许高's First Interview: Plaud, the World's Fastest-Growing AI Hardware

2025/09/28 · Source link · AI Summary archive

Summary

Plaud has evolved from an “AI recorder” into “the next generation of intelligence infrastructure and interface.” 许高’s roadmap runs Consumer → Teams → Enterprises → developer platform → vertical solutions, replacing the keyboard with “sensors + foundation models” as the medium through which humans digitize intelligence. The company could move from serving millions to “billions of people”; one colleague’s comment that “we can become a generation company” left him “deeply shaken.”
The growth data overwhelms the skeptics: “Since the launch of Apple Intelligence, our business has grown another 10x.” Revenue had already grown 10x in both 2023 and 2024, and 许高 believes the category still has a chance to grow 100x within 5 years. Call recording is only “a very small portion” of its use cases; Plaud currently serves just 1 million users. The host also said its revenue exceeded that of China’s six leading AI-model startups combined—even before Plaud had entered China.
The strategic target is 100 million of the 1 billion global knowledge workers who sit at the intersection of 3 traits: high knowledge density, high dependence on conversation, and high decision leverage. Plaud is giving up on the lower end of the market in the near term. His examples include 5 million lawyers, 5 million doctors, 20 million salespeople, 10 million consultants, and 3 million investment bankers. The commercial flywheel is blunt: best design, most expensive materials, highest prices, highest gross margins, best talent, and most expensive models. Shenzhen imitators may have shipped only 5%-10% of Plaud’s volume in aggregate, with no real impact on its share or business performance.
His answer on defensibility is unusually candid: “There isn’t much of a moat right now”; taste is the moat. The real moats are still being built: a memory center for individuals and enterprises, a connectivity “intelligent highway” linking 10,000 SaaS products, and an intelligence flywheel scaling from 3,000 templates to 30,000. The contrarian premise is that “conversation is a form of intelligence”: 99.99% of people still do not capture real-world conversations, and multiple advantages that each look like only 0.01% can compound into Plaud’s rapid growth.
The reusable operating doctrine was set in the summer of 2021: global from day one, build only products that can become No. 1 in the world, and fuse software, hardware, and AI deeply. The underlying assumption is that “a company is born to die”: a No. 2 player may have only 80% of No. 1’s performance, 50% of its pricing power, and 10% of its profitability, eventually getting eliminated by No. 4 or No. 5 while chasing the leader. After ChatGPT launched, Sam Altman’s comment that “It’s best for summaries” convinced 许高 that “this thing was made for us.”
All 4 ventures were funded with his own money and taken all-in, connected by the same search for the optimal solution to efficiency and the value of context. The first 3—graduate-school matching, a startup recommendation engine, and food-delivery lockers—failed: they converted instant logistics into scheduled logistics and cut per-order delivery costs from RMB8 to RMB2. He says the core assumptions behind each product still hold; what was missing was monetization instinct, capital, or other critical inputs. Asked about a giant offering an enormous term sheet, 许高 said he would consider the capital markets at the right time, but the company has no near-term financing need.
The key moment in March was real but less dramatic than the retelling: after China CEO 莫子皓 experimented with Gemini 2.5, Plaud decided to go all-in on foundation-model-related R&D, though that was an acceleration of an AI investment already underway rather than a decision triggered solely by Gemini 2.5. His experience at a model company showed 莫子皓 the ceiling on one-shot intelligence, while Gemini 2.5 Pro crossed an “emergence moment” in long-context processing. The model has memory and attention span that humans do not: “It is a 100-point omniscient, omnipotent god, but most people are using it to do 30-point work.” The people who create value through conversation rather than code have “never had anyone build an IDE for them”—a major contrarian opportunity.
Facing giants such as DingTalk, 许高 says he “felt pressure for a while, but has let it go,” guided by a belief in building “an organization that enables people.” “If we could become 10x bigger but make employees unhappy, I would rather we not become 10x bigger.” He entered China wanting to “relieve people’s suffering,” but agreed with the host that he did not need to liberate everyone; Plaud should still target the 3-high segment and “respect the market.” He is betting that Plaud can be a neutral routing platform: “Feishu will never route to others, and DingTalk will never route to Feishu—but we can.”

Deep dive

1. Opening Frame: The World’s Fastest-Growing AI Hardware Company, and the Beauty of Designing a Company

卫诗婕 sets the frame: by combining foundation models with software and hardware faster than anyone else and delivering a highly polished AI experience, Plaud has become “the world’s fastest-growing AI hardware company,” with revenue “greater than the six leading Chinese AI-model startups combined”—all before entering China. This is 许高’s first detailed public account of the 4 startups he has built.
She flags the episode’s central question in advance: defensibility. Recording cards that began appearing across China in 2024 seem to show that hardware barriers are easy to cross. Her summary theme is “the beauty of designing a company.” 许高 repeatedly returns to beauty, arguing that “scale does not always represent the best return” in business. Creating greater operating leverage, making the journey enjoyable for everyone involved, and continually breaking through the ceiling on the value of an existing market are, in her framing, the beauty of entrepreneurship—and may themselves become a moat.

2. Spending RMB1.5M to Take the Entire Team to Saipan: Entrepreneurship as a “Large-Scale Social Experiment”

许高’s first-principles logic is simple: if you want to hire the best people, “the answer is simple: offer the best compensation and benefits.” Yet many profitable, well-funded companies “do not really follow through on this.” So he chose the more extreme path: an overseas company trip every year, to Phuket or Saipan, while building a workforce spread across San Francisco, Seattle, Tokyo, Singapore, Beijing, and Shenzhen, with a European office coming soon.
He calls this part of the beauty of entrepreneurship: “You can think of it as a large-scale social experiment.” “I should explore interesting, imagination-breaking things more aggressively. So many companies in Shenzhen have high profits; logically, they could all support this. They choose not to, so I choose to do it.” After employees returned, they named offices after islands and built an “island pecking order”; another floor was named after foundation models.
He turns a skydiving trip in Saipan into a company metaphor: more than 100 people watched one another, then encouraged one another, discovering that “people can expand their own boundaries and do things they might initially fear or worry they cannot do.” He jumped too: “You start looking at the ground, the ocean, the sand, the greenery, and the mountains in a way you have never seen the world before… I think that is beauty.”

3. Multiplication Tables and a Red Ribbon at the Finish: The Pain and Endurance Behind Winning

His first childhood moment of brilliance came when his parents made him memorize the multiplication table. He volunteered to write it on the blackboard, and “after class, so many kids chased after me… When you have the ability to do something nobody else has done, people admire it.”
He uses the full story of a middle-school distance race as a metaphor for entrepreneurship: he started last among 50 runners, passed more than half after 1 lap, moved into the top 3 after 3 laps, then overtook a much taller runner in the final 50 meters to break the red ribbon. “When you enjoy the happiness of winning, there is another side behind it: pain, persistence, endurance, and oxygen deprivation. You have to accept it and keep carrying it.”
That remains his current state of mind. Even though the company’s revenue scale “should be one of the highest among AI technology companies globally” and its growth is in “the top 0.5% globally,” he still finds the work difficult because “our minds are filled with galaxies and oceans… we are still far from touching the red ribbon.” Every morning, he builds a red ribbon for that day.

4. First Startup: Graduate-School Matching—The Product Was Right, but He Did Not Know How to Do Business

He started his first company before graduating from college and lost his graduate-school tuition. The product scraped every successful overseas-study case it could find, structured applicant profiles—school, GPA, IELTS, TOEFL, GMAT, GRE—and matched them with the advisers best suited to help: “Let someone who wants to study abroad find the adviser who can help them most.” The product grew, senior executives at agencies approved of it, and some wanted to invest or partner. But he “realized that I did not know how to do business.”
His hindsight is brutally clear: “I should have built a telesales center and hired 200 people to call one by one… and it would have worked. But my mindset then was that telesales felt so uncultured—not elegant, not tech enough.” He still believes the basic assumptions, product thinking, and execution approach behind each startup were correct.

5. Second Startup: A Recommendation Engine—The Missing Ingredient Was Fundraising

His ambition as a young VC was “unrealistic”: “If I have to work 20 years before I can surpass Sequoia, I do not want to wait 20 years.” So he built a startup recommendation engine that scraped founding-team backgrounds and scored them—“Tencent gets 8, Baidu 6; a vice president gets a coefficient of 1.8, a director 1.5”—then added weighted factors such as hiring volume and website traffic to produce scores of 70 or 78.
His description of investing’s essence is worth preserving: “In a market with inherent information asymmetry, how do you gain priority access to the possibility of information symmetry—or uncover profit opportunities in information gaps?” But the product’s leverage came from being able to invest in the companies it identified. “I had only been working for 1 or 2 years. I had no fundraising ability, and I did not have RMB100M or $100M.” He says he is not very good at imposing limits on himself, but if the critical elements do not materialize, a company may fail or be forced to pivot.

6. The Value of Context: How People Work, and 4 Types of Loss

The host connects the first 2 startups to Plaud: he had recognized early that value could be extracted from the information behind public data. 许高 agrees: “I have particularly good intuition and talent for how to efficiently derive intelligence from context and derive value from context.”
His core framework is a chain of losses. People perceive with their organs, understand, analyze, and remember with their brains, and use keyboards to digitize their context—but the chain is highly lossy. “After a conversation, people can remember only 5%-10%” (memory loss); they get distracted in class or meetings (attention loss); working hours are limited (energy loss); and they “cannot understand the thing” (intelligence loss).
Plaud’s paradigm is different: the microphone acts like an ear, the camera like an eye, the speaker like a mouth, and the button like a hand. With sensors to perceive and foundation models to understand, “we can achieve 100% memory, zero attention loss, work 24/7, and run multiple threads in parallel… We are still only at the starting point of this paradigm.”

7. Third Startup: Food-Delivery Lockers—Turning Instant Logistics into Scheduled Logistics

The insight behind the project, planned in 2019, was that instant logistics cost RMB8 per order, while scheduled logistics providers such as ZTO and YTO charged RMB1.2. The gap came from the absence of route planning and order aggregation: point-to-point delivery consumed substantial resources while carrying products of limited value. The hardest orders were “a RMB12 cup of CoCo milk tea or a RMB15 lunch,” for which the merchant subsidized RMB3, Meituan subsidized RMB3, and the courier risked his life speeding to deliver within 30 minutes.
Inspired by Luckin’s dense office-building locations for customer acquisition and fulfillment, his solution was to build a local economic model around a 2,000-person office building and 50 ground-floor merchants. Customers would order 1 day ahead; 3-5 three-wheelers would carry 50 orders each; and consolidated deliveries would be loaded into smart lockers. “We could cut the delivery cost of every order from RMB8 to RMB2, with one-third of the efficiency gain going to the merchant, one-third to the customer, and one-third to the company.”
The project failed as the pandemic arrived. “My thought was to keep going—that was it, just keep going.” But installing lockers required approvals, while signing restaurants meant moving from staff to store managers, regional managers, area managers, and the head of Greater China. Integration could take 6 months after that, so supply variety never reached critical mass. A single meal subsidy was not enough; users needed “3 to 5 free meals” to form a habit. The company eventually burned cash and moved away from the original model.

8. A Side Critique of the Delivery War: Everyone Is Chasing Efficiency, but Delivering Goods with Too Little Value Density

He believes today’s food-delivery war “has not solved anything fundamental.” “Platforms are encouraging consumers to buy a very low-value-density product at a very cheap price. In theory, the price should be several times higher for the system to function and for all sides to benefit.” He sees Pinhaofan as “just a form of light group-buying”; the real opportunity remains turning instant delivery into scheduled delivery.
The result is resentment throughout the chain: merchants are furious, couriers face greater danger, and Meituan complains that “I am too difficult to deal with.” The entire model emerged from the era of heavy subsidies and carries a historical burden. Asked whether he could have succeeded had he started 10 years earlier, he admitted that “on that battlefield, the most brutal approach could work”—at a certain point in history, the most efficient way to win share was to spend money. “Between recognizing the optimal solution and getting a result through entrepreneurship, many inputs are required. Entrepreneurship means slowly evolving yourself.”

9. The Reality of 3 Straight Failures, Sea Fishing, and Delayed Gratification

He does not hide the reality of repeated failure: “You will definitely doubt yourself, lose heart, feel sorry for the brothers who started with you, and feel that you failed to deliver on what you promised and wasted their time.” Which failure hurt most? “Maybe every one of them was painful.” He was also comfortable taking a job after failing: “Entrepreneurship and employment are both vehicles. What is truly rewarding is whether you are doing the right thing.” The host relays a teacher’s description of him as “non-stop.”
Sea fishing is his mental model. He has spent dozens of days at sea, catching at most 1 fish per day on average; “99% of the time, you catch nothing.” Lure fishing requires constant casting and reeling in: “You keep working without necessarily seeing results, and then suddenly you win.” Borrowing a phrase from 张一鸣, he says: “The satisfaction in delayed gratification is enormous, almost fanatical, and that is what gives people the persistence and mental strength to keep going.”
The host invokes The Old Man and the Sea, where Santiago hypnotizes himself after failing by saying, “I am an unusual old man,” and asks whether 许高 tells himself, “I am an unusual young man.” The answer is immediate: “Yes, of course.”

10. Why He Went All-In with His Own Money Every Time, and the Giant’s Enormous Term Sheet

There are 2 reasons. The first is personality: “If I am going to do something, I have thought it through. It is worth doing, and I believe I can do it well, so I do it… If you believe, you believe; if you do not, do not move. If you are going to start a company, you have to put your own savings all-in.” The second is candid: “I was not particularly good at raising money. My starting point was not whether investors liked the idea or whether it could make money in the short term…”
The host mentions that a giant recently approached him with an enormous term sheet. 许高 says he will engage with and consider the capital markets at the right time, but “the company itself is financially very strong, and has no financing need in the near term. It is better to spend the time on the company and the business.”

11. The Starting Point in Summer 2021: Fatigue with Frenzy and a 3-Part Framework

The starting point was a values conflict. While working as an FA at Xiaofanzhuo, he was operating in the most frenzied phase of the capital market: “I am someone who seeks truth, rationality, depth of thought, and clarity on who is right. Things that were wrong were being described by everyone as right… I was not good enough at playing that game. I had to do what I wanted to do.”
Framework 1 was global from day one. “The prelude to deglobalization had already begun.” The value spillover from Chinese manufacturing was “overwhelming and a massive leverage point,” compounded by the capabilities of Chinese engineers and the commercial judgment trained by Chinese companies. “The leverage here was higher, and globalization naturally meant more possibilities.”
Framework 3 was deep integration of software and AI. The essence of a pure hardware company is to make the best invention in the world, but no company can do that consistently for 10 years. Software supplies data feedback and continuous iteration. AI was also a matter of commercial instinct: U.S. SaaS was “a beautiful business model, with high retention and high continuity,” while he had been building crawlers, NLP, and recommendation algorithms more than 10 years earlier—“the AI of the previous era. I have always believed in AI.”

12. “A Company Is Born to Die”: Why It Has to Be No. 1 in the World

The complete logic of framework 2 is: “The birth of any company is the beginning of its journey toward death.” So a founder must identify what is essential to sustainable survival. “If you are No. 2, you may have only 80% of No. 1’s performance, 50% of its pricing power, and 10% of its profitability.” The leader enters a positive loop of talent, product, and returns; No. 2 competes fiercely with No. 3 while chasing the leader, then is killed by No. 4 and No. 5.
Being No. 1 also creates a survival buffer. “If your pricing strategy is wrong or you are selling too expensively, you can lower the price because you make the best product in the world. You always have a way to adjust into a state where the company can survive.” The practical conclusion is severe: “If someone else is doing it and doing it well, and I cannot explain why my team can do it better, then do not do it.”

13. Narrowing Thousands of Ideas Down to Recording: The Contrarian Signal of 1 Billion Downloads

Applying the 3 frameworks, he narrowed more than 1,000 product ideas down to 2 or 3 and chose AI recording. “It was counterintuitive, because traditional recording devices were not a very large market”—the era of Sony and Olympus had already been eroded by smartphone apps. But the view changed when he saw that everyone had a recording app on their phone and that Google’s Live Transcribe and Notifications had 1B downloads. “It is rare to see an AI application exceed 1B downloads. That shows the demand is genuinely strong and effective.”
His reading of Google’s product is pointed. It offered real-time transcription but did not allow users to save audio or text, leaving them frustrated and complaining every day. “Google is trading data with users: I give you transcription capabilities, and you give me the data to train on. Why should I let you save it? If I save it, I have to charge you and consume server resources.” Google treated it as a killer feature for Pixel, but that itself was the signal.

14. The Challenge from iFlytek and Sogou: The “T” in TPMF Was Not There Yet

His diagnosis of the earlier products is that independent recording devices solved the phone’s unreliable experience—cumbersome operation, inadequate battery life, and interruptions from incoming calls. iFlytek and Sogou added ASR for real-time transcription, which was “actually very good.” But a 1-hour conversation generated 10,000 Chinese characters, with 5% transcription errors, and users still had to spend substantial time organizing and summarizing. Efficiency remained low; in his terminology, technology fitness was not enough. The T in Technology Product Market Fit was missing.
The host says the idea came from “王汉川,” then refers to him as “小川.” 许高 claims it on the spot: “This was my original idea… If he shared it publicly after 2021, I may have been the first to share it. It is fine.” The transcript does not further clarify the person’s name.

15. The First Product, i z rack: A $100 Price and Immediate Category Leadership

He started with his own money in 2021. “You cannot hire a group of AI engineers on day one,” so the team began with IoT. Small recording products on the market were so poorly designed that turning noise cancellation on or off required connecting to a computer and changing a 0 or 1 in a TXT file; deleting the wrong line would brick the device. With limited R&D investment, they built i z rack—easy recording—with recording controlled through a mobile app.
The result validated the premise: “The industry average was $30-$40. We could retail at $100 and immediately become No. 1 in the category.” Ad ROI was very high and word of mouth in the community was excellent. The sequence was to use IoT to deliver a step-change in experience, test market acceptance, and add AI afterward.

16. Charles Joins: A Year of Collaboration, the Beauty of an Ice Maker, and a Comprehensive Blueprint

Charles was then the founder and CEO of a company that worked closely with Sogou on AI recorders and translators—“closer than a supplier.” The 2 worked together as partners for 1 year, meeting monthly or every other month. Charles was patient through every discussion, from microphone fundamentals to acoustic algorithms to display modules. 许高 describes him as “one of China’s first generation of outstanding mobile software engineers, a pioneer of first-generation AI smart hardware 10 years ago, and also an operator and a businessman.”
The pitch was the 3-part framework, plus a string of small ideas that could all work. The ice-maker story was the most vivid: it took 10 minutes from startup to ice, while wanting ice in a Coke is an impulse need. “At least build an app for remote or scheduled ice-making. It is such a simple thing, but no ice maker has that capability… People with serious technical ability may look down on it, but I think this is beauty: using technology to deliver an extremely polished experience for small things.” The company was established in 2022. When a hardware expert first saw Plaud Note, his reaction was: “This must have been made by a very mature hardware team.”

17. The ChatGPT Moment: “This Thing Was Made for Us”

Once foundation models appeared, it became clear that Plaud was “definitely not just a recorder company.” The evidence arrived almost for free. Less than 2 days after ChatGPT launched, a reporter asked Sam Altman about the most useful API use case. He said, “It’s best for summaries.” “We already had the recordings and the transcriptions, so wasn’t summarizing all that was left?” Across all product formats, Plaud was “the most direct and efficient beneficiary of the foundation-model breakthrough.”
At the time, Plaud Note had just entered industrial design and the company was preparing to hire NLP engineers. After the pivot, “the earliest version was simply to add a foundation model, and it immediately took off because the experience jumped, and we were the first in the world to offer that capability.” Plaud put Playground on its website and published Whisper there; hundreds of people queued online in real time to try it. That was a powerful growth signal.
Growth was 10x in 2023 and another 10x in 2024. “It deeply reshaped the founding team’s mindset—from survival mode to growth mode.” Before profitability, the team thought only about “making sure the company did not die.” After 2 years of 10x growth, it began to understand concretely why it could claim to be building a world-class AI technology company and what roadmap could achieve that.

18. The June 28, 2023 Launch Speech: “Today, We Are the No. 1 Brand”

Plaud Note launched with a team of only a dozen-plus people. 许高 told them: “Although we have only a dozen-plus people, at the moment we launch this product today, we are the No. 1 brand in the world of AI audio hardware.” He says the word may have been “giant” at the time. “How does something become possible? Only people who dare to imagine it have a chance. If you do not even dare to imagine it or say it, you have no chance.”
He does not romanticize his colleagues’ reaction: “Their first instinct may have been, ‘Isn’t he exaggerating?’ But the fact that you can say it means you are different. At least you dared to say it, so it is possible.” Every launch had ceremony: he and Charles each shared a few words, ate cake, and ordered in catering. “We have a very particular style.”

19. The Key Call at the End of 2023: Plaud Had to Build a Wearable AI Device

The prior reasoning proved accurate: “It was very likely that phone makers would soon make meeting summaries a standard phone capability.” That happened with Apple Intelligence in 2024. Plaud therefore had to define its differentiation: its target users work in environments where they are particularly dependent on something other than a phone.
2 use cases made the point. A veterinarian wearing gloves to diagnose a dog cannot conveniently touch a phone; “the phone simply does not fit the situation.” Japanese construction users were unexpectedly numerous: a project manager in a hard hat may be climbing stairs without guardrails, needing both hands to find something to hold or risk falling. NotePin launched in August 2024. His meta-observation: “We only have 1 or 2 truly important decisions to make each year. The rest of the time is high-efficiency execution.”

20. Product Philosophy: Amplify Your Intelligence and Put People at the Center

The mission is “amplify your intelligence,” with an explicit trade-off. Some companies say they are building agents to do the work and replace people. “Our philosophy is to build product capabilities around enhancing human intelligence and making people extraordinarily efficient. That determines why many of our product-design choices are different.”
He compares the product with a CEO assistant. In a meeting, you can tell an assistant, “I find this interesting. Look into it and give me a report next Wednesday,” or ask someone next to you, “Please make a note of this.” The hardware button lets users transmit that intent to AI in real time.
The philosophy ultimately comes down to a question of purpose: after becoming more efficient, should users have more time with their spouse and children, or after doubling their income, take their children to Antarctica to see penguins and Africa to see elephants? “Otherwise, you finish the day’s work in 4 hours and your boss says to double the workload. What was the point?” The goal is abundant freedom, a better life, and a happier experience.

21. “Forgetting Is a Beauty”: A Product Design Philosophy Against Information Overload

His moment of realization directly contradicted Plaud’s own sales pitch: “We say the product helps people remember everything—but you have to ask why people forget. Forgetting is a beauty. If someone still remembers something 2 years later, it was the most valuable content in that conversation. If they slap their thigh and say, ‘What was that again?’ then it was probably not that important. The brain can choose to forget.”
The design principle follows: “Should we help users remember and display everything? That is wrong; it can create information overload. We need to solve the problems of incomplete, weak, and poor memory while avoiding the burden of overload. That balance matters. The human brain is a particularly beautiful thing; we need to respect and appreciate its underlying assumptions.”

22. Press to Highlight: Real-Time Human-AI Alignment Through Sensors

The best feature in the new Plaud NotePin Pro is Press to Highlight. When a user presses the card’s button while discussing an important point, the algorithm weights either the 30 seconds before the press or the 15 seconds on either side. When the foundation model processes the full context, it gives additional weight to the 3-5 points highlighted by the user.
His formal name for the concept is “sensor-enabled, real-time human-AI alignment.” The design still starts with the question: how would you interact with a real CEO assistant? “Please make a note of this.” The host’s experience confirms the shift: “I am learning to treat this intelligence the way I would treat another human. That is a major step.”

23. The Vision Expands: From AI as a Service to “the Infrastructure and Interface of Next-Generation Intelligence”

The older positioning came from a contrast: “Everybody is innovating AI with digitized data; only Plaud is processing data from real life.” It emphasized software-first AI, with hardware serving as the infrastructure for data collection. The new vision is driven by 3 developments arriving together: shipment growth validated the company’s ability to deliver; foundation models advanced from GPT-3.5 to GPT-4 to GPT-5, revealing “some dawn of AGI”; and the industry developed agent paradigms such as tool use, browser use, and phone use. “We are not limited to summarizing anymore. AI can now do the work for people.”
The user base expands in parallel. “People work as teams, inside companies and organizations.” The roadmap moves from Plaud for consumers to teams, enterprises, and then a developer platform. Vertical SaaS companies in real estate, finance, healthcare, and legal services “all need context. They can use Plaud’s sensors as leverage—a context operating system—to build their own agent solutions.” The conclusion is expansive: “We may have the opportunity to serve not just millions or tens of millions, but hundreds of millions or billions of people… Every technology paradigm creates world-class companies within that paradigm. We have the chance to become one.”
A colleague’s comment left him “shaken,” and the host returned to it: “We can become a generation company.” The execution plan has 3 keywords: saturated investment—“pay the highest salaries in the industry to hire the smartest, most native, most expensive talent”—occupy the high ground, and execute to the extreme. The goal is to become a truly globally influential, world-class AI technology company by 2030.

24. The Developer Platform’s Stack and What Plaud Explicitly Will Not Do

The infrastructure stack is sensors collecting data → SDK → API in the cloud → ASR and CV modality conversion → intelligence extraction, such as summaries, action items, or a patient’s chief complaint and contraindications in healthcare → delivery into downstream workspaces, SaaS products, or CRMs. Plaud may also provide model customization for specific scenarios and industries.
The boundaries are equally clear. “If a developer thinks one of our capabilities is not good enough, they have the right to choose another specialized API.” Plaud will not build heavy-industry CRMs, hospital HIS systems, or Feishu-style multidimensional spreadsheets. “Users can distribute context from us into the workflows they need and further develop their agents.” The standard for choosing what to do remains the intersection of what the team wants to do, can do, and can do well.

25. Data, Privacy, and Model Strategy: Build ASR, Never Pretrain

His data stance is conservative to the point of being contrarian: “The first right of ownership over all data belongs to the user. Without authorization, it will not be used for training.” To improve model capability, “we have to find additional training datasets ourselves—whether through outsourcing or dedicated collection. That is the more practical solution.”
The build boundary is clear: “The transcription algorithm is self-developed.” Plaud is assembling a speech-algorithm team to build proprietary ASR for different languages and industries, targeting SOTA accuracy. But it “definitely will not build a pretrained foundation model from scratch.” Foundation-model intelligence is already improving; the question is how to use it efficiently and fully, not how to pursue technical SOTA. “If you build it and cannot reach world SOTA, it is not a particularly good business decision.”
AI’s processing of context has 2 leverage points. Memory is a strength: “An agent is like your chief of staff. It knows what you know, so it does not repeatedly research the same thing.” Reasoning depends on more than a dozen AI product managers and a more-than-dozen-person algorithms team focused on workflow orchestration. Reinforcement learning helps the agent “understand you better and better—whether you are fact-oriented, opinion-oriented, or boss-oriented.”

26. 3 Potential SOTA Domains: Sensors, Human-Computer Interaction, and Systems Engineering

许高 believes Plaud has achieved or could achieve world-leading performance in 3 areas: the sensors themselves, including design, interaction, feel, and reliability; the human-computer interaction paradigm, where “not everything has been released yet, and many frontier, breakthrough product offerings should gradually come to market”; and systems engineering learned from earlier companies. “We may not have the world’s best compute or foundation model, but if we identify the user workflow and optimize every link in the experience chain, coupling everything together, the experience is the best.”

27. Another 10x Since Apple Intelligence: The Leverage Difference Between 100M and 1B Users

To the argument that recording cards will eventually be replaced by phones, he responds first with the facts: “Since Apple Intelligence, our business may have grown another 10x.” Call recording is only “a very small portion” of the use cases. “When the product has the opportunity to serve billions of people, and we currently serve 1 million, there is no reason for our growth rate to slow.”
His market ledger starts with 1B knowledge workers globally, of whom roughly 100M have the 3-high profile. Entrepreneurs “spend 80% of their time listening to reports and thinking through decisions.” Professional support workers may account for roughly 50%, including doctors, lawyers, consultants, and salespeople selling high-ticket, high-decision products while working on the go. He describes the research-company use case this way: “Go to Walmart and ask someone what they think of my packaging. They are not going to hold up a phone; it is not natural.”
He argues structurally that Apple will not enter the category in the same way. “For a company like Apple, every innovation has to hit 1B people at once to create leverage.” Plaud is thinking about the 5 million lawyers, 5 million doctors, 20 million salespeople, 10 million consultants, and 3 million investment bankers within that 100M. Its device, transmission, and AI capabilities are vertical solutions—for example, building API integrations to send data to Salesforce, HubSpot, or Epic. “We can connect to 10,000 SaaS products. Would Apple do that work? It does not make sense.”
What if on-device models catch up? His hedge is conditional: “Based on current trends, it could happen, but it will take a long time. We are also improving and can do on-device processing. All of these are solvable.” His view of the device is combinatorial: local compute, sensors, battery life, connectivity, screens, and cloud compute can be packaged across glasses, earbuds, phones, or other forms. “Start from the sensors and ask which form is best.” Glasses are not part of the offering for now, but buttons and speakers are also sensors; the button is now an “intention activator.”

28. Dissecting the Contrarian Thesis: “Conversation Is a Form of Intelligence”

A partner joined after reading the financials, saying: “There must be a contrarian insight inside beautiful PMF.” 许高 breaks down the thesis. The popular reaction is that recorders are old-fashioned, phones already work, and “there is nothing special here.” Plaud’s assumption is that “conversation is a form of intelligence. Conversations carry the knowledge and experience we have accumulated over the years, our current preferences, pain points, expectations, and the content we exchange with one another.” Yet “99.99% of people have not captured conversations from the real world.”
He believes there may be several 0.01% insights: “Recording, transcription, and summarization were the previous generation’s product concept. You could put 6 or 7 people together with a solution provider and build it; it looks simple. But we now have nearly 200 engineers optimizing every step. In the end, everything multiplies together, and people cannot understand why the product is so good. That is why Plaud has grown so fast.” The industry’s learning curve is also revealing: “A guy in Shenzhen was making AI recorders, doing nothing special, and making a lot of money… Then people saw there was something there, and the big companies entered. But the level of consensus is still very low.”

29. Feature or Product? More iPad Than iPhone

Borrowing Tony Fadell’s maxim that “you have to examine whether these devices are features or products,” 许高 gives an unusually self-critical answer: “If you look only at the Plaud Note device itself, it is a feature. It is a hardware sensor for collecting sound. Some people say a phone can do that, which is true. But from my perspective, we deliver a product that helps people solve problems.” Asked whether it is an iPad or iPhone, he says: “I would be more inclined to say it is still an iPad.”
The product line will be complementary rather than self-replacing. “Just as the iPhone has AirPods and Apple Watch, we can have Plaud, Note, Pro, X, Six, and other products, and they can interact with one another.” The product should not be viewed as a single-dimensional device, but as a portfolio of sensors, connectivity, local and cloud compute, and AI.

30. Capture, Extract, Utilize—and a Possible Fourth Step: Route

The 3-step framework starts with capture: use every sensor to collect context, with phones and computers also treated as usable sensors, and ultimately build a context operating system. Extract first converts modalities—speech, images, actions, and keystrokes—into text, then uses foundation models to retrieve structured summaries, action items, and mind maps. Utilize covers deep research, email drafting, presentation creation, and pushing intelligence into other workflows. The newly launched Plaud Intelligence 3.0 expands capture beyond voice into images and actions.
Route, or distribution, is a new term under consideration, aimed directly at the Chinese ecosystem. “The 3-high users may interact not only with DingTalk, but also Feishu, Tencent Meeting, and WeChat. We can capture conversations across all their scenarios, then automatically decide what goes to DingTalk, Feishu, SAP, Oracle, or SalesYi. We will build an independent, neutral context operating system.”
His view is supported by a comparison of overseas and Chinese software ecosystems. “The capabilities of one major Chinese company’s workplace-efficiency software can equal all of the hundreds of U.S. SaaS companies with market caps of $10B combined. But those companies are worth several trillion dollars combined. One company here owns everything, yet creates only 1% of the value created by all those companies. There is deep commercial thinking behind that.” Notion, Confluence, Monday, Asana, Linear, Jira, Slack, and Zoom are independent overseas products. “In that world, the word route becomes very meaningful.”

31. What Is Intelligence? The Ice-Pop Vendor and Personalized Extraction

Asked for a fundamental definition of intelligence, he refuses abstraction: “The best intelligence is intelligence that helps each person do their work and accomplish what they need to accomplish better.” His example is a street vendor selling ice pops. A brilliant biology PhD might offer a sophisticated analysis, but “that is a super-intelligent model, and it cannot help me. I need someone who knows how to sell ice pops: carry a broader selection, put the milk pops in front because they sell faster, and price them cheaply. Good intelligence fits you and gradually leads you to improve; it does not tell you something you understand but cannot do.”
The product implements this through multidimensional summary templates. Users can add tabs: some want every summary to identify the CEO’s view so they can stay aligned with the boss; others select a “critical thinking” template so AI flags questionable views and missing reasoning in each discussion. The objective is to teach AI “to extract intelligence in a way that fits the user like a human would.”
Can agents do this? His discipline comes before ambition: “We are testing it. If you do something, you have to reach SOTA. If you cannot reach SOTA, do not build it yourself—use someone else’s. That is our commercial value system.” There are already “exciting, explosive early signs,” but nothing has entered production and he has no conclusion yet. The warning is against the chaotic instinct to “do everything ourselves.”

32. Roadmap and Acquisitions: 4-5 $10B Business Lines

The roadmap in one sentence is Consumer, teams, enterprises, developer platform, and vertical solutions, with enterprise features including elevated privacy management, permission controls, and private-cloud deployment. The acquisition of StarJam, a healthcare AI startup announced in April but completed in January, was “often driven by talent.” The goal is to build “4 or 5 independent business lines worth $10B each” across healthcare, legal consulting, sales, and other areas. The challenge is finding world-class people who combine leadership, business sense, and technology.
His view of Silicon Valley’s scribe ecosystem is harsh: “There are hundreds or thousands of them, and many are identical—record with a phone app, transcribe, summarize. They do not have much of their own, and do not truly own the scenario. I think they will soon go through a phase of rapid birth and rapid death.” Sensors are difficult to build, so Plaud can acquire or invest in excellent companies and help them become No. 1 in their categories. The reference point is newly public ServiceTitan, with a market cap “in the tens of billions of dollars.” In industries where conversation accounts for 70%-80% of the workflow and the digital infrastructure is fragile, “you have an opportunity to enter.”

33. Competition and Moats: “There Isn’t Much of a Moat Right Now”; Taste Is Everything

His response to imitators is quantitative: “All the Huaqiangbei players combined may account for less than 5%-10% of our shipments or sales.” They have not materially affected Plaud’s market share or business performance. The strategy is segmentation: “Break through and dominate the mid-to-high end with extreme strength and aggression. In the lower-end market, you can understand our strategy as giving up, at least in the short term.” The commercial loop is explicit: best design, most expensive materials, highest prices, highest gross and net margins, best people, most expensive models, and best results. “Only by being expensive can I continue investing.” The hiring reflects that positioning: industrial design talent comes from luxury jewelry, and marketing talent includes people currently at LV.
Taste becomes an identity proposition. “Wearables represent your identity. Entrepreneurs meet and say, ‘You use Plaud—great,’ and everyone resonates. If you suddenly pull out something else, the atmosphere is different. I have a brooch; at least it is Dior. I would not hang a $3 street-market item on myself.” His judgment of followers is harsher: “When you make a product that looks a lot like someone else’s, the underlying cultural values are mediocre. A company like that cannot recruit the best talent in the world.”
Asked directly about the moat that exists today, he answers in 2 parts, unusually candidly: “There is not much of a moat right now. Taste is the moat.” The 3 moats under construction are becoming the memory center for individuals and enterprises; connectivity—“linking 10,000 SaaS products and building an intelligent highway network, because those APIs cannot simply be built with anyone”; and the intelligence flywheel—“we already have 3,000 templates. Nobody can replicate 3,000 templates. Soon it will be 30,000: an intelligence network effect.” The company that most resembles Plaud’s future? “Apple.”

34. Giants Entering the Market and “An Organization That Enables People”: Better Not to Be 10x Bigger

Did the arrival of players backed by giants create pressure? “It did for a while. Now I have let it go. We only need to control our own commercial ambition.” The underlying value is clear: “If our company could become 10x bigger but make employees unhappy, I would rather it not become 10x bigger.” He contrasts a 500-person company with high profit per employee and high employee happiness against a 10,000-person company where everyone is overworked, the market attacks its products, and profits remain thin. Bigger scale does not necessarily mean better business returns.
His view of industry-wide overwork evolved from indignation to understanding: “At first I was indignant. Later I came to understand that it is not anyone’s deliberate choice; it is shaped by the social environment, and it is very difficult.” The way out starts with controlling one’s desires. Plaud began overseas because payment habits and the business environment were friendlier, enabling faster growth and stronger returns that could fund continued investment. “Commercial choices matter too.”

35. Entering China: “I Realized I Did Not Need to Liberate Everyone”

His honest explanation of the original motivation is that after the company began doing well, he wondered whether he had reached a higher level and should pursue meaning. “I could see a lot of suffering—people were busy, inefficient, and stuck in endless meetings. I genuinely wanted to help everyone.” The host says that over the past 6 months he realized “I do not need to liberate everyone; someone else will liberate them.” 许高 agrees. Plaud will still target the 3-high segment in China, while remaining disciplined about the outcome: measure ROI and operating leverage, respect the market and objective results, and maintain a calm mindset.
The host poses a sharper scenario: Feishu and DingTalk have both workplace software and physical sensors, allowing them to connect context end to end. 许高 counters with route: “To what extent will Chinese entrepreneurs and senior executives rely on a single workplace SaaS? Feishu will never route to others, and DingTalk will never route to Feishu, but we can.” On DingTalk’s low-priced competitor A1, he takes the user’s perspective: “If someone works 100% inside the DingTalk ecosystem, then using DingTalk’s product is the right choice.” Will B2B and B2C conflict? “No. B2B is still extremely early.”

36. The Gemini 2.5 Moment and 莫子皓: “A 100-Point God Doing 30-Point Work”

The host asks him to verify the story: after Gemini 2.5 launched at the end of March, China CEO 莫子皓 ran experiments with Gemini, was surprised by the model’s intelligence, and the company then decided to go all-in on foundation-model R&D. 许高 confirms it was true, but reduces the drama: “It did not necessarily have to be Gemini 2.5 to create that surprise. We had expected for a long time that the company would invest in these capabilities.”
莫子皓 describes the shift in his thinking in full: “At this time last year, I was still building a sales agent and thought it was great that it could make another RMB40M. That was AI helping people work. But after the end of last December, one-shot intelligence could no longer improve, because there was not enough data or no way to improve further through data. Everyone chased multimodality, instruction following, tool use, and multistep planning. But to me, the ceiling of one-shot intelligence had never been used well. You can think of it as a 100-point omniscient, omnipotent god, while most people are using it to do 30-point work. That makes no sense.”
His technical judgment is precise on versions: people began to believe Claude could code after Claude 3.5, while pure-text processing over extremely long contexts produced an “emergence moment” only after Gemini 2.5 Pro. The models have memory and attention span that humans do not. Another contrarian insight: “People who work in internet companies always think the whole world is writing code. But many people create and transmit value through conversation. Nobody has ever built an IDE for them, and nobody has built tools to make them more efficient.”

37. Organization: Alternating Intensity, AI-First Internal Discipline, and Talent Red Lines

He has changed his mind on operating tempo. “For a long time we were sprinting through the business, with high intensity and low comfort. Going forward, we will spend more time on foundational work such as strategic co-creation, roadmap co-creation, and SOPs. We will pull back the aggression in the business by one notch. We will not say every quarter that we need to double again. We can be more patient; that is becoming a stronger conviction for me this year. We need to enter a rhythm of tightening and loosening, not slackness.” The culture is “tough but rewarding”; it was never meant to be an easy place to work.
Internal AI practices are being pushed to the limit. Every internal meeting will require each participant to have a Plaud summary; every workflow must begin with Plaud. Any new hiring request must answer why AI cannot do the job. Anyone seeking a decision must first ask AI, attach the summary, and then ask the question. The goal is to turn the company into a laboratory: produce 8 hours of output without requiring 8 hours of work.
His 4 talent criteria are strong learning ability, intelligence, entrepreneurial spirit, and being AI-first. The red line is low openness: “If you trust only yourself and not your colleagues, you cannot work back-to-back with them. People like that are dangerous; they quickly become the organization’s central bottleneck.” His management reference is Google or Microsoft: pay the best people the highest salaries, then leave them alone—“nobody cares how many hours you work.” He understands Tencent-style light performance ranking, but his initial view is still to prioritize talent density. No time clock does not mean no struggle: “Then struggle.”

38. Tough and Love: The Tiger and Koala as Two Sides of One Person

In a DISC assessment, his 2 highest scores were the tiger—dominance, nearly off the charts—and the koala, which seeks harmony. He explains how the 2 fit together: “What is the deepest part of the tiger? Making sure everyone is happy. You can be unhappy in many local areas, but the whole must have a foundation of happiness. If the company is failing, you cannot be happy no matter how pleasant daily life is. To do the right thing, you have to set a higher bar.”
The source of his generosity is one sentence: “Make other people’s happiness your own happiness.” The practical investments include an in-house café with a professional barista, catering at supplier quality, and an office directly above a subway station. The host compares it with 张一鸣’s long-standing insistence on locating offices near Beijing’s West 3rd Ring Road: genuine people-centered management.

39. Charles’s Other Testimony: From Advising Against the Company to Being Convinced by 5,000 Comments

Charles’s initial reaction was discouraging: “Hardware cannot be built on enthusiasm alone. I have seen too many cases. 95% of the time, you are throwing money into the water when you invest in hardware. Do not do it—figure it out first.” When Nathan returned for a second meeting, he brought Amazon sales data for the recorder category and user experience comments. “Some were within my expectations; some exceeded them. I verified them, and they were true.” Nathan had personally gone through 5,000 recorder reviews.
What convinced Charles was the combination of “passion and enthusiasm, plus focus on data and pragmatism.” “If you want to persuade me, you need enough data. We believe data, not the hype people say out loud. He had limited funds and not many chips to throw around, so he had to find the people who could genuinely solve the problem for him. When we ate a simple lunch at a random restaurant and talked, I felt he had those qualities.” The key character judgment was that Nathan “did not seek absolute certainty, but needed 80%-90% confidence before acting.” He shared Charles’s pragmatism while bringing youth and enthusiasm—qualities Charles felt he himself lacked to some degree.

Daniel Ek, Spotify | David Senra 29.4% ARC-AGI-2 🤯 (TOP SCORE!) - Jeremy Berman