DeepSeek’s reported push to raise at least $12 billion at a valuation of about $74 billion ahead of a listing on Shanghai’s STAR Market will be the first real test of how the China discount applies to a frontier AI lab. That discount is structural. Alibaba and Baidu serve hundreds of millions of customers on large and growing revenue bases, yet trade at valuation multiples well below their US counterparts, and we see little reason a mainland listing of an AI lab will escape the same overhang. A STAR Market debut should find strong domestic demand, but US investors will likely be limited to indirect access through China-focused ETFs, and we do not expect the hundred-billion-plus outcomes we anticipate from US AI leaders. The more important story is the model itself. DeepSeek’s open-source models are competitive and far cheaper than closed alternatives, and they can run locally or on US inference platforms such as Together AI and Baseten with no data flowing back to China. One recent estimate attributes roughly 7% of all debt created over the past seven years indirectly to OpenAI and Anthropic, so the real pressure point for the frontier labs is not a bubble in AI demand but a shift of that demand toward open models. That shift would not strand the infrastructure: if Anthropic’s reported $518 billion in compute commitments went underused, providers like Lambda would rationally release the capacity and resell it to other buyers rather than litigate for years. Our house view is that if AI demand holds, compute finds a buyer whether workloads run on closed or open models, which makes the neoclouds a natural hedge across both outcomes.
Brett Adcock, CEO of Figure AI and founder of publicly traded eVTOL maker Archer, has launched Hark Pro into a personal-agent market forming quickly around Meta’s Muse, OpenAI’s Dots, and Instinct. The launch demo shows where the category is heading: a single request to plan a date night produced concert tickets matched to both partners’ Spotify playlists, a reserved parking spot, and a calendar entry timed to the drive. Tasks that once took an hour of planning collapse into one sentence, and as these agents become proactive they will increasingly suggest the plan before the user asks. Adcock’s record of multiple exits makes Hark worth taking seriously, and Meta’s context across WhatsApp and Instagram gives it a strong position as well. Our house view, however, is that the long-term consumer winners will be Apple and Google, which already hold users’ email, messages, photos, payments, and app usage across iOS and Android and can deliver an exceptional AI personal agent the moment they choose to switch it on. For that reason we are cautious on private B2C agent startups, which may carve out niches but lack the data advantage, while B2B agents remain an open field. In the enterprise, agents are likely to follow the data, favoring incumbents like Microsoft Copilot and Google Workspace, where firms such as RIAs can deploy an agent without moving client information to a new vendor. We expect widespread personal-agent adoption within 6 to 12 months.
Anduril’s Arsenal-2, a $3.7 billion shipyard at Sparrows Point, Maryland, backed by a Navy contract worth up to $2.9 billion, takes the company from existing undersea work into a full-scale maritime push and puts it in direct competition with Saronic and Huntington Ingalls. The facility will build Virginia-class submarine components and undersea systems, and it reinforces how quickly Anduril converts new products into production capacity, even as talk of a defense bubble grows around richly valued peers such as Shield AI. Palmer Luckey, alongside Elon Musk, operates on a different level in real-world manufacturing, and it is hard to bet against him winning more mandates of this kind. The key structural question is concentration. With the US military as essentially its only customer, Anduril’s revenue is likely to follow an S-curve, accelerating sharply before eventually plateauing, which raises questions about the revenue multiple it can sustain absent a move into commercial markets. That need not be a problem, since a public listing followed by a healthy dividend once growth matures would be a successful outcome for early investors in what becomes a value stock. We also expect a meaningful founder premium at IPO, with demand that could resemble SpaceX’s but potentially without the same post-listing comedown, since private-market access to Anduril has been far tighter than it was for SpaceX.
Etched is reportedly fielding offers at $40 billion or more, roughly double the $21 billion valuation it set in September, despite not yet having a chip in market. The company has taped out its chip, reportedly holds about $1 billion in customer orders, and counts a strong investor roster, but paying that price for a business without revenue looks rich to us. We prefer Positron, last priced at $5 billion, which already has systems deployed in data centers and is generating revenue before its next-generation chip, combining memory and inference, has even shipped. The design differences matter for deployment: Etched’s chip is liquid-cooled, while Positron’s is air-cooled and built to drop into virtually any data center. The closest parallel is Groq, which later drew a deal from Nvidia, and which did not offer the most groundbreaking technology but delivered fast, US-fabricated, high-volume inference with clear customer value. Positron fits that same profile of solid technology meeting overwhelming demand. More broadly, we view AI inference semiconductors as a sector to watch, with real money to be made among the companies racing Nvidia.