AI News AI资讯 7d ago Updated 7d ago 更新于 7天前 45

Sail Research Closes $80M in Funding to Build Max-Efficiency Infrastructure for AI Agents Sail Research完成8000万美元融资,旨在构建AI代理的最大效率基础设施

Sail Research secured $80 million in Seed and Series A funding at a $450 million valuation, backed by Kleiner Perkins and Sequoia. The company offers infrastructure specifically designed for long-horizon AI agents, focusing on sustained throughput rather than low-latency single requests. Their platform achieves up to 10x lower cost per token through a custom inference stack and a pay-for-active-compute sandbox called Sailboxes. Sail topped the BrowseComp-Plus deep research benchmark with 90.72% Sail Research完成8000万美元融资(种子轮+A轮),估值4.5亿美元,由Kleiner Perkins和Sequoia领投,专注构建面向长周期自主AI代理的基础设施。 平台核心包含从头构建的高吞吐量推理栈及Sailboxes沙盒环境,通过仅对活跃计算时间收费实现高达10倍的Token成本降低。 在BrowseComp-Plus深度研究基准测试中以90.72%准确率登顶,证明其在处理数十亿Token的长期任务中具备极高的效率与准确性。 创始团队来自Nvidia、Apple和Together AI,已获Alphabet、Intel高管天使投资,并服务于Parallel Web Sys

65
Hot 热度
60
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Sail Research secured $80 million in Seed and Series A funding at a $450 million valuation, backed by Kleiner Perkins and Sequoia.
  • The company offers infrastructure specifically designed for long-horizon AI agents, focusing on sustained throughput rather than low-latency single requests.
  • Their platform achieves up to 10x lower cost per token through a custom inference stack and a pay-for-active-compute sandbox called Sailboxes.
  • Sail topped the BrowseComp-Plus deep research benchmark with 90.72% accuracy, demonstrating superior performance in complex, multi-step tasks.
  • Early production adoption includes companies like Parallel Web Systems and Detail.dev, validating the economic viability of autonomous agent workloads.

Why It Matters

This development signals a critical shift in AI infrastructure from optimizing for human-in-the-loop chat interfaces to supporting autonomous, long-running agent workflows. For practitioners, it highlights that current cost structures and latency optimizations are misaligned with the emerging needs of agentic AI, creating a market opportunity for specialized backend solutions. The significant funding and early enterprise adoption suggest that scalable, cost-efficient inference for multi-hour tasks is becoming a primary bottleneck and priority for the industry.

Technical Details

  • Custom Inference Stack: Built from scratch to prioritize high throughput and token efficiency over single-request latency, utilizing deep customization of open-source engines and strategic use of underutilized compute resources.
  • Sailboxes Environment: A stateful sandbox architecture that charges only for active compute time, allowing agents to run for hours or days without incurring costs during idle periods, significantly reducing overall operational expenses.
  • Benchmark Performance: Achieved 90.72% accuracy on the BrowseComp-Plus benchmark, a rigorous deep research evaluation, while maintaining costs up to 10 times lower than leading competitors.
  • Model Compatibility: The API remains compatible with existing OpenAI-based workflows and supports major open-source models including DeepSeek, Gemma, GLM, Kimi, and Nemotron.
  • Workload Distribution: Employs intelligent routing across multiple providers to maximize resilience and optimize for both cost and performance during sustained, high-volume token usage.

Industry Insight

  • Infrastructure Specialization: General-purpose LLM APIs may become insufficient for advanced agentic applications; companies should evaluate specialized inference providers that offer better economics for long-context, multi-step reasoning tasks.
  • Cost Structure Evolution: The shift toward paying for active compute time rather than total token count or uptime could redefine pricing models in the AI industry, making long-running autonomous agents economically feasible for broader use cases.
  • Talent and Expertise: The success of founders with backgrounds in hardware and large-scale systems (Nvidia, Apple) underscores the importance of deep systems engineering expertise in building the next generation of AI infrastructure.

TL;DR

  • Sail Research完成8000万美元融资(种子轮+A轮),估值4.5亿美元,由Kleiner Perkins和Sequoia领投,专注构建面向长周期自主AI代理的基础设施。
  • 平台核心包含从头构建的高吞吐量推理栈及Sailboxes沙盒环境,通过仅对活跃计算时间收费实现高达10倍的Token成本降低。
  • 在BrowseComp-Plus深度研究基准测试中以90.72%准确率登顶,证明其在处理数十亿Token的长期任务中具备极高的效率与准确性。
  • 创始团队来自Nvidia、Apple和Together AI,已获Alphabet、Intel高管天使投资,并服务于Parallel Web Systems等生产级客户。
  • 旨在解决现有基础设施为短时交互设计导致的速率限制和扩展瓶颈,支持OpenAI兼容API及DeepSeek等多开源模型。

为什么值得看

本文揭示了AI基础设施正从优化“低延迟单次请求”向优化“高吞吐长周期任务”转型的关键趋势,这对开发自主代理(Agents)的企业至关重要。Sail Research通过重构推理栈和创新的计费模式,解决了长期运行AI代理时面临的成本高昂和算力瓶颈问题,为万亿Token规模的Agent工作负载提供了可行的经济模型。

技术解析

  • 专用推理栈架构:Sail Research构建了从头开始的推理堆栈,专门针对吞吐量(Throughput)和Token效率进行优化,而非传统的最小化单次请求延迟。该架构通过深度定制开源推理引擎、智能跨提供商工作负载分布以及利用未充分利用的计算资源,将GPU性能推向前沿水平。
  • Sailboxes沙盒环境:这是一种专为数小时至数天运行设计的状态保持沙盒环境。其核心创新在于计费模式——仅对代理实际执行工作的“活跃计算时间”收费,从而大幅降低了长时间后台任务的总体拥有成本(TCO)。
  • 基准测试表现:在代表深度研究能力的BrowseComp-Plus基准测试中,Sail的推理服务达到了90.72%的最高准确率,同时其每Token成本比主要竞争对手低多达10倍,验证了其在复杂、长上下文任务中的性价比优势。
  • 兼容性与模型支持:API完全兼容现有的OpenAI工作流,确保迁移成本低;同时支持DeepSeek、Gemma、GLM、Kimi和Nemotron等主流开源模型,为开发者提供了灵活的选择。

行业启示

  • Agent基础设施的经济性成为竞争壁垒:随着AI支出预计达到2.5万亿美元,能够以极低边际成本支撑“长周期”、“高并发”Agent任务的底层基础设施将成为关键差异化因素。企业需关注那些能解决Token消耗过大问题的新型推理平台。
  • 从“交互式”到“自主式”的技术范式转移:当前主流云服务多针对人类交互的低延迟场景优化,而未来的AI应用将大量依赖无人值守的后台代理。行业需要重新评估算力架构,优先选择支持长时间状态保持和高吞吐量的解决方案。
  • 垂直领域落地加速:代码审查(如Detail.dev)和网页系统(如Parallel)等垂直领域的早期采用表明,高效的基础设施正在释放Agent在复杂工程和工作流自动化中的巨大潜力,建议企业在构建Agent时优先考虑此类专用基础设施合作伙伴。

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。

Agent Agent Funding 融资 Inference 推理 Product Launch 产品发布