The Cambricon Siyuan 590 and Siyuan 690 (Chinese: 思元590 / 思元690) are data-center artificial intelligence accelerators designed by Cambricon, a partially state-owned Chinese chip company often called "the NVIDIA of China." The two chips sit at the top of Cambricon's accelerator range and are positioned as domestic alternatives to NVIDIA hardware for training and running large neural networks. The Siyuan 590, already shipping, is benchmarked by analysts at roughly the level of NVIDIA's A100, while the next-generation Siyuan 690, still in testing as of late 2025, is reported to target NVIDIA H100-class performance [1][2][3]. In December 2025, Bloomberg reported that Cambricon planned to more than triple its output in 2026 to about 500,000 AI accelerators, including as many as 300,000 of these advanced parts, a ramp constrained mainly by foundry capacity at SMIC and by limited supply of high-bandwidth memory [1][4][5].

The chips are central to China's push to build a domestic AI compute base after United States export controls restricted access to NVIDIA's most capable processors. Because much of the technical detail comes from analyst estimates and unnamed sources rather than Cambricon's own datasheets, several specifications below are reported or estimated rather than officially confirmed.

Naming and the MLU product line

Cambricon markets its accelerators under the Siyuan brand, where 思源 (siyuan) is a classical phrase meaning roughly "to remember one's source." The same products are identified in technical documentation by the prefix MLU, short for Machine Learning Unit, the name of Cambricon's underlying processor architecture and its accelerator cards. The marketing and product numbers line up directly: the Siyuan 290 is the MLU290, the Siyuan 590 is the MLU590, and the Siyuan 690 is the MLU690 [2][6]. This article uses the Siyuan names, which dominate recent English-language reporting.

The Siyuan family grew out of Cambricon's pivot away from licensing NPU intellectual property toward selling its own chips. The MLU100, launched in 2018, was the company's first cloud chip. The Siyuan 290 (MLU290) followed in 2021, built on a TSMC 7-nanometer process with 32 GB of HBM2 memory in the OAM (Open Accelerator Module) form factor [2]. The 590 and 690 are the next two steps up that ladder, and they arrive under very different manufacturing conditions than the TSMC-built 290.

The Siyuan 590

The Siyuan 590 is Cambricon's current flagship accelerator for cloud computing. It is aimed primarily at inference and at training mid-sized models, and reporting describes deployments in server clusters on the order of 1,000 to 3,000 cards [3]. According to TechRadar, the 590 delivers about 80 percent of the performance of NVIDIA's A100 while being built on a domestic 7nm process [3]. It carries roughly 80 GB of high-bandwidth memory, enough to hold sizable language models on a single card, although Cambricon has not published a full official specification and some figures vary by source [3][7].

The 590 was originally a TSMC-era design, but Cambricon's placement on the U.S. Entity List in 2022 cut its access to TSMC's leading-edge nodes and pushed volume production onto SMIC. Industry trackers report that the 590 is manufactured on SMIC's N+2 process and entered volume production in the third quarter of 2024, with mass shipment beginning in early 2025 [2][6]. Estimates of 2025 shipments range from roughly 100,000 to 200,000 units, with ByteDance identified as the dominant buyer [7].

Analysts are blunt about where the 590 stands against NVIDIA. It lags the A100 in raw speed, memory bandwidth, and software maturity, and the A100 is itself two generations behind NVIDIA's newest parts. The case for the 590 is availability rather than peak performance: for Chinese buyers cut off from NVIDIA's high-end chips, the relevant comparison is not the 590 against an H100 but the 590 against nothing [7].

The Siyuan 690

The Siyuan 690 is the unreleased successor and the part that has drawn the most attention. TechRadar and other outlets report that it is designed to be positioned against NVIDIA's H100, the GPU that became the workhorse of the generative-AI boom [2][8]. If the 690 reaches anywhere near that target on a domestic process, it would close much of the gap between Chinese accelerators and the Western high end, at least for chips that Chinese firms can actually buy.

That is a large "if." As of late 2025 the 690 was still in the testing phase, and reporting suggested that large-scale mass production could slip to the second half of 2026 [2]. The H100 comparison is a design goal attributed to sources and analysts, not a measured result, and no independent benchmarks of shipping 690 silicon were available at the time of reporting [2][8]. Several analysts add that even an H100-class domestic chip would still trail NVIDIA's current generation on throughput, memory bandwidth, and software maturity [8].

Reported specifications

The table below summarizes the figures reported for the two chips. Most are estimates or design targets from journalism and analyst commentary rather than official Cambricon specifications, and they are marked as reported throughout.

Attribute	Siyuan 590 (MLU590)	Siyuan 690 (MLU690)
Product type	Cloud AI training and inference accelerator [3][7]	Next-generation cloud AI accelerator [2]
Process node	SMIC N+2 (7nm class), reported [2][6]	SMIC N+2 (7nm class), reported [1][2]
Memory	About 80 GB high-bandwidth memory, reported [3][7]	Not disclosed
Performance target	About 80% of NVIDIA A100, reported [3]	NVIDIA H100-class, design goal [2][8]
Status	Volume production from Q3 2024; mass shipment early 2025 [2][6]	In testing; mass production possibly H2 2026 [2]
Main customer	ByteDance (largest buyer) [4][7]	Not disclosed

Manufacturing and supply constraints

Cambricon's 2026 ramp depends on two scarce inputs, and both are downstream of U.S. export controls. The first is advanced foundry capacity. With TSMC off-limits, Cambricon relies on SMIC's N+2 7-nanometer node, which is built entirely around older deep-ultraviolet (DUV) lithography rather than the extreme-ultraviolet (EUV) systems that TSMC and Samsung use, because the United States and its allies have blocked EUV tool sales to China [4][5]. Bloomberg reported that yields on Cambricon's largest dies run at roughly 20 percent, meaning about four of every five chips coming off the wafer fail to meet the target, which sharply limits how many good 590 and 690 dies SMIC can actually produce [1][4][5].

The second constraint is high-bandwidth memory. AI accelerators need large pools of HBM to keep their compute units fed, and HBM3 and HBM3E supply is dominated by South Korea's SK Hynix and Samsung, with no fully competitive domestic Chinese alternative yet [4][5]. ByteDance already accounts for more than half of Cambricon's orders, with Alibaba expected to expand its purchases as it scales its domestic AI clusters, so demand is not the bottleneck [4][5]. Supply is.

Production plans for 2026

The figure that pushed these chips into headlines came from Bloomberg on 4 December 2025: Cambricon planned to more than triple its AI chip output in 2026 to roughly 500,000 accelerators, including as many as 300,000 units of the Siyuan 590 and 690 [1][9]. Tom's Hardware and other outlets, citing a Goldman Sachs estimate, put 2025 output at about 142,000 chips, which makes the 2026 figure more than a tripling [4][5]. The plan was described as Cambricon's bid to fill the gap left by NVIDIA's retreat from China and to compete with Huawei, whose Ascend line is the leading domestic accelerator [5].

Whether Cambricon can hit those numbers is the open question running through the coverage. At a 20 percent yield, producing 300,000 good advanced dies means pushing a very large volume of wafers through a foundry whose most advanced capacity is already in heavy demand, and it is not clear the capacity exists [4][5]. The targets are best read as ambitions reported by sources rather than committed shipments.

Geopolitical and market context

The Siyuan chips exist because of export controls. U.S. restrictions from 2022 onward barred NVIDIA from selling its top accelerators to China, then tightened around the cut-down H20 part NVIDIA designed for the Chinese market, leaving Chinese AI developers short of compute [4][5]. Beijing responded by pushing self-reliance, with reporting describing an estimated $98 billion in government and corporate spending on the effort in a single year and moves to favor homegrown chips in domestic data centers [4]. Cambricon itself was added to the U.S. Entity List in December 2022, which severed its TSMC access and forced the shift to SMIC [3][8].

The financial turnaround has been steep. After nine years of losses, Cambricon reported its first profit on the strength of demand for the 590, with first-half 2025 revenue of about 2.88 billion yuan, more than 40 times the year-earlier figure, and net income of roughly 1.04 billion yuan [3][7]. Yet the company held only about 1 percent of China's AI chip market in 2024, ranking behind NVIDIA, Huawei Ascend, and AMD, so its position rests on a small number of large customers and on chips that still trail the global frontier [7]. The Siyuan 590 and 690 are the test of whether a DUV-constrained domestic supply chain can deliver AI accelerators at the scale China now wants. The ambition is clear. The yields, the memory supply, and an unproven 690 are what stand in the way.

References

Bloomberg, "Cambricon Aims to Triple Output to Replace Nvidia in China," 4 December 2025. https://www.bloomberg.com/news/articles/2025-12-04/cambricon-aims-to-triple-chip-output-to-replace-nvidia-in-china
TrendForce, "[Insights] Cambricon Remains China's Top AI Chip Startup; Rumored 2026 Triple Output Faces SMIC Limits," 15 December 2025. https://www.trendforce.com/news/2025/12/15/insights-cambricon-remains-chinas-top-ai-chip-startup-rumored-2026-triple-output-faces-smic-limits/
TechRadar, "Cambricon's Siyuan 690 is designed to rival Nvidia's H100 but questions remain," 2 September 2025. https://www.techradar.com/pro/could-cambricon-create-a-deepseek-moment-in-ai-hardware-the-rise-of-chinas-answer-to-nvidia-has-been-nothing-short-of-meteoric-but-is-it-too-good-to-be-true
Tom's Hardware, "Cambricon targets 500,000 AI chips in 2026 as China accelerates domestic hardware push," 12 December 2025. https://www.tomshardware.com/tech-industry/semiconductors/cambricon-targets-500000-ai-chips-in-2026-as-china-accelerates-domestic-hardware-push
Tom's Hardware, "Chinese tech firm Cambricon looks to step into Nvidia void, triple AI chip production next year," 4 December 2025. https://www.tomshardware.com/tech-industry/chinese-tech-firm-cambricon-looks-to-step-into-nvidia-void-triple-ai-chip-production-next-year-seeks-to-rival-huawei-but-production-remains-a-concern
StockCounterparts, "Cambricon: China AI chips, MLU series." https://www.stockcounterparts.com/companies/cambricon
Tech Buzz China, "Cambricon: China's Nvidia, or Nvidia Without the Profits?" https://techbuzzchina.substack.com/p/cambricon-chinas-nvidiaor-nvidia
TechRadar, "Cambricon aims to triple its AI chip production next year, but faces massive fabrication challenges and rivals like Huawei." https://www.techradar.com/pro/this-chinese-chip-giant-is-boosting-production-to-try-and-take-on-nvidia-but-how-will-huawei-feel
Reuters via TradingView, "Cambricon Technologies Plans To More Than Triple Production Of AI Chips In 2026 - Bloomberg News," 4 December 2025. https://www.tradingview.com/news/reuters.com,2025:newsml_FWN3XA00E:0-cambricon-technologies-plans-to-more-than-triple-production-of-ai-chips-in-2026-bloomberg-news/

Editing: Cambricon Siyuan 590/690