E200 | An In-Depth Investor Conversation: Core Moats and Investment Logic for AI Agents
Summary
Agent financing has entered the phase of $10B valuations and talent poaching by tech giants. But pricing increasingly reflects strategic scarcity, not proven startup moats. Cursor parent Anysphere raised $900M at nearly $10B; Google paid $2.4B to bring over Windsurf’s core team, while Cognition AI pursued the remaining team; Thinking Machines Lab raised $2B at a $12B valuation. 周炜’s warning is that global capital is crowding into a handful of quality assets, and “money isn’t money anymore”（钱都不是钱了）. Big Tech will build in-house and acquire at a premium.
The Agent boom is being driven by advances in models, open-source ecosystems, and the first real end-to-end products—not by a concept that suddenly appeared. 张璐 defines an Agent as a system that can make decisions autonomously, handle complex multistep tasks, and complete work end to end with support from tool libraries and vertical knowledge bases. Over the past 6 months, even traditional businesses running on little more than Excel and email have become addressable. She emphasizes that open-source models have lowered the barrier to vertical fine-tuning and small-model innovation; 周炜 believes the industry has finally moved from debating foundational technology to building Agents that help users achieve an actual end goal.
The most dangerous illusion in AI Coding is mistaking first-mover advantage for a durable moat. Claude is one of Cursor’s underlying engines and also competes directly through Claude Code. 张璐 observes that “70% to 80%” of research-oriented developers use Claude Code, with some unicorns migrating from Cursor or Windsurf. Cursor’s UI/UX, context integration, and existing integrations create migration costs, but 周炜 believes a large model can fill in the gaps with “one night of training.” Without proprietary data or a genuine technical moat, “sell while you can”（该卖就卖了）. 张璐 also classifies AI Coding as primarily to C, while Fusion Fund focuses mainly on to B opportunities sold directly to enterprises.
The investability of vertical to B Agents comes from high-value workflows, not from how large the industry sounds. 张璐 cites Walmart, which issues roughly $60B in commercial paper each year. A portfolio company with fewer than 7 employees automated a process previously dependent on people and banks, generating $6B of paper in a week and charging just “0.01 to 0.02”—still a substantial business. The investment thesis is to find workflows that are frequent, important, repetitive, and boring, because “many verticals are actually very large markets.”
In highly regulated industries, the technical goal is not zero hallucinations but an error rate within a controllable range, while making cost, privacy, and deployment work at the same time. 张璐 is particularly interested in combining large language models with reinforcement learning: LLMs handle understanding, analysis, and expression, while reinforcement learning adds decision-making, drive, learning, and memory. Repetitive tasks can also reuse verified answers instead of regenerating them every time. She strongly favors vertical small models because they require less inference, training, and energy and can run locally; one model described as having “under 100M to 1B tokens” can reportedly run on a Raspberry Pi with performance similar to GPT-4.
General-purpose Agents may be technically viable, but the economics and competitive dynamics are far harder for startups. They must call powerful models such as GPT-4, Claude, and Gemini, while broad cross-industry coverage drives up inference costs. If “AI labor costs more than human labor,” enterprises have little reason to adopt them. More fundamentally, foundation-model companies will build general-purpose Agents themselves. 周炜 is still willing to bet on the category because “without dreams, how are you different from a salted fish?”（没有梦想那跟咸鱼有什么区别）—but he admits the odds of success are “extremely slim.”
More durable investment moats include layered business processes, non-public data, and a clear willingness to pay. 周炜 uses the octopus machines from The Matrix as a metaphor: General AI may break through the outer layer quickly, but industry know-how, granular processes, and private data force Big Tech to break through “layer by layer.” These companies may never reach $10B, yet could still produce a cohort of $1B businesses. US enterprises pay once they see commercial value; Chinese customers are more likely to haggle and cap usage-based fees, so Chinese teams often need to target the global market from day one.
The next to C cycle may not be another general chat interface, but products that use local memory and emotional relationships to create real switching costs. 周炜 favors companion robots with local deployment and continuous training: cloud models “each rule for 100 days,” leaving users free to switch at any time. 泓君 imagines that if a $29 monthly subscription lapses and causes a dedicated AI to forget shared experiences, the emotional connection could itself become a reason to pay. 周炜 expects it to become common for the AI-native generation to fall in love with robots, and believes a lifelike product that crosses the uncanny valley could appear “within 2 years.” Meanwhile, this cycle’s high-priced acquisitions show that startups need not bet everything on the endgame; exiting midway is a viable path.
Deep dive
1. Financing Frenzy, Model Capability, and Product Adoption Are Accelerating in Parallel
泓君’s valuation map now extends well beyond the usual “unicorn” tier: Anysphere raised $900M in June 2025 at a valuation near $10B; after Windsurf’s talks with OpenAI collapsed, Google paid $2.4B to bring over its core team, while Cognition AI pursued the remaining team; Devin’s parent company was valued at roughly $4B.
Vertical categories are crowded as well: enterprise search Agent Glean raised $150M at a $7.2B valuation; legal Agent Harvey AI raised $300M at nearly $5B; Genspark and Manus raised $100M and $75M, respectively, with both valued above $500M.
张璐 argues that Agents are not a new concept—the industry has discussed them for nearly a decade. What has changed is the astonishing growth of vertical companies, rapid model iteration, and an open-source ecosystem that has lowered the threshold and cost of fine-tuning small models. Traditional industries running on little more than Excel and email are now becoming viable targets for Agent integration. 周炜 believes the industry has finally shifted from discussing foundational technology to building end-to-end products that let users actually complete their goals.
2. An Agent’s Boundary Is Autonomous Completion of Complex Tasks, Not a Single Model Call
张璐’s investment definition is clear: an Agent must handle complex rather than linear tasks, make decisions autonomously, and complete a multistep workflow from start to finish. Teams still need to provide a tool library, while vertical products also require a knowledge base; these are her basic screening criteria.
The forms she cites range from Cognition’s Devin, Rabbit OS, You.com, and Glean to the healthcare vertical Clarity. You.com recently became a unicorn. Asked about her personal experience, she says Anthropic’s Claude is her favorite. With the episode recorded around the launch of Grok 4, she found its parameter performance “extremely impressive” and was also looking forward to the yet-to-be-released Coding Model.
泓君 divides the market into 3 groups: AI Coding tools such as Cursor, Windsurf, and Devin; general-purpose Agents in the style of Manus; and vertical products such as Harvey, Cresta, and healthcare Agents. All 3 are called Agents, but they have completely different cost structures, moats, and investor pools.
3. Claude Code Shows That Foundation Models Can Move Downstream and Take Application-Layer Users at Any Time
张璐 sees Claude Code as a direct competitor to Cursor. Claude is one of Cursor’s built-in engines, but Anthropic’s own product is more deeply integrated, with stronger capabilities in complex code analysis, explanation, generation, and long-context processing. It can handle large projects at the 100K- and 200K-token scale.
The development environment is a direct migration trigger. Among a group of research-oriented developers she has spoken with, “70% to 80%” already use Claude Code. Some unicorns that previously relied heavily on Cursor or Windsurf have also begun switching, with its Slack and Notion integrations proving particularly smooth.
xAI illustrates another form of threat. 张璐 relays internal information that “70% to 80%” of xAI’s code is already written by its own internal coding model. If that model were opened to third parties, its native performance could surpass external tools that merely embed Grok 4.
New entrants include Google’s Gemini CLI, which enables natural-language interaction through the terminal. It may not compete head-on with a full IDE, but it gives AI Coding startups a new platform and further shows that model companies are extending upward across the entire toolchain.
4. Cursor Has Real Product Advantages, but Must Iterate at the Limit to Stay Alive
Cursor’s breakthrough over Copilot was not just model capability; it rebuilt the UI/UX for AI programming. Autocomplete, code explanation and modification, and natural-language context integration are all smoother. With a background in materials science, 张璐 was able to use it alongside ChatGPT or Gemini to build small applications, “dramatically lowering the barrier.”
Windsurf’s supporters place greater value on the fluency of multistep execution, which may make it better suited to building AI Agents. Teams that have already integrated Windsurf deeply will not immediately migrate just because a new model appears; this migration cost is a real but limited defense.
张璐 says she has watched interviews with Cursor’s founder and believes that after Anthropic’s July 2024 update, Copilot failed to turn the leap in coding capability into a timely product iteration, while Cursor seized the window. Startup growth is not linear: when models make a leap, the winners rapidly productize the experience gap.
张璐 summarizes the competition in one word: “involution”—卷. Roughly “70% to 80%” of her portfolio companies achieved 20x revenue growth within a year; the fastest went from zero to tens of millions of dollars, with one rising from $500K to more than $100M. But the faster the growth, the shorter the window: stop iterating and the moat is hard to maintain.
张璐 also classifies AI Coding as primarily to C. Fusion Fund has long preferred to B products sold directly to enterprises, so it has paid relatively little attention to general-purpose consumer Agents.
5. The Market for Vertical to B Products Is Hidden in High-Value, Boring Workflows
Fusion Fund has long preferred to B: beyond whether a product sells directly to enterprises, it distinguishes between a thin wrapper around a large model and a product capable of training a vertical model or fine-tuning a small one. 张璐 believes AI competition has expanded from model performance to “data and cost,” with enterprise customers carefully comparing the price of every call.
Commercial paper issuance is the clearest example. Walmart has a large global workforce and issues roughly $60B of commercial paper annually to supplement payroll cash flow. This frequent, high-value, important yet repetitive process was previously handled mainly by people and traditional banks.
A portfolio company with fewer than 7 employees automated the workflow, generated $6B of commercial paper in a week, and signed Walmart as a customer. 张璐 says it charges just “0.01 to 0.02”—the episode does not further specify the pricing unit—but her conclusion is clear: specialization does not mean a small market. Industry expertise can reveal enormous transaction volumes hiding beneath the surface.
6. For Highly Regulated Agents, the Goal Is a Controllable Error Rate, Not a Fictional 100% Accuracy
When 泓君 asks how to make results 100% accurate, 张璐 does not promise to eliminate hallucinations. Finance, insurance, healthcare, and legal services cannot tolerate arbitrary errors the way consumer products sometimes can. The realistic goal is to use a new architecture to keep the error rate within a defined, controllable range.
The “small-circle consensus” she observes is to combine LLMs with reinforcement learning: large language models provide understanding, analysis, and expression, while reinforcement learning handles decision-making, drive, learning, reinforcement, and memory. The 2 systems “fill each other’s gaps”; reinforcement learning may become an indispensable layer of Agent architecture.
Enterprise workflows have a natural advantage: across 10 calls, 5 may require the same function. The system need not ask a large model to regenerate an answer every time. It can retrieve an accurate result that has already been successfully verified, using memory and reuse to reduce randomness while cutting redundant inference calls.
7. Small Models Win on Accuracy, Energy, Privacy, and Deployment at the Same Time
张璐 has a “very strong preference” for fine-tunable vertical small models: the narrower the task boundary, the easier it is to optimize for precision. Smaller models also reduce training, inference, compute, and energy costs, at a time when electricity has become a real constraint on AI expansion.
Enterprises may not want to upload all of their core data to the cloud. Existing IT architecture, privacy, and compliance requirements will push customers in finance and healthcare toward deployment on local networks or devices. Large models may demand more compute and power than on-site equipment can provide; small models can complete the work at the edge.
She cites a portfolio company whose smallest model has “a token scale of under 100M to 1B”—the original wording is somewhat ambiguous—but can run directly on a Raspberry Pi, with performance described as similar to GPT-4. That deployment capability creates both cost and local-execution advantages.
8. General-Purpose Agents Have a High Capability Ceiling, but Startups Face a Tighter Economic Ceiling
Asked why the US has seen few Manus-style general-purpose Agents, 张璐 first corrects the premise: Mira Murati’s Thinking Machines Lab is also pursuing a similar general-purpose terminal. During the episode’s editing, the company was reported to have raised $2B led by a16z at a $12B valuation, showing that capital has not abandoned the direction.
The first obstacle is cost. Cross-industry, cross-task generalization requires powerful foundation models and extremely capable Agent optimization. If “AI labor costs more than human labor,” many industries will prefer to keep using people; vertical Agents can use small models and fine-tuning to achieve lower costs and higher precision.
The second obstacle is competition. General-purpose products need to call models such as GPT-4, Claude, and Gemini, while the companies behind those models will also build their own general-purpose Agents. 泓君’s question is direct: “Where is your competitive advantage against them?” Technical feasibility does not mean a startup has structural advantages.
9. Coding Is Strategic Infrastructure for Big Tech, and Negative Gross Margins Will Not Stop the Competition
周炜 explains why AI Coding commercialized first: code is itself a language, making it a natural fit for large models. It is also the foundation of all digital technology, creating a massive market; programmer costs have remained high, so the value of substitution can be measured directly. It is one of the first directions outside the foundation-model layer to show a clear revenue model.
泓君 mentions Google’s Gemini CLI, Anthropic’s Claude Code, Grok 4, and the upcoming Coding Model. 周炜 emphasizes that OpenAI, Microsoft, and other giants will not abandon the field. Windsurf’s $2.4B deal and Cursor’s nearly $9.9B valuation remind him of the dot-com bubble and the early mobile internet—only now acquisition prices have expanded from $100M to several billion dollars.
He believes large models have sharply compressed the first-mover window. In the past, a team needed 1 or 2 years to learn a new industry; now, “one night of training and it already has almost everything it needs.” If capability comes from public internet data, with no proprietary data or technical moat, Cursor-like companies should seriously consider that “it’s time to sell.”
Big Tech can also tolerate negative gross margins for strategic value. 周炜 uses the $20-per-month ChatGPT subscription to illustrate how high users’ inference and API costs can be. As in the subsidy wars between Didi and Kuaidi, giants with additional strategic value—such as payments—can absorb a business whose revenue does not cover its costs. Coding is foundational to “letting AI write its own code,” so model giants may continue investing regardless of margins.
10. A Truly Investable Agent Must Force Big Tech to Break Through Layer by Layer
周炜 believes the industry previously looked “a lot like the Blockchain circle”: too much discussion of foundational technology and too little effort to build usable products. Manus may not ultimately succeed, but it pointed in the right direction. Users care whether a task can be completed end to end—“who cares which AIs it directed behind the scenes?”
He compares the competition to the octopus machines attacking Zion in The Matrix. General AI may break through the outermost layer quickly; the second layer is industry-specific knowledge, and the third is finer-grained know-how and business processes. The more layers there are, the more time Big Tech must spend breaking through them; if the niche is not large enough, large companies will be even less willing to invest in the short term.
Private data is the moat 周炜 values most. Industries such as healthcare contain data unavailable on the public internet, making it difficult for large-model companies to find and use that data for training. These companies may never become $10B businesses, but reaching $1B is not difficult. As for General Agents, although “the hope of success is extremely slim,” he continues to bet because “without dreams, how are you different from a salted fish?”（没有梦想那跟咸鱼有什么区别）
to C products face high costs, entrenched giants, and low user loyalty at the same time. Foundation-model and elite-team projects are funded mainly by deep-pocketed investors such as Amazon, Google, Microsoft, a16z, SoftBank, and Rive Capital. Users are still in the novelty phase, and purchased traffic can migrate at any time, as 泓君 puts it: “each model rules for 100 days.”
11. Local Memory May Create Consumer Lock-In, While the US-China Market Gap Determines the Route to Monetization
周炜 is most excited about the next wave of to C products: companion robots with a local AI brain. Users should train them themselves, giving them unique memories and communication styles. 泓君 takes the idea further: if interrupting a $29 monthly subscription caused a dedicated AI to forget shared experiences, the slightly Black Mirror-like design might create a willingness to pay through emotional attachment.
周炜 uses a family experience to demonstrate the value of long-term memory. ChatGPT connected a child’s medical examination report from several days earlier and analyzed the trend; it also once inferred from blood-test and ultrasound materials that the child might have Epstein-Barr virus, after which the hospital tested for and confirmed it. In his argument, the value lies not in a single answer but in personalized historical data accumulating over time.
On human-machine relationships, 周炜 makes a clear prediction: “I completely believe that people can fall in love with robots, and I believe this is the inevitable future.” The AI-native generation, accompanied by intelligent toys from childhood, will accept humanoid robots as naturally as the iPad-native generation accepted e-books. He believes a product capable of crossing the uncanny valley could appear “within 2 years.”
Commercial adoption still differs by geography. US to B customers generally pay once they see the value, with both willingness to pay and pricing levels relatively clear. Chinese customers are more likely to push down prices, reject unlimited usage-based increases, and impose revenue caps, so Chinese teams need to globalize early. 周炜’s biggest change of view at midyear was recognizing that the market is now dominated by giants and giant enterprises, while end-to-end products have finally become a consensus. High-priced acquisitions also give founders an exit path without requiring them to “play all the way to the endgame.”