Can AI Applications Raise the Next Round? A Conversation with 马克汤

Summary

• 马克汤 was categorical that Anthropic would never, ever acquire Cursor: “It doesn’t need to acquire it; you’ll die on your own anyway.” The logic: Cursor is essentially an advanced, highly engineered wrapper, but it has to pay Anthropic the cost implied by Claude API’s roughly 75% gross margin; over time, it cannot beat Claude Code. On the same $40/$100 plan, Cursor offers far less Opus usage—“it’s not that it doesn’t want to give you more; it can’t afford to.” Raymond said Musk bought Cursor into SpaceX for $60B; 马克汤 believes the asset is the “intersection” for high-value tasks, and that connecting it to X’s data centers would make the gross-margin model work. Whether the deal uses SpaceX stock is only Raymond’s speculation; the $10B breakup fee is 马克汤’s description of an option-like structure.
• AI applications have three possible moats—data advantage, data-cleaning “dirty work,” and inference efficiency—and 马克汤 believes efficiency is what survives. Model labs “haven’t eaten them yet only because they haven’t had the bandwidth”: enterprise services are labor-intensive, while model companies pursuing AGI do not want the distraction. Vertical small models are essentially out of the conversation by 2026: “They’ll all be steamrolled by the emergent intelligence that comes with a sufficiently large parameter count.” Raymond’s analogy is that hiring generalists from Peking University or Tsinghua will beat specialists from lower-tier universities even with the same training time. He also cited OpenEvidence buying access to major U.S. journals as a reminder that data advantages can be weakened if model vendors simply buy the rights.
• Account pools and intermediaries form an AI-application supply chain, and they were also part of Seedance’s success. Users resell unused Codex 20x keys; unused enterprise token allocations may likewise be converted into pocket money. Cloud providers’ trial and credit balances flow into account pools, allowing intermediaries to undercut direct purchases from Amazon and Azure—the competition is supply-chain management. Volcano Engine generates RMB1B a month from Seedance alone and raised its 2026 target from RMB10B to RMB15B, with wrappers such as LibTV making an important contribution.
• AI-app gross margins structurally cannot match traditional software: application companies effectively pay model-company gross margins as costs. Traditional software routinely carries 80%-90% gross margins; for an AI application, getting above 50% or 60% is already exceptional. Many to C/prosumer products are negative-margin businesses, including Manus in 马克汤’s view; positive margins are concentrated among niche vertical users willing to pay a steep premium. The fate of companion products is even more explicit: “Borderline content is the easiest way to make money; without it, honestly, making money is pretty hard.”
• ByteDance’s Seed lineup is underappreciated, and Feishu is due for an AI-era re-rating. Doubao feels good but benchmarks poorly because its free product assigns users to different model tiers—“the only thing that can respond quickly is a small model.” The newly released Seed 2.1 is already benchmarking in the GPT-5.5 and Opus 4.7 tier, with particularly strong multimodal capabilities; its audio performance benefits from Douyin data. Feishu had already built comprehensive interfaces before the AI era and is “native MCP”; in 马克汤’s Codex/Claude Code workflow, its skills rank second in usage frequency. ByteDance rates itself “at least an 8” after 2 years, with its focus on Doubao as the winning choice.
• The most valuable AI application may not be an app at all, but the acquisition and retrofit of a traditional company. The Thrive Holdings-style roll-up and AI insurers that write their own policies and underwrite them follow the same logic: licenses, SOPs, and customer relationships already prove the business model; what is missing is a scaling engine. OpenAI and other model companies are also forming JVs with PE firms, potentially covering 500 portfolio companies across 5 PE firms. Raymond’s analogy is that the winner in real-estate internet was Beike, with physical stores—not a pure lead-generation app.
• Asked whether to spend RMB100 on an AI application or on storage, optical components, or PCBs, 马克汤 chose the application: “If applications don’t take off, all the hardware is a bubble.” He invoked Huang’s five-layer cake: “Consumers don’t pay for the packaging technology inside HBM… They pay for products that improve the experience, optimize efficiency, and make life better.” But pure-play AI-app names remain scarce in public markets; Meitu, BlueFocus, and Chinese Online do not qualify—“the product function has to be sufficiently AI-native, and the organization has to be sufficiently AI-native.”

Deep dive

1. Cooling Sector: The North Star Shifts from DAU to ARR, and Capital Is Concentrating

• Raymond set the tone at the outset: market attention has shifted toward foundation models and semiconductors, and “no one has talked about AI applications for a long time.” Many of the star companies that attracted bets in 2024-25 as candidates to become “the Douyin of the AI era” have not announced a financing round in years. He made his position clear: he is “extremely skeptical” of the AI-application market as a whole.
• 马克汤’s timeline: GPT-3.5 was unusable in early 2023; GPT-4.0 made many things doable; the theme gained momentum in mid-2023 and exploded in 2024-25. Large financings were still happening in 2026, but the bet had changed—from DAU and traffic to profitability, user mindshare, and “what its ecosystem position is in the competition with foundation models.”
• Silicon Valley’s capital allocation shows the same concentration. Roughly 40% of VC money invested in AI over the past year went to OpenAI and Anthropic, 30%-40% went into hardware, and the remaining 30% went to applications—but that last bucket was split across several thousand deals and was “extremely fragmented, small, and miscellaneous.”

2. Why Foundation Models Have Not Eaten Vertical Applications: Timing and Three Moats

• Raymond’s extreme scenario: model companies poach quants from Jane Street, spend heavily to label DCFs, codify best practices as skills, and put them in a marketplace—at which point vertical applications should disappear entirely. Why are U.S. applications still alive? 马克汤’s answer: model vendors “haven’t eaten them yet only because they haven’t had the bandwidth.” Enterprise deployments require on-site work and customized project management, while model companies pursuing AGI “don’t want to turn themselves into labor-intensive companies.”
• China’s B2B market is primarily project-based. Legal, finance, healthcare, and automotive are the main verticals, and related services generally depend on local deployment, typically handled by model companies or cloud vendors. 马克汤 cited DeepSeek’s optimization work for Huawei’s Ascend 950PR.
• The limits of the data moat are becoming clear. Nobody is training vertical models anymore: injecting vertical data into a general-purpose model produces more lift than training a small vertical model. 财搭子 was still working with Zhipu AI’s institute on LoRA and SFT for a financial model in 2024; “as of today in 2026, basically nobody is discussing this anymore.” Data is valuable only when it is private—meeting minutes, brokerage research, and Knowledge Planet-style alternative data. “If it exists in the public domain and ChatGPT and Claude can search it, it has no value.” Raymond again cited OpenEvidence buying access to major U.S. journals, arguing that even this kind of data moat can be purchased by model vendors.
• The second moat is dirty work. A vertical pipeline splits the same text into objective information—Nvidia’s share price at 3:10 p.m. is unambiguous—and subjective views, which require cross-checking and analysis of the distribution of opinions, then routes them through different inference paths. “This dirty work still matters.” Efficiency is where the economics lie: pre-classification and tagging reduce the money and time required for the same task. Real-time tasks need speed, but for long-running agents that can run for 1-2 days, it may not be worth cutting runtime from 4-5 hours any further.
• 马克汤 ranks efficiency first among the three moats as the one that will remain, while conceding that “the moat really isn’t that solid.” Brand and switching costs can keep companies alive—“just look at how people have started making fun of Cursor online over the past few months.” Asked whether heavy API users will be targeted, he went back to the first lesson in microeconomics: Bill Gates has absolute advantage at mowing lawns, but no comparative advantage.

3. Cursor’s Fate and Musk’s $60B Calculation

• Raymond said SpaceX had acquired Cursor for $60B. Cursor’s comparative advantage is engineering: Claude Code uses “progressive disclosure,” locating directories layer by layer at a high token cost, while Cursor uses grep extensively to jump directly to the target. Accuracy suffers somewhat, but efficiency improves sharply and gross margins are better. Opus and Sonnet remain cost items, which is why Cursor had to “wash” a Composer (K2.6) through Kimi to lower inference costs. Anthropic once reassured the founders that Claude Code was merely an internal tool and would never be released. Raymond’s response: “Who believes that?”
• The deeper issue is the data flywheel. Cursor had the highest call volume at the time, and Manus was similarly heavy. Their requests to closed-source models hit Amazon’s Bedrock or Anthropic’s own data centers, allowing model vendors to detect common patterns and decide how to train the next generation. Cursor has no pretraining or post-training capabilities and is therefore “at a disadvantage”; its current cooling-off period is already visible. 马克汤’s conclusion was unequivocal: “It will never, ever be acquired”（“打死都不会收购”）. “It doesn’t need to acquire it; you’ll die on your own anyway.”
• In Raymond’s account of the $60B transaction, he speculated that SpaceX stock might be used. 马克汤 explained Musk’s calculation: X still has a massive user base that can power a data flywheel for Grok, while coding and its high-value derivatives allow users to pay a higher premium. Once connected to X’s own data centers, the business would no longer need to pay Anthropic’s Opus premium, making the model work at the gross-margin level.

4. Consumer Product Philosophy: Layered Entry Points, Skill Ecosystems, and the Mysticism of PPT

• Efficiency tools are the biggest consumer category and the largest direction in the market. Raymond criticized Claude’s Chat, Cowork, and Code tabs as “essentially the same thing”—legacy left behind from different stages of the product’s history, not a design for the future. He now uses only Code. 马克汤 explained the layers as a cost architecture: the chatbot does not include skills, MCP, or memory, making it highly token-efficient. “Ask ‘hello’ in Claude Code and it costs $2.” Sam Altman has also asked users not to say thank you, since one extra message may trigger more rounds of conversation. Raymond’s rebuttal is that this pushes the cost of choice onto users: “Most people stop at the first layer.” Tencent’s CodeBuddy/WorkBuddy and Kimi Code/Kimi Work are both copying this path.
• How does a skill monetize? 马克汤’s ecosystem framework is blunt: a skill is just Markdown text. “I share the skill with people, and all my secrets are known to you,” leaving the creator with little more than a one-time fee. That means the platform has to monetize it—Manus and Lovable, for example, can use revenue sharing to activate the ecosystem. From GPTs to MCP to skills, the test is ultimately whether we understand ecosystems.
• The PPT category is “a little mystical.” AiPPT.com is backed by Zhipu and Visual China. Genspark started as a PPT tool, although Raymond had long mistaken it for a Manus copycat. Gamma reached a $2.1B valuation on this one product alone. But PPT cannot support hover interactions or dynamic charts, while HTML is much better suited to presentations in this era. 马克汤 even uses HTML when pitching investors. Raymond expects the slide format to be “completely overturned,” creating an opening for a new editor built around “PPT for H5.”
• Product stickiness is not yet obvious, particularly on the B2B side. 马克汤 sees two possible sources of retention: the emotional connection created by companion products, and memory that makes a product “understand you better the more you use it,” potentially extending to proactive agents that notify users based on past conversations.

5. The Companion Category’s Borderline-Content Fate and the Prevalence of Negative Gross Margins

• Character AI has largely disappeared from the conversation. MiniMax’s Talkie/星野 and ByteDance’s 猫箱 have shifted the playbook from avatar creation to gacha mechanics and otome products, taking cues from Paper Games’ Love and Deepspace. The central problem is monetization: the audience is young and has limited purchasing power. “Borderline sexual content is the easiest way to make money; without it, honestly, making money is pretty hard.” Raymond’s old internet memory—that Momo went public after 3 years—has not held up; today, the strongest performers are large games.
• The solution for avatar-creation products and UGC world-building platforms is to narrow the target audience. Casual users will not wait 20-30 seconds for each image; “they basically just leave.” The first step is to serve anime creators who are willing to wait and pay, then prove out the business model. Gross margins do not have to be negative; “it depends on your ecosystem position—you have to hit the people willing to spend money with you.”
• The gross-margin math is straightforward. Many to C/prosumer products are negative-margin businesses; 马克汤 suspects Manus could have high ARR and still ultimately have negative gross margins. Positive margins sit with niche vertical users: 财搭子’s high-net-worth traders, export-trade AI, and 信封 AI. The ceiling is structurally lower than in traditional software, where gross margins are often 80%-90%; for an AI application, 50% or 60% is already excellent because it must pay Claude Opus’s 75% model gross margin as its own cost.

6. Seedance at RMB1B a Month: Wrappers, Account Pools, and the Intermediary Supply Chain

• LibTV, a new product from the same product line as Lovart, was built on Seedance 2.0’s native 4K video generation. “Even Jimeng doesn’t do it as well as LibTV.” It has become a major revenue contributor for Volcano Engine: Seedance alone generates RMB1B a month, while the platform’s 2026 target has been raised from RMB10B to RMB15B. Companies are building scripts and storyboards around video models, but the core value being sold underneath is still Seedance.
• One source of LibTV’s low pricing may be account pools; the model may also involve gross-margin subsidies and account harvesting across different channels. Users resell unused Codex 20x keys, while unused enterprise token allocations can likewise be exchanged for pocket money. Cloud-provider trials and credits flow into account pools—Amazon’s example is $5,000, while Volcano Engine’s is RMB5,000. Intermediaries can therefore sell Opus more cheaply than Amazon and GPT more cheaply than Azure. “What you’re competing on is supply-chain management.”
• The structure is the same as Cursor’s. From Seedance’s perspective: “I see all your innovation; thanks for doing the work. We’ll come back and have a conversation another day.” 马克汤 confirmed that it is “very similar to Claude Code and Cursor.”

7. ByteDance Gets an 8: Seed Is Underestimated, and Feishu Gets Re-rated

• The mystery of why Doubao “feels good but benchmarks poorly” is that it is entirely free and does not let users switch models. A daily question is unlikely to receive the best model: “The only thing that can respond quickly is a small model.” Seed 2.1, released that day, was already benchmarking in the GPT-5.5 and Opus 4.7 tier. Its multimodal capabilities were particularly strong; the new audio model had “no detectable AI feel,” benefiting from Douyin data. Domestic coding discussions center on Kimi K2.6/K2.7 and GLM; Raymond’s Claude Code backup moved from GLM 4.7 to 5.1 and then 5.2, and he “wouldn’t consider connecting Doubao.” ByteDance rates itself “at least an 8” after 2 years, winning by focusing on Doubao, in contrast to Alibaba, which has “a few too many business units.”
• Feishu’s re-rating rests on infrastructure built before the AI era: multidimensional spreadsheets, IM bots, and API documentation that is “exceptionally clear.” It has recently been iterating rapidly on AI integration and ranks second in 马克汤’s skill usage, behind Super Power for coding. Use cases include having Codex review documents through a fixed workflow; having blogger 张杂娃 mark “cut this section” in the transcript’s comments so Claude Code can edit the spoken clip; agents calling product APIs on a schedule for regression testing; and digital avatars configured as Feishu bots, so colleagues can ask an agent “what has 马克汤 been busy with lately?” Raymond said on the spot that he planned to move back from Lark to Feishu.

8. The Endgame: Retrofitting Traditional Companies, and “If Applications Don’t Take Off, Hardware Is a Bubble”

• The opportunity beyond the app format includes Thrive Holdings acquiring traditional companies for AI roll-ups and a YC technology insurer obtaining a license, using AI to write policies and underwriting them itself. It “could grow 10x in a year.” 马克汤’s framework: the business model is already proven—licenses, customer relationships, and SOPs are in place—and only a scaling engine is missing. An AI-native retrofit “can absolutely build this company quickly.” When Raymond asked why 财搭子 does not launch a fund, the answer was simple: “It’s a licensing issue.” With a license and CSRC approval, it could; the company is already discussing the matter with licensed institutions.
• Raymond extended the point: foundation-model companies cannot directly enter hospitals, insurance, offline retail, or manufacturing, so they will form alliances or make acquisitions. These companies “may ultimately be the most core AI applications,” just as Beike—with offline stores—emerged as the winner in real-estate internet, rather than pure lead-generation apps such as Aiwujiwu or Fangdd. OpenAI and other model companies are also forming JVs with PE firms, potentially covering 500 portfolio companies across 5 PE firms, while deploying FDE engineers to carry out the retrofits.
• The RMB100 choice—application versus storage, optical components, or PCBs, excluding foundation models—goes to applications in 马克汤’s view. He first acknowledged that perspective is determined by one’s position: “The best quality in an entrepreneur is blind optimism.” Rationally, he returned to Huang’s five-layer cake: “Consumers don’t pay for the packaging technology inside HBM… They pay for products that improve the experience, optimize efficiency, and make life better. If AI ultimately fails to make applications work, all the hardware is a bubble.” For the loop to close, the economy needs social-scale GDP growth, potentially even the transition Musk describes from Universal Basic Income to Universal High Income.
• The closing warning concerns public markets: “AI applications” are largely a speculative trade. Meitu, BlueFocus, game companies, and Chinese Online are “not pure-play AI applications.” Legacy internet organizations lack the internal force to build AI: “To make an AI application work, the product function has to be sufficiently AI-native, and the organization has to be sufficiently AI-native.” That is also the subject of the next cross-show episode: why big tech companies cannot build AI.