Why Most AI Projects Will Fail — And How to Find the Companies That Won't

以下是内容的中文翻译： Bumi公司董事长兼首席执行官、拥有30年企业软件经验的史蒂夫·卢卡斯（Steve Lucas），对当前人工智能领域的真正受益者和面临的挑战提出了深刻见解。在接受Motley Fool分析师雷切尔·沃伦（Rachel Warren）采访时，卢卡斯断言，尽管许多公司正在构建出色的人工智能模型，但目前只有英伟达（NVIDIA）一家明确地从人工智能中赚取了丰厚的利润。 卢卡斯强调了一个显著的转变：在公司仅仅因为拥有人工智能战略而获得回报的初期阶段之后，华尔街现在要求可衡量的业务成果和清晰的投资回报率（ROI）。他指出，“投资回报率超越人工智能”，并提到许多由于董事会压力而仓促上马的早期人工智能项目未能带来预期的高回报。 人工智能开发的财政规模令人震惊。卢卡斯指出，美国四大超大规模云服务提供商在人工智能上的资本支出从去年的约4000亿美元猛增至今年超过7000亿美元——这是对更广泛市场而言的“煤矿里的金丝雀”（早期预警信号）。尽管人工智能和基础设施支出飙升，但软件应用的投资却在下降。这种巨额支出，加之缺乏可证明的投资回报率，标志着人工智能“空白支票”时代的终结。卢卡斯透露，他所咨询的首席执行官们迫切需要展示成果，董事会也很快将积极地追究高管的责任。 训练先进人工智能模型的成本已从GPT-2的大约5万美元暴涨，预计到2026年，前沿模型的成本可能超过10亿美元——卢卡斯将其比作从“二手车跳到航空母舰”。他指出，OpenAI据报道每月烧钱30亿美元，这在长期来看是不可持续的，暗示这些成本最终将不可避免地转嫁给消费者或企业。这种财务压力已经导致包括特斯拉（Tesla）和SpaceX在内的一些公司限制其内部人工智能支出。卢卡斯预测，在六个月内，大多数首席执行官将实施内部支出限制，并且董事会将在批准进一步重大人工智能投资前，要求明确的投资回报率。 卢卡斯认为，多达40%的企业人工智能项目可能会被放弃，这一数字得到了Gartner预测的支持——Gartner认为到2027年，许多“智能体项目”（agentic projects）将失败。他将这种失败率归因于人们轻易启动人工智能项目，却未充分考虑业务需求或战略成果，从而导致了缺乏生产性成果的广泛试验。 卢卡斯表示，项目成功的关键因素是信任。借鉴他在软件领域三十年的经验，他强调：“如果人类不信任某样东西，它就永远不会被使用。”他指出，许多商业智能项目因数据不准确导致信任破裂而失败。对于人工智能而言，挑战甚至更大，因为它不仅涉及数据，还可能涉及工作职能。他将“人工智能抢走工作”的说法斥为“无稽之谈和恐慌、不确定、怀疑情绪（FUD）”，坚称真正的问题是人们是否足够信任人工智能，以允许它执行关键任务。 放眼英伟达之外的下一波人工智能赢家，卢卡斯指出了提供“关键人工智能基础设施”的公司。例如，他自己的公司Bumi专注于连接人工智能模型、提升数据质量并安全地将数据传输给它们。他引用Snowflake、Databricks和Datadog为例，这些以数据为中心的组织正在获得“巨大的收益”和增长。他认为，下一波主要浪潮将是“面向B2B的企业人工智能”，它将严重依赖数据和基础设施。 对于投资者，卢卡斯建议寻找拥有“数据和图谱护城河”的企业——即那些独特的、不可复制的，能作为人工智能“能源”的数据。他还警告投资者警惕首席执行官们笼统的“我不仅仅是软件”的说法，这通常预示着他们试图抵御颠覆性担忧。他认为，目前亏损数十亿的前沿模型公司，如果无法显著渗透劳动力市场，其长期变现将举步维艰，而这正是他们最初希望实现但尚未实现的情景。 对于归因于人工智能效率的裁员，卢卡斯称之为“大量的炒作”，敦促投资者要求具体的、财务层面的数据和证据来证明这些效率，而不是接受那些方便的说辞。他表示，真正变革性技术最可靠的指标是加速的*新客户获取*，而不仅仅是向现有客户推销产品。 从更宏观的层面来看，卢卡斯对人工智能在医疗保健领域的潜力表达了深切的乐观，援引了他作为1型糖尿病患者的个人经历。他相信人工智能将彻底改变疾病管理，并最终在二十年内带来更长寿、更健康的生活。然而，他最大的担忧在于一个伦理困境：人工智能获取极其个人化的健康数据，同时又被用于产品和服务的营销——这在造福人类与仅仅是“卖你一杯咖啡”之间划出了一条细微的界限。

Steve Lucas, Chairman and CEO of Bumi and a 30-year enterprise software veteran, offers a sharp perspective on the true beneficiaries and challenges in the current AI landscape. Speaking with Motley Fool analyst Rachel Warren, Lucas asserts that while many companies are building impressive AI models, NVIDIA is currently the *only* one unequivocally making substantial money from AI. Lucas highlights a significant shift: after an initial phase where companies were rewarded for merely having an AI strategy, Wall Street is now demanding measurable business outcomes and a clear return on investment (ROI). He states, "ROI supersedes AI," noting that many early AI projects, rushed due to board pressure, are not delivering expected high returns. The financial scale of AI development is staggering. Lucas points to the four major US hyperscalers, whose CapEx on AI jumped from approximately $400 billion last year to over $700 billion this year—a "canary in the coal mine" for the broader market. While AI and infrastructure spending are soaring, investment in software applications is declining. This immense spending, coupled with a lack of proven ROI, signals the end of the "blank check" era for AI. Lucas reveals that CEOs he consults are desperate to demonstrate returns, and boards will soon aggressively hold executives accountable. The cost of training advanced AI models has ballooned from around $50,000 for GPT-2 to potentially over a billion dollars for frontier models by 2026—likening the jump from a "used car to an aircraft carrier." He notes that OpenAI reportedly burns $3 billion a month, unsustainable in the long term, implying these costs will inevitably be passed on to consumers or enterprises. This financial pressure is already leading some companies, including Tesla and SpaceX, to cap internal AI spending. Lucas predicts that within six months, most CEOs will impose internal spending limits, and boards will demand clear ROI before approving further significant AI investments. Lucas suggests that as many as 40% of enterprise AI projects could be abandoned, a figure supported by Gartner's predictions that many "agentic projects" will fail by 2027. He attributes this failure rate to the ease of starting AI projects without proper consideration for business requirements or strategic outcomes, leading to widespread experimentation that lacks productive results. A crucial factor in project success, according to Lucas, is trust. Drawing from his three decades in software, he emphasizes that "if humans don't trust something, it will never be used." He points out that many business intelligence projects fail due to inaccurate data leading to a breakdown of trust. For AI, the challenge is even greater, as it concerns not just data but potentially job functions. He dismisses the "AI taking jobs" narrative as "nonsense and FUD," asserting that the real issue is whether people trust AI enough to allow it to perform critical tasks. Looking at the next wave of AI winners beyond NVIDIA, Lucas identifies companies providing "critical AI infrastructure." His own company, Bumi, for instance, focuses on connecting to, improving the quality of, and securely delivering data to AI models. He cites Snowflake, Databricks, and Datadog as examples of organizations centering around data that are seeing "massive benefit" and growth. The next major wave, he believes, will be "enterprise AI for B2B," heavily reliant on data and infrastructure. For investors, Lucas advises looking for businesses with "data and graph moats"—unique, non-replicable data that serves as the "energy" for AI. He also cautions against generic "I'm not just software" messaging from CEOs, which often signals an attempt to ward off disruption fears. He argues that frontier model companies, currently losing billions, will struggle to monetize long-term without significant penetration into labor markets, a scenario they initially hoped for but hasn't materialized. Regarding layoffs attributed to AI efficiency, Lucas calls it "a whole lot of spin," urging investors to demand concrete data and financial proof of such efficiencies rather than accepting convenient narratives. The most reliable indicator of truly transformative technology, he says, is accelerated *new customer acquisition*, not just pushing products on existing clients. On a philosophical note, Lucas expresses profound optimism for AI's potential in healthcare, citing his personal experience as a Type 1 diabetic. He believes AI will revolutionize disease management and ultimately lead to longer, healthier lives within two decades. However, his greatest caution lies in the ethical dilemma of AI's access to deeply personal health data being simultaneously used for marketing products and services—the fine line between benefiting humankind and simply "selling you a cup of coffee."