AI News AI资讯 17h ago Updated 13h ago 更新于 13小时前 47

Inference startup Infinity raises $15M from Touring Capital, OpenAI and Anthropic researchers 推理初创公司Infinity从Touring Capital、OpenAI和Anthropic研究人员处筹集1500万美元

Infinity AI secured $15 million in funding at a $100 million valuation to develop software that democratizes access to diverse AI hardware beyond Nvidia’s ecosystem. The company utilizes an AI research agent named Ignition to automatically generate, test, and optimize low-level kernels for various chip architectures, aiming to replicate CUDA-level performance. Infinity’s business model is unique, charging based on performance gains and cost savings measured in tokens per second rather than tradi Infinity完成1500万美元融资,估值1亿美元,投资者包括OpenAI和Anthropic研究人员。 公司致力于开发通用推理库和内核软件,旨在打破Nvidia CUDA在AI芯片领域的垄断地位。 其核心产品Ignition是一个自优化的AI研究代理,能自动生成、调试并优化底层代码以适配不同硬件。 商业模式采用按性能提升和成本节省分成的方式,而非传统的前置许可费。

65
Hot 热度
70
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Infinity AI secured $15 million in funding at a $100 million valuation to develop software that democratizes access to diverse AI hardware beyond Nvidia’s ecosystem.
  • The company utilizes an AI research agent named Ignition to automatically generate, test, and optimize low-level kernels for various chip architectures, aiming to replicate CUDA-level performance.
  • Infinity’s business model is unique, charging based on performance gains and cost savings measured in tokens per second rather than traditional upfront licensing fees.
  • Founded by Jeremy Nixon, the startup leverages the concept of "automated invention" to accelerate hardware-software co-design, significantly reducing the time required to port models to new chips.

Why It Matters

This development highlights a critical shift in the AI infrastructure landscape where software abstraction layers are becoming as important as hardware performance. By enabling non-Nvidia chips to run state-of-the-art models efficiently, Infinity addresses the growing need for hardware diversity and reduced vendor lock-in among AI practitioners and enterprises seeking cost-effective scaling solutions.

Technical Details

  • Ignition Agent: An autonomous AI system that writes, tests, debugs, and optimizes low-level kernel code for AI inference across heterogeneous hardware, including SRAM, GPUs, phone chips, and Systolic Arrays.
  • Self-Optimizing Stack: The software continuously learns and adapts to proprietary chip designs, aiming to achieve a software stack comparable to Nvidia’s CUDA in terms of ease of use and performance.
  • Automated Inference Library: Designed to allow chips to automatically replicate state-of-the-art research results, removing the barrier of manual kernel optimization for most application-level startups.
  • Performance Metrics: Optimization is driven by measurable improvements in inference speed, specifically tracking tokens per second to quantify efficiency gains over human-led development processes.

Industry Insight

  • Democratization of AI Hardware: Startups and researchers without deep expertise in low-level systems programming can now leverage diverse AI accelerators, potentially breaking Nvidia’s monopoly and fostering innovation in specialized chip designs.
  • New Economic Models in AI Infra: The shift toward value-based pricing (taking a cut of performance gains) aligns vendor incentives with customer success, encouraging deeper integration and long-term partnerships between software providers and chip manufacturers.
  • Acceleration of Hardware Adoption: By reducing the time to port models from months to hours, Infinity lowers the friction for adopting new AI chips, accelerating the deployment of next-generation hardware in production environments.

TL;DR

  • Infinity完成1500万美元融资,估值1亿美元,投资者包括OpenAI和Anthropic研究人员。
  • 公司致力于开发通用推理库和内核软件,旨在打破Nvidia CUDA在AI芯片领域的垄断地位。
  • 其核心产品Ignition是一个自优化的AI研究代理,能自动生成、调试并优化底层代码以适配不同硬件。
  • 商业模式采用按性能提升和成本节省分成的方式,而非传统的前置许可费。

为什么值得看

对于AI从业者而言,Infinity展示了通过自动化手段解决异构计算兼容性的新路径,有望降低非Nvidia芯片的使用门槛。该案例揭示了AI基础设施领域从单纯追求硬件算力向软件定义硬件效率转变的关键趋势。

技术解析

  • Ignition代理:一个自优化的AI系统,能够针对Nvidia以外的芯片(如SRAM、手机芯片、脉动阵列)自动生成低层内核代码,并通过测试、调试和自动重写来持续优化性能。
  • 通用推理库:旨在构建一种跨硬件平台的软件栈,使开发者无需为每种芯片编写专用代码即可运行最新的研究成果。
  • 自动化发明理念:创始人Jeremy Nixon基于其早期算法Omega的经验,将“自动化发明”应用于硬件层面,让AI系统自主生成和优化执行代码。
  • 性能评估机制:通过测量每秒token数(tokens per second)的变化来量化性能增益,作为商业结算的依据。

行业启示

  • CUDA护城河面临软件层挑战:Nvidia的统治力不仅源于硬件,更源于CUDA生态;类似Infinity这样的初创公司正试图通过通用软件层削弱这种锁定效应。
  • AI驱动的基础设施优化:利用AI本身来优化AI运行的底层代码(即“AI写AI代码”),将成为提升硬件效率和降低部署成本的重要方向。
  • 新型商业化模式兴起:按效果付费(Performance-based pricing)而非一次性授权费的模式,可能成为AI基础设施服务的新标准,降低客户尝试新技术的风险。

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。

Funding 融资 Chip 芯片 GPU GPU Deployment 部署 Inference 推理