Quantum computing is waiting for its own "NVIDIA". - 36 Kr

Quantum computing is waiting for its own "NVIDIA".36氪的朋友们2026-01-28 12:08If a general-purpose quantum computer is an all-powerful CPU, what Bose Quantum wants to do is the GPU in the quantum era.At the beginning of 2024, when formulating the annual OKR, the management of BosonQ Psi set a seemingly rather "conservative" goal: to sell at least one quantum computer. In the context of the industry at that time, quantum computers were still sophisticated scientific research instruments entangled with cables and lying in laboratories. Even in the eyes of many peers, it was already not easy to meet the laboratory indicators, and it seemed too early to talk about "selling products". However, the subsequent market feedback caught the team off guard. The number of orders was several times higher than expected, and the "hand - made" prototypes in the laboratory could no longer meet the real commercial needs. At the end of 2025, this startup invested in and established a factory in Shenzhen, which is also the first large - scale dedicated optical quantum computer manufacturing factory in China. This production line hastily extended from the laboratory to the factory reflects a possibility that has long been ignored by the mainstream narrative: in the NISQ (Noisy Intermediate - Scale Quantum) era before the birth of a perfect general - purpose quantum computer, the not - so - perfect dedicated quantum computer may have already opened the door to commercialization. The Non - Consensus Path Among Chinese quantum computing enterprises, BosonQ Psi is an atypical example. Origin Quantum is backed by the University of Science and Technology of China, Turing Quantum relies on Shanghai Jiao Tong University, and Guodun Quantum has state - owned capital background. This is almost the standard configuration in the industry. However, BosonQ Psi is a pure private enterprise founded by Wen Kai, who graduated from Stanford University, and Ma Yin, who has worked in the aerospace field for several years. The early development process of BosonQ Psi is also very different from the industry mainstream. At the beginning of its establishment in 2020, when most peers were still struggling to choose between technical paths such as superconducting, ion trap, or optical quantum, BosonQ Psi made a more fundamental "anti - consensus" decision: at the crossroads of general - purpose and dedicated, it bet on the latter. Ma Yin admitted in an exclusive interview with Jiemian News that the logic behind this decision was extremely simple - "there is an opportunity to use it immediately". In the field of quantum computing, developing a practical general - purpose quantum computer is regarded as the mainstream ultimate goal. It attempts to construct physical qubits by manipulating natural elementary particles (such as Josephson junctions in the superconducting route and ions in the ion trap route) to achieve quantum computing. The number of physical quantum bits is the core indicator to measure the technical level. The problem is that quantum bits are extremely error - prone. Often, a large number of physical bits are needed to correct a stable "logical bit" to achieve relatively consistent and accurate calculations. There is a consensus in the industry that there is at least a hard threshold of 10 to 20 years for application implementation. Even industry giants are struggling - IBM has postponed its plan to achieve 2000 quantum bits to 2033; although Google has a chip with 105 quantum bits, there is still no definite timetable for reaching its set milestone of 1000 bits. "Quantum computers are currently still in the stage of building devices," a technician from Guoyi Quantum told Jiemian News. "Currently, the number of bits that domestic players can achieve may be around ten, and it's also very difficult to reach a fidelity of 99% - that is, there is a 1% probability of error every time a two - bit gate is performed." The real feedback Ma Yin got from customer data is that at least 1000 calculated bits after error correction are needed to form a quantum advantage over classical computers. Facing the long technical climbing period, BosonQ Psi chose a more vertical path - the coherent optical quantum computer. As a dedicated optical quantum computer, it does not pursue the construction of complex logic gates. Instead, it uses the physical properties of optical quanta to specifically solve problems such as combinatorial optimization. The key difference is that a dedicated quantum computer does not need to construct complex logic gates, thus bypassing the error - correction problem faced by general - purpose quantum computers. A technician engaged in general - purpose quantum computing research told Jiemian News that the parameters of a general - purpose quantum computer can be changed arbitrarily, and different functions can be achieved by changing the software, but the technical difficulty is extremely high; the devices of a dedicated quantum computer are relatively fixed and can only target one type of problem, but they do not require very precise operations. Instead, they exchange degrees of freedom for accuracy and other capabilities. Ma Yin used a more intuitive analogy to summarize this difference: if a general - purpose quantum computer is an all - powerful CPU, what BosonQ Psi wants to do is the GPU in the quantum era. The GPU achieves large - scale parallel computing by integrating thousands of computing cores. Although its control logic is relatively simple, it is extremely good at handling highly parallel tasks such as graphic rendering. The computing advantage of the coherent optical quantum computer is based on the laws of statistical physics. When solving complex combinatorial optimization problems, classical computers often need to traverse all possible solutions, and the computing power consumption and time cost increase exponentially. The coherent optical quantum computer constructs a heterogeneous optical path system architecture of "spatial optical path + fiber optical path", which automatically collapses to the lowest energy state of the Hamiltonian during the evolution process; this physical state exactly corresponds to the "optimal solution" in the mathematical problem, allowing the answer to emerge automatically. Learning from NVIDIA Avoiding the problems of error correction and noise does not mean that there are no thresholds on the path of dedicated quantum computing. The hardware development of quantum computers is extremely difficult. Different from semiconductor chips that pursue nanometer - scale processes, the difficulty of optical quantum chips lies not in process accuracy but more like a "semi - scientific, semi - engineering" problem. Lithography, etching, parameter adjustment... There is a lack of mature industrial standards for each process, and one can only explore step by step. Fortunately, the industrial foundation of optical communication and optical module chips can be used for reference. Beyond the hardware, the ecosystem of quantum computing in previous years was almost a wasteland: there was no compiler, no framework, and no developers to develop algorithms. Ma Yin thought of NVIDIA as a reference. The rise of the GPU was not only due to the hardware but also relied on the CUDA compiler, the Transformer framework, and thousands of algorithm developers in the wave of artificial intelligence. "The two - way interaction between hardware and algorithms," he repeatedly emphasized this sentence. "Talking only from the hardware perspective can only be a one - sided view." BosonQ Psi's response is to replicate NVIDIA's structure. It builds the Kaiwu SDK compiler on top of the hardware, which is comparable to CUDA; above that is the quantum AI framework based on the Boltzmann machine; and the top layer consists of application algorithms for various industries. The core design logic of the compiler is to greatly lower the usage threshold for developers, allowing AI developers to log in to PyTorch and develop quantum algorithms based on Python according to the standard framework without learning the technical principles of quantum physics, so as to call the real - machine computing power of quantum computers. "The developer's interface doesn't even change," Ma Yin said. This is the core of his so - called "migratory ecosystem". At this stage, the quantum computing ecosystem is unable to train too many new people, so it can only strive to get more existing AI developers involved. According to Ma Yin, the cooperation with customers is essentially based on the successful implementation of algorithms - because there is a demand for algorithms, there is hardware deployment. It is worth noting that at this stage, neither general - purpose nor dedicated quantum computers can operate independently. Dedicated quantum computers are good at solving combinatorial optimization problems, but a complete computing task still requires the participation of classical computers. Ma Yin calls it "collaborative cooperation": first, the task is decomposed at the upper level. Some steps are suitable for classical computers, and others are suitable for quantum computers. "There may also be alternating calculations in the middle." This is actually also the mainstream choice in the industry. NVIDIA's previously released CUDA - Q platform allows the collaborative use of GPU, CPU, and QPU resources for computing in a single quantum program. The NVQLink released in October 2025 is an interconnection technology designed to connect quantum processors with the AI supercomputers they need to operate effectively. Schematic diagram of NVIDIA's NVQLink technology. Image source: NVIDIA's official website Commercialization Trials Although the ecosystem is not yet fully formed, commercialization attempts have already begun. Supercomputing centers have become the first group of customers that quantum computing enterprises are vying for. Quantum annealing company D - Wave once sold a dedicated quantum computer to the Jülich Supercomputing Centre in Germany for $12.64 million, which was regarded as a landmark event for the commercialization of the dedicated route. And several practitioners in the field of general - purpose quantum computing told Jiemian News that they have also made many deliveries and achievements in the path of local supercomputing. BosonQ Psi targeted the same market. In 2025, it won the bid for the Chengdu Supercomputing Center project and deployed a self - developed coherent optical quantum computer in the supercomputing cluster, becoming the first dedicated quantum computer deployed in a national supercomputing center in China. Model of a 1000 - qubit coherent optical quantum computer (1:10). Image source: Photographed by a Jiemian News reporter Ma Yin revealed that deployment plans for supercomputing centers in other regions are also on the agenda. However, there are different opinions within the industry about how much practical value the cooperation with supercomputing centers can bring. A general - purpose quantum computing enterprise also delivered a quantum computer to a local supercomputing center and built a hybrid quantum - supercomputing cloud platform. Its technician told Jiemian News that "overall, it still meets the city government's expectations for the supercomputing center", but also admitted that there is still a long way to go before real commercialization, and it is more of a "last - resort choice" at this stage when real application scenarios cannot be found. The cloud platform is another way. The quantum cloud platform is similar to buying AI computing power through cloud - calling servers. It is carried out through public cloud platforms such as China Mobile Cloud, Alibaba Cloud, and Huawei Cloud, as well as self - developed cloud platforms. Consumers register and apply for accounts online and are charged according to the computing power and computing tasks they call each time. According to BosonQ Psi's statistics, more than 39% of the users on its platform are from the biomedical field, and it has reached cooperation with institutions such as the Guangzhou National Laboratory, Jingtai Technology, and BGI. However, a considerable proportion of users are from universities and use it for free. This path has also given rise to other quantum enterprises focusing on the application layer. For example, Keda Guochuang, which focuses on the quantum computing cloud platform, does not manufacture quantum computers itself. Instead, it cooperates with hardware manufacturers such as Jiuzhang Quantum and Guodun Quantum to provide an operating environment for classical - quantum hybrid algorithms for developers and research institutions and opens algorithm application interfaces to government and enterprise customers. "The whole industry is still in its very early stage," a staff member of Keda Guochuang admitted in a previous interview with Jiemian News. Their strategy is to jointly build quantum software with industry customers and explore application scenarios in fields such as transportation and communication. Whether it is general - purpose or dedicated, the profit dilemma is a common reality faced by the industry. A staff member of Guizhen Chip, an integrated optical quantum chip enterprise, told Jiemian News that, as far as he knows, most domestic quantum computing enterprises have not achieved profitability at present. "Just in terms of the whole machine, we only provide services and create some cases, which cannot solve the company's commercialization problem," he revealed. Most enterprises still rely on financing to survive. This division - of - labor model reflects the reality of the quantum computing industry: hardware manufacturers focus on improving the scale and stability of quantum bits, while platform manufacturers try to lower the usage threshold and connect the supply and demand sides. During the window period when general - purpose quantum computers are not yet mature, each layer of the ecosystem is exploring the possibility of commercialization on its own. BosonQ Psi has not stopped at dedicated quantum computing. Public information shows that it completed its A++ round of financing in October 2025, and the funds will continue to be used for the research and development of general - purpose optical quantum computers. Ma Yin revealed that the company will start with the research and development of general - purpose optical quantum computing chips and then gradually move towards the whole machine. "In 2026, the company will gradually release the roadmap for its products." The international competition is also fierce. In the dedicated track, Japan's NTT company has long claimed to have reached a scale of 100,000 quantum bits in the laboratory stage; D - Wave, which originally focused on quantum annealing, recently released the world's first scalable on - chip cryogenic control technology for gate - based quantum computing at CES 2026, which also marks its strategic transformation into general - purpose quantum computing. The deeper gap lies in the ecosystem. According to Ma Yin, there are about 50 quantum computing whole - machine manufacturers in the United States, and there are five or six hundred algorithm companies downstream. It is a typical "garage culture", and hundreds of VCs are investing. In contrast, there are no more than 10 whole - machine manufacturers in China, and there are "almost no" downstream innovative enterprises. The industry mainly relies on state - owned enterprises such as China Mobile and China Electronics Technology Group. "To be honest, internationally, the best in the whole industry are definitely IBM and Google," a person from China Telecom Quantum told Jiemian News. As one of the few domestic enterprises operating a quantum cloud platform relying on a state - owned enterprise background, China Telecom Quantum has currently connected multiple superconducting quantum computers, including the "Tianyan - 287". The person believes that the advantage of domestic enterprises lies more in relying on group support and "expanding quickly", but they are still in a state of catching up in terms of core technology. "Releasing a product is a step ahead, but then it's about who can build the ecosystem faster and attract more developers," Ma Yin said. "In the next five years, it should be a period of vigorous growth for the downstream." For quantum computing enterprises, finding their own areas of expertise may only be the starting point of commercialization. How far this path can go and whether any enterprise can truly grow into the "NVIDIA" of the quantum era still remains to be tested by the market and time. This article is from "Jiemian News", author: Zhou Mo, editor: Wen Shuqi. Republished by 36Kr with permission.