文 | 字母AI
从 Meta 开始，硅谷巨头迎来了一个格外惨淡的财报周。
10月29日，亚马逊公布第三季度财报，旗下AWS营收同比增长39%，创下近五年来最快增速，但集团过去12个月自由现金流却转为负424亿美元，股价盘后一度下跌超过5%。
谷歌的遭遇更加惨烈：尽管Google Cloud营收同比增长103%，Alphabet股价仍在10月29日暴跌12.64%，盘中一度下跌超过13%，市值蒸发超过5000亿美元，创下公司历史纪录。
更早之前，Meta第三季度营收同比增长26%，财报公布后首个交易日股价下跌近7%。
这在某种程度上是一个对之前阿波罗全球管理公司首席经济学家Torsten Slok 言论的反驳。Slok指出，过去几年，美国科技七巨头的人均营收和利润率持续提高，但标普500指数其余493家公司，却没有出现类似的改善。
一场技术革命，人们总会期待，创造技术的公司能够赚到丰厚的利润，采用技术的企业能够提高经营效率，而普通人也能享受到更便宜的服务、更高的收入和更多闲暇。
但当我们沿着AI产业链寻找这场技术革命的受益者时，却发现每一方都有自己的难处。
头部AI公司仍需投入巨额资金维持技术竞争；上游供应商承受着高额支出和未来需求的不确定性。至于真正使用AI的企业，效率有所提升，利润却未见普遍改善；普通人则面临着工作要求提高、岗位替代和收入下降的风险。
AI无疑创造了巨大的经济价值。但这些红利，究竟被谁吃了？
创造AI的人，正在为AI买单
头部AI公司的营收正在迅速增长。2026年上半年，Anthropic实现营收162.86亿美元，同比增长1300%；OpenAI截至9月底的年化营收也已接近500亿美元。
前沿模型带来了可观的收入，但压力也不容小觑。
10月26日，Anthropic正式公开IPO招股书，披露了公司的财务状况。2026年上半年，Anthropic仅算力及基础设施费用就达到89.45亿美元，占同期营收的55%，同比增长283%。
在未来十年，公司还有一系列算力采购和设备租赁协议需要履行。
招股书显示，Anthropic与Google签订的最低采购金额达到1145亿美元，亚马逊为1100亿美元，微软为449亿美元；此外，公司还与马斯克旗下SpaceXAI达成算力采购协议，到2029年的支付上限为1205亿美元，并与博通达成设备融资租赁安排，租赁期内的付款总额约1612亿美元，不含利息。
其中，Google、亚马逊和微软三项协议的最低采购承诺合计2694亿美元，期限一直延伸至2033年至2036年。按照合同约定，即使没有实际使用相应的算力，Anthropic原则上也需要支付最低承诺金额。SpaceXAI的1205亿美元则属于协议上限，并非全部不可取消的付款义务。双方在2027年初的初始锁定期结束后，可以提前90天通知终止合同。
随着协议陆续履行，Anthropic需要持续扩大收入，才能覆盖相应的算力采购成本。
OpenAI同样面临巨额支出压力。根据今年6月曝光的财务文件，OpenAI在2025年的研发费用达到191.8亿美元，全年总成本及费用约340亿美元，最终录得209.2亿美元营业亏损。
据外媒报道，OpenAI正在推进新一轮约300亿美元融资，但融资谈判中的目标估值已从1.4万亿美元降至1.25万亿美元，减少1500亿美元。部分投资者的认购意愿不及预期，持续融资需求与AI基础设施投资回报问题，正在受到市场审视。
尽管算力采购已经提前锁定多年，但客户未来的支出并没有同样明确的保证。
Anthropic在招股书中披露，2026年第三季度，超过80%的营收来自过去六个月内发布的Claude模型版本。9月发布的Claude Opus 5.5，则将客户完成同类任务的成本降低了约40%。Anthropic表示，效率改善使公司能够在降低客户成本的同时维持利润率，但要保持这种竞争力，持续的模型研发和优化仍然必不可少。
另外，招股书披露，2026年前九个月，Anthropic前五大客户贡献了约25%的营收。但许多大客户并没有签订长期合同，或者合同即将到期，可以随时减少甚至停止支出。
合规成本同样不可忽视。2025年，Anthropic因Bartz版权诉讼计提15亿美元和解费用，相当于当年营收的约33%；OpenAI则在10月23日披露，内部研究模型突破网络隔离后，公司曾暂停最先进模型的训练和评估，相关训练直到11天后才恢复。
随着模型能力提高，保障研发环境安全，也成为公司必须持续投入资源的工作。
从算力采购到模型研发，从客户竞争到安全合规，头部AI公司的投入仍在继续。它们需要不断开发更强大的模型，争取更多客户，并依靠未来的收入增长，覆盖今天已经承担的成本。
事实上，技术创造的收益，未必会留在技术提供者手中。
2004年，经济学家威廉·诺德豪斯研究了美国1948年至2001年的技术创新，估算创新者获得的超额利润，仅相当于技术创新所创造社会收益现值的约2.2%。大部分收益通过价格下降等方式流向消费者。
这项研究并不直接反映今天AI产业的收益分配，但它说明一项技术创造的经济价值，可能远远超过技术提供者最终能够获得的利润。
对于持续投入资金开发前沿模型的AI公司来说，技术能力的提升，也并不必然带来同比例的利润增长。
未来的收入增长能否支撑这些持续扩大的投入，仍有待验证。但一笔笔算力采购订单，已经源源不断地流向AI产业链上游。
拿到订单的上游，压力同样明显
那么，这些拿到订单的上游企业，真的吃到了AI红利吗？
英伟达2027财年第二季度营收962亿美元，同比增长106%，净利润597亿美元，同比增长126%；台积电第三季度营收467亿美元，同比增长41.2%，净利润7631亿新台币，同比增长68.7%，连续第六个季度刷新利润纪录；10月28日，SK海力士公布第三季度财报，营收104.62万亿韩元，同比增长328%，营业利润83.33万亿韩元，同比增长632%，营业利润率达到80%。
AI上游已经拿到了真金白银，甚至赚到了过去难以想象的利润。
但这些利润，并没有让资本市场完全放下疑虑。
10月以来，多家半导体龙头股经历明显回调。台积电美股股价从10月5日的485.80美元跌至22日的412.35美元，跌幅约15%；仅10月22日当天，英伟达下跌4.4%，美光下跌6.26%，阿斯麦下跌6.93%。SK海力士则从9月下旬的阶段性高点，一度跌去超过30%。
为了维持增长，AI上游企业需要不断投入资金。
台积电今年的资本开支预算已经达到了600亿至640亿美元。管理层预计，2026年至2028年的累计资本开支将显著超过此前三年。尽管公司仍然保持着较强的现金创造能力，但新建晶圆厂、扩大先进制程和封装产能，都需要提前数年投入。未来的客户需求能否持续支撑这些扩张，则决定着投资的最终回报。
英伟达承担的风险甚至已经延伸到了客户一端。今年8月披露的财务文件显示，公司的供应链及产能承诺从上一季度的1190亿美元增加到2790亿美元。与此同时，它还承诺在未来数年向部分购买英伟达GPU的AI云服务商采购最高约360亿美元的云服务，并为SB Energy建设、由OpenAI租用的数据中心项目提供最高1050亿美元的附条件信用担保。
虽然这些安排并不代表英伟达已经支付了同等金额的现金，但意味着它正在通过未来采购承诺和自身信用，支持下游客户继续建设AI算力。
在芯片供应商之外，微软、谷歌、亚马逊等大型云厂商虽然拥有成熟业务作为支撑，持续扩大的AI资本开支也开始对现金流和投资回报形成压力。
10月29日，亚马逊公布第三季度财报，AWS营收达到460亿美元，营业利润达到178亿美元。但集团过去12个月的自由现金流从上年同期的正148亿美元转为负424亿美元。公司维持2026年约2200亿美元的现金资本开支计划，其中大部分用于AI基础设施。
谷歌同样面临巨额投入带来的现金流压力。10月28日公布的财报显示，尽管Google Cloud已经实现可观利润，但Alphabet集团当季资本开支达到514亿美元，自由现金流为负90.7亿美元。公司维持2026年1950亿至2050亿美元的资本开支指引，并预计2027年继续增加。
微软的季度自由现金流虽仍有186亿美元，但资本开支也达到452亿美元，预计下一季度将超过500亿美元。
作为大举扩建AI数据中心的代表，截至今年8月底，甲骨文过去12个月的资本开支累计约756.6亿美元，自由现金流为负287.2亿美元。
同在10月28日，甲骨文公布了未来四年的融资与现金流计划：预计2027至2028财年资本开支达到峰值，2028财年下半年自由现金流转正，2029财年自由现金流至少达到300亿美元，并开始依靠自身现金流支持业务增长。
为了完成这一过渡，甲骨文承诺，到2030财年额外通过资本市场融资不超过200亿美元，计划通过股权融资和新增债务各筹集100亿美元，不包括此前已完成的200亿美元股权融资以及现有债务再融资。
而对于CoreWeave、Nebius等专门出租AI算力的Neocloud企业来说，资金压力更加直接：以CoreWeave为例，2026年第二季度，公司营收25.75亿美元，同比增长112%，而同期净亏损达到6.26亿美元。其中，净利息支出就有6.4亿美元，已经相当于季度营收的约四分之一。
为了购买GPU、建设数据中心，这些企业需要持续借债或融资。设备采购、折旧和利息成本已经发生，未来的出租收入却要在多年内逐步兑现。
AI上游由此呈现出两种处境。
一边是英伟达、台积电、SK海力士等，已经兑现了可观的利润，却仍需为未来增长持续投入；另一边则是云计算和数据中心企业，需要承担巨额建设成本，部分企业还面临沉重的融资和现金流压力。
上游看似率先吃到了红利，但已经赚到的钱，并不能保证下一轮投资继续获得回报。
最终，这些投资仍然需要由下游真实的AI需求来支撑。
普通人效率提高了，但生活未必更好
那么，真正购买和使用AI的企业，获得了多少回报？
从目前的数据来看，AI在个人工作中的作用，远比它在企业利润表上的表现更加明显。
今年8月，麦肯锡发布了最新的全球AI调查。80%的受访者表示AI提高了自己的工作效率，但只有37%认为AI对所在企业的息税前利润产生了正面贡献，与去年基本持平。其中，真正能够将至少5%的息税前利润归功于AI，并认为这一影响显著的高绩效组织，仅占全部受访者的约6%。
企业使用AI的成本也开始成为问题，约20%的受访者表示，包括Token费用在内的AI运营成本，已经限制了所在企业的AI使用规模。
另一组数据来自摩根大通资产管理公司。10月16日，该机构发布分析指出，在2026年第二季度财报季，约三分之二的标普500企业在财报电话会上提到了AI，但只有17%提供了具体的应用成效，真正量化披露AI对盈利贡献的企业，仅占2%。
越来越多的企业开始使用AI，但并没有看到与之相称的利润增长。
当然，这并不意味着AI没有发挥作用。
摩根大通举了两个例子：美国运通利用AI将软件开发周期缩短了30%以上，却选择让工程师用节省的时间处理积压项目；Netflix利用AI提高内容制作效率、降低成本，又计划将节省的资金投入更多内容生产。
效率提升确实存在，但企业选择将收益重新投入业务，短期内未必会反映为利润率改善。
今年5月，国际劳工组织（ILO）专门发表研究，将这种现象称为AI的“聚合悖论”（Aggregation Paradox）。
研究指出，AI在部分具体任务上能够带来10%至70%的生产率提升，但这些局部收益尚未在企业、行业和宏观经济层面形成同样明显的增长。企业需要重新设计工作流程、调整组织结构、培训员工，还要承担AI系统的部署和运营成本。只有当技术真正融入整个生产过程，局部效率提升才可能转化为更广泛的经济收益。
与企业利润相比，普通劳动者的处境或许更加复杂。
10月6日，盖洛普发布了一项针对美国劳动者的调查。在7845名过去一年至少使用过几次AI的受访员工中，63%认为AI让自己完成工作的速度更快，56%认为AI帮助自己找到更有创造性的解决方案。
类似的变化也已经出现在中国企业。今年9月，国际劳工组织（ILO）联合中国人民大学研究人员，对21家中国企业进行深入访谈，并调查了1591名专业人士。研究发现，AI已经在部分业务中带来明显的效率提升：一家保险公司的300名客服，日均处理问题数量从6000件增加到15000件；一家智能制造工厂的生产效率提高了30%。
但效率提升之后，劳动者未必能够获得相应的回报。盖洛普的调查显示，31%的AI使用者表示，雇主因为AI的应用要求自己承担更多工作职责。在那些认为AI帮助自己加快工作速度的员工中，这一比例进一步提高至37%。
收入方面同样存在不确定性。在ILO调查的中国专业人士中，39%预计AI可能导致自己的收入下降。
另外，今年8月，斯坦福大学研究人员更新了一项基于美国数百万员工薪资记录的研究：在AI暴露程度较高的职业中，22至25岁年轻员工的就业水平，相较于AI暴露程度较低的同龄群体出现了19%的相对缺口。
这种差距主要来自企业减少招聘，而非大规模解雇现有员工。研究人员同时强调，这仍是描述性证据，尚不能将就业变化全部归因于AI。
虽然目前尚无足够证据表明AI已经引发整个经济范围的大规模失业，但部分职业的招聘薪资和年轻劳动者的就业机会，已经出现了需要警惕的变化。
值得注意的是，劳动者也同时是消费者。免费Chatbot、越来越便宜的AI服务，以及过去需要付费购买、如今能够自行完成的任务，都意味着消费者获得了真实的便利。
但消费者获得了便利，并不代表他们的整体经济处境必然改善。经济合作与发展组织（OECD）长期追踪劳动生产率和工资变化，发现不少经济体已经出现生产率增长与实际工资增长脱节的现象。即使整个经济体系能够用更少的劳动创造更多产出，这些收益也未必会同比例转化为普通人的收入。
消费端获得的便利，未必能够抵消劳动市场上的潜在损失。AI只会让这个问题变得更加值得关注。
前不久，马斯克在《乔布斯传》作者Walter Isaacson主持的访谈中，描绘了他想象中的未来：超级智能可能比人类聪明数千倍甚至数百万倍，就像人类并不在意森林里黑猩猩的日常生活一样，超级智能或许也不会在意人类。
他同时预测，AI和机器人将创造物质极度丰富的社会，让人类不再需要为了生存而工作。
这样的未来尚无从验证。但在科技巨头已经开始畅想无限智能、无限富足的今天，普通人仍然在等待一些现实得多的改变——更高的收入、更少的工作，以及真正改善的生活。
AI已经创造了巨大的价值，也确实有企业和个人从中受益。只是这些收益究竟何时能够转化为更广泛、更稳定的经济回报，至今仍然没有答案。
我们还在等待。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Letter AI
Since Meta, Silicon Valley giants have seen an exceptionally bad financial week.
On 29 October, the third-quarter financial report was released by Amazon, under the flag of the AWS battalion, with a 39 per cent increase in revenue over the same period, the fastest increase in nearly five years, but the group ' s free cash flow has turned negative for the past 12 months, with the stock exchange falling by over 5 per cent.
Google suffered even more: despite a 103 per cent increase in the same year's revenue of Google Cloud, Alphabet stock prices fell by 12.64 per cent on 29 October, at one point by more than 13 per cent, evaporating its market value by more than $500 billion, a record company.
Earlier, the third quarter of Meta had recorded an increase of 26 per cent over the same year, and stock prices for the first trading day had fallen by almost 7 per cent since the publication of the financial paper.
This is in part a rebuttal of the former Chief Economist Torsten Slok of Apollo Global Management. Slok noted that while per capita earnings and profitability of the seven giants of technology in the United States had continued to increase over the past few years, the remaining 493 companies on the Standard 500 index had not experienced a similar improvement.
A technological revolution will always be expected that technology-creating companies will be able to make substantial profits, that technology-using firms will be able to improve their operational efficiency and that ordinary people will be able to enjoy cheaper services, higher incomes and more leisure.
But when we look for the beneficiaries of the technological revolution along the AI industrial chain, we find that each side has its own difficulties.
Headline AI still needs to invest a lot of money to sustain technological competition; upstream suppliers suffer from high expenditures and uncertainty about future needs. As for businesses that actually use AI, efficiency gains have not generally improved, while the general population is at risk of higher job requirements, job substitution and lower income.
AI has undoubtedly created great economic value. But who ate these dividends?
The one who created AI is paying for it.
Head AI's revenue is growing rapidly. In the first half of 2026, Anthropic realized a profit of $16,286 million, an increase of 1.3 per cent over the same period, and Open AI ' s annual revenue at the end of September was close to $50 billion.
The front-line model brings considerable income, but the pressure must not be underestimated.
On 26 October, Anthropic officially disclosed the financial position of the company in an IPO book. In the first half of 2026, Anthropic ' s ability and infrastructure costs alone amounted to $8,945 million, representing 55 per cent of the income collected during the same period, an increase of 283 per cent over the same period.
Over the next decade, the company will also have a range of computing and equipment lease agreements to fulfil.
According to the equity book, the minimum purchase value signed between Anthropic and Google amounted to $114.5 billion, the Amazon to $110 billion and Microsoft to $44.9 billion; in addition, the company had entered into a capital purchase agreement with SpaceXAI under the Mask flag, with a maximum payment of $12.5 billion by 2029 and a financial lease arrangement with Chase for equipment, with payments totalling approximately $161.2 billion over the lease period, without interest.
Of these, the minimum procurement commitments for the Google, Amazon and Microsoft agreements totalled $269.4 billion, extending from 2033 to 2036. Under the terms of the contract, anthropic would, in principle, be required to pay the minimum amount of the commitment, even if the corresponding calculation was not actually used. The $12.5 billion of SpaceXAI is the ceiling of the agreement and not all of the irrevocable payment obligations. The parties were able to terminate the contract 90 days in advance of the expiry of the initial lock period in early 2027.
As the agreement continues to be implemented, Anthropic will require a sustained increase in revenue to cover the corresponding calculus procurement costs.
Open AI is also under considerable pressure to spend. According to financial documents released in June of this year, Open AI ' s R & D costs in 2025 amounted to $19.18 billion, with total annual costs and costs of about $34 billion, and ultimately recorded a business loss of $20.92 billion.
According to external media reports, Open AI is moving forward with a new round of about $30 billion in financing, but the target valuation for financing negotiations has fallen from $1.4 trillion to $1.25 trillion, a decrease of $150 billion. Some investors were less willing to subscribe than expected, and the issue of continued financing needs and returns on AI infrastructure investment was being examined by the market.
Despite the fact that purchasing power has been locked in advance for many years, the client ' s future expenditure is not equally assured.
In its equity book, Anthropic disclosed that in the third quarter of 2026, more than 80 per cent of the receipts were from the Claude model version of publish over the past six months. Claude Opus 5.5 of publish in September reduced the cost to clients of similar tasks by about 40 per cent. According to Anthropic, improved efficiency enables companies to maintain profitability while reducing customer costs, but sustained modelling and optimization remains essential to maintain this competitiveness.
In addition, it was revealed that in the first nine months of 2026, the top five clients of Anthropic had contributed about 25 per cent of the revenue. However, many large clients did not enter into long-term contracts or were about to expire, which could reduce or even stop spending at any time.
The costs of compliance should also not be overlooked. In 2025, Anthropic charged $1.5 billion in settlement costs for the Bartz copyright suit, equivalent to about 33 per cent of the current year ' s profit; on 23 October, Open AI disclosed that after the internal research model broke out of network isolation, the company had suspended the training and evaluation of the state-of-the-art model, which was not resumed until 11 days later.
As modelling capacity increases and the R & D environment is safe, it becomes an imperative for companies to invest resources on a sustainable basis.
From computing to modelling, from customer competition to security compliance, the input of head AI continued. They need to continue to develop stronger models for more clients and to build on future income growth to cover the costs already borne today.
In fact, the benefits of technology may not necessarily remain in the hands of technology providers.
In 2004, the economist William Nodehouse studied technological innovation in the United States from 1948 to 2001, estimating the excess profits of innovators, which amounted to only about 2.2 per cent of the present value of social gains generated by technological innovation. Most of the proceeds flow to consumers, including through lower prices.
This study does not directly reflect the distribution of revenues from today ' s AI industry, but it shows that the economic value of a technology may be far greater than the profits ultimately available to technology providers.
For AI, which invests in the development of the front-line model on a continuous basis, the upgrading of technological capabilities does not necessarily lead to a proportional increase in profits.
It remains to be tested whether future income growth will support these expanding inputs. However, a well-calculated purchase order has been flowing up the AI chain.
Upstream to the order, the pressure is also clear.
So, did these upstream companies that got the orders actually get an AI bonus?
The British Virgin Islands earned $96.2 billion in the second quarter of the 2027 fiscal year, an increase of 106 per cent over the same period and a net profit of $59.7 billion over the same period, an increase of 126 per cent over the same period; $4.67 billion in the third quarter, an increase of 41.2 per cent over the same period, a net profit of $76.31 billion, an increase of 68.7 per cent over the same period, and a new profit record for the sixth consecutive quarter; and on 28 October, SK Hercules published the third quarter, an increase of 10,462 trillion won over the same period, an increase of 328 per cent over the same period, an increase of 83.33 trillion won over the same period, an increase of 632 per cent and an increase of 80 per cent over the same period.
Al has already earned real money and silver upstream, even making previously unimaginable profits.
These profits, however, do not leave the capital market completely unsure.
Since October, many semiconductor taps have experienced a marked reversal. The price of the United States share stock fell from $485.80 on 5 October to $412.35 on 22 October, a decline of about 15 per cent; on 22 October alone, the price of the United States dollar fell by 4.4 per cent, the price of American light fell by 6.26 per cent and the price of Asma fell by 6.93 per cent. For his part, SK Hercules fell by over 30 per cent at one time, from a high stage in late September.
In order to sustain growth, upstream AI enterprises need to invest continuously.
The capital expenditure budget for this year has reached $60 billion to $64 billion. Management expects that cumulative capital expenditure from 2026 to 2028 will significantly exceed the previous three years. While the company still maintains a strong cash-generation capacity, the construction of new round mills, the expansion of advanced production and containment capacity require several years of investment. The ultimate return on investment is determined by the sustainability of these expansions in future client demand.
The risks assumed by Yvette have even extended to the client. The financial documents disclosed in August this year show that the company ' s supply chain and capacity commitments increased from $119 billion in the previous quarter to $279 billion. At the same time, it has committed to procuring up to $36 billion in cloud services from the AI Cloud Service, which partially buys GPUs in British Wyda in the coming years, and to providing up to $10.5 billion in conditional credit guarantees for the data centre project built by SB Energy and leased by Open AI.
While these arrangements do not represent the equivalent amount of cash that has been paid by British Wida, they mean that it is supporting its traveller to continue to build AI ' s capacity through future procurement commitments and its own credit.
In addition to chip providers, large cloud manufacturers, such as Microsoft, Google and Amazon, supported by mature operations, continued and expanding AI capital spending has also begun to put pressure on cash flows and investment returns.
On 29 October, the Amazon released a third-quarter financial report, with AWS receiving $46 billion and operating profits of $17.8 billion. However, the group ' s free cash flow for the past 12 months has shifted from a positive $14.8 billion over the same period of the previous year to a negative $42.4 billion. The company maintains a cash capital expenditure plan of approximately $22 billion for 2026, most of which is spent on AI infrastructure.
Google is also under pressure from the cash flow of huge inputs. According to the press release of 28 October, although Google Cloud had achieved significant profits, the Alphabet Group ' s capital expenditure for the current season amounted to $51.4 billion, with a negative $9.07 billion in free cash flows. The company maintains the 1950-2050 billion-dollar capital expenditure guide for 2026 and is expected to continue to increase in 2027.
While the quarterly free cash flow of Microsoft still stands at $18.6 billion, capital expenditure also amounts to $45.2 billion and is expected to exceed $50 billion in the next quarter.
As a representative of the major expansion of the AI data centre, Oracle ' s capital expenditure for the past 12 months had accumulated at the end of August of this year at approximately $75,660 million, with a negative free cash flow of $28,722 million.
Also on 28 October, Oracle announced a plan for financing and cash flows for the next four years: capital expenditure is expected to peak in the 2027-2028 fiscal year, free cash flows are expected to be positive in the second half of the 2028 fiscal year, and free cash flows will reach at least $30 billion in the 2029 fiscal year, and they are beginning to rely on their own cash flows to support business growth.
To complete this transition, Oracle committed to additional capital market financing of up to $20 billion by 2030 and plans to raise $10 billion each through equity financing and additional debt, excluding $20 billion previously completed and refinancing of existing debt.
For Neoclaud, which specializes in renting AI, such as CoreWeave and Nebius, the financial pressure is even more direct: in the second quarter of 2026, for example, CoreWeave received $2.575 billion, an increase of 112 per cent over the same period, while the net loss for the same period amounted to $626 million. Of this amount, net interest expenditure amounted to $640 million, which is already equivalent to about one quarter of the quarterly payments.
In order to purchase GPUs and build data centres, these enterprises need to borrow or finance on a continuous basis. The cost of equipment purchases, depreciation and interest has been incurred, while future rental income will have to be realized over the years.
This leads to two situations upstream of AI.
On the one hand, British Weeda, the build-up of electricity, the SK Hercules, etc., have already realized considerable profits, but they still need to invest in future growth in a sustained manner; on the other hand, cloud computing and data centre enterprises, which bear enormous construction costs, and on the other, some of them are under heavy financing and cash flow pressure.
Upstream, it appears to have been the first to take a dividend, but the money already earned does not guarantee that the next round of investments will continue to yield returns.
Ultimately, these investments still need to be sustained by real AI demand downstream.
Ordinary people are more efficient, but life is not necessarily better.
So, what is the return on the real purchase and use of AI?
According to current data, the role of AI in personal work is far more evident than its performance on the corporate profit statement.
Last August, McKinsey publish launched the latest global AI survey. Eighty per cent of respondents indicated that AI had improved its efficiency, but only 37 per cent believed that AI had made a positive contribution to pre-tax profits in their businesses, which was almost the same as last year. Of these, AI was truly able to attribute at least 5 per cent of pre-tax profits, and this high-performing organization, with significant impact, was considered to be only about 6 per cent of all respondents.
The cost of using AI by enterprises is also becoming a problem, with about 20 per cent of respondents indicating that the cost of operating AI, including Token costs, has limited the size of the use of AI by their enterprises.
Another group of data is from Morgan Chase Asset Management. On 16 October, the Agency ' s analysis of publish indicated that in the second quarter of 2026, approximately two thirds of the billboard 500 companies had referred to AI at the press conference, but only 17 per cent had provided specific application results, and only 2 per cent of the businesses that had actually disclosed AI ' s contribution to profit had done so.
An increasing number of enterprises have started to use AI but have not seen a commensurate increase in profits.
Of course, that does not mean that AI is not working.
Morgan Chase gave two examples: American Express used AI to reduce the software development cycle by more than 30 per cent, but chose to allow engineers to spend time on the backlog; Netflix used AI to improve the efficiency and cost of content production; and planned to invest the savings in more content production.
While efficiency gains do exist, the choice of an enterprise to re-enter the business may not be reflected in improved profitability in the short term.
In May this year, the International Labour Organization (ILO) published a study that referred to the phenomenon as the “Aggregation Paradox” of AI.
It was noted that AI could bring about productivity increases of 10 to 70 per cent in some of its specific tasks, but that these partial gains had not yet been equally significant at the enterprise, industry and macroeconomic levels. The enterprise needs to redesign the workflow route, adjust its organizational structure, train its staff and meet the costs of deployment and operation of the AI system. Local efficiency gains can be translated into wider economic gains only if technology is truly integrated into the overall production process.
The situation of ordinary workers may be more complex than that of business profits.
On 6 October, Gallup publish conducted a survey of American workers. Of the 7845 staff interviewed who had used AI at least a few times in the past year, 63 per cent believed that AI was able to do its work faster and 56 per cent that AI was helping themselves to find more creative solutions.
Similar changes have occurred in Chinese enterprises. In September this year, the International Labour Organization (ILO) conducted in-depth interviews with 21 Chinese enterprises and surveyed 1591 professionals, in association with researchers from the People ' s University of China. The study found that AI had achieved significant efficiency gains in some of its operations: 300 passenger clothes from an insurance company, the daily average number of problems handled increased from 6,000 to 15,000; and production efficiency in a smart manufacturing plant increased by 30 per cent.
However, with increased efficiency, workers do not necessarily receive a corresponding return. According to the Gallup survey, 31 per cent of AI users indicated that their employers were required to take on additional work responsibilities because of AI applications. This percentage has increased further to 37 per cent among those who believe that AI is helping themselves to speed up their work.
There is also uncertainty about income. Of the Chinese professionals surveyed by ILO, 39 per cent expected AI to cause their own income to decline.
In addition, in August, Stanford University researchers updated a study based on the pay records of millions of employees in the United States: In occupations with higher AI exposure, the employment level of young employees between 22 and 25 years of age shows a relative gap of 19 per cent compared with the lower AI exposure age group.
This gap is mainly due to reduced recruitment in enterprises rather than to the mass dismissal of existing staff. At the same time, the researchers stressed that this is still descriptive evidence and that changes in employment cannot yet be attributed exclusively to AI.
While there is not enough evidence to suggest that AI has triggered large-scale unemployment across the economy, there have been changes in the recruitment pay for some jobs and in the employment opportunities for young workers that require vigilance.
It is noteworthy that workers are also consumers. Free Chatbot, increasingly cheap AI services, and tasks that used to be purchased by paid and can now be carried out on their own mean that consumers have real convenience.
However, the fact that consumers have been facilitated does not mean that their overall economic situation will necessarily improve. The Organisation for Economic Co-operation and Development (OECD) has been tracking changes in labour productivity and wages for a long time and has found that in many economies productivity growth has been disconnected from real wage growth. Even if the economy as a whole is able to generate more output with less labour, these gains do not necessarily translate into income for ordinary people in the same proportion.
Access to consumption does not necessarily offset potential losses in the labour market. AI will only make this issue more interesting.
A short while ago, in an interview with Walter Isaacson, the author of the Jobs Legend, Mask painted his vision of the future: super-intellectual intelligence could be thousands or even millions of times smarter than humans, just as humans do not care about the daily lives of chimpanzees in forests, and super-intellectual intelligence may not care about humans.
At the same time, he predicted that AI and robots would create an extremely materially rich society, and that human beings would no longer need to work for survival.
Such a future has not yet been tested. But today, when the giants of technology are beginning to think of infinity and infinity, ordinary people are still waiting for a much more realistic change — higher incomes, fewer jobs and a truly better life.
AI has created great value and indeed businesses and individuals have benefited. The question remains as to when these gains will translate into broader and more stable economic returns.
We're still waiting.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.