CITIC Construction Investment: China's optical fiber export ratio has significantly increased, continuing to be optimistic about the AI computing power sector

CITIC Construction Investment Research Report points out that in February of this year, China exported 3,779.9 tons of optical fibers, totaling 790 million yuan, a year-on-year increase of 63.6% and 126.8%. Converted to kilometers, in February, China exported approximately 25.2 million core kilometers of optical fibers, accounting for about 65% of China’s monthly effective production of optical fibers. If we include the optical fibers in the optical cable exports, the export proportion of optical fibers is even higher. From the export amount, it is inferred that the impact of optical fiber price increases on performance is not expected to be significant in the first quarter. In summary, overseas markets are scrambling to purchase optical fibers produced in China, and Chinese optical fiber suppliers are in a “no worries about sales” state. Therefore, we believe the market does not need to be overly concerned about the optical fiber centralized procurement by domestic telecommunications operators. Overall, the demand for optical fibers is being driven by overseas telecommunications networks, AI, drones, and other factors, pushing prices to continue rising, placing the industry in a high prosperity cycle, and we continue to recommend the optical fiber sector. Google’s TurboQuant compression algorithm achieves near-lossless compression of AI inference memory, significantly reducing the cost of long context inference, with applications such as edge AI and AI video generation expected to benefit, and we remain optimistic about the AI industry chain.

The full text is as follows

CITIC Construction Investment: The proportion of China’s optical fiber exports has significantly increased, and we remain optimistic about the AI computing power sector

In February of this year, China exported 3,779.9 tons of optical fibers, totaling 790 million yuan, a year-on-year increase of 63.6% and 126.8%. Converted to kilometers, in February, China exported approximately 25.2 million core kilometers of optical fibers, accounting for about 65% of China’s monthly effective production of optical fibers. If we include the optical fibers in the optical cable exports, the export proportion of optical fibers is even higher. From the export amount, it is inferred that the impact of optical fiber price increases on performance is not expected to be significant in the first quarter. In summary, overseas markets are scrambling to purchase optical fibers produced in China, and Chinese optical fiber suppliers are in a “no worries about sales” state. Therefore, we believe the market does not need to be overly concerned about the optical fiber centralized procurement by domestic telecommunications operators. Overall, the demand for optical fibers is being driven by overseas telecommunications networks, AI, drones, and other factors, pushing prices to continue rising, placing the industry in a high prosperity cycle, and we continue to recommend the optical fiber sector.

Google’s TurboQuant compression algorithm achieves near-lossless compression of AI inference memory, significantly reducing the cost of long context inference, with applications such as edge AI and AI video generation expected to benefit, and we remain optimistic about the AI industry chain.

In February of this year, China exported 3,779.9 tons of optical fibers, totaling 790 million yuan, a year-on-year increase of 63.6% and 126.8%, a growth of 17.6 times compared to the 203.5 tons of optical fibers exported in February 2018 (the historical peak of domestic demand). If we convert based on approximately 0.15 kg of optical fiber per kilometer (including packaging, etc.), in February, China exported approximately 25.2 million core kilometers of optical fibers, accounting for about 65% of China’s monthly effective production of optical fibers. If we include the optical fibers in the optical cable exports, the export proportion of optical fibers is even higher. From the export amount, it is inferred that the impact of optical fiber price increases on performance is not expected to be significant in the first quarter. Overall, overseas markets are scrambling to purchase optical fibers produced in China, and Chinese optical fiber suppliers are in a “no worries about sales” state. Therefore, we believe the market does not need to be overly concerned about the optical fiber centralized procurement by domestic telecommunications operators.

From the perspective of countries exporting optical fibers in February, the top ten exporting countries are Côte d’Ivoire, Burkina Faso, Poland, the Philippines, Argentina, Russia, Nigeria, the United States, Panama, and Australia. Among them, the export volume from the three African countries has increased significantly, which is expected to be mainly related to local network construction demands. In addition, the demand from Russia is expected to be primarily for drones, while the demands from the United States, Australia, and the Philippines are expected to be related to AI. Overall, the demand for optical fibers is being driven by overseas telecommunications networks, AI, drones, and other factors, pushing prices to continue rising, placing the industry in a high prosperity cycle, and we continue to recommend the optical fiber sector.

The TurboQuant compression algorithm released by Google Research can reduce the memory usage of large language models (LLMs) while increasing operational speed and maintaining accuracy. TurboQuant can compress the “working memory” of AI runtime, that is, the key-value cache (KV cache), by at least 6 times. On H100 graphics cards, compared to the 32-bit baseline, the computation attention speed of 4-bit has surged 8 times, significantly lowering AI operating costs. Meanwhile, the most critical highlight of TurboQuant is: zero loss in precision, no need for fine-tuning, and no need for training data.

The optimization goal of TurboQuant is to reduce the volume of the key-value cache. The core of TurboQuant is a sophisticated two-stage process. The first stage: PolarQuant looks at the world from a different coordinate system. Traditional quantization operates in a Cartesian coordinate system (X, Y, Z axes), where the value range of each axis is not fixed and requires additional normalization parameters to align. The second stage: QJL eliminates residual errors using 1-bit. The second step of TurboQuant applies the Johnson-Lindenstrauss transformation to the residual errors from the first stage, compressing each error value into a sign bit: +1 or -1. Then, in conjunction with a special estimator—using high-precision query vectors and low-precision compressed keys for joint computation—only the last 1 bit is consumed to smooth out the systemic bias left from the first stage. These two steps allow TurboQuant to achieve near-lossless compression effects with just a total budget of 3 bits, with no additional overhead throughout the process.

Google has rigorously validated TurboQuant on five long context benchmark tests: LongBench, Needle In A Haystack, ZeroSCROLLS, RULER, and L-Eval, covering models like Gemma, Mistral, and Llama-3.1-8B-Instruct. In comprehensive tasks such as Q&A, code generation, and text summarization on LongBench, the TurboQuant with a 3-bit configuration outperformed baseline methods like KIVI across the board, even approaching the performance of full-precision models. At a 4x compression ratio, TurboQuant’s retrieval accuracy maintained at 104,000 tokens, which is completely consistent with full-precision models. In high-dimensional vector searches, TurboQuant outperformed the two leading methods, PQ and RabbiQ, on the GloVe dataset (200 dimensions), achieving the best recall rate. Google’s TurboQuant compression algorithm achieves near-lossless compression of AI inference memory, significantly reducing the cost of long context inference, with applications such as edge AI and AI video generation expected to benefit, and technological iteration driving continuous upgrades in the computing power industry chain.

On March 23, Liu Liehong, Director of the National Data Bureau, officially announced at the China Development High-Level Forum 2026 that the standard Chinese name for the core term of AI, “Token,” has been determined to be “词元”. Liu Liehong stated that “词元” is not only a value anchor in the intelligent era but also a settlement unit connecting technological supply and commercial demand, and revealed that a model enterprise set a record by earning more than the total revenue of 2025 in just 20 days. 100 billion, 100 trillion, 140 trillion, this is the three-level jump of China’s average daily token usage within two years.

The current telecommunications industry is in a dual dividend period driven by AI technology and supported by new infrastructure policies, with the computing power industry chain still being a core line with high prosperity. Computing power and chips are the core foundation for the development of the AI industry and also the current investment main line of high prosperity and high growth in the telecommunications industry, which is recommended for close attention.

Changes in the international environment have impacted the safety and stability of the supply chain, affecting the pace of related companies’ overseas expansion; tariff impacts exceeded expectations; the development of the AI industry has not met expectations, affecting the demand for companies in the cloud computing industry chain; intensified market competition has led to a rapid decline in gross profit margins; exchange rate fluctuations affect the foreign exchange gains and gross profit margins of export-oriented enterprises, including those in the ICT equipment and optical module/optical device sectors; the development of the digital economy and digital China construction has not met expectations; the development of telecommunications operators’ cloud computing businesses has not met expectations; operators’ capital expenditures have not met expectations; cloud vendors’ capital expenditures have not met expectations; the demand in the communication module and intelligent controller industries has not met expectations.

(Source: Financial Associated Press)