Inference startup Infinity raises $15M from Touring Capital, OpenAI and Anthropic researchers
Inference startup Infinity raises $15M from Touring Capital, OpenAI and Anthropic researchers
AI infrastructure company Infinity announced a $15 million raise at a $100 million valuation on Monday from investors including Touring Capital, Principal VC, and researchers from companies such as OpenAI and Anthropic. AI 基础设施公司 Infinity 周一宣布完成 1500 万美元融资，估值达到 1 亿美元。本轮投资方包括 Touring Capital、Principal VC，以及来自 OpenAI 和 Anthropic 等公司的研究人员。
The startup is building software to make it easier for AI chips to run AI models. One big reason Nvidia became the top player is not just its high-performance chips, but also its CUDA software (Compute Unified Device Architecture), which allows its GPUs (originally designed to run graphics) to act as general-purpose processing CPUs. 这家初创公司正在开发软件，旨在让 AI 芯片运行 AI 模型变得更加容易。英伟达之所以能成为行业领头羊，一个重要原因不仅在于其高性能芯片，还在于其 CUDA（统一计算设备架构）软件。该软件使原本为图形处理设计的 GPU 能够充当通用处理 CPU 使用。
The largest AI development frameworks PyTorch and TensorFlow have been built on top of CUDA. This allows developers to write their apps in popular languages like Python, use those major AI frameworks and their apps will, by default, run on Nvidia chips. 目前最大的 AI 开发框架 PyTorch 和 TensorFlow 都是基于 CUDA 构建的。这使得开发人员能够使用 Python 等流行语言编写应用程序，并利用这些主流 AI 框架，从而使他们的应用默认就能在英伟达芯片上运行。
Most of these app-level startups wouldn’t have the resources or know-how to write their own kernels — the low-level software that operates chips — and port their apps to other AI chips. So Infinity is trying to build CUDA-alternative kernel software that works with any type of chip, like SRAM, GPUs, phone chips, and Systolic Arrays. 大多数应用层初创公司并没有资源或技术能力去编写自己的内核（即操作芯片的底层软件），并将应用移植到其他 AI 芯片上。因此，Infinity 正试图构建一种 CUDA 替代型内核软件，使其能够适配任何类型的芯片，如 SRAM、GPU、手机芯片和脉动阵列（Systolic Arrays）。
Infinity is part of a new wave of startups that are attempting, product by product, to chip away at Nvidia’s market dominance. Infinity is attempting to build a universal inference library to run on all chips, allowing these chips to automate replicating state-of-the-art research results. Infinity 是新一波初创公司中的一员，它们正试图通过一个个产品，逐步削弱英伟达的市场主导地位。Infinity 旨在构建一个通用的推理库，使其能在所有芯片上运行，从而让这些芯片能够自动复制最前沿的研究成果。
Infinity was launched last year by Jeremy Nixon, once a researcher at Google Brain and creator of the hacker network community AGI House. Nixon told TechCrunch he decided to launch this company because he was obsessed with the idea of “automated invention” — the belief that “AI systems can actually be a meta technology.” Infinity 由 Jeremy Nixon 于去年创立，他曾是 Google Brain 的研究员，也是黑客网络社区 AGI House 的创始人。Nixon 告诉 TechCrunch，他决定创办这家公司是因为他痴迷于“自动化发明”的理念，即相信“AI 系统实际上可以成为一种元技术”。
He himself had invented a machine learning algorithm called Omega, he said, which essentially created new machine learning algorithms and automatically evaluated them in a feedback loop. That success got him thinking about other cases where this approach could work, and he turned to hardware, believing that automated systems could also generate the low-level code, like the kernels and so forth, needed to help run chips more effectively. 他表示，他曾发明过一种名为 Omega 的机器学习算法，该算法本质上可以创建新的机器学习算法，并在反馈循环中自动对其进行评估。这一成功让他开始思考这种方法在其他领域的应用，于是他转向了硬件领域，坚信自动化系统同样可以生成运行芯片所需的底层代码（如内核等），从而更有效地驱动芯片。
Infinity’s AI research agent Ignition is intended to write the low-level code needed for AI inference on Nvidia-alternative chips. It tests, debugs, and measures how fast the hardware performs with the code, and automatically rewrites the code if needed to improve performance. The system is self-optimizing, meaning it continuously learns and improves itself. It also adapts to different chip architectures, regardless of proprietary designs, Nixon says. Infinity 的 AI 研究智能体 Ignition 旨在为英伟达替代芯片编写 AI 推理所需的底层代码。它能够测试、调试并衡量硬件在代码运行下的性能，并在必要时自动重写代码以提升表现。Nixon 表示，该系统具有自我优化功能，意味着它能持续学习并自我改进。此外，无论芯片采用何种专有设计，它都能适应不同的芯片架构。
The result is what Infinity claims is a CUDA-level software stack. Customers include the AI chip maker (and would-be Nvidia challenger) D-Matrix, and Infinity is in talks with other big chip and cloud companies, Nixon said. Nixon 称，其成果是一个堪比 CUDA 级别的软件栈。客户包括 AI 芯片制造商（也是英伟达的潜在挑战者）D-Matrix，目前 Infinity 正在与其他大型芯片和云公司进行洽谈。
Humans are in the loop, however, providing high-level direction while the agent does more of the tedious grunt work. In one case study, the startup found the agent works much faster than a human alone, reducing what could have been a years- or months-long process to hours or days. 不过，人类仍然参与其中，负责提供高层指导，而智能体则负责处理更多繁琐的底层工作。在一项案例研究中，该公司发现该智能体的工作效率远超人类，将原本可能需要数年或数月的过程缩短至数小时或数天。
Infinity doesn’t charge an upfront license fee; instead, it takes a cut of performance gains and cost savings, measuring changes in tokens per second. Right now, Infinity has 26 employees, including those in design, operations, and engineering. Infinity 不收取预付许可费，而是通过衡量每秒 Token 处理量的变化，从性能提升和成本节约中分成。目前，Infinity 拥有 26 名员工，涵盖设计、运营和工程等部门。