A pair of AI glasses typically carries the battery of a smartwatch and a much worse thermal problem, because it sits on your face in the sun. The camera, the microphone array, the Wi-Fi radio, and the model all want power at once. The company that solves this is unlikely to be a GPU designer. It is more likely to be a company that has spent ten years making earbud chips that sip microwatts. That is the bet Bestechnic (恒玄科技, 688608.SH) has been making since 2015.
The company's 2025 results show both the opportunity and the awkwardness of that transition.
From earbuds to eyewear
Bestechnic's core skill is not raw compute. It is a specific and unfashionable discipline: keeping a chip alive, listening, and processing audio on a battery smaller than a coin. Bluetooth earbuds forced the company to master always-on voice detection, ultra-low-power audio codecs, and aggressive power gating — the art of shutting down ninety percent of a chip and waking it in microseconds.
AI glasses (智能眼镜) need exactly that discipline, plus a little more. The product category that took off after 2024 asks for wake-word detection, spoken interaction with a cloud or on-device model, a camera for photos and short video, and a Bluetooth or Wi-Fi link to a phone. Every one of those features competes for the same milliwatt budget.
The company's flagship for this market is the BES2800, a 6nm chip that combines multiple CPU cores, a GPU, and an NPU with an audio subsystem. The company's 2025 annual report places the chip in Xiaomi's AI glasses, Alibaba's Quark AI glasses (夸克), and Li Auto's Livis glasses (理想); DJI's Mic 3 wireless microphone (大疆) also uses the platform. That is not proof of volume, but it is proof of passing qualification at companies that ship consumer hardware at scale.
What the financials reveal
Bestechnic's full-year 2025 results, reported in early 2026, are a case study in how a component company changes shape without changing much of its revenue base.
- Revenue: RMB 3.525 billion (≈US$490 million / HK$3.83 billion), up 8.02%.
- Net profit attributable to shareholders: RMB 594 million (≈US$82.5 million / HK$645 million), up 29.00%.
- Gross margin: 38.69%, up 3.98 percentage points.
- Bluetooth audio chips: 53% of revenue, down nine points year on year.
- Smartwatch and fitness band chips: 35%, up three points.
- Smart hardware and other chips: 12%, up six points, with the segment's revenue at RMB 423 million (≈US$58.8 million), up 116.1%.
Then the number that keeps the story honest: fourth-quarter 2025 revenue was RMB 591 million (≈US$82 million), down 25.15% year on year and 40.58% quarter on quarter. The decline coincided with fading consumer subsidy programmes and rising memory prices that squeezed demand for the devices its chips go into.
So the transition is real but not yet decisive. The glasses business is growing fast from a small base, inside a company whose revenue still depends on earbuds and watches, in a market that is cyclical and subsidy-sensitive.
Why the next chip is the actual test
The BES6100, which the company is developing for glasses, is where the strategy becomes legible. The company's briefing material on the chip describes a 6nm hybrid architecture split into two domains:
- A low-power domain with Wi-Fi 6, a next-generation Bluetooth spec it says it supports, and multi-core Cortex-M processors, intended for always-on tasks such as voice wake, audio processing, and environmental sensing.
- A high-performance domain whose compute the company says is significantly above the BES2800, capable of running Linux or Android and of on-device inference (端侧推理).
- Integrated ISP and VPU blocks for real-time image processing and video recording.
That two-domain structure is the correct answer to the physics. Glasses cannot afford a single high-performance die idling all day, and they cannot afford a low-power die that cannot handle a frame of video. Splitting the silicon lets the always-on part stay in a state measured in milliwatts while the vision and inference part sleeps.
But it also moves Bestechnic onto ground it has never held. An ISP and a video pipeline are camera-company problems. Running Android on a wearable is a phone-company problem. On-device inference is a model-compression problem, where the chip vendor is expected to supply quantization tools, operator coverage, and a runtime that model authors will actually target. Being good at milliwatt audio does not automatically transfer to any of those.
The competitive frame nobody states plainly
AI glasses are an odd market because the bill of materials is dominated by things that are not the main chip: the optical module, the frame, the battery, the camera. The SoC is a modest share of cost and a large share of the user experience, which is precisely the profile of a component that gets squeezed on price once volumes rise.
Two pressures follow. First, Qualcomm holds the reference design position for premium glasses, and any Chinese SoC has to win on cost, customization, or supply security rather than on brand. Second, the Chinese smart-glasses market has many brands and short product cycles, which rewards fast turnarounds but punishes suppliers who cannot amortize tooling across large runs.
There is a third pressure that is easy to overlook: memory prices. Broker commentary on the fourth quarter pointed to rising memory costs reducing demand for consumer devices. A glasses SoC vendor does not control DRAM pricing, but its customers' willingness to build a product in a given quarter depends on it.
Honest limitations
- Design wins are not shipments. Xiaomi, Alibaba, and Li Auto are large customers, but none of them discloses unit volumes attributable to a specific chip, and glasses volumes remain small relative to phones.
- The BES6100 specifications are the company's own description in its briefing material, not an independently measured benchmark. No third-party power or performance figures exist yet.
- Battery-life claims in this category are notoriously dependent on test conditions — brightness, camera duty cycle, network, and how often the assistant is invoked. I have not repeated any.
- The 2025 revenue split depends on how the company classifies products. "Smart hardware and other" is a catch-all that includes wireless microphones and conference devices as well as glasses.
- This is not investment analysis. Market capitalisation, valuation, and share-price performance are outside the scope of what I verified.
What readers can do now
- If you are designing a wearable, start from the power budget, not the TOPS number. Ask vendors for always-on current draw with the wake-word engine running and the radio connected. That single figure decides whether your product ships with a day of battery or four hours.
- If you are buying AI glasses, look at the chip generation. A device built on a phone-class application processor will run hotter and heavier than one built on a purpose-designed wearable SoC, and the difference shows up in the second hour of use, not the first.
- If you are tracking the supply chain, watch whether the BES6100 reaches volume production. A vision-capable, dual-domain wearable SoC that actually ships would mark the point where Chinese audio-chip makers became wearable platform suppliers rather than component vendors.
