Anker Does Chips: From the Five-Series to the Seven-Series, a Consumer Electronics Company Catching Up on Its Homework
Anker's first 11 years went smoothly, but in 2022 it found that the product lines it had built couldn't beat the unicorns outside. Yang Meng's solution wasn't to swap people out, but to transform the company from a five-series product company into a seven-series innovation company — at the cost of catching up on the values of first principles, pursuing excellence, and growing together.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The first 11 years went smoothly because they chose easy mode
Yang Meng defines the company's first decade as a ‘five-series’ company: good quality, some micro-innovation, priced 20% to 30% above the market average, using Amazon's star-rating mechanism to force products to reach 4.5 stars or above. This model had extremely high certainty — demand was visible on the left, good products could be customized on the right, and the bridge in between was built by themselves. But in 2022 he found that the product lines he had built couldn't beat the unicorns in various industries in real competition, and he once felt ‘very depressed, even despairing.’ His retrospective: it wasn't that the processes and methods were wrong, but that the company hadn't clearly articulated its mission, vision, and values — and no matter how good the processes were, in the end they would still fight a losing battle.
— Yang MengThe power bank category is certain to disappear
In 2016, while studying at Hupan Business School, Yang Meng was asked ‘Will your category become a hundred-billion category?’ His answer was frightening: not only would power banks not become a hundred-billion category, they would most likely die in a few years. Because consumer electronics is a fast-birth, fast-death industry — MP3 players, tape players, CD players went from starting to buy to stopping buying in a span of about ten years. At the time he judged the power bank's life cycle to be ten to fifteen years; in hindsight that was a bit filtered, but he believed the category would definitely disappear. This fear of categories directly gave rise to the later search for ‘categories that won't die’ — he once looked at e-cigarettes, because nicotine is most likely a human need that won't disappear, but found it was a heavily regulated category, unsuitable for a startup.
— Yang MengThe chip effort started because call noise cancellation wasn't good enough
Anker established an acoustic algorithm team in 2021, initially using deep learning models to solve sub-problems of noise cancellation. After seeing ChatGPT in 2023, Yang Meng judged that the future was the era of end-to-end problem solving — computers spent their first 80 years solving problems by decomposition, breaking big problems into small ones, each sub-problem solved by a group of people; end-to-end is data learning plus reinforcement learning. But to run a million-parameter model in earbuds, the von Neumann architecture's separation of memory and compute became a bottleneck: every frame, every second, all the model's parameters had to be moved from memory to the NPU, and the movement itself accounted for the largest power consumption. So they turned to a compute-in-memory architecture, starting the project in August 2023, with products carrying the chip launching in May 2026.
— Yang MengThe chip is only for internal use, not for sale
Yang Meng explicitly says this compute-in-memory chip is ‘our own little pants,’ completely customized by Anker's models, with operators and model structures adapted to its own models. He compares it to Apple's A-series and M-series chips and Huawei's Kirin chips — the best chips are all used in one's own hardware devices. The logic is a closed loop: one's own devices can run the chip to its best performance, producing the best user value; for the chip team, the commercial value generated by selling the chip is far less than putting it in one's own seven-series products to make the products sell for more. He judges that compute-in-memory is more competitive on the inference side, starting from a million parameters, and in the coming years we will see applications at the scale of billions or tens of billions of parameters.
— Yang MengToday's smart home is essentially adjustable
Yang Meng divides products into three stages: non-adjustable, adjustable, and self-adjusting. Today's smart home is essentially ‘adjustable, presettable’ furniture — a dozen buttons on a toilet isn't smart, an ergonomic chair has a dozen adjustable parts but you neither know how to adjust them nor remember to. A truly smart stool should sense you coming and move back, sense you playing a game and recline the backrest. Why isn't the home truly smart in 2026? Because to achieve true intelligence requires three capabilities — perception, planning, and control — with sensors and better controllers below, and models and software above, all of which are gradually maturing. He judges that this set of capabilities first matured in autonomous driving cars, and now it's time for them to spill over into home products.
— Yang MengTechnological advantage will always be erased by the limits of human perception
Yang Meng proposes three layers of moat. The first layer is uniquely leading technology, but technology encounters ‘editorial benefit diminishing’ — human perception has limits, and when technology hits the limits of human perception, there isn't much room to go further, and others can catch up. He gives two examples: Apple's Retina display, where pixels are so dense that a normal person can't see the difference, and going further, people can't perceive it; a 146-watt charger made as small as a card, half the size again would still be meaningful, but even smaller would be meaningless. So excess technological advantage will definitely be erased over a long enough time. The second layer is brand, i.e., trust in users' hearts, but brand can also be consumed. The deepest moat is mission, vision, and values, and a group of people who truly believe in them.
— Yang MengUnlimited tokens for all employees, internal usage has exceeded 10 trillion
Anker started building an AI middle platform in September 2025, connecting various models, providing front-end access via chat, local command line, web pages, etc., with the whole company using AI on the same platform. By the time the episode aired, internal token usage had exceeded 10 trillion, about 150 billion per day, with 6,000 employees, equivalent to 25 million tokens per person per day, five to six hundred million per month. About one-third is the highest-tier model, over 40% mid-tier, and 20% entry-level. Yang Meng says management had no objections at the time and didn't strictly measure output; the logic was ‘believe in order to see’ — the future organization must be one strengthened by AI, and it's best to start early. This year the cost of employees using tokens internally is already a few hundred million.
— Yang MengAI eliminates junior positions, but senior positions won't be cut off
Regarding the claim that ‘AI eliminates junior positions so future senior positions can't grow,’ Yang Meng uses architecture as a counterexample: ancient architects rose slowly from frontline craftsmen, and before becoming designers they probably spent a dozen years laying bricks and plastering; modern architectural design is already an established category, and you don't need to have laid bricks to become a designer. He believes the future will most likely be the same — you won't need to have studied programming and debugged many bugs to be an architect at the top. Anker has non-programming-background campus hires who, in clear business domains, use AI to deliver complete systems supporting business operations, completely skipping traditional junior positions. He also proposes the concept of ‘N-time efficiency worker,’ divided into big cows (people who build AI agents) and small cows (people who use AI agents), and the organization will consist of a few big cows and more small cows.
— Yang MengIn their own words · checked verbatim
This world is very fair. Just now we talked about how we didn't go through financing difficulties, right? And things went smoothly. But in recent years, we've actually had a pretty painful time.
这世界是特别公平的 刚刚讲就是我们前面 其实也没有经过融资的困难 对吧 然后过得也很顺 那这几年其实我们过的还是挺痛苦的
Yang Meng30:19
So actually, today's so-called smart furniture is essentially adjustable, presettable furniture.
所以其实 今天的所谓的智能家具 本质上是叫可调节 可预设的 的家具
Yang Meng1:27:02
When you understand today that technology will always encounter the limits of human perception, when you hit the limit, you find there isn't much room to go further, and at that time others will gradually catch up.
当你今天理解说 技术 它总会遇到人的感知的极限的时候 当你遇到极限的时候 你发现你在往上做的空间不大了 而这个时候别人会逐步的追上来
Yang Meng2:15:38
The company's deepest moat is its mission, vision, and values, and a group of people who truly believe in the mission, vision, and values.
这公司最深的户层盒是他的使命愿景价值观和一群真正的相信使命愿景价值观的人
Yang Meng2:17:39
Believe in order to see. Actually, up to this point, we haven't strictly measured what all the tokens we've allocated have specifically produced today.
因为相信所以看见 其实在一直到这个点 我们并没有去严格的衡量 不过说我们划出去的所有的token 它具体今天产出了什么
Yang Meng2:32:59
If this company didn't have me, right? It might not be as good as it is today — getting better and better, becoming more challenging, more complex and more fun. And if this company didn't have Dongping, it would already be dead.
这公司如果没有我 对吧 这公司可能不一定像今天这么 就是越来越好像 就是变得好像更有挑战 对吧 更更复杂更好玩 然后 这公司如果没有东平的话 应该已经死了
Yang Meng3:20:33
Figures
| Anker employee count | 6,000 | 2:28:52 |
| Anker AI-related engineer count | about 270-300 | 1:31:05 |
| Anker compute-in-memory chip parameter count | total four trillion, single model max about two trillion | 1:24:02 |
| Anker earbud chip model parameter count | less than 2 million parameters | 1:10:53 |
| Anker internal AI token cost | a few hundred million this year | 2:32:59 |
| Anker power bank global recall count | several million | 2:26:52 |
| Anker mutual-use energy storage product average order value | 10,000 USD, over 1,000 customers ordered directly online in the first month | 2:20:43 |
Glossary
- Compute-in-Memory
- Placing storage and computation in the same unit, avoiding repeatedly moving parameters between memory and compute units, reducing power consumption.
- von Neumann architecture
- A computer architecture with separated storage and computation, where programs and data are placed in memory and moved to the CPU for execution.
- N-time efficiency worker
- A term coined by Yang Meng, referring to employees who use AI to multiply their personal efficiency by N times, divided into big cows who build agents and small cows who use agents.
- SackKit
- A set of process agents for software development, defining how product managers, developers, and testers work, which inspired Anker's AI organizational change.
- Conway's Law
- The shape of an organization determines the results and effectiveness of its output.
How to listen
Consumer electronics founders, hardware investors, CEOs undergoing AI organizational change, and engineers concerned with the path to deployment for on-device models and compute-in-memory chips.
The growth history and early startup stories from 1:13 to 31:13 can be fast-forwarded; the core judgments are in the second half.