Huawei and Cambricon lead domestic chips in rapidly capturing Nvidia's China market share

New players are quietly entering the top domestic GPU market.

According to recent media reports, ByteDance is negotiating with Daysci to purchase at least 50,000 AI chips, primarily for inference tasks. The chips in question are mainly Daysci's Zhixia series cloud inference GPUs, while the Tiangai series is used for training scenarios.

The news sent shockwaves through the market. After all, ByteDance is the largest domestic AI computing power procurement entity—planning to increase capital expenditure by over 20 billion yuan in 2026.

However, as of now, neither ByteDance nor Daysci has responded.

If this deal materializes, Daysci would become ByteDance's third-largest domestic GPU supplier, after Huawei and Cambricon.

For Daysci, which listed in Hong Kong in January this year, this concerns not only a massive order but also securing "certification from major tech companies."

Yet what merits more attention than the order itself is the signal behind this event—domestic AI chips are genuinely transitioning from "policy-driven procurement and industrial pilots" into real-world applications at major internet companies, shifting from alternative options to essential computing power support.

The U.S. has currently approved certain Chinese companies to purchase Nvidia H200 under controlled conditions. However, the H20 "backdoor incident" has added compliance and security review pressures to procurement decisions for Chinese buyers. For major domestic tech companies, avoiding putting all eggs in one basket has become standard practice.

More importantly, ByteDance's computing power demand is undergoing structural changes.

According to QuestMobile data, as of March 2026, ByteDance's AI assistant Doubao has reached 345 million monthly active users. The pressure from user growth comes not only from model training but also from ongoing inference costs post-launch. In inference scenarios, chips face less stringent requirements for interconnect bandwidth, VRAM, and ecosystem maturity compared to training scenarios. Domestic chips have reached usable levels on the inference side.

As of March 2026, Doubao's large language model has exceeded 1.2 quadrillion daily average token calls, representing a thousandfold increase from its initial launch. Based on Volcano Engine pricing and user behavior projections, daily average computing power consumption costs have reached tens of millions of yuan—not including one-time investments in intelligent computing centers and chips.

Although Doubao 2.0 achieved a 43% improvement in inference efficiency, with per-10,000-token costs at only 38% of overseas leading models' compliant pathways, it remains difficult to bridge the loss gap through computing cost reductions when serving 345 million monthly active users for free.

Under this pressure, ByteDance has embarked on "ambitious" investments.

According to multiple media outlets citing South China Morning Post, ByteDance's AI infrastructure capital expenditure budget increased by approximately 25% to 20 billion yuan in 2026. This increase is primarily driven by two factors: the company's continued increased investment in artificial intelligence and rising memory chip costs.

There are also reports that ByteDance is considering raising its 2026 spending cap to $7 billion. Meanwhile, in 2025, the company's net profit declined over 70% year-over-year. With such a massive gap between profits and expenditures, Zhang Yiming's bold bet on computing power likely aims at securing the company's position over the next five years.

ByteDance's computing power supply chain strategy is clear: using Huawei Ascend and Cambricon's high-end training cards for training, introducing Daysci's Zhixia series for inference, with three pathways running in parallel. This "dual-legged approach to training and inference, with dual backup for domestic and imported chips" is becoming the "standard configuration" for major internet companies.

However, just as ByteDance's plans to procure domestic chips are making headlines across the board, the actions of another domestic GPU manufacturer warrant even closer attention.

On the evening of June 16, Zhipu officially open-sourced its next-generation flagship large language model GLM-5.2. The following day, Birei Technology and Moore Threads respectively announced completion of "Day-0" adaptation. Birei's Bili 166 series achieved adaptation optimization based on the vLLM inference framework, providing developers with rapid deployment solutions. Following the announcement, Birei's stock price rose 7.09% that day.

"Day-0 adaptation" is key to understanding the domestic GPU competitive landscape—it means a model can run on the chip on its release day. This indicates that chip manufacturers must not only have excellent hardware but also keep pace with software stacks, toolchains, and developer ecosystems. In this regard, Birei Technology has already established a clear first-mover advantage.

Over 20 leading domestic large language models, including Tencent's Hunyuan Hy3 preview, Alibaba's Qwen3.6, the entire DeepSeek series, MiniMax M3, Zhipu's entire GLM series, and Kimi, have all achieved Day-0-level synchronized adaptation with Birei's chips. DeepSeek is particularly noteworthy—Birei reportedly completed full-series adaptation in just hours, setting a record for domestic chip response speed.

Overlaying this adaptation list with ByteDance's supplier list reveals a clear signal: Birei Technology now stands on equal footing with Huawei and Cambricon.

Huawei has long held a leading position. Ascend's ecosystem depth and 10,000-card cluster capabilities remain benchmarks that other domestic manufacturers struggle to match. Cambricon entered early and has consistently supplied chips to ByteDance, serving as a core player in the computing power supply chains of major tech companies. As a new force, Birei Technology has achieved parity with the preceding two through national-level certification, capital backing, and large language model ecosystem positioning.

In May 2026, the state established a dedicated AI chip category in its security and reliability assessment for the first time. Nine domestic chip enterprises, including Huawei HiSilicon, Alibaba Pingtouge, Birei Technology, Hongmeng Information, Daysci, Muzhai Information, and Moore Threads, all achieved the highest security and reliability rating of Level 1. Within the national certification framework, Birei Technology stands shoulder-to-shoulder with Huawei and Alibaba Pingtouge.

Capital markets voted more directly: Birei Technology listed in Hong Kong on January 2, 2026, with opening stock price surging 82%, and market capitalization briefly exceeding 100 billion Hong Kong dollars, becoming the first GPU stock listed on the Hong Kong Stock Exchange.

The value of this "circle of friends" lies in creating a positive feedback loop: the more models running on Birei chips, the more mature its software stack becomes; the more mature the software stack, the faster new model adaptation proceeds; the faster the adaptation, the more model developers will choose Birei. This is the "flywheel effect" of ecosystem development.

Of course, Birei Technology is not fighting alone. The entire domestic GPU sector is engaged in an arms race centered on large language model adaptation.

As mentioned earlier, the ecosystem depth of Huawei Ascend is difficult for peers to match. This time, Zhipu GLM-5.2's inference adaptation with Ascend was completed on Day 0. Cambricon completed Day-0 adaptation on the day DeepSeek-V4 was released. As one of ByteDance's two largest GPU suppliers, its NeuWare software stack's influence continues to expand.

Since June, Moore Threads has continuously achieved same-day adaptation for MiniMax M3 and Zhipu GLM-5.2, with its MTT S5000 response speed matching any competitor.

Enflame Technology focuses on the cluster direction. It jointly released with Tencent Cloud the commercial version 3.0 of the "Liaoyuan" intelligent computing cluster, which has adapted to mainstream large models including DeepSeek, Tencent Hunyuan, and Zhipu AI, completing thousand-card and ten-thousand-card cluster deployments.

It's also worth noting that Enflame Technology received approval on June 15. Should it successfully list, the "four rising stars of domestic GPUs"—Moore Threads, Muzhai Information, Birei Technology, and Enflame Technology—will gather on capital markets for the first time.