Summary

An earbud product 马啸 worked on at iFlytek sold just 40,000 units and fell short of expectations; today, Future Intelligence has millions of users, several hundred million yuan in revenue, and has just turned profitable. The company raised 3 rounds this year, with Ant Group leading the latest and Qiming Venture Partners oversubscribing its follow-on investment; valuation is approaching RMB1B. 马啸 sees the journey as the product of more than a decade in AI and audio, not a last-minute bet on a hot trend: “3 rounds in 1 year are backed by 3 years of no financing.”
马啸 is not betting on one device replacing the smartphone, but on earbuds, glasses, cards, and speakers collectively becoming the foundation model’s distributed “ears and eyes.” Earbuds’ advantage is not exclusive control of the entry point, but that users already accept them, can carry them everywhere, and can use them to turn unstructured audio from meetings and other settings into data streams. Smartphone usage will be gradually dispersed: “Earbuds are the consensus; what kind of earbuds to build is not.”
Future Intelligence is tackling AI hardware’s “impossible triangle” through business and office use cases, sacrificing some appearance, size, and sound quality to prioritize AI functions, battery life, and business-grade reliability. Standard earbuds typically deliver 5–6 hours of single-ear calling; the company is targeting 9–10 hours of continuous calls without charging and 30–40 hours with the charging case, enough to cover 6–8 hours of calls in a workday. The premise remains that sound quality, design, battery life, noise cancellation, and comfort—the “five plus X”—must come first; even the strongest AI cannot rescue a substandard consumer product.
Its most distinctive technical asset today is stable, real-time voice transmission over BLE GATT, a low-bandwidth link vulnerable to 2.4G interference. The host’s experience: the earbuds automatically prompt users to record and transcribe calls across Feishu, WeChat, Zoom, and other apps, keeping information from fragmenting across platforms. But 马啸 rejects the idea of turning patents into mythology: “Technology is never the ultimate moat.” The longer-term barriers are user mindshare, accumulated data, and sustained innovation.
Large companies will prioritize general-purpose intersections shared by billions of users, such as translation, while Future Intelligence is betting on meeting records and end-to-end office workflows they are unwilling to pursue deeply. 马啸 expects foundation models eventually to resemble “water pipes”: Grok, Claude, ChatGPT, and others may differ, but none has a 5x or 10x lead. More value will accrue to “what you do with the water.” The real competition is over which device stays with users longer, captures information more accurately, and feeds vertical data back into the service.
The company has deliberately rejected internet-bubble logic that uses inflated valuations to stockpile capital, because hardware expansion follows the “short-board principle,” not software’s “long-board principle.” 马啸 values the company on revenue and PS multiples, raising only enough to support a scale it is confident of delivering over the next 1–2 years. Even if peers are valued at 5–10x its level, he will not follow. One hardware mistake can hit R&D, tooling, inventory, channels, and distributor confidence at once: “You can successfully make 3 generations of products, but if 1 generation fails, you’re finished.”
Earbuds are only the starting point: Future Intelligence will continue launching companion-oriented office hardware from the end of this year into early next year, while staying focused on the 8-hour workday. 马啸 does not rule out 24-hour always-on becoming the eventual trend, but believes privacy, battery life, content understanding, and social acceptance are not ready. The company’s organizational bet is similar: find AI-native talent who can fuse scenarios, coding, and algorithms, and build human-machine collaboration products where “AI is not your opponent.”

Deep dive

1. The Company Has Cleared the Revenue and Profitability Hurdles Behind Its RMB100M-Plus Raise

Future Intelligence was formally established in 2022. Its core product is the iFlytek AI Earbuds, now used by millions; this year it launched the overseas brand VM and plans to expand into more office-focused software and hardware products.
The latest round was in the RMB100M-plus range, led by Ant Group, with existing shareholder Qiming Venture Partners oversubscribing its follow-on investment. It was the company’s 3rd fundraise this year, but 马啸 cautioned that the current heat followed several consecutive years of capital-market indifference.
Operationally, 马啸 disclosed that the company now generates “several hundred million yuan in revenue” and has just become profitable. The team has about 150 people across Hefei, Beijing, Shanghai, Hangzhou, and Shenzhen, and has operated remotely across cities since day one to broaden access to talent.
His background also explains why the company did not make a last-minute pivot into AI hardware. Before founding it, he ran a business unit at iFlytek; his final project there was an earbud product. “It wasn’t particularly successful, so I left to start a company.”

2. “Integrated Software and Hardware Office Assistant” Is the Result of Deliberate Convergence After Years of Trial and Error

马啸 believes “every piece of hardware can be re-engineered with AI,” but a startup cannot do everything. It must choose a focus where users can immediately feel real value. Future Intelligence therefore narrowed its scope to AI, integrated software and hardware, and office assistance—“every word was paid for in blood and tears.”
That convergence followed his experience of watching the previous AI boom move from noisy excitement back to silence. Even when a technology trend is directionally correct, that does not mean a product works today. The company therefore starts by asking whether users can immediately gain efficiency, rather than leading with a grand theory of the next platform.
“Office assistant,” rather than general-purpose assistant, is itself a product constraint. Every hardware, software, and model capability must serve the collection, understanding, and downstream processing of work information, avoiding the fate of an assistant that can do everything but does nothing particularly well.

3. The Next Entry Point Will Not Reproduce the Smartphone’s Single-Winner Structure

The host placed AI earbuds in the race to become the “next-generation entry point,” but 马啸 said the choice of earbuds was partly accidental. The team did not first decide to build an AI gateway; after years of making earbuds, it discovered the potential of natural-language interaction and wearable form factors.
His central view is that the future may not bring another device that directly defeats the smartphone. More likely, earbuds, glasses, and other software and hardware entry points will form a matrix sharing a cloud-based “brain,” gradually taking time away from the smartphone rather than causing users to throw it away overnight.
After 8 hours of sleep, people have 8 hours of work and 8 hours of life left. Today, the smartphone occupies most of those 16 hours. Once AI becomes smart enough that some tasks no longer require picking up a screen, earbuds could capture part of that time because they are “always portable and available for communication anywhere, anytime.”
马啸 also acknowledged the constraints. Without a screen, earbuds are weaker at data access and interaction efficiency. They are simply one unavoidable medium among others—“a way for everyone to interact with AI on equal terms.”

4. AI Earbuds Must First Clear the “Five Plus X” Consumer-Product Bar

马啸 agrees with the logic that “AI glasses must first be good glasses.” AI earbuds must first be good earbuds. One of the team’s early mistakes was overemphasizing AI while failing to build the underlying hardware properly.
The team summarizes the foundation as “five plus X”: sound quality, appearance, battery life, noise cancellation, and wearing comfort. Users compare these 5 factors first when buying earbuds. Even if AI drives the initial purchase, poor fundamentals kill long-term usage, recommendations, and word-of-mouth.
The “X” comes from the fact that AI earbuds listen and sense the outside world alongside the user. They must hear the wearer clearly while recording surrounding voices when needed, converting unstructured speech in the environment into data streams AI can process—“hearing what you hear.”
That, in turn, raises the difficulty of microphone pickup, power consumption, and wearability. The earbuds cannot merely play music; they must become AI’s ears in meetings, calls, and in-person conversations, while letting the user decide whether they are listening to the wearer or the entire room.

5. Battery Life, Weight, and Compute Form an Unavoidable Hardware Triangle

马啸 defines the structural contradiction in AI earbuds as an “impossible triangle”: battery life, the wearing comfort represented by weight, and processing performance are difficult to optimize simultaneously. Even with the best chips and components, the product requires trade-offs; no amount of willpower can produce a perfect all-rounder.
Future Intelligence’s choice is dictated by business users. Mainstream earbuds typically support 5–6 hours of continuous calls and 7–8 hours of music. The team wants 9–10 hours of calling and recording without charging, rising to about 30–40 hours with the charging case.
The target reflects real workloads. A business user may have 2 conference calls in the morning and 2 in the afternoon, for 6–8 hours of calls in a day. The earbuds must keep AI accompanying and recording continuously, rather than forcing users to swap left and right earbuds to recharge.
The trade-off is a product that does not chase the most “sexy” or lightweight form factor. It leans businesslike, rugged, and reliable. Sound quality also requires compromise: dynamic drivers, dual dynamic drivers, and hybrid dynamic-and-balanced-armature designs all change power consumption, leaving the team to keep balancing music quality against recording time.

6. Earbuds Are the Public Opportunity; Vertical Use Cases Are the Underwater Non-Consensus

马啸 calls AI earbuds “an open card on the first layer.” Smartphone makers and startups can all see the opportunity; the difference lies in what kind of earbuds they build. Smartphone companies must serve their existing mass user bases, so they will push general-purpose qualities such as sound, appearance, and battery life toward the limit.
Startups cannot directly replicate Apple or Huawei. They must seek a vertical wedge. Some are building AI sports products; others focus on lifestyle, companionship, or spoken-language practice. Future Intelligence chose the serious, data-heavy meeting and office market: “Earbuds are the consensus; what kind of earbuds to build is not.”
The host asked why the product also offers translation, but 马啸 rarely brings it up. His answer: translation solves cross-language barriers and is part of the largest common denominator that platform companies such as Apple, Huawei, and Samsung must cover. Meeting records and closed-loop office workflows serve only tens of millions of vertical users; for large companies, going deep would be “using a sledgehammer to crack a nut.”
The relationship between platform companies and vertical players is therefore closer to that between an operating system and an App. Platforms solve the common intersection shared by billions of users; Future Intelligence concentrates limited resources on deeper unmet workflows among business users.

7. The Earbuds’ First Office Value Is Unifying Data Across Platforms

Feishu can record Feishu meetings, while Tencent and other platforms have their own tools. WeChat and other settings may not offer complete capabilities at all. Users switch across multiple Apps every day, and information fragments with them. 马啸 believes earbuds, as peripherals, can bypass platform boundaries to collect and manage data centrally.
The host’s actual experience was that when a call began on Feishu, WeChat, Zoom, or another platform, the earbuds prompted him to start recording. Text was generated in sync during the call, and a usable record was ready when it ended. The “start anytime” experience comes from a wearable device that stays near the ear without requiring the user to hold up a phone.
Future Intelligence wants to provide recording, documentation, and closed-loop office workflows as a vertical capability, rather than merely building earbuds with a record button.

8. BLE GATT Real-Time Voice Transmission Is the Core Capability, Not a Permanent Patent Wall

The underlying challenge in real-time transcription is Bluetooth bandwidth. Once a classic Bluetooth channel is occupied by a call, transmitting other data at the same time becomes difficult. The team turned to protocols such as BLE GATT, originally used for small-byte control commands, and reworked them into a link capable of continuously transmitting voice streams.
Bluetooth operates in the 2.4G band, which is vulnerable to Wi‑Fi and other signal interference. Interference can cause slowdowns, packet loss, or data interruptions; compressing too aggressively for stability sacrifices voice detail and ultimately lowers recognition accuracy.
The team has iterated on this for more than 3 years. 马啸 condenses the objective into: “Maintain quality under low bandwidth and achieve stable data-stream transmission while preserving real-time performance.” This is what he describes as the company’s core capability.
Asked whether it could become a moat like Shokz’s patents, 马啸 gave a restrained answer. The company has filed patents, but competitors may work around them by changing the encoding format or implementation. “Technology is never the ultimate moat.” Technology’s role is to make the experience work; the long-term barriers remain user mindshare, recognition, data, and continuous innovation.

9. Top-Tier Supply Chain Support for Small Orders Came From a Verifiable Path to Scale

Luxshare is one of the best ODM manufacturers in Future Intelligence’s earbud supply chain and an important manufacturing partner in 马啸’s view. In the early days, the company’s individual orders were only tens of thousands of units; for factories of this type, volumes in the millions are needed to keep a production line running over time.
The team approached roughly 20–30 companies across the industry, working down from the leaders to the second and third tiers, and found almost nobody willing to take the business. In 2021, ChatGPT did not exist, “AI earbuds” had no market consensus, and a small company lacked orders, brand power, and certainty at the same time.
The turning point came through repeated visits to Luxshare’s chairman. 马啸 did not present a sudden dream; he repeatedly walked through the current product, existing data, future targets, and the path for getting from today to “A.” The other side was also looking beyond traditional earbud manufacturing toward small but potentially expanding sources of incremental growth.
This “winning people over with sincerity” approach was later used with upstream chip partners. 马啸 believes what truly persuades the supply chain is not an exaggerated end state, but whether the team can describe the future clearly enough for partners to judge the path difficult yet feasible.

10. Failure of the Previous Product Provided the Most Important Evidence of Demand

At iFlytek, the team launched an earbud priced at RMB799. Total sales reached only 40,000 units, causing a loss for its former employer. Yet monthly active users on the companion App remained around 80% of total registered users—roughly 80 out of every 100 buyers were still using it each month.
There was no foundation-model summarization at the time, but users still recorded meetings and converted speech to text. A 2-hour recording would take nearly 2 hours to listen to; in text form, it could be read in roughly 20–30 minutes. ASR was far weaker than today’s, but it already generated an efficiency gain.
The strongest signal was that users “complained and kept using it.” They criticized the earbuds’ appearance, build quality, and sound, yet still depended on the recording function. 马啸 concluded that demand did exist; the hardware was holding back distribution. Users would use the product themselves, but were “embarrassed to recommend it to others.”
This also explains why the product struggled to continue inside iFlytek. Office tablets, voice recorders, and translation devices were vertical blue oceans not yet transformed by AI; after an upgrade, they could become category leaders with pricing power. Earbuds were a red ocean from day one, contested by Apple, Huawei, and established audio brands, so the organization naturally preferred the more familiar blue-ocean products.

11. A Red Ocean Is Not Monolithic; The Breakthrough Is a New Use Case, Not a Copy of the Leaders

Even before generative AI, 马啸 believed in the earbud direction. Part of the inspiration came from scenes in Her and The Wandering Earth where people interacted through earbuds at any time. His view was not that the technology would mature immediately, but that “these things will happen sooner or later”; the entrepreneurial question was whether the company could survive until the timing was right.
Shokz offered the first piece of evidence. Even after the mainstream earbud battle appeared settled, sports could support a distinct vertical category. Users who already owned Apple earbuds might still buy a separate pair for exercise. New demand keeps emerging inside red oceans; the key is to identify it and occupy it in time.
Insta360 offered a second piece of evidence. GoPro was already strong, yet a later entrant could still improve the experience, showing that incumbents’ innovation does not cover every direction. 马啸 therefore believes a mature category does not mean the entrepreneurial window is closed.
The host worried that Apple might add AI functions to earbuds. 马啸 agreed that giants would certainly add general-purpose features such as translation, but argued that they would not go deep on records, summaries, and office applications in the near term. Competitors can study today’s public features without necessarily understanding or agreeing with the team’s judgments 6 or 12 months from now.

12. AI Hardware Is Competing to Become the Foundation Model’s Closest Senses

马啸 compares the foundation model to an upgraded brain. Transformer brought reasoning, logic, and problem-solving; pretraining, post-training, and Agents form a capability chain that expands outward layer by layer. Application companies have little chance of making the foundation brain stronger, but they can decide how it reaches the physical world.
Smartphones and laptops are already closed loops controlled by giants, but many situations are not convenient for opening them. The lapel microphone in the interview, the recording card attached to the back of a phone, glasses worn on the face, and speakers placed on a desk are all converting physical information into bits—the foundation model’s “ears and eyes.”
Earbuds, recording cards, glasses, and speakers therefore compete and cooperate. The metric is not which form looks most futuristic, but which can accompany users more naturally for longer and capture information more clearly. The longer the device stays with users, the more precise the vertical data becomes, creating a reinforcing loop between AI processing and service value.
Products such as AI Pin, in 马啸’s view, first ask users to understand and accept an entirely new form, then wait for it to deliver enough value. When adoption costs and functional value are both too weak, the loop cannot close. Earbuds, by contrast, begin as mature, useful consumer products and then add meeting-assistance capabilities.

13. Future Intelligence Wants to Own ThinkPad-Style “Reliability” in Users’ Minds

When the host said several AI earbuds looked almost identical, 马啸 did not dodge the criticism: “Maybe we haven’t worked hard enough and don’t have enough differentiation.” He believes brand character ultimately binds to the founding team, target users, and problems being solved, rather than being generated by advertising alone.
The company will not pursue a fashion positioning. Its ideal prototype is closer to an early ThinkPad: engineers and finance professionals trust it to remain stable and rugged in complex environments, and know it will work the moment they pick it up. Future Intelligence wants business users to reach the same conclusion—the product need not be beautiful, but it must be highly accurate and rarely fail.
That reliability extends to model output. Business users cannot accept summaries that invent facts, so the team uses its own model technology to minimize hallucinations and keep generated content grounded in original records. The brand promise is not that “AI knows everything,” but that it will first execute a limited set of tasks properly.
马啸 acknowledged that the product is still approaching this state. He expects products launched at the end of this year and next year to reflect the character more clearly. Reliability is not a packaging term; it is a shared constraint on hardware stability, recording accuracy, and functional boundaries.

14. 24-Hour Recording May Be the End State, but Privacy and Acceptance Still Stand in the Way

The host mentioned 24-hour recording, modified Apple Watches used for recording, and GoPro’s exploration of capturing the environment every few minutes. 马啸 called this direction “fairly aggressive.” He does not rule out eventual mainstream adoption, but considers it extremely early and far from mainstream today.
He illustrated the social cost with a story about glasses at a dinner. Once the wearer demonstrated that the device could record continuously, everyone at the table immediately stopped speaking: “The dinner was over.” Even if technology can listen and see, it still must clear the privacy, ethical, and social-acceptance hurdles imposed by everyone around the wearer.
Privacy is not the only unresolved issue. Battery life, clarity, and whether content understanding is actually useful also remain open questions. Future Intelligence therefore starts with the parts of the 8-hour workday where users explicitly need records, without trying to capture the 8 hours of personal life, much less the 8 hours of sleep.
The company’s route is “one small step forward”: first make users feel the value immediately after a meeting, then gradually approach Silicon Valley’s always-on future, rather than asking the market to accept the final form all at once.

15. The Real Bottleneck for General-Purpose Assistants Is Not Conversation, but the Lack of End-to-End Commercial Services

马啸 once worked on one of the first Siri-like products in the Chinese-speaking world at iFlytek. At the time, the system might fail to understand 8 out of 10 sentences. Today, he believes language understanding is “70–80%” solved, but service loops still have not been automatically connected by model capabilities.
He cites booking a flight by voice. The team went to enormous lengths to connect to an online service platform, only for the pilot interface to be shut down soon afterward because the platform did not want an external assistant deep inside its transaction flow. Food delivery, flights, and other industries each have their own giants; a general-purpose Agent must penetrate commercial boundaries one by one, and those interfaces will not be opened casually.
The host mentioned MCP and screen-control approaches, but 马啸 still believes they may be blocked and may not be stable or effective enough. A natural-language interface does not constrain the user’s path like an App button does; many requirements remain underwater. If even one open-ended question is answered poorly, the experience falls off a “cliff.”
Vertical convergence changes what is feasible. An in-car assistant may focus only on navigation, vehicle settings, content playback, and communications, with the relevant interfaces already connected in the vehicle system; voice interaction can then become genuinely useful. Future Intelligence follows the same approach, starting with records, constraining the office boundary, and extending gradually.

16. New Hardware Will Stay Within Three Constraints: Office Use, Companionship, and AI Reconstruction

马啸 confirmed that the company is developing hardware beyond earbuds in secret, with related products expected from the end of this year into early next year. He disclosed no specific form factor, but said the goal is not to expand the SKU count; it is to continue building an “integrated software and hardware office assistant.”
New products must satisfy 3 conditions simultaneously: they must relate to office work; they must address a need existing products do not serve well and create a visibly different experience through AI; and they must offer a degree of companionship, allowing users to solve work problems anywhere, anytime.
Earbuds are therefore only the first carrier. The company’s long-term direction is not to become a single-category audio brand, but to place more suitable sensing entry points in different office situations and feed the data they collect into the same software and AI service stack.

17. Capital Heat Is Being Treated as a Window, Not a Reason to Raise the Valuation

“3 rounds in 1 year are backed by 3 years of no financing.” The team began preparing in the second half of 2021 and became fully independent on January 1, 2022. At the time, investors often associated AI with losses, and AI hardware was even colder. Industry attention was not reignited by foundation models until 2023.
In 2023, many people advised a team with iFlytek roots to build a foundation model. 马啸 judged that Transformer had already been public for years, and that the leap in capability was being driven mainly by ballooning compute. Training required resources the company could not afford and did not fit the team’s DNA, so it stayed focused on the application layer.
He expects foundation models gradually to resemble “water pipes.” Data and algorithms differ, but Grok, Claude, ChatGPT, and Chinese models have not opened up a 5x or 10x gap. Users will ultimately care more about what they do after turning on a particular pipe—“what they do with the water.”
The company discusses valuation in terms of revenue and PS multiples. Even if companies in the same sector are quoted at 5–10x its valuation, it will raise only enough to support a scale it is confident of delivering over the next 1–2 years. 马啸’s principle is: “I can deliver what I have the ability to deliver.” Capital is shareholders’ investment in the future, not a cost-free war chest.

18. Hardware Follows the Short-Board Principle; More Money Cannot Skip the Industry’s Pace

The host repeatedly asked: if the market is willing to assign a higher valuation, why not take more money, accelerate the products, or secure the capital in advance? 马啸 explicitly rejects the strategy. “Fake it until you make it” may work for some software products, but not necessarily for integrated AI software and hardware.
Internet products follow the long-board principle: one compelling feature can attract users while other weaknesses are patched quickly in the 2nd and 3rd versions. Hardware follows the short-board principle: one fatal flaw affects R&D, tooling, trial production, sales, and inventory at the same time. “You can successfully make 3 generations of products, but if 1 generation fails, you’re finished.”
He has experienced the downward spiral after a product failed to sell. Weak sales make channels even less willing to stock it, leaving no resources for further promotion. Hardware may ultimately need to be cleared at roughly one-tenth of the original price, so no link in the team, supply chain, channel, or brand can leapfrog the industrial cycle through financing alone.
Distributor confidence is particularly difficult to restore. Distributors buy inventory with real cash, and a sudden price cut directly damages their stock and expectations. An internet team can kill a failed product and start over; a hardware company may damage its reputation across the industry. 马啸 says hardware is “more like farming”: sow in spring, harvest in autumn, replenish the soil, and accumulate crop by crop.

19. A Long-Run Culture of “Doing One’s Part” Shapes the Founder’s Choices and Hiring Bar

马啸 does not see himself as a genius who arrived at a sudden insight. He entered one circle and gradually became above average. Moving from China Mobile to iFlytek, he adapted to increasingly stronger peers and confirmed that he is “relatively good at slowly beating others over a long race,” building a compounding feedback loop through daily, repeated reasoning.
Before deciding to start the company, he spent 6 months trying to persuade iFlytek to continue making earbuds, repeatedly hitting walls and wondering whether he was the fish swimming upstream. 胡玉, iFlytek’s co-founder and later chairman of Future Intelligence, did not judge the project during a roughly 30-minute call. He told 马啸 that following his inner conviction was the only way to unlock his nature, commit fully, and find happiness. “Follow your heart” became an executable choice rather than a slogan for the first time.
The company’s culture follows the same line: it does not pursue狂 or showmanship, but creates tangible user value in a grounded way. AI should amplify human capability and make work and life easier—“AI is not your opponent”—rather than frame the product around replacing people or taking their jobs.
The people 马啸 most wants to hire are AI-native talents who can fuse scenarios, coding, and algorithms. The previous generation often separated algorithms, product, and deployment; younger people now integrate AI tools, scenarios, and coding into development from university onward. He even says that “the previous generation of AI experts may already have been淘汰,” while “100x” versus “1.5x, 2x, or 3x” is only a description of his felt experience—the point is that the way work gets done has undergone a generational break.

Trader Joe’s: Hawaiian shirts and counter positioning (Audio)How Brad Jacobs Built 8 Billion-Dollar Companies