AI News AI资讯 2d ago Updated 1d ago 更新于 1天前 52

Liangkun Technology Secures Hundreds of Millions in Angel Round Funding, AI4S Needs Quantum-Level Precision Data | 36Kr Exclusive 量坤科技获数亿元天使轮融资,AI4S急需量子级精度数据 | 36氪独家

Another "water seller" has entered the quantum computing track, but this time it's not selling shovels— it's selling an operating system. Quantenkun has just completed an angel round of several hundred million RMB. The founder, Lü Dingshun, has a clear logical chain: the model capabilities of AI for Science (AI4S) are bottlenecked by data precision, while quantum computing is inherently a high-precision simulator of the microscopic world, capable of providing "quantum-level" clean data for AI. I 量子计算赛道又多了一个“卖水人”,但这次卖的不是铲子,是操作系统。量坤科技刚拿完数亿人民币天使轮,创始人吕定顺的逻辑链条很清晰:AI for Science(AI4S)的模型能力上限被数据精度卡了脖子,而量子计算天然是微观世界的高精度模拟器,能为AI提供“量子级”的干净数据。听起来完美,但这条“硬件不足,软件先行”的路,走得通吗?

75
Hot 热度
80
Quality 质量
70
Impact 影响力

Analysis 深度分析

Another "water seller" has entered the quantum computing track, but this time it's not selling shovels— it's selling an operating system. Quantenkun has just completed an angel round of several hundred million RMB. The founder, Lü Dingshun, has a clear logical chain: the model capabilities of AI for Science (AI4S) are bottlenecked by data precision, while quantum computing is inherently a high-precision simulator of the microscopic world, capable of providing "quantum-level" clean data for AI. It sounds perfect, but can this path of "hardware lags, software leads" actually work?

Lü Dingshun's resume is indeed impressive—a PhD from Tsinghua University's first cohort in quantum computing, with stints at Huawei and ByteDance, and even leading a team to challenge Google's "quantum supremacy" narrative on a GPU cluster of hundreds of cards. This operation itself is quite "exciting," but he claims not to be overly thrilled. He is clear-headed: 53 quantum bits can still be勉强 chased by classical computers, but when the scale reaches 100 or 200 bits, classical computers will fall behind. The real excitement lies in using this computational power to solve "real-world big problems" like high-temperature superconductivity and new material design. During his time at Huawei, he used quantum computers as a hammer looking for nails; at ByteDance, his approach shifted to combining AI, quantum computing, and classical algorithms, all with one goal: solving scientific problems more accurately and efficiently.

Thus, Quantenkun was born, positioned as the "operating system for quantum computing power." They don't build hardware, nor do they directly serve end-user researchers; instead, they occupy the middle ground—packaging quantum algorithms, AI models, and industry workflows into "scientific intelligence units." This idea is both clever and pragmatic. Currently, quantum hardware is still in the "scarce resource" phase; a slight inefficiency in algorithms could mean the difference between "runs" and "doesn't run"—a binary outcome. What Lü Dingshun's team does is squeeze as much value as possible from the limited quantum bits, using algorithmic optimization to compensate for hardware shortcomings. They even convert chemical problems requiring tens of thousands of bits into solvable problems using just 20 quantum bits through quantum embedding methods. This "leveraging the small to achieve the great" approach truly embodies the value at the software level.

However, stepping back, the risks here cannot be ignored. The advancement of software ultimately depends on progress in hardware. If the scaling of quantum hardware doesn't meet expectations, or if there's a disruptive shift in technological routes (e.g., among ion traps, superconducting qubits, or topological qubits), could the algorithmic moat built in advance suddenly find itself facing no battlefield on the other side? Lü Dingshun says the algorithmic barrier lies in "design and optimization," making the path as short as possible. But if future hardware provides enough raw computational power, will such extreme optimization still be as critical? It's like designing the optimal route for a horse-drawn carriage, only to have a railway built to your doorstep the next day.

Another intriguing point is the founder's mention that "the team is a reflection of the founder's inner cognition"—they recruit people with "strong ambition." This is good, but cross-disciplinary teams (quantum, AI, high-performance computing) inherently have high coordination costs. Aligning top quantum physicists, algorithm engineers, and industry experts toward the somewhat vague goal of "solving real-world big problems" likely poses management challenges as formidable as the technical hurdles.

Returning to the investment logic, institutions like Inno Angel and Baidu Ventures are betting on the long-term narrative that "AI4S needs quantum computing." What Quantenkun offers is a solution that can generate value even during the current immature phase of hardware—using a hybrid heterogeneous computing platform to orchestrate classical computing, AI, and quantum computing, prioritizing "bottleneck" fields like strongly correlated materials and chemical simulations. If they can continuously demonstrate advantages of a few decimal places over traditional methods, as they have with high-temperature superconductivity models, this path will only widen.

Ultimately, Quantenkun is like a competitor in a vast but fog-shrouded arena, choosing not to directly participate in the race to "build the most advanced machine," but instead to first set the rules and create the training software. The bet is that once the fog lifts, all competitors will need the "navigation system" they provide. The idea is exciting, but it also tests patience and resolve. After all, in the long-termist track of quantum computing, excitement is important, but surviving and proving one's indispensability is even more so.

量子计算赛道又多了一个“卖水人”,但这次卖的不是铲子,是操作系统。量坤科技刚拿完数亿人民币天使轮,创始人吕定顺的逻辑链条很清晰:AI for Science(AI4S)的模型能力上限被数据精度卡了脖子,而量子计算天然是微观世界的高精度模拟器,能为AI提供“量子级”的干净数据。听起来完美,但这条“硬件不足,软件先行”的路,走得通吗?

吕定顺的履历确实硬核——清华第一批量子计算博士,在华为、字节都干过,还带团队在百卡GPU集群上把谷歌“量子霸权”的叙事给扬了。这操作本身就挺“exciting”,但他自己却说不怎么兴奋。因为他清醒:53个量子比特还能用经典计算机勉强追一追,等到了100个、200个比特的规模,经典计算机就跟不上了。真正的兴奋点,在于用这些算力去解决高温超导、新材料设计这种“真实世界的大问题”。在华为时,他拿量子计算机当锤子找钉子;到了字节,思路变成了AI、量子计算、经典算法一起上,目标只有一个:把科学问题算得更准、更快。

于是量坤科技诞生了,定位是“量子算力的操作系统”。他们不造硬件,也不直接面对最终科研用户,而是卡在中间——把量子算法、AI模型和行业流程打包成一个个“科学智能体”。这个想法很妙,也很务实。因为当下量子硬件还处于“稀缺资源”阶段,算法效率差一点,可能直接从“能跑”变成“跑不动”,这是0和1的区别。而吕定顺团队做的,就是尽可能把有限的量子比特“榨干”,用算法优化来弥补硬件的不足。他们甚至把上万比特的化学问题,通过量子嵌入方法,转换成只需要20个量子比特就能求解。这种“四两拨千斤”的思路,是软件层价值的真正体现。

但冷静下来想,这里的风险也不容忽视。软件的先进性,终究要寄生于硬件的进步。如果量子硬件的scaling速度不及预期,或者技术路线发生颠覆性变更(比如离子阱、超导、拓扑等路线之争),那么提前构筑的算法护城河,会不会突然发现河对岸根本没有战场?吕定顺说算法壁垒在于“设计和改造”,让路径足够短。可如果未来的硬件算力足够暴力,这种极致优化是否还会那么关键?这就像为马车设计了最优路线图,结果第二天铁路就铺到了家门口。

另一点值得玩味的是,创始人提到“团队是创始人内心认知的映射”,他们招的是“心气儿要足”的人。这很好,但跨学科融合的团队(量子、AI、高性能计算)本身协调成本就极高。让一群顶尖的量子物理学家、算法工程师和行业专家拧成一股绳,朝着“解决真实世界大问题”这个有点模糊的目标前进,管理上的挑战恐怕不亚于技术攻关。

回到投资逻辑,英诺、百度风投等机构押注的,是“AI4S需要量子计算”这个长期叙事。量坤科技提供的,是一个在当下硬件不成熟期也能产生价值的解决方案——用混合异构计算平台,把经典算力、AI和量子算力调度起来,优先服务强关联材料、化学模拟等“卡脖子”领域。如果他们真能像在高温超导模型上那样,持续做出比传统方法小数点后几位的优势,这条路就越走越宽。

说到底,量坤科技像是在一个巨大但雾气弥漫的赛场里,选择不直接参与“造最先进机器”的比赛,而是先做赛事规则和训练软件的制定者。赌的是雾散之后,所有选手都需要他们提供的“导航系统”。这想法性感,但也考验耐心和定力。毕竟在量子计算这个长期主义的赛道上,兴奋感很重要,但活下来,并证明自己不可或缺,更重要。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

科学研究 科学研究 融资 融资 大模型 大模型
Share: 分享到: