科技的未来：智能体AI收入TAM的探讨

Future of Tech: Debating the TAM for agentic AI revenue A quantitative framework for agentic AI adoption.We’ve felt for a while that the realworld usefulness of AI models was being held back by users’ ability to prompt effectively. Assuch, we’d expect AI agents that can plan tasks and formulate more complex responses - loop workflows… and user ARPU based on regional wage norms.Replacing human labour is the real prize.If the task completion duration of AI agentscontinue to double every 6-9 months, per METR’s observation, then AI agents capable of handling weeks or months-long tasks could become available either side of 2030. As thishappens, we’d expect agentic AI services that displace humans to unlock a much largeropportunity - where pricing models shift from subscription fees to capture of the value of displaced human labour. Compared with a $100bn estimated revenue TAM by 2030, mainlyreflecting AI agents deployed as part of human-in-the-loop workflows, we think the 2035TAM (factoring in some labour displacement) could be multiples of this number. Nearer term, are we close to a local maximum for agentic AI sentiment?In practice,many issues complicate the adoption of AI agents that displace humans, like organisationalbureaucracy, accountability systems, tech validation and maintenance, etc. The demo-to- product gap for agentic AI bears watching, with benchmaxxing and hallucination remaining changes into the very distant future has felt notable. For what it’s worth, when we asked arange of SOTA AI models to plot recent AI innovations on a Gartner-style hype cycle chart,“AI agents” were unanimously assigned to the peak of inflated expectations. OpenClaw raises the floor for agentic AI capabilities.From an investing perspective,one of the more notable recent developments (apart from Minimax and Z.ai hitting $80bn ofcombined market cap) was Tencent’s WorkBuddy announcement. We’ve always considered the view that Tencent might somehow be locked out of an agentic AI future somewhatshort-sighted… in an environment where open source continues to make progress, and tech diffusion remains rapid. Tencent’s decision to allow users to use multiple models topower WorkBuddy workflows felt notable… and raises interesting questions about relativevalue capture between the model and application layers within AI stack. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS The step change in agentic AI capabilities year to date has been impressive to observe, and has driven both massive interest inthe new tech (including Minimax and Z.ai as pure play stocks), and terminal value fears for the Internet platforms we cover. From a technology standpoint, the ability of AI agents to plan and complete more complex tasks should help to unlock a wider rangeof AI use cases. Beyond the initial discovery phase driven by early adopters among developers and investors, we expect the use of AI agents to augment human-in-the-loop workflows to proliferate globally in the coming years. Beyond 2030 we’d expectthe expansion of task completion horizons to support the emergence of AI agents capable of replacing human labour acrosssome jobs. Adoption of such services could expand the AI agent market by a factor of multiples… compared with a $100bn orso revenue TAM by 2030. Nearer-term, we consider OpenClaw in a somewhat similar light as DeepSeek a year ago, in that the new open source techhelps to help raise the floor for agentic AI capabilities across the Internet platforms, while catalysing usage among developersand users. For Tencent, we think the WorkBuddy launch - which interestingly allows users to use a range of AI models - argues WeChat’s Mini Program ecosystem remains one of the most intriguing potential uses of agentic AI.For Alibaba, we think the recent decline in the company’s shares chiefly reflects concerns on the state of the e-commerce VALUATION COMPS TABLE DETAILS SIZING THE AGENTIC AI OPPORTUNITY Year to date, agentic AI has dominated the discussion among internet sector investors globally. The development directionallyaligns with our expectations at the end of last year that the prevailing AI discourse would need to move from model performance to action-based AI, and has driven both massive excitement (helped by a low free float, etc.) around the shares of listed AI labsMinimax and Z.ai, and disruption fears for many Internet sector incumbents. For a measure of the fluidity of investor views on this topic, see Tencent’s shares getting sold because of concerns the company had lagged on agentic AI development… thenbought in response to Tencent’s roll-out of its version of OpenClaw.In response to a recurring question in our discussions with investors, this note outlines our attempt to build an estimate for the TAM of agentic and broader AI services. Rather than pretend we can come up with some kind of high conviction revenueprojection, our aim here was to build an intellectually defensible framework based on bottom-up d