A Shanghai-based AI chip startup called Fangqing Technology just closed a $1.5 billion funding round without having manufactured a single chip or announced a manufacturing partner. The company's bet is simple but ambitious: the hardware that powers AI inference, the continuous running of trained models at scale, requires a completely different architectural approach than the chips used for training. If Fangqing is right, it could reshape how the world deploys artificial intelligence.

What Makes Inference Hardware Different From Training Hardware?

To understand why investors are backing a pre-silicon startup with a $1.5 billion valuation, you need to understand how transformer models, the architecture behind every large language model in use today, actually work during deployment. Transformers have two distinct computational layers that behave very differently: the attention mechanism and the feedforward network (FFN).

The attention mechanism is what makes AI models contextually intelligent. To generate each new word, it reads and writes a key-value cache that grows with every token produced. This makes attention heavily memory-bandwidth-bound, meaning the bottleneck is not how fast the chip can do math, but how quickly it can move data from memory. On state-of-the-art graphics processing units (GPUs), attention typically achieves only 30 to 40 percent of the chip's theoretical computing capacity because it spends most of its time waiting for data rather than performing arithmetic.

The feedforward network is the opposite. It performs the same dense mathematical operations for every token, independent of context. This makes it compute-bound, meaning if the arithmetic units are fast and densely packed, FFN runs efficiently. GPUs handle this workload well.

The problem is that standard GPU hardware makes a single engineering compromise that serves neither workload optimally. A chip designed to maximize memory bandwidth is not maximally efficient at dense arithmetic, and vice versa.

How Does Fangqing's Decoupled Architecture Work?

Fangqing's proposed solution is to break this compromise by treating the two sublayers as genuinely separate hardware concerns. In the company's framing, attention is "context-aware" and memory-intensive, while the feedforward network is "context-free" and compute-intensive. The architecture decouples these two functions and routes each to a dedicated hardware module optimized for its specific computational profile, rather than running both through the same general-purpose GPU pipeline.

A chip architecture expert who reviewed the approach described it as a "promising R&D direction" that "may achieve parallel processing between modules and optimal matching of hardware resources, improving computing efficiency, throughput, and scalability of the model, and reducing costs and power consumption at the same time".

However, that endorsement comes with a crucial caveat: the expert confirmed the architectural direction is plausible, not that Fangqing has demonstrated it in silicon. The unresolved challenge is what engineers call the communication wall. In a distributed system where attention and FFN computations are routed to separate hardware modules, the inter-module interconnect must move intermediate activations between those units. If the latency and bandwidth cost of that transfer exceeds the efficiency gain from hardware specialization, the architecture produces worse end-to-end performance than a single GPU. No public benchmark from Fangqing has quantified where that tradeoff lands for its specific design.

Who Is Behind Fangqing, and Why Does That Matter?

The case for Fangqing's valuation rests heavily on its founder's track record. Liang Jun, the company's CEO, joined Huawei in 2000 and became chief architect of the HiSilicon Kirin system-on-chip (SoC), the processor that powered Huawei's flagship smartphones for nearly a decade and represented the most technically ambitious consumer chip ever designed to production quality inside China.

In 2017, Liang left Huawei and joined Cambricon as chief technology officer, where he oversaw the launch of the company's first 7-nanometer AI chip, the Siyuan 290, and subsequent models. His departure from Cambricon in February 2022 was contentious. Liang said the company failed to provide working conditions stipulated in his labor contract. Cambricon described it as a mutual disagreement. The day news of his departure spread, Cambricon's market capitalization fell by nearly 6 billion yuan, approximately $889 million.

Liang joined Fangqing as CEO in August 2024, becoming the public face of a company incorporated in September 2022. The two largest shareholders hold their stakes through partnerships co-managed by Liang Jun and Li Kaipu, the company's founding legal representative, with Liang holding an 80 percent controlling interest in the managing entity.

Why Is the Inference Chip Market Suddenly So Hot?

Fangqing's fundraising reflects a structural shift in global artificial intelligence hardware economics. Inference spending, the cost of running trained models continuously at scale, has grown to rival or surpass training spending as AI services move from experimental to operational. Custom AI accelerators, also called application-specific integrated circuits (ASICs), are projected to grow at 22 percent in 2026, outpacing general-purpose GPU growth at 19 percent.

Bloomberg Intelligence estimates the custom AI ASIC market will reach $118 billion by 2033 at a 27 percent compound annual growth rate.

In the United States, the inference chip race has already produced landmark deals. Groq's Language Processing Unit technology, which achieved the highest sustained throughput for real-time voice and interactive agent applications on the market, was acquired by NVIDIA for approximately $20 billion in December 2025. Cerebras, whose wafer-scale chip broke 1,000 tokens per second for large-scale model inference, signed a $20 billion-plus framework agreement with OpenAI.

What Strategic Factors Are Driving Investment in Chinese Inference Chips?

China has no domestic incumbent at comparable production scale or contractual position to Groq or Cerebras, a gap that investors appear to believe Fangqing is positioned to eventually address. The driving strategic context is US export controls, which have progressively limited China's access to advanced semiconductor manufacturing and cutting-edge AI hardware.

Fangqing's funding round was led by Xu Hui Capital, the state-owned investment platform of Shanghai's Xuhui district, and Zhuhai Sci-Tech Industrial Group, a state-backed industrial fund, with CICC Capital and Guotai Haitong Creative Investment, two of China's largest financial institutions, also participating. Existing backer Junshan Capital made a substantial follow-on commitment.

Steps to Understanding the Inference Chip Opportunity

• Recognize the architectural mismatch: Current GPUs compromise between memory bandwidth and compute density, optimizing for neither the attention mechanism nor the feedforward network in transformer models.
• Understand the market timing: Inference spending now rivals training spending as AI moves from experimental to operational deployment, creating demand for specialized hardware.
• Evaluate the execution risk: Fangqing's decoupled architecture is theoretically promising but unproven in silicon; the communication wall between specialized modules could eliminate efficiency gains.
• Consider the geopolitical context: US export controls on advanced semiconductors create strategic incentives for China to develop domestic AI chip capabilities independent of American suppliers.

Fangqing completed three funding rounds within roughly six months, a Pre-A, a Pre-A+, and now the A1, accumulating institutional conviction before the company has taped out a single chip or named a manufacturing partner. The milestone reflects investor confidence in both the architectural thesis and the founder's ability to execute on one of the most technically demanding challenges in semiconductor design.

"The next competition in intelligent computing will be systematic innovation at the architecture level," stated Liang Jun.

Liang Jun, CEO at Fangqing Technology

Whether Fangqing can bridge the communication wall and deliver on its architectural promise remains an open question. But the $1.5 billion valuation signals that the inference chip market has become the real battleground in AI hardware, and that investors believe specialized silicon designed for deployment, not training, will define the next generation of AI infrastructure.