Tianshu Zhixin's Inference Card Margins Decline — Is the Price-for-Volume Trade-Off Worth It?

Stock Price Plunges on Earnings Release as Market Expectations Fall Short

Tianshu Zhixin, which listed on the Hong Kong Stock Exchange just over two months ago, recently delivered its 2025 annual report card. The filing showed full-year revenue of 1.034 billion yuan, up 91.6% year over year. Yet this seemingly impressive performance failed to win over capital markets — Tianshu Zhixin's stock price plummeted 10 percentage points on the day the results were announced. Prior to that, shares had already pulled back sharply from a high of HK$342, with market sentiment under sustained pressure.

Lin Yun, a primary-market investor, was blunt: "This performance didn't meet people's psychological expectations." Comparing growth rates of several other domestic GPU companies over the past two years, the market had set a revenue floor expectation of 1.2 billion yuan for Tianshu Zhixin. The nearly 200-million-yuan gap between actual and expected revenue became the pivotal factor shaking investor confidence.

Inference Card Gross Margin Drops Over 7%: The Cost of Cutting Prices for Volume

More concerning to the market than the revenue miss was the notable decline in Tianshu Zhixin's inference card business gross margin. Financial data showed that gross margins on its inference card products fell by more than 7 percentage points year over year, pointing directly to the company's ongoing price-for-volume strategy.

Against the backdrop of increasingly fierce competition in China's domestic GPU space, price cuts have become virtually a mandatory move for startups looking to crack open the market. Tianshu Zhixin's logic in trading margin for shipment volume and broader customer coverage is not hard to understand — during the window of exploding inference demand, capturing customer mindshare and supply chain positioning matters more than short-term profit.

But the risks of this strategy are equally apparent. Sustained margin erosion means the company needs ever-larger shipment volumes to maintain profit levels. If volume ramp-up falls short of expectations, it could face a double whammy of declining both prices and volumes. More critically, cutting prices is easy while raising them is hard — once customers form low-price expectations, the pricing headroom for subsequent product iterations gets further compressed.

From an industry cross-comparison, players across the domestic GPU landscape broadly face similar pricing dilemmas. Cambricon achieved its scale breakthrough through major clients like ByteDance and Alibaba, backed by the bargaining power equilibrium that large orders bring. The core challenge Tianshu Zhixin currently faces is precisely the lack of such "anchor" orders.

Fragmented Orders: A Collective Pain Point for Domestic Chip Startups

"The orders are too fragmented." Lin Yun hit the nail on the head when summarizing why Tianshu Zhixin's revenue fell short of expectations.

Just as Cambricon's revenue growth story is inseparable from large-scale purchases by ByteDance and Alibaba, internet giants are the critical battleground for domestic chips seeking scale. But the lack of large orders is precisely the collective pain point of domestic chip startups today, and Tianshu Zhixin is no exception.

A fragmented customer structure brings multiple hidden risks: first, elevated sales costs, as every small order requires significant technical support and adaptation resources; second, greater difficulty in revenue forecasting, given the volatility in small customers' procurement cycles and renewal rates; third, a weak brand endorsement effect — without benchmark clients to vouch for the product, it is difficult to create a demonstration effect in the industry.

However, this situation may be turning a corner. According to industry insiders, Tianshu Zhixin began entering the supply chain systems of major tech companies starting in 2025. By 2026, leading internet companies had initiated product testing, with some clients reportedly placing orders in the hundreds of millions of yuan. If this information proves accurate, Tianshu Zhixin's revenue structure could undergo a qualitative transformation within the next one to two years.

The Training-Inference Separation Approach: Can a Three-Year-Old Bet Pay Off?

Notably, Tianshu Zhixin established its "training-inference separation" product roadmap as early as 2022, designing dedicated chips for training and inference scenarios respectively rather than pursuing a unified training-inference solution.

This decision was quite controversial at the time. The prevailing industry view then was that unified training-inference chips offered greater market versatility and simplified customer purchasing decisions. But as demand for large-model inference surged explosively from 2024 onward, the performance requirements, power consumption demands, and cost sensitivity of inference scenarios diverged significantly from training scenarios, and the foresight of the training-inference separation approach gradually gained market recognition.

From a technical standpoint, inference-specific chips hold inherent advantages in cost per unit of computing power and energy efficiency. As large models move from laboratories to large-scale commercial deployment, compute consumption on the inference side will be several times — even tens of times — greater than on the training side. This is precisely Tianshu Zhixin's window of opportunity.

However, having the right roadmap does not guarantee commercial success. Competition in the inference card market is equally intense, with NVIDIA's inference product line continuing to push downmarket and Huawei Ascend accelerating its inference ecosystem buildout. Whether Tianshu Zhixin can leverage its first-mover advantage and differentiated positioning to gain a firm foothold in the inference market remains to be seen.

Valuation Trough or Value Trap?

Despite the short-term earnings disappointment, some analysts hold a different view. Analyst Ethan noted that capital markets believe Tianshu Zhixin's current price-to-sales valuation sits below the median for the domestic GPU sector, representing a relatively undervalued position. He expects the company's 2026 revenue and order book to both exceed market expectations.

The logic behind this judgment rests on three pillars: first, the landing of major client orders will significantly improve revenue structure and growth trajectory; second, the inference market boom will deliver structural dividends to companies focused on inference chips; third, as shipment volumes scale up, chip manufacturing costs should be amortized, creating room for margin recovery.

But other investors have sounded warnings. Continued margin erosion under the price-for-volume strategy could put pressure on the company's cash flow in the near term. As a recently listed company, Tianshu Zhixin needs to find the balance between land-grabbing expansion and healthy operations.

Outlook: 2026 Is the Critical Validation Year

Taking a comprehensive view, Tianshu Zhixin stands at a critical strategic crossroads. The short-term pain of price-for-volume is intertwined with the medium-term upside of major client orders, while the training-inference separation technical roadmap is being validated by industry trends.

2026 will be Tianshu Zhixin's pivotal year of validation. If major client orders land at scale on schedule and inference card shipments achieve breakthrough growth, then the current margin decline will be proven a wise "retreat to advance" maneuver. Conversely, if progress in securing major clients falls short of expectations, continued price cutting could drag the company into a vicious cycle of low margins.

For the domestic GPU space at large, Tianshu Zhixin's case also poses a thought-provoking question: during the technology catch-up phase, should startups prioritize profit quality, or seize market share at all costs? The answer to that question may have to wait until the inference era truly arrives.