He Doesn't Need to Acquire You — You'll Die on Your Own

• The frontier labs haven’t spared the wrapper apps out of mercy — they just haven’t gotten around to them. Enterprise AI runs on bodies, and AGI-chasing labs refuse to become staffing companies.
• Of a vertical app’s three moats — proprietary data, dirty work, inference efficiency — only efficiency survives the next model generation.
• Anthropic will never buy Cursor: it shipped the better product and reads Cursor’s API traffic to train the next model. SpaceX paid $60 billion for an on-ramp to high-value tasks — the one thing X never had.
• Cursor rations Opus because it must: at Anthropic’s roughly 75% API gross margin, matching Claude Code’s quota leaves Cursor no gross margin at all.
• A “hi” inside Claude Code costs two yuan. The labs split cheap chat from expensive agents on purpose — and dump the cost of choosing onto the user.
• Companion AI drifts toward suggestive content by economic law: its users are young and broke, so the menu is gacha or skin.
• Gray-market “account pools” and relay resellers — flipped subscription keys, surplus corporate quotas, free cloud credits — undercut official API prices everywhere, Seedance included, which books about ¥1 billion a month for ByteDance’s Volcano Engine.
• A 50-60% gross margin now counts as exceptional; traditional software did 80-90. And the endgame may not be an app at all but AI roll-ups of licensed traditional businesses — where product and organization must both be AI enough.

1. Why did everyone stop talking about AI applications?

Mark Tang — co-founder of Caidazi AI (an AI finance companion for Chinese retail investors), host of the podcast Facai Dazi (roughly “Get-Rich Buddy”), and, scare quotes his own, one of “China’s first AI product managers” — has built AI applications since 2023. Attention has drifted to foundation models and semiconductors — memory, optics, PCB, everything silicon-based life eats — while the startups that raised in 2024-25 as “the TikTok of the AI era” have gone months without news or rounds.

His timeline: early 2023 offered only GPT-3.5 — Chinese models unusable, GPT-3.5 itself barely usable in hindsight. GPT-4 made things doable; ferment through mid-2023, explosion in 2024-25. Large rounds still close in 2026, but the bet moved from DAU and traffic to profit, user mindshare, and the company’s niche against the foundation model — a brutal test: you must first understand where the model eats everything and where it leaves applications alive.

Raymond’s read: the north star flipped from DAU to ARR — in early 2023 people were still asking GPT how many R’s are in “strawberry” — and understanding both layers may now be two different people. Tang agrees; even with “agent” a household word, few grasp how agents personalize or break workflows. The money concurs: of last year’s VC dollars into AI, roughly 40% went to OpenAI and Anthropic, 30-40% to hardware like inference chips, the last 30% to applications — thousands of small, scattered deals.

2. How different is enterprise AI in China from the US?

Split consumer from enterprise, and enterprise into SaaS versus project work — in China, projects dominate. The money sits in a few fist verticals — legal, finance, healthcare, autos — exactly where Zhipu (known internationally as Z.ai) and the other domestic labs aim. Deployments are on-premise for data security: clients buy their own GPUs, sometimes build a small data center. Mostly model companies and clouds do the work — nobody squeezes inference out of a model like its maker; DeepSeek optimized for Huawei’s Ascend 950PR, so DeepSeek deployed by DeepSeek on Huawei silicon is the best version. The US skews SaaS — Harvey in legal. Raymond’s read: the model companies are running forward-deployed engineering. Exactly — and this year’s narrative has Claude eating it all, a fear deepened once OpenAI and Anthropic fielded their own FDEs.

3. Why haven’t the labs eaten the wrapper apps yet?

Raymond’s bear case: legal, medical, and finance are wired to money and life, and the labs are already feeding — OpenAI hired Jane Street quants; Anthropic pays veteran Wall Street analysts to annotate DCF models, assumption by assumption. Ecosystem they’ll never lack: distill best practice into a skill, drop it in a marketplace, one click on “finance” or “legal” makes anyone an expert. Taken to the extreme, the verticals should be dead. Why are they alive in America?

Timing, mostly. The labs hold nearly half the industry’s funding and could do it all — they haven’t eaten the verticals only because their hands aren’t free. Enterprise service is intensely human — engineers on the client’s premises, bespoke project management — and labs carrying the AGI dream refuse to become labor-intensive companies. Raymond’s read: against the AGI mainline’s exponential curve, doing Harvey’s job is a distraction.

The verticals also have walls. First, data. Nobody trains vertical models anymore — Harvey won’t train one for law, because injecting vertical data into a frontier general model beats injecting it into a purpose-built small one. Raymond’s analogy: hire the Tsinghua generalist over the third-tier specialist — fundamentally smarter wins at equal training time. Especially on ceiling, Tang agrees: in 2024 Caidazi worked with Zhipu’s AI institute on LoRA and SFT for a vertical finance model; by 2026 nobody discusses the approach — scale-driven emergence rolled over it. And data only counts if private — alternative data: meeting minutes, sell-side research, the paid Zhishi Xingqiu circles (private subscription research communities) Chinese investors read. If ChatGPT or Claude can search it, it’s worthless. Hence Raymond’s doubt about Harvey — private clouds per firm, Skadden sees only Skadden’s cases, plus all the world’s Common Law precedent as an exclusive, extensible corpus: persuasive in 2023-24, but law is pure text, paragraphs, not the complex structured data that might excuse a moat claim.

Second, dirty work. Caidazi spends heavily on data cleaning: same RAG, but are its ten retrieved chunks more accurate than a generic product’s web search? Its pipeline splits one text into objective information and subjective opinion for different reasoning paths — Nvidia’s share price at 3:10 this afternoon has no ambiguity; opinions need cross-validation. This still matters.

Third, inference efficiency: enormous for real-time tasks — the meeting is in ten minutes, the deck isn’t done — though overnight long-running agents don’t need a four-or-five-hour task cut to two or three. Cost-wise, efficiency is money: with statutes pre-categorized and conclusions pre-digested, Harvey skips real-time classifying and tagging. Same task, same result, less money and time.

4. Which of the three moats actually survives?

Raymond’s picture: the model as an ocean wave washing again and again over a sandcastle. The data moat already looks thin — labs can buy access: OpenEvidence bought rights to major American medical journals and re-indexed them for doctors; later Anthropic, or someone, bought the same. Not expensive, not exclusive — OpenEvidence was merely first.

Tang’s ranking: efficiency is the last one standing. And if Anthropic notices its biggest API consumer is Caidazi? Attack requires comparative, not absolute, advantage — Bill Gates might mow lawns better than you, but you still do the mowing. The funding gap is the crux — the giants’ cash exceeds every vertical combined, and money buys manpower — but that’s an organizational problem: OpenAI carving off hundreds of people each for finance, law, and medicine doesn’t necessarily work; small companies run faster. Honest version: the walls aren’t sturdy, so accumulate user mindshare while you can. Products squarely in Codex and Claude Code’s lane still grow users and ARR — people haven’t caught on, and brand plus switching costs are real. Which is why the internet started roasting Cursor these past months.

5. Why won’t Anthropic acquire Cursor?

A viral skit has a boss reviewing his AI tools: “Cursor, outstanding last year; this year outshone by Claude Code and Codex — canceling your subscription next month.” Plausible. Yet Tang’s own programmer still uses it — the VS Code surface feels like home, while Codex and Claude Code sell never looking at code: speak the requirement, inspect the deliverable. Path-dependent trades don’t switch cheaply; long term, the generalist products are too fierce.

Raymond put his position on the table: he is skeptical of the entire AI application market “to the greatest possible degree” — and recorded this episode anyway, because survival in a shakeout has reasons, and those reasons can produce differentiated returns. His template: China’s vertical e-commerce died to the last company, yet the thesis missed Pinduoduo (a niche-market attack, not vertical e-commerce), Dewu — formerly Poizon — and Weipaitang, a live-auction antiques app that grew from the scrap heap into a multi-billion-dollar company. Hold “vertical e-commerce doesn’t work” and you miss Pinduoduo.

Hence the deal of the season: SpaceX bought Cursor for $60 billion, apparently fully closed — after a16z offered a $50 billion round two or three months earlier and Musk said no, I’ll pay 60. Cursor sits in Anthropic’s main shipping lane, on Anthropic’s models — Raymond suspects nobody actually uses Composer, the in-house model. Why doesn’t Anthropic just buy it?

Cursor’s comparative advantage: Composer plus engineering. Claude Code uses progressive disclosure, walking the directory tree level by level, which burns tokens; Cursor leans on grep — think Ctrl+F — jumping straight to candidate spans: some accuracy loss, big efficiency gain, hence faster in many scenarios with slightly better margins. Why “launder” a Composer out of Kimi’s K2.6? Claude is expensive: Opus and Sonnet are Cursor’s cost of goods at Anthropic’s roughly 75% API gross margin, and paying that premium forever means never beating Claude Code. Cursor is a wrapper — sophisticated, superbly engineered, genuinely valuable — but it needs Composer to cut inference cost. Why does the same $40 or $100 plan buy far less Sonnet and Opus on Cursor than on Claude Code? Not won’t — can’t: at equal quota Cursor has no gross margin; Claude Code does. Cursor’s founder has said Anthropic promised early on that Claude Code was an internal tool that would never ship. And Cursor was the biggest API caller of its era — Manus too. Claude is closed: every request lands on Amazon Bedrock or Anthropic’s own data centers, where the vendor inspects traffic and decides how to train the next generation — the new data flywheel. Cursor has zero pre- or post-training capability, permanently downwind, visibly cooling. Would Anthropic acquire it?

“Not in a million years. He doesn’t need to acquire you — you’ll die on your own. He’s already built a better product than Cursor.”

The old story of vertical social apps courting Tencent: why buy you — good idea, let’s copy it. So why does Musk buy? First, the deal is option-shaped: no acquisition after SpaceX’s IPO means a $10 billion breakup fee; completing it costs 60 — Raymond guesses payment in SpaceX stock, at a three-trillion-dollar valuation. Second, the flywheel: put the model on X and use X’s still-enormous user base to spin up Grok’s data loop. Grok’s traffic comes mainly from X, whose tasks are low-value; Codex and Claude Code won by aiming at coding’s high-value tasks, where users pay premiums and the loop closes. Raymond’s read: Musk is buying an on-ramp — Cursor to high-value tasks, X to whatever people doomscroll daily. Third, margins: Cursor stops paying the 75% Opus premium, plugs into X’s own data centers — currently rented to three or four large buyers — and the model stands, same tier as Codex and Claude Code. Google bought Windsurf and shipped Antigravity in very little time; precedent exists.

6. Why is typing into a chat box a waste of the AI’s intelligence?

Real stickiness, beyond switching costs? Nothing obvious yet, especially enterprise-side. Two candidates: emotional attachment, for companions; and memory — the quote-unquote “knows you better the more you use it,” extending into proactive agents that ping you about things you mentioned weeks ago. You stop being the only one who starts the conversation. Enterprise is mostly productivity anyway, and demands more reliability: near-zero hallucination tolerance, risk control, compliance, audit, fixed SOPs that can’t be swallowed because “the AI is smarter.” Learning them takes stacked human labor — the biggest reason on-premise players keep their seats.

Consumer — anything an enterprise isn’t signing checks for, prosumer and Claude Code included — is led by productivity tools, and Chinese giants are cloning the Claude Code and ChatGPT playbook: Kimi shipped Kimi Code with Kimi Work; Tencent has CodeBuddy and WorkBuddy — inside Tencent, Buddy is a very senior product line, not a name anyone gets. The Code/Work splits map directly off Anthropic’s product path.

Raymond thinks that path is itself the problem: after a while on Claude Code he stopped touching the first two tabs — chat, Cowork, and Code are one thing, sediment of the product’s history, legacy rather than future. Tang’s explanation: the chatbot, like ChatGPT, is one-shot Q&A — no skills, no MCP, no memory, extremely token-cheap. Say “hi” inside Claude Code and that’s two yuan — cue the meme video where one “hi, how’s the weather” sets a data center roaring. Sam Altman once asked people to stop telling GPT hello and thank-you — wasted intelligence, millions of dollars a day; a thank-you tempts an “and you?”, another multi-turn session, context growing. So vendors split the entrances — a general-knowledge question needs no local files or memory, one cost, one settlement — and having paid, you’re steered toward the cheap tier. WorkBuddy versus CodeBuddy, Kimi Work versus Kimi Code: same layering. Raymond pushed back: that dumps the cost of choosing onto users, most of whom stop at layer one forever. In his parallel universe everything goes into Code — smarter user behavior is what upgrades a $20 plan to $200.

And Skills, which Claude Code made famous this year — monetization? Tang’s ecosystem test: who supplies, who consumes, how does the supplier charge, who is the platform? A Skill is a Markdown file; shared once, secrets out, a one-time fee at most. Only platforms can make it work — Manus and Lovable are building Skill ecosystems with creator revenue share. From GPTs through MCP to Skills, the test has always been ecosystem literacy.

7. Is AI slideware a $2.1 billion superstition?

Within productivity, Tang finds the PPT sub-track a bit of a superstition. The ur-ancestor is AiPPT.com — excellent templates, shareholders including Zhipu and Visual China Group (China’s Getty Images). Genspark was also born in PPT — Raymond, who met Manus first, had assumed it was a Manus copycat — running growth marketing through PPT touchpoints. Gamma turned AI slides into an art form, with a valuation Tang recalls at $2.1 billion on this one trick.

The superstition: a deck is made for human eyes and projectors — animations fire one click at a time, nothing responds to hover, no live charts. HTML is plainly this era’s reporting medium. Tang’s company runs internal reporting and research on HTML and pitches investors with it: externally sent materials go as images and PDF, the information being sensitive, but the live demo is HTML — simply better. Raymond concurs: before recording, Tang sent two pages better-looking than his own, and Raymond now puts all his research on Cloudflare Pages. The slides form factor may be swept away entirely; H5 pages (mobile-friendly interactive web pages) are what agents natively produce well. Somebody will build the PPT-for-H5, a new kind of editor — a real opening.

8. Why does companion AI always end up selling skin?

The time-wasting tracks, as Raymond calls them. Companions: Character AI, the earliest, is barely audible now; MiniMax runs Talkie abroad and Xingye at home; ByteDance has Maoxiang (“cat box”). 2024-25 bred the culture of nie zai — “molding your own companion”: persona, prose style, parentheses for action — the text says “that coffee in your hand looks delicious,” the parenthetical says “(reaches over to take it).” Maoxiang bolted on gacha, styled like an anime game; others went otome, following Papergames’ Love and Deepspace.

Raymond tried every one and watched them slide irresistibly toward soft porn — in the US, Character AI and its kin all ended in controversy, drawing younger, sometimes underage users while prose and imagery grew less appropriate. He had wondered whether the track held a Momo-shaped opening — the location-based social app went public barely three years after founding — but the winners turned out to be big games. Tang’s diagnosis is payment.

“The suggestive stuff is the easiest money. If you won’t go there — honestly, it’s quite hard to get paid at all.”

The audience skews young, paying power weak, so monetization collapses into game mechanics: gacha, quests. Some pair frontier image and video generation with world-building — Nieta, a very good company in MoSu Space (Shanghai’s flagship large-model incubator): you build a world, a community co-builds it — characters, items, plotlines, the way Dungeon & Fighter or World of Warcraft accreted content. Charging stays hard: beyond the clientele, image latency — an advanced creator tolerates twenty or thirty seconds per image; a first-timer waiting thirty seconds for the first aha moment just leaves. So narrow: serve anime creators, not fans — they wait and pay; close the loop with them first. Despite pricey video and image tokens, Nieta isn’t necessarily negative-margin — creators pay. It comes down to niche and model performance: hit the people willing to spend on you. Caidazi is the same — a large share of its users are high-conviction retail whales, and they pay.

9. What gray market sits under Seedance’s billion-yuan months?

Lovart — also in MoSu Space — just raised a large round at around $2 billion. Earliest product LibLib; the breakout was Lovart, an agent for creators and designers with very strong growth and ARR. This year’s runaway success is the newer LibTV, video generation — because ByteDance’s Seedance 2.0 is astonishing, generating native 4K. ByteDance’s own Jimeng (Dreamina overseas) wires Seedream for images and Seedance for video — and still isn’t as good as LibTV, a major revenue contributor to Volcano Engine (ByteDance’s cloud). Volcano books about ¥1 billion a month from Seedance alone, and its 2026 target, ¥10 billion at new year, has been raised to ¥15 billion. Many users never call Volcano directly — they reach Seedance through LibTV or similar sites layering scripts and storyboards on top. The core value sold underneath is still Seedance.

Raymond’s read: this smells like Cursor and Anthropic again. Why does LibTV beat Seedance’s own channel? Maybe not better — maybe cheaper. LibTV subsidizes margin and sources accounts across channels. Enter the relay stations, gray-market API resellers — you can’t officially use OpenAI’s or Anthropic’s models in China. (Raymond, for the record: Mossfire is truly not in the relay business; any praise of a foreign model is hearsay.) The relay trade has been on fire since last year, its prices beating Opus on Amazon or GPT on Azure, and the root cause is the account pools: someone buys a Codex 20x plan, can’t burn the quota, resells the key; employees flip surplus corporate token allowances for pocket money. The same dynamics hit Seedance. Cloud vendors are in a knife fight: Amazon dangles $5,000 in credits; Baidu, Tencent, and Huawei clouds run the identical playbook. Volcano hands you ¥5,000 in trial credit; flip it to a pool and the pool quotes below official price — you got it free. The coupons are pool inventory; what relays really compete on is supply-chain management.

From Seedance’s seat, the message to LibTV writes itself: all your innovation is duly noted, thank you for your contribution, let’s have a chat someday. Exactly Claude Code and Cursor — whose earliest reassurance from Anthropic was that Claude Code was merely an internal product.

10. Why can’t AI applications match traditional software margins?

If many products beyond Lovart are Seedance wrappers, are they negative-margin? Very possibly. Many consumer and prosumer AI companies are — Tang’s show has discussed Manus and suspected the same: ARR very high, gross margin possibly still negative. Clearly positive? Some prosumer businesses charging real premiums: Caidazi; a friend’s AI for the export trade; Xinfeng AI — fixed, niche vertical audiences willing to pay up, plus heavy dirty work making the analysis genuinely more efficient. But AI application margins can’t stand next to traditional software’s 80-90%, readable right off the filings.

“In AI applications, getting above 50 or 60 percent already makes you exceptional. We’re application companies, not model companies — Claude’s Opus carries about 75% gross margin of its own, and the application company pays the model company’s margin as its cost.”

11. Why does Doubao fail the benchmarks and win the vibe check?

On to the most terrifying company on Earth, ByteDance — right now it’s ByteDance doing the dancing, and unclear how long anyone else keeps step. Tang’s headline: the real strength is the ecosystem, but the foundation models have not fallen behind. Raymond half-agrees: Doubao (ByteDance’s consumer AI assistant) is extremely good hands-on — he gives it real research tasks, and few models do research to his standard — yet it benchmarks poorly. Tang’s correction: the very day of recording, Seed 2.1 dropped with scores basically at the first tier — the GPT-5.5 and Opus 4.7 tier. The real complaint is that Doubao agrees with whatever you say. The joke: day one, you ask whether to buy a stock — solid financials, capital inflows, buy. Day two: “it crashed” — “Ah, I was wrong yesterday, let me re-summarize as fast and concisely as possible.” It apologizes daily.

Which tells you Doubao-the-product isn’t necessarily running the best model. Seed 2.0 performs well in Tang’s internal use, but models are tiered: Doubao is completely free, every saving counts, and unlike ChatGPT it won’t let free users pick a model. A daily question almost certainly doesn’t get the best one — only a small model answers that fast and cheap. And for high-value complex tasks, people aren’t reaching for ByteDance: domestic web coding runs Claude Code wired to local models — all Kimi, K2.6 and now 2.7, and GLM from 4.7 on. Raymond’s fallback in Claude Code since late last December has been GLM — 4.7, then 5.1, now 5.2 — and it’s good; Doubao’s model wouldn’t cross his mind. Doubao has coding on Volcano Engine too, but the share of voice isn’t there — why, Tang doesn’t know. He thinks Seed is underrated, and ByteDance’s fearsome edge is multimodal: audio, video, image. The audio model released that same day makes voice cloning table stakes — it imitates the sound of any scene, so your voiceover comes out scored with background music and zero AI tell. That’s the Douyin data corpus. The base models crossed the threshold — and once intelligence crosses it, applications get easy: Trae for coding, Doubao, even Maoxiang hold decent comparative advantages. Breakouts: Doubao first, Jimeng probably second, maybe Maoxiang. (CapCut predates GPT — AI-empowered, doesn’t count.)

Tang grades ByteDance’s past two years at least eight out of ten: execution, speed, no ornamental side quests — the essence is focus, folding the ecosystem into Doubao, image and video generation included. A dig at Alibaba? Less focused — a few too many business units. And ByteDance holds an overlooked pre-GPT pair, CapCut and Feishu — both staring at a massive AI-era revaluation.

12. Is Feishu the most underrated dark horse of the AI era?

Feishu (ByteDance’s workplace suite; Lark overseas) has complete interfaces underneath — infrastructure finished before the AI era needed it. Base, its multidimensional-table product, absorbs heavy customization; the IM bots shipped very early; every interface has documentation written with actual clarity, presumably by humans, back then. Recent months: furious AI iteration. In Tang’s Codex and Claude Code, the Feishu skill ranks second by usage — behind only Super Power, his pure coding skill.

His uses: first, documents. Before a spec goes to engineering, he pastes the link to Codex for a fixed audit — omissions, misaligned context, risks, weak selling points. It reads comments too: Zhang Zawa, a creator Tang admires, records meetings in Feishu, exports the transcript, marks comments — “cut this part” — and hands the link to Claude Code, which cuts exactly that segment into a voiceover video. Raymond’s read: a born MCP. Second, writing: research with Claude Code or Codex used to be trapped in chat logs; now the agent writes straight into Feishu docs, even spreadsheets. Third, monitoring: agents hit Caidazi’s product Q&A endpoints several times a day checking for breakage — regression testing, fully automated. And you can mount Code inside Feishu as a bot, a digital twin. Tang keeps two: one answers only to him, read-write on his local machine; one read-only, for colleagues — before asking him anything, they ask his agent what Mark Tang has been busy with lately. (Terrifying, Raymond noted — a question bosses ask. Tang whitelists what it sees; the boundary stays in his hands.)

Raymond’s takeaway was immediate: he’s migrating his Lark setup to Feishu — historical reasons put him on Lark, and many AI features simply don’t exist there. Note to Feishu: this episode is open for sponsorship; quite enough advertising has been done. Everyone, just tell Claude Code — install the Feishu skill for me.

13. What if the most valuable AI application is a retrofitted old company?

Jumping out of the application track: Thrive Capital set up Thrive Holdings to buy traditional businesses and run AI roll-ups. Thrasio did a version with top Amazon sellers — your product sells but you can’t do TikTok, so I acquire you: you did zero-to-one and one-to-ten, I’ll do ten-to-a-hundred. And a YC company Raymond finds delightful does tech-enabled insurance: an actual license, AI-written policies, in-house underwriting, its own paper — growing absurdly fast, the ten-x-a-year kind. His lesson: the AI application may not ship in the shape of a mobile app. It may ship in the shape of a company.

Many traditional businesses deserve an AI retrofit, Tang agrees, along two axes: product and organization. Organization deserves its own episode — AI-native organizations are painfully scarce; the iteration speed, communication cost, and transparency of shared context are nothing like a traditional company’s, and there’s no hierarchy. On product: many traditional winners won on business model — an insurer eats off the license, client relationships, SOPs, proven economics missing only a growth engine. An AI-native retrofit can stand that company up fast.

Raymond flipped it concrete: why doesn’t Caidazi become a fund — an asset manager? The license — the whole answer. Given money to buy one and CSRC (China’s securities regulator) sign-off, Tang would do it; Caidazi is already in talks with licensed institutions. Every business leaning hard on its model gets re-empowered: organizational efficiency, plus product power reaching much further down the long tail. It used to take hard-coded rules and back-office SOPs; now acquisition may need no humans — AI sends the emails and makes the calls at scale, AI service and pre-screening handle intake — and leverage per head explodes, margins compounding fast.

This rhymes, in Raymond’s view, with PE buyouts — acquire Starbucks China because you hold better know-how for Chinese F&B. Same essence — and the model companies are pairing with PE directly: OpenAI is forming JVs with private-equity firms, five PEs holding maybe 500 portfolio companies, all standardized onto OpenAI’s models with forward-deployed engineers sent in. Raymond’s read: there are places on Earth the model cannot reach. Cursor it can reach — software, native to the digital world. Hospitals, insurers, offline retail, manufacturing — it can’t, yet the intelligence would help them enormously. So: alliance, or acquisition. Flip it around and this class of company may be the truest AI application of all. China’s biggest real-estate platform today is Beike — KE Holdings, previously Lianjia, the brick-and-mortar brokerage chain — while the store-less listing apps like 5i5j and Fangdd that promised pure traffic lost. The survivor had the stores.

14. If the applications never take off, is all the hardware a bubble?

Last question: $100 today — a basket of AI applications, name it, Cursor, whatever, or a basket of memory, optics, and PCB? Applications are consumed by humans; optics and memory are consumed by the AI. (Ruling: applications exclude the model companies — Claude Code can count, Opus cannot.)

Tang takes the applications. First, he builds them and is obligated to believe — a founder’s finest quality is blind optimism. Second, the rational case: Jensen Huang’s five-layer cake, where applications are the layer he keeps stressing — the layer that educates the market, where the consumer finally pays.

“Consumers will not pay for the packaging technology inside HBM. They will not pay for the light source inside an optical module, or for ABF glass substrates. They pay for products that improve their experience, their efficiency, their lives. If AI ends with the applications never taking off — then all of the hardware is a bubble.”

For the loop to close, GDP must rise at the level of the whole society — even, as Musk puts it, robots carrying us from Universal Basic Income to Universal High Income. Either nothing takes off, applications and hardware alike — or everything does, and applications are the layer facing the customer, digesting those earnings directly. Raymond’s coda: if an enormous AI bubble really pops, beautiful things will be lying on the ground — some AI application companies might be exactly what’s worth watching, or buying.

Tang’s parting point: AI applications are nearly absent from public markets, and what trades there is mostly speculation — not only Meitu but BlueFocus, some game companies, COL Digital. None are pure-blood, and an old internet organization doing AI applications is a different species — it lacks the propulsion of an AI-native organization. For an AI application to work, the product has to be AI enough and the organization has to be AI enough — which may be why some giants have failed, and is the topic of Tang and Raymond’s next crossover: why big companies can’t do AI.

One note Raymond saved for the end that belonged at the top: Tang trained as an electrical engineer, and his recent Facai Dazi series on optics, memory, and PCB is deep material told plainly — Raymond learned a lot listening, and made a little money because of it. Links in the show notes.