Chinese company Tian Shu (天数智芯) revealed 50,000-unit GPU procurement, indicating enterprise shift toward domestic AI chips.

On June 24, Doubao Professional Edition officially launched. Just days before, ByteDance released plans to make a major purchase of at least 50,000 AI inference chips from Chinese GPU enterprise Tianshu Intelligent Technology for Doubao's terminal inference needs.

The news of this major procurement is still unfolding. If it comes to fruition, it will be a major breakthrough for Tianshu Intelligent Technology's operations, with its stock price surging as a result, which seems to confirm the relevant reports.

On China's capital markets, while Tianshu Intelligent Technology's prominence is less than Moore Threads and Muxi Shares, which went public later, it was established earlier and was China's first mass-produced training and inference GPU, the first company to achieve 7nm mass production.

After founder Li Yunpeng's departure, Tianshu Intelligent Technology experienced two additional CEOs, with the company forming a management combination of "capital-backed chairman + executives with AMD and NVIDIA technology backgrounds." After listing on the Hong Kong stock exchange, Tianshu Intelligent Technology, Moore Threads, Muxi Shares, and Birentech form the "four giants of domestic GPUs" in the secondary market.

With explosive downstream demand and high-end chip imports blocked, Tianshu Intelligent Technology didn't choose to directly confront NVIDIA on high-performance training chips, but instead cleverly opted for a "three-legged" approach spanning training, inference, and edge computing.

Zhipu founder Tang Jie, who just released GLM-5.2, immediately responded: "That won't take that long." He also claimed the company aims to train models this year to match Claude Fable 5. In the following days, Zhipu's total market capitalization broke through 1 trillion Hong Kong dollars.

Today, AI industry chain stocks have evolved into an extremely frenzied state, far exceeding traditional valuation frameworks of PE/PS/ROE and growth stocks alike. Any effort to measure AI industry chain enterprises by revenue or profit would prove futile.

As domestic large models catch up with world-leading standards, domestic computing power capabilities are also rapidly closing the gap. With the important cooperation with ByteDance completed, Tianshu Intelligent Technology is destined to become a landmark enterprise in this "technology war."

On June 18, a user asked Elon Musk: When will Chinese large models catch up with Anthropic's Fable? Musk replied: "Probably the first quarter of 2027."

In 2027, the Tianquan architecture will emerge, establishing the goal of surpassing NVIDIA's Rubin architecture, and planning to advance breakthrough computing chip architecture development. If successfully realized, that would mark a new phase for Tianshu Intelligent Technology and China's GPU in shifting from technological following to independent leadership.

Last year, the company achieved revenue of 1.034 billion yuan, with a compound annual revenue growth rate of 76.21% from 2022-2025. If it reaches cooperation with ByteDance and Alibaba, it will further solidify its high-growth trajectory. Clearly this will become the cornerstone for achieving its medium and long-term plans.

Tianshu Intelligent Technology also released a "three-year (2025-2027) four-generation architecture" plan, clearly outlining a path of successive surpassing.

In 2025, the Tianshu architecture lands; this year plans to introduce Tianxuan and Tianjing architectures, with Tianxuan architecture matching NVIDIA's Blackwell; Tianjing architecture aims for comprehensive surpassing of Blackwell, striving to achieve full-scenario AI and accelerated computing coverage.

NVIDIA's CUDA ecosystem, after more than a decade of cultivation, has over 4 million developers. Based on CUDA, with compilers, debuggers, math libraries, communication libraries, and deep learning frameworks, a vast and flourishing ecosystem has grown.

Like WPS's compatibility with Microsoft Office back then, Tianshu Intelligent Technology's strategy is to build a full-stack self-developed software system from bottom-layer drivers and compilers to upper-layer AI frameworks, while maintaining deep CUDA compatibility, to solve the problems of domestic chips being "difficult to use and having high migration costs."

Industry insiders say that theoretically, over 80% of code used in products with Tianshu Intelligent Technology chips can migrate smoothly. Compared with Ascend and Cambricon, adaptation cycles and costs are significantly reduced.

In comparison, this is a relatively pragmatic path—not starting from scratch and directly competing head-to-head with NVIDIA's CUDA, but instead choosing the path of minimized migration costs, allowing developers to "get on board" with minimal resistance.

However, just like past disputes between Office and WPS, Tianshu Intelligent Technology's current choice of "CUDA compatibility" rather than "CUDA replacement" still shows insufficient long-term competitive moats.

"The core problem of domestic AI chips is inadequate ecosystem; if the ecosystem is good, performance at 60% would still have users," said Chen Weimin, an academician of the Chinese Academy of Engineering, hitting the core pain point of domestic GPUs.

NVIDIA's ascension to AI chip king precisely relies on CUDA, this software ecosystem deeply ingrained in developers' bones. And this is exactly the weak point of domestic GPUs.

Gai Lujiang is another major shareholder of Tianshu Intelligent Technology, a core investor of Shanghai Xishi. Unlike the previous two leaders, Gai Lujiang lacks chip technology background, with a professional trajectory primarily in finance and investment.

On the technical research and development side, CTO Lü Jianping previously worked at world-class chip enterprises including Samsung, NVIDIA, MediaTek, and Intel, accumulating deep expertise in general-purpose GPU fields; VPs Sun Yile and Shi Jiasheng both come from AMD, respectively responsible for chip R&D, mass production implementation and software development; VP Song Yu previously worked at IBM.

Similar to investment in Luckin Coffee, as early as 2017, Holtzbrinck Ventures completed early-stage investment in Tianshu Intelligent Technology through affiliated enterprises, continuously accompanying and deeply supporting the company thereafter, even leading the completion of key financing rounds, shareholding restructuring and board reorganization.

In 2022, when Lü Jianping was interviewed, he emphasized: "Tianshu Intelligent Technology has always approached graphics from the perspective of general-purpose computing rather than starting from traditional graphics pipelines and struggling to catch up; we will be leaders in this field."

With the strategic course corrections by several helmsmen and a powerful technical team, the company ultimately bet on the difficult track of domestic general-purpose GPU and achieved "three firsts"—China's first mass-produced training GPU, first mass-produced inference GPU, and first to achieve 7nm mass production.

After days of substantial gains, Tianshu Intelligent Technology's total market capitalization continued to surge, being defined by the market as the key force in breaking through international AI and chip covert warfare and blockade. Holtzbrinck Ventures not only became the biggest winner, but through this also entered the domestic GPU and AGI tracks.

After listing, Tianshu Intelligent Technology launched the "Tongyang Series" of edge computing products and, leveraging Holtzbrinck's connections, landed them in scenarios like Luckin Coffee. Thus becoming one of China's few chip enterprises with full-scenario general-purpose computing capabilities.

At that time, NVIDIA was just formally entering the artificial intelligence field. Li Yunpeng, then serving as research director of the database division at Oracle, sensed the potential of China's technology industry and resolutely resigned to start his own business back in China.

He founded Tianshu Intelligent Technology in Shanghai's Minhang District in December 2015, at a time 5 years earlier than Moore Threads and 4 years earlier than Birentech.

At that time, AGI's prospects were unclear. Li Yunpeng focused on edge-side AI inference chips and in 2019 released the first AI chip "Iluvatar CoreX I," targeting visual intelligence algorithm scenarios.

However, less than a year after this chip's release, investors believed the focus should be on general-purpose GPU direction, while Li Yunpeng insisted on his vision, ultimately departing. Subsequently, Liao Shijing, former co-president of Unisplendour and chairman of Unisplendour Guoxin, took over.

In 2021, Tianshu released the first mass-produced general-purpose GPU "Tiangai Gen 1," filling the gap in China's high-performance AI training chips; in July 2022, after Liao Shijing went missing, Gai Lujiang became the new helmsman, serving as chairman and CEO. The company released the "Zhikai Gen 1 Inference Acceleration Card" at year-end, completing the comprehensive layout of both training and inference product lines, thereby pioneering the "training + inference" dual-line layout domestically.

It's worth noting that Tianshu Intelligent Technology's largest institutional shareholder is Holtzbrinck Ventures—holding about 22.92% before IPO. Another notable investment case that gained fame in the industry was Luckin Coffee.

In other words, the volume of ByteDance's procurement has already exceeded all of Tianshu Intelligent Technology's GPU shipments from last year. If this order materializes, it's not only a leap in scale but also proof that it has obtained "major manufacturer certification."

According to calculations by Guoxin Securities, Tianshu Intelligent Technology's 2025 inference chip unit price was 11,700 yuan. If calculated at this price, after this order lands, the company will receive approximately 600 million yuan from ByteDance, equivalent to 60% of last year's entire annual revenue.

As of March 2026, China's daily Token consumption had surpassed 1.4 quadrillion, a more than 1,000-fold increase compared to early 2024. After AGI's large-scale landing, inference demand is showing faster growth than training. Morgan Stanley released a report at the end of last year indicating that China's AI chip industry is undergoing significant structural adjustment, with inference computing demand gradually replacing mere computing power accumulation and becoming the core driving force.

With both explosive market demand and industrial application promotion, Tianshu Intelligent Technology welcomes another critical moment since its IPO.

Since entering the GPU track, Tianshu Intelligent Technology has adhered to the "training and inference" two-legged strategy, with the Zhikai series of inference chips under its umbrella. In 2025, this series achieved revenue of 339 million yuan, a year-on-year increase of 238%, with growth speed double that of training chips.

Among these, Zhikai Gen1/Gen1X (released in 2022) continue to see strong volume growth, Gen2 was released in the fourth quarter of 2025 and has begun small-batch delivery, with Gen3 to enter mass production in the latter half of this year.

According to IDC statistics, in 2025, Huawei's Ascend shipped 812,000 units, ranking first domestically in GPUs, followed by Alibaba's Pingtouge, Cambricon and others, with Tianshu Intelligent Technology at 45,000 units.

Doubao is currently the leading AI-native application in the market, with monthly active users in May reaching 368 million, up 182% year-on-year. As of March this year, Doubao's daily average Token invocation surpassed 1.2 quadrillion, with continuous geometric growth plus high-frequency invocation, Doubao urgently needs to supplement resources on the inference side.

If the transaction ultimately succeeds, Tianshu Intelligent Technology will become ByteDance's third-largest domestic GPU supplier after Huawei's Ascend and Cambricon. Around the same time, there were also reports that ByteDance was considering using Baidu's Kunlun chip products, and Alibaba was also reaching out to Tianshu Intelligent Technology, likely becoming its next customer.

Compared with the ultra-high difficulty of training, the task of inference chips is much lighter. Its working environment is the deployment stage after model training, primarily responsible for executing model inference tasks, with high requirements for real-time performance, needing to maintain accuracy while possessing fast response and low power consumption.

Training has an endpoint, but as AI large models move toward application, as long as a product is online for a day, inference never stops for a second. It can be foreseen that future inference demand will rise exponentially.

According to Reuters, the core reason ByteDance made large-scale procurement is to respond to the enormous computing power invocation pressure generated by Doubao AI—this application's user base is simply too large.