In the first quarter of 2026, China’s domestic AI chip industry delivered a remarkable scorecard. Homegrown chip companies led by Cambricon achieved dual breakthroughs in both revenue and profit, while Goldman Sachs issued starkly contrasting rating adjustments for Cambricon and Inspur Information on the same day — raising the target price for the former to RMB 2,406 (≈ USD 334.17) and downgrading the latter from Buy to Neutral. Does this simultaneous upgrade and downgrade signal the comprehensive rise of domestic AI chips, or the beginning of a structural industry upheaval?
Performance Explosion: The Logic Behind Cambricon’s Hundred-Billion Market Capitalization
In 2025, Cambricon achieved full-year profitability for the first time: annual revenue reached RMB 6.497 billion (≈ USD 902.36 million), up 453.21% year-on-year; net profit attributable to shareholders came in at RMB 2.059 billion (≈ USD 286.00 million), representing a historic swing from losses to profitability compared with the previous year. Entering the first quarter of 2026, this growth momentum accelerated further: single-quarter revenue reached RMB 2.885 billion (≈ USD 400.69 million), up 159.56% year-on-year; net profit was RMB 1.013 billion (≈ USD 140.69 million), up 185.04% year-on-year; and operating cash flow was RMB 834 million (≈ USD 115.83 million), turning positive on a quarterly basis for the first time since the company’s listing.
The core driver supporting this performance was the explosive growth of the cloud product line. In 2025, Cambricon’s cloud product line contributed revenue of RMB 6.477 billion (≈ USD 899.58 million), accounting for more than 99% of total revenue. At the customer level, leading internet enterprises such as ByteDance, Alibaba, and Tencent became its primary purchasers. According to company announcements, in December 2025 Tencent signed a framework contract with Cambricon for tens of thousands of MLU590 units for 2026, while ByteDance provided procurement expectations for hundreds of thousands of MLU580 units.
It is worth noting that Cambricon is not the only beneficiary in this space. According to the China Academy of Information and Communications Technology’s 2025 Cloud Computing White Paper, China’s public cloud IaaS market grew 32% year-on-year in 2025, with AI server procurement exceeding general-purpose server procurement for the first time, creating vast market space for domestically produced chips. Based on corporate earnings reports, Haiguang Information’s 2025 revenue from its Shensuan series DCU reached RMB 4.63 billion (≈ USD 643.06 million), up 187% year-on-year. Although Enflame Technology is not publicly listed, according to market research institution data, its 2025 cloud inference chip shipments exceeded 180,000 units with a market share of approximately 7%. Overall, against the backdrop of constrained supply chain capacity for Huawei’s Ascend, from the second half of 2025 to the first quarter of 2026, the combined share of public tender awards won by Cambricon, Haiguang, and Enflame in domestic AI chip public procurement rose from less than 15% in 2024 to approximately 38% (source: public tender statistics from the China Procurement and Bidding Network and estimates by Guosheng Securities Research Institute).
However, the market’s reaction to this scorecard was far from unanimous applause. Following the release of the full-year 2025 profitability data, Cambricon’s share price opened high but trended downward, with an intraday swing exceeding 15% — a stark contrast to the moment of capital frenzy last year. The reasons behind this may be more complex than the surface-level performance figures suggest.
Concerns Emerge: The Three Layers of Pressure Beneath the Spotlight
Beyond the earnings figures, multiple underlying concerns have surfaced in Cambricon’s fundamentals.
The first is the risk of in-house chip substitution by major customers. According to a February 2026 report by Semiconductor Industry Observer, ByteDance’s first AI inference chip, “ByteStone-V1,” was planned to receive its first samples in the first quarter of 2026. The chip is based on a 28nm process and is specifically optimized for recommendation systems and large language model inference. Alibaba’s subsidiary T-Head also launched its high-end AI chip “Zhenwu 810E” in January 2026, with a peak computing power reportedly reaching 256 TOPS. Once leading internet enterprises complete the volume production transition to their in-house chips, the revenue Cambricon derives from these core major customers faces significant uncertainty. According to Cambricon’s 2025 annual report disclosures, combined revenue from its top two customers accounted for more than 70% of total revenue, a proportion that has continued to climb over the past three years.
The second concern is persistently elevated inventory. As of the end of 2025, Cambricon’s inventory book value stood as high as RMB 4.944 billion (≈ USD 686.67 million), representing 36.79% of total assets. In the first quarter of 2026, according to Cambricon’s Q1 2026 earnings report, despite continued revenue growth, inventory remained at the elevated level of RMB 4.497 billion (≈ USD 624.58 million), and the company recorded an inventory impairment loss of RMB 246 million (≈ USD 34.17 million). In accordance with semiconductor industry convention, the lifecycle of AI chips is approximately 18 to 24 months, and inventory turnover days exceeding 180 days carry a significant impairment risk. A research report issued in March 2026 noted that Cambricon’s inventory turnover days had risen from 96 days at the end of 2024 to 214 days at the end of 2025. While stockpiling to some extent reflects an optimistic outlook on future demand, the rapid pace of product iteration in the chip industry means that once inventory builds up, its value deteriorates quickly.
The third concern is customer concentration. Over the past three years, the combined sales amount to Cambricon’s top five customers as a proportion of operating revenue has been 92.36%, 94.63%, and 88.66% respectively. This means the company’s performance is highly dependent on a small number of giants. By comparison, Haiguang Information, which is also engaged in the AI chip business as a listed company, had its top five customers accounting for 57.2% of revenue in 2025 — a relatively more dispersed customer structure. If the procurement strategies of core customers such as ByteDance, Alibaba, and Tencent shift or their in-house chips go online, Cambricon’s revenue will face severe volatility. Additionally, according to industry reports from March 2026, Tencent’s in-house “Zixiao” chip demonstrated inference efficiency comparable to Cambricon’s MLU590 in internal testing; if progress continues smoothly, it is expected to partially replace externally purchased chips by 2027.
Demand Scan: The Computing Power “Hunger” of ByteDance, Tencent, and Alibaba, and the Supply Bottleneck
Cambricon’s performance explosion is not an isolated case but a microcosm of the surge in demand across the entire domestic AI chip industry. According to an IDC report, China’s AI chip shipments grew 47% year-on-year in 2025, reaching 4 million units. Among these, the proportion of domestically produced chips rose from 30% in 2024 to 41%. At the same time, Gartner’s January 2026 forecast indicated that China’s AI chip market would surpass USD 18 billion in 2026, up 38% year-on-year, with the domestic production rate expected to further increase to 52%.
From the perspective of downstream demand, domestic AI inference computing power demand in 2026 is showing explosive growth, with the computing power gap continuing to widen compared with 2025. The core drivers come from two directions: first, overseas chip supply continues to be constrained — the U.S. Department of Commerce’s Bureau of Industry and Security further tightened export controls starting in January 2025, making it impossible to legally sell NVIDIA’s H20 and AMD’s MI300 series domestically; second, domestic AI chip supply capacity remains insufficient — according to estimates by the China Electronic Special Equipment Industry Association, the domestic advanced packaging (Chiplet, HBM integration) capacity shortfall in 2026 exceeds 30%, and the combination of multiple factors has further exacerbated the supply-demand imbalance.
Looking at the demand scale of major vendors specifically, according to reports from relevant research institutions in February 2026, ByteDance’s inference chip shortfall is approximately 1.3 million units, Alibaba’s inference chip shortfall is approaching 1 million units, and Tencent is accelerating domestic substitution with procurement volumes that are highly likely to reach a scale comparable to Alibaba and ByteDance. However, domestic chip manufacturers’ production capacity also faces bottlenecks. Huawei’s Ascend 910B, as shown in multiple domestic test reports, has inference performance of approximately 70% to 80% that of NVIDIA’s A100, but constrained by yield rates and capacity, its estimated total 2025 shipment volume was less than 500,000 units; the Ascend 910C has poor cost-effectiveness and scarce shipment volumes. On the Kunlunxin side, according to Baidu’s 2025 annual report disclosures, its AI chip capacity is primarily used for internal supply to Baidu Intelligent Cloud, with external sales accounting for less than 15%. Alibaba T-Head’s in-house chips are currently manufactured by TSMC (Nanjing), with a monthly capacity of only approximately 20,000 units. Additionally, a research institution’s in-depth report from December 2025 noted that ByteDance’s actual final order volume for Cambricon’s MLU580 did not reach the previously anticipated 800,000 units, amounting to only approximately 420,000 units, resulting in approximately half the capacity for that chip model sitting unused — this detail illustrates to some extent that both the supply and demand sides of the current domestic AI chip market still contain matching uncertainties, while also reflecting the fact that leading internet enterprises are strategically diversifying their chip suppliers in order to control and reduce single-source risk.
There is also a dimension that is easy to overlook: power consumption and cluster efficiency. According to internal test data from a Chinese cloud computing vendor, Cambricon’s MLU590 achieves a linear speedup ratio of approximately 0.78 when building a thousand-card cluster, while an NVIDIA A100 cluster at the same scale can achieve 0.85. This means that even when single-card computing power is comparable, there remains a certain efficiency gap in actual data center deployment, which is also the reason why cloud vendors need to invest additional effort in software stack optimization when promoting domestic substitution.
Landscape Reshaping: What Does Goldman Sachs’ “One Up, One Down” Signal?
On May 4, 2026, Goldman Sachs released two research reports on the same day covering Cambricon and Inspur Information. The core judgment in both reports points in the same direction: domestic AI chips are rapidly expanding their share of China’s AI server market, and their competitiveness relative to chip giants such as NVIDIA is improving rapidly.
This judgment about the trend, however, leads to two entirely different conclusions. For Cambricon, as a manufacturer of domestic AI chips, domestic substitution represents a direct business increment; Goldman Sachs therefore raised its target price to RMB 2,406 (≈ USD 334.17), representing approximately 28% upside from the closing price on the day before the report was published, and significantly raised its profit forecasts for 2026 to 2030, projecting that Cambricon’s full-year 2026 revenue will reach RMB 9.5 billion (≈ USD 1.32 billion) and net profit RMB 3.2 billion (≈ USD 444.44 million). For Inspur Information, the situation is more complex: as an assembler of AI servers, when the chips inside servers are switched from NVIDIA GPUs to lower-cost domestic chips, the selling price of the servers falls accordingly and the profit margin is compressed. Goldman Sachs projects that the average selling price of Inspur’s AI servers will decline from approximately USD 120,000 (≈ RMB 864,000) in 2025 to approximately USD 72,000 (≈ RMB 518,400) in 2028, a cumulative decline of 40%.
At the same time, the procurement concentration of cloud service providers is also reshaping bargaining power. Industry research reports show that ByteDance, Alibaba, and Tencent together accounted for 67% of the total AI server procurement volume in 2025. This class of large customers has strong bargaining power and typically places orders through annual framework tenders with multi-vendor price comparisons, leaving thinner profit margins for server ODM/OEM manufacturers. Inspur’s gross margin has already declined from the 10% to 15% range of 2015 to 2023, to 6.8% in 2024 and 4.9% in 2025. Goldman Sachs accordingly downgraded Inspur Information from Buy to Neutral, and simultaneously cut its earnings per share forecasts for 2026 to 2028 by an average of 12%.
What this set of rating adjustments reveals is not simply the rise and fall of two companies, but a structural transfer in value distribution across the entire AI computing power supply chain. As computing power demand grows explosively, the segments of the supply chain that possess core chip design capabilities and proprietary technical barriers are capturing a greater share of incremental value; while the assembly segment — in a position of “doing the heavy lifting” — continues to expand in business volume yet sees its profit margin continuously compressed. Also worth noting is that Goldman Sachs also lowered its expectations for NVIDIA’s revenue in the Chinese market, projecting that its China revenue for 2026 will decline 45% year-on-year, with domestic substitution proceeding faster than previously modeled assumptions.
According to Morgan Stanley’s March 2026 report calculations, China’s AI computing chip market will reach USD 67 billion by 2030, with a compound annual growth rate of as high as 23% from 2024 to 2030; the domestic localization rate of AI chips will accelerate dramatically from 41% in 2025 to 86% by 2030. In this race for computing power autonomy, Cambricon — as the second-largest domestic third-party AI chip supplier after Huawei’s Ascend — occupies an advantageous position. But the time window the market has left for it may not be as wide as imagined: the maturation cycle of customers’ in-house chips, the speed at which competitors ramp up production capacity, and the continuously tightening U.S. restrictions on advanced process equipment (in March 2026, BIS further restricted immersion DUV exports to China) will all jointly determine the final form of this competitive track.
Looking back at the overall picture of China’s domestic AI chip industry in the first quarter of 2026, one sees a market that is rapidly expanding yet whose structure has not been settled. Goldman Sachs’s optimism on Cambricon and pessimism on Inspur Information are essentially a repricing of the logic of value distribution across the supply chain. In this structural upheaval, for companies with core chip design capabilities, the test lies not only in whether they can seize the demand window to achieve revenue growth, but also in whether they can hold their ground under the dual pressure of customer in-house substitution and peer competition for market share. For the assembly segments downstream of the supply chain, how to maintain profit margins in the squeeze between increasing bargaining power from large customers and declining end-product prices is equally an unavoidable practical challenge. As multiple semiconductor industry analysts reached consensus at the first-quarter 2026 strategy conference: the narrative of domestic AI chips has moved out of the laboratory and the press release and into the real-world contest of production capacity and order volumes, and the decisive factor in this contest will depend on who can simultaneously solve the impossible triangle of performance, scale, and cost within the next 18 months. The story of domestic AI chips is still being written.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.