美国新兴互联网：市场护城河与代理商业的颠覆风险

Nikhil Devnani, CFA+1 917 344 8425nikhil.devnani@bernsteinsg.comNathan Gee+1 917 344 8573nathan.gee@bernsteinsg.com THE MARKET IS SCRUTINIZING MARKETPLACES A year ago, we co-authored a thought piece on the risks AI wouldpose to the incumbent Internet businesses we cover — USInternet Weekender: Could AI agents kill the Internet? For much of2025, the debate felt like a longer-term consideration and not animminent risk factor. OTAs aside, the rest of the Internet basket wasgetting credit for partnering with OpenAI and Gemini (particularly We underestimated how quickly the narrative could flip. The pace ofinnovation and product development out of the AI labs has forcedinvestors to re-think baseline assumptions around competitionand moats. Software stocks have been particularly hard-hit by this GOOD BUSINESSES GETTING DISCOUNTED Asset-light, inventory-light, efficient working capital cycles,two-sided network effects...previously lauded for scalability,marketplace business models have now come under scrutinyas investors question who the ultimate aggregator is going tobe several years down the road. The US Internet landscape hasbeen divided into fiefdoms — distinct, scaled aggregators across We appreciate that the world is moving quickly, and the nature ofsoftware and technology development is changing. This createsuncertainty, and uncertainty handicaps multiples. We’re left fightinga bear case that cannot be disproven outright. Sentiment is likely to Nonetheless, the Street’s expectations of risks and opportunitiescould ebb and flow — we liken it to the debate we have been livingwith on Uber and the Autonomous Vehicle narrative.And thependulum on Agentic Commerce feels like it’s swung too farto the negative at this stage.Delays on AGI timelines, limiteduptake from consumers, capacity/compute/ financial constrains The possibility of personalized, centralized Agents that can handlethe full gamut of tasks from scheduling your dentist appointmentto ordering you groceries to booking your next vacation to helpingyou buy a home is a sci-fi dream that now feels within reach — on Signs are starting to emerge with OpenAI’s evolving go-to-market and reprioritization given Anthropic’s success.Consumer behavior does not change that quickly, newchannels can be incremental (online penetration rates are There are additional considerations at play as well — do you end upwith more competition if the barriers to writing software diminishconsiderably? Vibe coding already appears to be ushering in anew era of accelerated app development...but is anybody usingthese apps? Separately, what do all these advancements mean CONSUMER BEHAVIOR IS HARD TO CHANGE We have observed limited impact to date on Consumer Internetmarketplaces. AI tools are certainly being used in the searchand discovery process, but are not yet acting as closed loopecosystems. Observed referral traffic from AI chatbots to thetop eCommerce marketplaces remains under 1% of total web There continue to be hurdles, be it around the checkout flow,basket building, or simply trusting that the AI tools have made thebest recommendation possible. Is ChatGPT’s suggestion more Agentic Commerce bulls can argue these variables aresurmountable, and we remain in the very early stages of thistechnology shift. But, trust and reliability are aspects of themarketplace moat that go under-appreciated. The onus is on new The Middle East war and its knock-on effects for consumerspending have thrown another variable into the equation.And thus we are still looking for the bottom on some of thesestocks given the possibility of estimate cuts. However, themedian Internet P/E is now at a compelling ~20x GAAP The best business models in the sector have proven tobe resilient through ups and downs. DASH is the namewe’d highlight in our coverage, but we could make similararguments for ZG, UBER, and others. We expect investors to EXHIBIT 5:Consumer behavior is hard to change — newexperiences have to be markedly better. Our non-technical definition of Agentic is that personalized AIassistant that can do things for you with increasing complexity overtime: (1) complete assigned research and information gatheringtasks → (2) curate the results based on personal preferences → (3)transact on your behalf → and (4) and eventually become proactiveand anticipate future needs (is this AGI?). To work, AI agents need We are starting to see progress towards this end-state, but thesetools are far from becoming mainstream. OpenClaw sparked theimagination. You can now dispatch desktop tasks to Claude viatext. Gemini’s latest offerings on Android devices will allow you tocall a car or order food, though early reviews also acknowledgethe experience is both amazing and clunky at the same time — 9 LONGER-TERM CONSIDERATIONS MAY BE DIFFERENT, Existing apps and websites are built for humans — not agents— which adds a layer of friction. Hence, the need for structuredcommerce protocols if this is all going to