最近几周，软件股因担忧人工智能会使软件商品化并空心化传统商业模式而持续下跌；问题是：如果任何人都可以“vibe code”自己的应用，Salesforce、Intuit、Workday 或 Snowflake 会发生什么？尽管 Nvidia 上周报告了创纪录利润并给出强劲预测，其股价仍下跌，显示市场预期极高。该公司大部分营收来自最大的云计算提供商，但其增长正在越来越多地来自企业业务。围绕这一背景，Nvidia 首席执行官 Jensen Huang 认为已到达“agentic AI”拐点——即嵌入企业工具并实际执行任务的软件代理，这与 Salesforce、ServiceNow 等公司正在大力推动的产品方向一致，信息非常明确：若 AI 使软件更有价值，嵌入 AI 的公司（并由 Nvidia 芯片提供支撑）应与其共同增长。

在 Huang 的框架下，经济核心不再是训练大语言模型，而是推理（inference），即在真实世界中持续运行 AI。每次模型运行都把数据转成 token，进而驱动代码助手、客服机器人和流程自动化。如果客户愿意为输出支付的金额高于芯片和云基础设施成本，那么就形成“可盈利 token”；计算使能推理，推理生成 token，token 产生收入。由此，AI 并不会毁掉软件，而是成为嵌入其中的新营收层。平台则变为 AI 代理的交付系统，持续在核心运营中生成 token。

有早期迹象支持这一观点。Snowflake CEO Sridhar Ramaswamy 表示，过去四周（约 28 天）完成的内部项目，如今可在 40 分钟内完成，客户愿意高额购买 token 以换取此类效率；其公司预计产品收入年增约 27%，好于预期。Salesforce 同时指引到 2030 年年收入达 630 亿美元，并称其 AI 产品已超过 8 亿美元年度经常性收入。此前 Bloomberg Intelligence 分析师 Anurag Rana 就认为，长期里 AI 将强化微软、SAP 和 Workday 这类深度嵌入企业的大型平台，而较小公司更易受冲击，但近期这些领头公司仍可能受压，因为它们仍按用户数计费且大型企业招聘受限。尽管如此，怀疑仍是必要的：Nvidia 以销售用于推理的 GPU 为生，直接受益于 AI 需求持久且扩张的叙事，而“计算=推理=收入”恰好符合其激励结构。市场困境在于，AI 是在中空化软件公司，还是作为更高算力的增长引擎嵌入其中；Nvidia 坚持后者，但近期抛售表明投资者更倾向于前者。

Software stocks have been sold off in recent weeks because investors fear AI will commoditize software and hollow out traditional business models; the central question is whether companies like Salesforce, Intuit, Workday, or Snowflake can survive if anyone can “vibe code” their own apps. Despite Nvidia reporting record profits and strong forward guidance last week, its shares still fell, signaling that expectations had become extremely elevated. Most of Nvidia’s revenue still comes from the largest cloud providers, yet its growth is increasingly coming from enterprise. Against this backdrop, CEO Jensen Huang argued that the market has reached an inflection point with “agentic AI”—software agents embedded in enterprise tools that actually perform tasks—a direction actively promoted by Salesforce, ServiceNow, and peers; the message is that if AI makes software more valuable, the firms embedding it (and powered by Nvidia chips) should grow along with Nvidia.

In Huang’s framing, the real economic engine is not model training but inference, the continuous operation of AI in real-world contexts. Each inference run transforms data into tokens, which then power code assistants, customer-service bots, and workflow automation. If customers are willing to pay more for that output than it costs in chips and cloud infrastructure, those are “profitable tokens.” Compute enables inference; inference generates tokens; tokens drive revenue. In this framing, AI does not destroy software, but adds a new revenue layer embedded inside it, with platforms becoming delivery systems for AI agents that continuously generate tokens as part of core operations.

There are early signs supporting this view. Snowflake CEO Sridhar Ramaswamy said internal projects that once took four weeks (about 28 days) can now be completed in 40 minutes, and customers are willing to spend heavily on tokens to unlock that productivity; Snowflake expects product revenue to rise about 27% annually, better than expected. Salesforce guided toward $63 billion in annual revenue by 2030 and said its AI product has surpassed $800 million in annual recurring revenue. Earlier, Bloomberg Intelligence analyst Anurag Rana argued that AI should strengthen deeply embedded enterprise platforms like Microsoft, SAP, and Workday, while smaller firms are more likely to be disrupted, though near-term pressure remains for leaders because software is still priced by user counts and hiring at large firms is constrained. Skepticism is necessary: Nvidia sells the GPUs that power inference and benefits directly from the assumption that AI demand is durable and expanding, and the compute-inference-revenue chain aligns with the chipmaker’s incentives. The market dilemma is whether AI is hollowing out software firms or becoming a more compute-intensive growth engine embedded within them—Nvidia argues for the latter, while the selloff suggests investors currently lean toward the former.