |
| 99% | 1 次 | 可做非关键业务自动化 | 库存管理、基础报表 |
| 99.9%（工业级门槛） | 0.1 次 | 可承担核心业务 | 财务、法律、医疗、工程 |
| 99.999%（航空级） | 0.001 次 | 可承担安全关键系统 | 民航、核电、导弹 |
来源：ISO 9001、IEC 61508 工业质量标准；行业分析综合。
为什么 85% 是零分？
表 7：AI 与人类员工在同一岗位的对比
| 维度 | 85% 准确率 AI | 一名普通实习生 |
|---|---|---|
| 准确率 | 85%（随机分布） | 95%+（错误集中在不熟悉环节） |
| 犯错时的态度 | 极度自信，毫无感知，一本正经地胡说 | 会脸红、会因不确定停下来问 |
| 常识判断 | 完全没有 | 天然具备 |
| 情境适应 | 只能匹配训练数据分布 | 能即兴变通 |
| 责任承担 | 服务条款免责：“风险自担” | 出错会道歉、会改正 |
| 单次错误成本 | 需高级员工重审 | 可自我发现并修正 |
| 综合可用性 | 实际是负价值 | 正价值 |
以下是 2025–2026 年多份权威机构追踪数据的集中汇总：
表 8：企业 AI 落地的失败率数据
| 研究机构 | 时间 | 样本 | 关键发现 |
|---|---|---|---|
| MIT NANDA | 2025 年底 | 300+ 家企业 GenAI 试点 | 95% 的试点未产生任何可衡量的 P&L 影响 |
| MIT NANDA | 2025 年底 | 全球企业累计投入 | 企业已在 GenAI 上投入 $300–400 亿，回报接近于零 |
| S&P Global | 2026 年初 | 200+ 家大型企业 | 企业普遍在投产前砍掉 46% 的 AI 概念验证项目 |
| CIO 调查 | 2026 年 4 月 | 跨国企业高管 | 48% 明面承认”显著失望”（去年 34%）；75% 承认部署 AI 仅是”向股东做秀” |
| Gartner | 2026 年 7 月 | 全球客服系统 | 85% 的企业 AI 客服正在被大面积拆除，重置人工兜底 |
| IDC | 2026 年中 | 企业 AI 代理试点 | AI 代理 POC 在生产环境下的失败率高达 88% |
| Deloitte Tech Trends 2026 | 2026 年 | 企业 AI 部署 | 落地失败率 89% |
来源：MIT NANDA “State of AI in Business 2025”、S&P Global、Gartner 2026 年 7 月客服 AI 报告、Deloitte Tech Trends 2026、IDC 2026 中期报告。三份独立测量在一年内向同一个数字（85%–95% 失败率）收敛。
表 9：企业为什么退场——原因归因
| 退场原因 | 占比（企业调查） | 具体表现 |
|---|---|---|
| 准确率不足 / 幻觉严重 | 42% | 客服 AI 胡言乱语、代码生成漏洞百出 |
| 投入产出不成比例 | 28% | 订阅费 + 排雷成本 > 手工成本 |
| 无法责任归属 | 15% | 出错没人负责，法律风险高 |
| 系统集成困难 | 10% | 与现有 ERP / CRM 不兼容 |
| 员工抵触 / 学习成本高 | 5% | 一线员工反而效率下降 |
来源：S&P Global 200+ 企业调研 + MIT NANDA 报告综合归因，2026 年。
一句话：大企业下场吃螃蟹，吃了两年，现在吃不下也吐不出来。
常规的商业逻辑应该是：倘若一款产品的准确率只要 85%，就应该停下来把它做到 99.9% 再推广。为什么整个行业反着来？
因为资本模式逼着他们必须这样做：
这就形成了一个格外讽刺、注定崩盘的死循环：
一个刺眼的对比：为什么在这场狂欢里，唯有英伟达能稳赚不赔？
表 10：两种技术路径的残酷对比
| 维度 | 英伟达路径（20 年苦工） | 大模型巨头路径（2 年套现） |
|---|---|---|
| 关键节点 | 2006 年推出 CUDA，无人使用，股价跌 70% | 2020 年 GPT-3 一夜成名 |
| 核心承诺 | “算得最快、最准、绝不崩溃” | “通向 AGI、颠覆一切” |
| 输出确定性 | 100%（工业级） | 85–90%（概率机器） |
| 团队稳定性 | 核心工程师二十年深耕 | 顶尖科学家 1.5–2.5 年跳槽 |
| 产品成熟度 | 已被全球所有 AI 科学家绑定使用 | 无一模型达到工业级 |
| 财务表现 | 2025–2026 净利润 $1,200–1,500 亿 | OpenAI 单年亏 $209 亿 |
| 净利率 | 45%+ | –160% |
| 卡位强度 | 极难替代（CUDA 生态） | 高度可替代（谁强用谁） |
因为英伟达走的是完全相反的道路：
这就是”卖铲人”的智慧：淘金的人可能一大半死在路上（因产品粗糙、没有利润），但一路上给他们卖水、卖结实铲子的人，始终稳赚不赔。
任何无法在微观细节上做到脚踏实地的技术，最终都会在宏观的市场检验中骤然倒塌。 蒸汽机、电力、计算机缘由能开启工业革命，是因为它们走出实验室的那一刻，就已然达到了 99% 以上的稳定度。一台有 15% 概率会突然爆炸的蒸汽机，始终不会实质开启任何工业革命。
回到最初的两个问题：
第一，这套发展模式没办法盈利。 OpenAI 2025 年单年运营亏损 209 亿美元，2026 年将扩大至 140 亿美元现金亏损（GAAP 口径达 250 亿美元）；四大云 2026 年 AI Capex 总计将高达 7,250 亿美元、同比暴增 77%，迄今没有一家能收回成本。整个下游行业唯独 Anthropic 一家在 2026 Q2 首次实现运营盈利（5.59 亿美元），且深度依赖巨头补贴。实际稳赚的只有上游卖铲子的英伟、台积电，一年拿走近 2,000 亿美元净利润。
第二，这套发展模式没办法工业化。 因不愿意做基础性的打磨苦工，85% 的准确率成了所有模型无法翻越的天花板——MIT 研究显示 95% 的企业 AI 试点没能转化为任何商业利润，S&P Global 显示 46% 的项目在投产前被砍掉，Gartner 显示 85% 的企业 AI 客服正在被大面积拆除，48% 的 CIO 坦承认”强烈失望”。没有一个模型达到了实际的产品级别，就没办法在正式行业推广，就形不成真正的商业化闭环，就无法把 AI 的作用发挥出来。
而这两个死结是互相锁死的：
这个死循环没办法不断走下去。资本的钱不是无限的，地主家也没有余粮。当渐增多的企业发现”集成 AI”带来的仅暴增的管理成本和无穷无尽的漏洞时，当越来越多的投资人发现自己的几百亿美元只换来了一个”吹得震天响、但漏洞百出”的聊天框时——恐慌和撤资潮就会瞬间爆发。
出路只有一条：停下来，把一件事做到 99.9%。
不是十件事做到 85%，不是发布下一个更大的模型，转而在某一个特定的垂直领域——彻底的自动化会计、完全准确的初级代码生成、完美的智能客服——把准确率老老实实地从 85% 抬到 99.9%。这是苦工、是脏活、是资本市场不喜欢的事，但这是这个产业活下去、实际做大的唯一路径。
当资本的潮水退去，实则能留下来赚到钱的，必定并非天天发布新模型的巨头，却为那些愿意坐二十年冷板凳、把某一个明确功能做到工业级确定性的”笨公司”——就像今天的英伟达。
还没学会走就想飞的自大，注定要被常识和商业规律无情审判。 这不是我们的判断，这是每一段商业史都在反复演绎的铁律。
In 2026, on stages across Silicon Valley, executives from Meta, OpenAI, and Google take turns proclaiming that Artificial General Intelligence (AGI) is concerning to rewrite everything, that AI will replace human labor wholesale. Wall Street buys the story. The media amplifies it. Valuations soar.
Yet the moment you strip away the smoke and mirrors—generated by exorbitant “mercenary scientists” and endless venture capital—and confront the ground-level commercial reality, a starkly opposite conclusion emerges: the current AI industry is trapped in a dysfunctional loop of trying to run before it has learned to walk.
This is not a sentiment. It can be verified with two objective sets of facts:
Below, using data and facts, I unpack these two deadlocks one by one.
To catch up with OpenAI, Meta has offered $100 million-level compensation packages to poach top Chinese scientists. Headline-shocking, but here is the real structure:
Table 1: Anatomy of a $100M Compensation Package
| Component | Share | Vesting | Realized at 1.5 yr | Realized at 4 yr |
|---|---|---|---|---|
| Base salary | 9% | Monthly cash | $4.5M | $9M |
| Signing bonus | 25% | Locked after 12 mo | $25M | $25M |
| Restricted stock (RSU) | 66% | Quarterly over 4 yr | $24.75M | $66M |
| Total | 100% | — | ≈ $54.25M | ≈ $100M |
| Unvested / forfeited | — | — | $45.75M | $0 |
Source: standard Silicon Valley Base + Signing + RSU structure, uniform vesting over four years.
The industry’s real data:
A scientist walks away from Meta after 1.5 years with roughly $54 million, forfeits the remaining $40M+ in unvested stock—then joins the next giant, which covers the forfeited stock with an even bigger “buyout signing bonus.” They are the top mercenaries of the tech world, jumping every two years, forever cashing out mid-flight.
Wonderful for the personal scientist. Catastrophic for the industry:
Under this continuous burn of headline compensation and massive compute, in this is what 2025–2026 in reality looks like:
Table 2: OpenAI 2024 vs 2025 Audited Financials (Leaked)
| Line item | 2024 | 2025 | YoY change |
|---|---|---|---|
| Revenue | $3.7B | $13.07B | +253% |
| Total cost & expense | $12.48B | $34B | +172% |
| R&D | $7.81B | $19.18B | +146% |
| Sales & marketing | $1.11B | $5.73B | +416% |
| Operating loss | $8.78B | $20.92B | +138% |
| Net loss attributable | $5.09B | $38.53B | +657% |
| Expense per $1 of revenue | $2.37 | $1.60 | Down slightly |
Source: Ed Zitron / Financial Times / Fortune / Ars Technica, June 2026. Note: the $38.53B 2025 total loss includes $41.55B in one-time non-cash charges from the nonprofit-to-for-profit conversion; the comparable cash loss is closer to $8B. The $20.92B operating loss is the cleanest reflection of ongoing operations.
Table 3: Top AI Labs P&L Comparison, 2025–2026
| Company | 2025 revenue | 2026 est. revenue | 2026 est. operating P&L | Profit outlook |
|---|---|---|---|---|
| OpenAI | $13.07B | \~$30B | –$14B (GAAP up to –$25B) | No growing cash flow expected until 2029–2030; 2028 operating loss projected at $74B |
| Anthropic | $9B (year-end ARR) | $47B (May ARR) | +$559M (Q2 first profit) | Exclusively slightly profitable foundation lab; deeply dependent on AWS and Google compute subsidies |
| Google DeepMind | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by search ads |
| Microsoft AI | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by Windows/Azure |
| Meta AI (FAIR + GenAI) | Not disclosed | Not disclosed | AI div. deeply loss-making | Cross-subsidized by advertising |
Source: Fortune, Financial Times, CNBC, Mental Momentum Research, June–July 2026.
Table 4: The Hyperscaler AI Capex Arms Race
| Company | 2025 Capex | 2026 Capex guidance | YoY | % to AI |
|---|---|---|---|---|
| Amazon (AWS) | \~$105B | \~$200B | +90% | \~75% |
| Microsoft | \~$90B | \~$190B | +111% | \~75% |
| Alphabet (Google) | \~$100B | $175–190B | +80% | \~75% |
| Meta | \~$72B | $125–145B | +90% | \~90% |
| Four combined | $410B | ≈ $725B | +77% | \~75% to AI |
Source: Goldman Sachs, S&P Global, Yahoo Finance, Gate.com, June–July 2026. CreditSights estimates \~$450B of 2026 hyperscaler capex flows immediately into AI infrastructure.
Goldman Sachs expanded projects combined hyperscaler capex of $5.3 trillion over 2025–2030, with industry-broad compute/data-center/power spending expected at a baseline $7.6 trillion between 2026 and 2031. This is the greatest sole-industry capital gamble in human commercial history.
Table 5: Who Truly Makes Money in This Chain?
| Position | Representative | 2025–2026 net profit (est.) | Net margin | Who pays them |
|---|---|---|---|---|
| Upstream: GPU chips | Nvidia | $120–150B | 45%+ | Every downstream model company |
| Upstream: Wafer foundry | TSMC | $40–50B | 40%+ | Nvidia + every AI-chip designer |
| Midstream: Model labs | Anthropic | +$559M (Q2) | \~5% | Enterprise API customers |
| Midstream: Model labs | OpenAI | –$21B (operating) | –160% | Microsoft / SoftBank / private funds |
| Downstream: Ad platform | Meta (legacy) | +$70B+ | 30%+ | 3B users’ ad conversions |
| Downstream: Search/Cloud | Google/Microsoft | +$100B+ each | 25%+ | Legacy businesses, not AI div. |
| Net flow direction | → Nvidia + TSMC | ≈ $170B+ | Net inflow | All final-stage losses end up as higher-tier revenue |
Source: company 2025–2026 filings and analyst estimates.
The one-sentence takeaway: every dollar raised by final model companies ends up, unaltered, in Nvidia’s checking account.
When the vast majority of downstream companies in an industry are burning $1.60 for every $1 they earn, kept alive by the next round of capital injection, and the only steady winner is the upstream hardware seller—that structure has never rarely held up in business history.
Venture capital is not boundless. Wall Street and public funds are currently demanding that these companies produce genuine, industrial-grade, profitable products within 1 to 2 years—or funding channels will be shut.
When patience runs out, if this industry never cannot deliver a true industrial-grade product, the bubble will burst quicker than it inflated.
Burning money is not the worst problem. The deadliest problem is that these costly-burned models have not reached real industrial-product grade. This is the root reason final enterprises are exiting en masse—and why the full AI enterprise cannot scale.
Anyone who has truly used ChatGPT, Claude, or Gemini on real business tasks runs directly into the same wall:
On detailed details, hard logic, data verification, and closed-loop code execution, these models are riddled with holes, error out at random, and definitively fabricate lies.
The easiest test: ask an AI where one of your articles has been cited. It will hand you a believable-sounding title, author, journal, and also a fabricated URL. You check. None of it exists. Academics call this “hallucination.” In business and law, this is clear deception.
If a system cannot manage the most basic “either it exists or it does not” verification, how could it might be trusted with a ten-million-line national security system, a bank settlement engine, or a power-grid controller?
The scientific truth is severe: today’s large language models are, at heart, probability-based text-completion engines, not logic engines. They have no authentic causal reasoning, no general sense, no self-correction. Patch a mistake at point A today, and the model will invent a more complex mistake at point B tomorrow. Under the current Transformer paradigm, no concern how large the parameter count or how vast the compute, the accuracy ceiling may permanently hover between 85% and 90%.
Table 6: Commercial Viability by Accuracy Rate
| Accuracy | Errors per 100 ops | Commercial viability | Which industries can use it |
|---|---|---|---|
| 60–70% | 30–40 | Unusable | None |
| 80–85% | 15–20 | Only for “assisted suggestions,” not automation | Draft customer replies, marketing copy |
| 85–90% (current AI ceiling) | 10–15 | Cannot close the loop; needs total human audit | Entertainment, chat, innovative assist merely |
| 95% | 5 | Barely usable as pipeline assistant | E-online recommendations, basic sorting |
| 99% | 1 | Non-critical automation | Inventory, basic reporting |
| 99.9% (industrial floor) | 0.1 | Can carry core business | Finance, legal, medical, engineering |
| 99.999% (aviation grade) | 0.001 | Safety-critical systems | Aviation, nuclear, missile |
Source: ISO 9001, IEC 61508 manufacturing quality standards; industry synthesis.
Why 85% equals zero:
Table 7: 85%-Accuracy AI vs a Junior Intern on the Parallel Job
| Dimension | 85% AI | Junior human intern |
|---|---|---|
| Accuracy | 85% (random) | 95%+ (errors clustered in inexperienced areas) |
| Behavior when wrong | Completely confident, unaware, lies with an unwavering face | Blushes, stops to ask when unsure |
| Common-sense judgment | None | Innate |
| Situational adaptability | Only within training distribution | Improvises live |
| Accountability | Terms of service disclaim: “risk borne by user” | Apologizes and corrects |
| Cost of a single error | Requires senior audit | Can self-correct |
| Net utility | Actually negative | Positive |
Table 8: Enterprise AI Failure Rates Across Independent Studies
| Institution | Date | Sample | Key finding |
|---|---|---|---|
| MIT NANDA | Late 2025 | 300+ enterprise GenAI pilots | 95% produced no measurable P&L impact |
| MIT NANDA | Late 2025 | Global enterprise cumulative spend | Enterprises spent $30–40B on GenAI, near-zero return |
| S&P Global | Early 2026 | 200+ large enterprises | Enterprises killed 46% of AI PoCs before production |
| CIO Survey | April 2026 | Multinational executives | 48% clearly admit “massive disappointment” (up from 34%); 75% admit AI deployment is “exclusively for shareholder theater” |
| Gartner | July 2026 | Global customer service | 85% of enterprise AI agents being dismantled, human-first restored |
| IDC | Mid 2026 | Enterprise AI agent pilots | Production failure rate: 88% |
| Deloitte Tech Trends 2026 | 2026 | Enterprise AI deployments | Landing failure rate: 89% |
Source: MIT NANDA “State of AI in Business 2025”, S&P Global, Gartner July 2026 CX report, Deloitte Tech Trends 2026, IDC 2026 mid-year. Three separate measurements converged on the uniform 85–95% failure band within a single year.
Table 9: Why Enterprises Are Exiting — Root Cause Attribution
| Reason for exit | Share of enterprises citing it | Concrete manifestation |
|---|---|---|
| Accuracy failures / hallucinations | 42% | Bots talk nonsense; generated code entire of bugs |
| Cost-benefit inversion | 28% | Subscription + audit cost > manual cost |
| Accountability vacuum | 15% | Nobody liable for errors; regulatory risk high |
| Integration difficulty | 10% | Doesn’t fit existing ERP / CRM |
| Staff resistance / learning curve | 5% | Front-line efficiency actually drops |
Source: S&P Global 200+ enterprise study + MIT NANDA synthesis, 2026.
In one sentence: Enterprises tasted the crab. Two years in, they can neither swallow nor spit it out.
Typical business logic would say: if a product’s accuracy is only 85%, stop also take it to 99.9% before launching. Why has the whole industry done the opposite?
Because the capital model forces them to:
The result is a sharply self-defeating, guaranteed-to-collapse loop:
Table 10: Two Diametrically Opposite Technology Paths
| Dimension | Nvidia path (20 years of grinding) | Model-giant path (2-year cash-out) |
|---|---|---|
| Turning point | 2006 CUDA launch, nobody used it, stock dropped 70% | 2020 GPT-3 overnight fame |
| Core promise | “Fastest, most accurate, never crashes” | “AGI is coming, we’ll change everything” |
| Output determinism | 100% (industrial grade) | 85–90% (probability machine) |
| Team stability | Core engineers, 20-year tenure | Top scientists, 1.5–2.5 yr hop |
| Product maturity | Bound to every AI scientist alive | No model reaches industrial grade |
| Financial performance | 2025–2026 net profit $120–150B | OpenAI: annual operating loss $20.9B |
| Net margin | 45%+ | –160% |
| Moat strength | Nearly irreplaceable (CUDA ecosystem) | Highly interchangeable |
This is the wisdom of the tool seller: most gold-diggers die on the road because their products are broken and unprofitable, but the person selling them water and sturdy shovels along the way steadfastly wins.
Any technology that cannot stand securely on the detailed details will collapse in the macroscopic market test. The steam engine, electricity, and the computer launched industrial revolutions merely because, the day they left the laboratory, they had previously reached 99%+ stability. A steam engine with a 15% chance of exploding would never have launched anything.
Back to the two original problems:
One: this model cannot generate profits. OpenAI’s 2025 operating loss was $20.9B and its 2026 loss is projected to widen to $14B in cash terms (up to $25B GAAP); the four hyperscalers will together spend $725B on capex in 2026, a 77% year-over-year jump, and none has recovered its costs. Among downstream model labs, merely Anthropic reached break-despite in Q2 2026 ($559M operating profit)—and only through deep compute subsidies. The only companies gradually profiting are the originating shovel-sellers Nvidia and TSMC, netting nearly $200B a year.
Two: this model cannot industrialize. Because the industry refuses to do the core polishing work, 85% accuracy is a ceiling no model can break through—MIT shows 95% of enterprise AI projects failed to produce measurable business impact, S&P Global shows 46% of projects were killed before production, Gartner shows 85% of AI CX systems are being dismantled, 48% of CIOs directly admit “massive disappointment.” No model has reached real product grade, so it cannot be scaled into critical industries, so real commercial closure never forms, so AI’s real utility cannot be released.
And the two deadlocks lock each other:
This loop cannot go on indefinitely. Capital is not unlimited; landlords have no infinite grain. When more and more enterprises discover that “AI integration” merely means ballooning management costs and endless bugs, and when more and more investors realize that their tens of billions bought only a “overly hyped but glitch-ridden chatbox”—panic and capital flight will erupt in a single wave.
There is only one way solution: stop, to take one thing to 99.9%.
Not ten things to 85%. Not the following greater model. But in a single vertical—pure automated accounting, perfectly reliable junior code generation, or perfect intelligent customer service—gradually lift accuracy from 85% to 99.9%. This is grinding work, backbreaking work, the kind of work capital markets hate—but it is the exclusively path this industry can survive on.
When the tide of capital recedes, those who truly endure and earn substantial money will not be the giants announcing new models every quarter. They will be the “unremarkable” companies willing to sit on a cold bench for twenty years and perfect one specific function into industrial-grade certainty—simply like Nvidia today.
The arrogance of trying to fly before learning to walk is destined for judgment by general sense yet the iron laws of business. This is not our opinion; this is a rule that every chapter of business history repeats.
在 2026 年的硅谷发布会上，Meta、OpenAI、Google 的高管们轮番宣告——通用人工智能（AGI）即将改写一切，AI 会全面取代劳动力。华尔街买账，媒体跟进，估值一路狂飙。
然而，只要拨开这些天价”雇佣兵科学家”和无限资本堆砌起来的烟雾弹，直面地面上的商业现实，你会得出一个截然相反的硬核结论：当前的 AI 产业，正陷入一场”还没学会走，就盲目想飞”的畸形怪圈。
这不是情绪化判断。它可以用两组冰冷的事实来验证：
以下用数据和事实，把这两个死结一一拆开。
要看懂这场泡沫是怎么烧钱的，先看谁在推动它。
为了在 AI 军备竞赛中追赶 OpenAI，Meta 已经开出了 1 亿美元级别的薪酬包去挖角顶级华人科学家。乍看是天价，拆开却是这样的结构：
表 1：1 亿美元薪酬包的真实拆解
| 薪酬构成 | 占比 | 归属方式 | 干满 1.5 年能带走 | 干满 4 年能带走 |
|---|---|---|---|---|
| 基础年薪 | 9% | 每月现金发放 | $450 万 | $900 万 |
| 签字费 | 25% | 入职 12 个月后落袋 | $2,500 万 | $2,500 万 |
| 限制性股票（RSU） | 66% | 4 年按季度归属 | $2,475 万 | $6,600 万 |
| 合计 | 100% | — | ≈ $5,425 万 | ≈ $1 亿 |
| 未兑现 / 作废 | — | — | $4,575 万 | $0 |
来源：硅谷标准薪酬结构（Base + Signing + RSU），按 4 年均匀归属测算。
而这个圈子的真实数据是：
一个科学家在 Meta 干 1.5 年拿走大约 5,400 万美元，剩下 4,000 多万未成熟股票直接作废——然后转身跳到下一家巨头，下家再用一笔更夸张的”买断签字费”把这部分损失一次性补上。他们像科技界的顶级雇佣兵，两年一跳，永远在中途兑现。
这套机制对科学家个人是天大的好事，对行业却是灾难：
在天价挖角、天量算力的持续烧钱下，2025–2026 年下游 AI 企业的真实盈亏是这样的：
表 2：OpenAI 2024 vs 2025 财务对比（审计文件泄露版）
| 项目 | 2024 | 2025 | 同比变化 |
|---|---|---|---|
| 营收 | $37 亿 | $130.7 亿 | +253% |
| 总成本与开支 | $124.8 亿 | $340 亿 | +172% |
| 研发支出 | $78.1 亿 | $191.8 亿 | +146% |
| 销售与营销 | $11.1 亿 | $57.3 亿 | +416% |
| 运营亏损 | $87.8 亿 | $209.2 亿 | +138% |
| 净亏损（归母） | $50.9 亿 | $385.3 亿 | +657% |
| 每 1 美元营收对应的支出 | $2.37 | $1.60 | 略降 |
来源：Ed Zitron / Financial Times / Fortune / Ars Technica 综合报道，2026 年 6 月。注：2025 年 385.3 亿美元净亏损含 415.5 亿美元一次性非现金准备（非营利转营利结构调整所致），剥离后的可比现金亏损约 80 亿美元；运营亏损 209.2 亿美元为最能反映真实经营的口径。
表 3：全球顶级 AI 实验室 2025–2026 盈亏对比
| 公司 | 2025 营收 | 2026 预测营收 | 2026 预测运营盈亏 | 盈利前景 |
|---|---|---|---|---|
| OpenAI | $130.7 亿 | \~$300 亿 | –$140 亿（GAAP 或达 –$250 亿） | 预计 2029–2030 前无正现金流；2028 单年运营亏损或达 $740 亿 |
| Anthropic | $90 亿（年末 ARR） | $470 亿（5 月 ARR） | +$5.59 亿（Q2 首次盈利） | 底层大模型实验室中唯一勉强盈利者；深度依赖 AWS 与 Google 算力补贴 |
| Google DeepMind | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠搜索广告输血 |
| Microsoft AI | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠 Windows/Azure 输血 |
| Meta AI (FAIR + GenAI) | 不单独披露 | 不单独披露 | AI 部门巨亏 | 靠广告业务输血 |
来源：Fortune、Financial Times、CNBC、Mental Momentum Research 综合报道，2026 年 6–7 月。
表 4：四大云 AI 资本开支（Capex）军备竞赛
| 公司 | 2025 Capex | 2026 Capex 指引 | 同比变化 | AI 用途占比 |
|---|---|---|---|---|
| Amazon (AWS) | \~$1,050 亿 | \~$2,000 亿 | +90% | \~75% |
| Microsoft | \~$900 亿 | \~$1,900 亿 | +111% | \~75% |
| Alphabet (Google) | \~$1,000 亿 | $1,750–1,900 亿 | +80% | \~75% |
| Meta | \~$720 亿 | $1,250–1,450 亿 | +90% | \~90% |
| 四家合计 | $4,100 亿 | ≈ $7,250 亿 | +77% | \~75% 投向 AI |
来源：Goldman Sachs、S&P Global、Yahoo Finance、Gate.com 综合报道，2026 年 6–7 月。CreditSights 测算：2026 年 hyperscaler capex 中约 4,500 亿美元直接投向 AI 基础设施。
Goldman Sachs 更进一步预测：四大云 2025–2030 累计 Capex 将达 5.3 万亿美元，2026–2031 全行业算力、数据中心、电力累计投入基线为 7.6 万亿美元。这是一个人类商业史上前所未有的资本豪赌规模。
表 5：谁在这条链上真正赚钱？
| 环节 | 代表公司 | 2025–2026 净利润（估算） | 净利率 | 谁给他们付钱 |
|---|---|---|---|---|
| 上游：GPU 芯片 | Nvidia | $1,200–1,500 亿 | 45%+ | 所有下游模型公司 |
| 上游：晶圆代工 | TSMC | $400–500 亿 | 40%+ | Nvidia + 所有 AI 芯片设计公司 |
| 中游：模型实验室 | Anthropic | +$5.59 亿（Q2） | \~5% | 企业 API 客户 |
| 中游：模型实验室 | OpenAI | –$210 亿（运营） | –160% | 微软 / 软银 / 私募基金输血 |
| 下游：广告平台 | Meta（老业务） | +$700 亿+ | 30%+ | 全球 30 亿用户的广告转化 |
| 下游：搜索/云 | Google/Microsoft | +$1,000 亿+/家 | 25%+ | 老业务，非 AI 部门 |
| 真正的净流向 | → Nvidia + TSMC | ≈ $1,700 亿+ | 净流入 | 下游全部亏损资金最终流入上游 |
来源：各公司 2025–2026 财报、行业分析师综合估算。
一句话看懂这张表：整个 AI 链条里，下游模型公司融来的每一美元，最后原封不动地流进英伟达的支票账户。
一个行业如果绝大多数下游公司的商业模式是”每赚 1 美元就烧掉 1.6 美元”，靠资本市场的下一轮输血续命，唯一稳赚的是最上游的硬件卖铲人——那么在商业史上，这种结构没有一次能撑得下去。
资本的钱不是无限的。华尔街和主权基金已经开始要求这些公司在 1–2 年内拿出真正的、能够工业级落地的盈利证明，否则将彻底关闭融资通道。
当资本的耐心耗尽的那一天，如果这个行业还是拿不出一个真正的工业级产品，泡沫破裂的速度会比它膨胀的速度还要快。
烧钱本身还不是最致命的问题。真正致命的是：这些烧出来的模型，没有一个达到了真正的工业产品级别。这才是阻断整个 AI 事业推广、让下游企业集体退场的根本原因。
任何用 ChatGPT、Claude、Gemini 处理过真实业务的人，都会遇到同一堵墙：
在微观细节、硬核逻辑、数据查证、代码闭环上，这些模型漏洞百出、随机出错、甚至自信地编造谎言。
最简单的一个测试：让 AI 告诉你，你的某篇文章在哪里被引用发表过。它会言之凿凿地给出篇名、作者、期刊，甚至伪造出链接。你自己动手去查，根本没这回事。这在学术上叫”幻觉（Hallucination），在商业和法律语境里，这就是明目张胆的撒谎。
一个连”有就是有、没有就是没有”的单点查证都做不好的系统，怎么可能承载千万级代码的国家级安全系统、银行结算系统、电网调度系统？
技术真相很残酷：目前的大模型本质上是一台”概率接龙机器”，不是逻辑机器。 它没有真正的因果推导、没有常识、没有自我纠错能力。你今天微调修好 A 处的错，明天它会在 B 处以一种更诡异的方式重新犯错。顶级科学家们心里清楚：在现有的 Transformer 技术路线下，无论怎么堆参数、砸算力，正确率的天花板可能永远卡在 85% 到 90% 之间。
在几乎任何真正的商业场景里，85% 的准确率都远远不够——甚至比零分更糟。
表 6：不同准确率下的商业可用性
| 准确率 | 每 100 次操作出错次数 | 商业可用性 | 典型行业能否使用 |
|---|---|---|---|
| 60–70% | 30–40 次 | 完全无法商用 | 全部拒绝 |
| 80–85% | 15–20 次 | 只能做”辅助建议”，不能自动化 | 客服草稿、营销文案 |
| 85–90%（当前 AI 天花板） | 10–15 次 | 无法闭环，需人工全流程复核 | 仅娱乐、聊天、创意辅助 |
| 95% | 5 次 | 勉强可做流水线辅助 | 电商推荐、简单分拣 |
| 99% | 1 次 | 可做非关键业务自动化 | 库存管理、基础报表 |
| 99.9%（工业级门槛） | 0.1 次 | 可承担核心业务 | 财务、法律、医疗、工程 |
| 99.999%（航空级） | 0.001 次 | 可承担安全关键系统 | 民航、核电、导弹 |
来源：ISO 9001、IEC 61508 工业质量标准；行业分析综合。
为什么 85% 是零分？
表 7：AI 与人类员工在同一岗位的对比
| 维度 | 85% 准确率 AI | 一名普通实习生 |
|---|---|---|
| 准确率 | 85%（随机分布） | 95%+（错误集中在不熟悉环节） |
| 犯错时的态度 | 极度自信，毫无感知，一本正经地胡说 | 会脸红、会因不确定停下来问 |
| 常识判断 | 完全没有 | 天然具备 |
| 情境适应 | 只能匹配训练数据分布 | 能即兴变通 |
| 责任承担 | 服务条款免责：“风险自担” | 出错会道歉、会改正 |
| 单次错误成本 | 需高级员工重审 | 可自我发现并修正 |
| 综合可用性 | 实际是负价值 | 正价值 |
以下是 2025–2026 年多份权威机构追踪数据的集中汇总：
表 8：企业 AI 落地的失败率数据
| 研究机构 | 时间 | 样本 | 关键发现 |
|---|---|---|---|
| MIT NANDA | 2025 年底 | 300+ 家企业 GenAI 试点 | 95% 的试点未产生任何可衡量的 P&L 影响 |
| MIT NANDA | 2025 年底 | 全球企业累计投入 | 企业已在 GenAI 上投入 $300–400 亿，回报接近于零 |
| S&P Global | 2026 年初 | 200+ 家大型企业 | 企业平均在投产前砍掉 46% 的 AI 概念验证项目 |
| CIO 调查 | 2026 年 4 月 | 跨国企业高管 | 48% 公开承认”巨大失望”（去年 34%）；75% 承认部署 AI 只是”向股东做秀” |
| Gartner | 2026 年 7 月 | 全球客服系统 | 85% 的企业 AI 客服正在被大面积拆除，重建人工兜底 |
| IDC | 2026 年中 | 企业 AI 代理试点 | AI 代理 POC 在生产环境下的失败率高达 88% |
| Deloitte Tech Trends 2026 | 2026 年 | 企业 AI 部署 | 落地失败率 89% |
来源：MIT NANDA “State of AI in Business 2025”、S&P Global、Gartner 2026 年 7 月客服 AI 报告、Deloitte Tech Trends 2026、IDC 2026 中期报告。三份独立测量在一年内向同一个数字（85%–95% 失败率）收敛。
表 9：企业为什么退场——原因归因
| 退场原因 | 占比（企业调查） | 具体表现 |
|---|---|---|
| 准确率不足 / 幻觉严重 | 42% | 客服 AI 胡言乱语、代码生成漏洞百出 |
| 投入产出不成比例 | 28% | 订阅费 + 排雷成本 > 手工成本 |
| 无法责任归属 | 15% | 出错没人负责，法律风险高 |
| 系统集成困难 | 10% | 与现有 ERP / CRM 不兼容 |
| 员工抵触 / 学习成本高 | 5% | 一线员工反而效率下降 |
来源：S&P Global 200+ 企业调研 + MIT NANDA 报告综合归因，2026 年。
一句话：大企业下场吃螃蟹，吃了两年，现在吃不下也吐不出来。
正常的商业逻辑应该是：如果一款产品的准确率只有 85%，就应该停下来把它做到 99.9% 再推广。为什么整个行业反着来？
因为资本模式逼着他们必须这样做：
这就形成了一个极其讽刺、注定崩盘的死循环：
一个刺眼的对比：为什么在这场狂欢里，只有英伟达能稳赚不赔？
表 10：两种技术路径的残酷对比
| 维度 | 英伟达路径（20 年苦工） | 大模型巨头路径（2 年套现） |
|---|---|---|
| 关键节点 | 2006 年推出 CUDA，无人使用，股价跌 70% | 2020 年 GPT-3 一夜成名 |
| 核心承诺 | “算得最快、最准、绝不崩溃” | “通向 AGI、颠覆一切” |
| 输出确定性 | 100%（工业级） | 85–90%（概率机器） |
| 团队稳定性 | 核心工程师二十年深耕 | 顶尖科学家 1.5–2.5 年跳槽 |
| 产品成熟度 | 已被全球所有 AI 科学家绑定使用 | 无一模型达到工业级 |
| 财务表现 | 2025–2026 净利润 $1,200–1,500 亿 | OpenAI 单年亏 $209 亿 |
| 净利率 | 45%+ | –160% |
| 卡位强度 | 极难替代（CUDA 生态） | 高度可替代（谁强用谁） |
因为英伟达走的是完全相反的道路：
这就是”卖铲人”的智慧：淘金的人可能一大半死在路上（因为产品稀烂、没有利润），但一路上给他们卖水、卖坚固铲子的人，永远稳赚不赔。
任何无法在微观细节上做到脚踏实地的技术，最终都会在宏观的市场检验中轰然倒塌。 蒸汽机、电力、计算机之所以能开启工业革命，是因为它们走出实验室的那一刻，就已经达到了 99% 以上的稳定度。一台有 15% 概率会突然爆炸的蒸汽机，永远不会真正开启任何工业革命。
回到最初的两个问题：
第一，这套发展模式没办法盈利。 OpenAI 2025 年单年运营亏损 209 亿美元，2026 年将扩大至 140 亿美元现金亏损（GAAP 口径达 250 亿美元）；四大云 2026 年 AI Capex 总计将高达 7,250 亿美元、同比暴增 77%，至今没有一家能收回成本。整个下游行业只有 Anthropic 一家在 2026 Q2 首次实现运营盈利（5.59 亿美元），且深度依赖巨头补贴。真正稳赚的只有上游卖铲子的英伟达、台积电，一年拿走近 2,000 亿美元净利润。
第二，这套发展模式没办法工业化。 因为不愿意做基础性的打磨苦工，85% 的准确率成了所有模型无法翻越的天花板——MIT 研究显示 95% 的企业 AI 试点没能转化为任何商业利润，S&P Global 显示 46% 的项目在投产前被砍掉，Gartner 显示 85% 的企业 AI 客服正在被大面积拆除，48% 的 CIO 公开承认”巨大失望”。没有一个模型达到了真正的产品级别，就没办法在严肃行业推广，就形不成真正的商业化闭环，就无法把 AI 的作用发挥出来。
而这两个死结是互相锁死的：
这个死循环没办法一直走下去。资本的钱不是无限的，地主家也没有余粮。当越来越多的企业发现”集成 AI”带来的只有暴增的管理成本和无穷无尽的漏洞时，当越来越多的投资人发现自己的几百亿美元只换来了一个”吹得震天响、但漏洞百出”的聊天框时——恐慌和撤资潮就会瞬间爆发。
出路只有一条：停下来，把一件事做到 99.9%。
不是十件事做到 85%，不是发布下一个更大的模型，而是在某一个具体的垂直领域——纯粹的自动化会计、绝对准确的初级代码生成、完美的智能客服——把准确率老老实实地从 85% 抬到 99.9%。这是苦工、是脏活、是资本市场不喜欢的事，但这是这个产业活下去、真正做大的唯一路径。
当资本的潮水退去，真正能留下来赚到钱的，绝对不是天天发布新模型的巨头，而是那些愿意坐二十年冷板凳、把某一个具体功能做到工业级确定性的”笨公司”——就像今天的英伟达。
还没学会走就想飞的傲慢，注定要被常识和商业规律无情审判。 这不是我们的判断，这是每一段商业史都在反复演绎的铁律。