kaifaai.app AI 快讯 · 2026-10-14

无问芯穹单轮23亿，累计43亿刷新国产AI Infra纪录

无问芯穹单轮融资23亿元，累计融资43亿元，创国产AI基础设施赛道融资金额双重纪录。

无问芯穹在最新一轮中获23亿元人民币融资，叠加此前融资，总融资额达43亿元，成为国产AI Infra领域单轮及累计融资额双料第一。其核心能力覆盖算力调度（支持万卡集群异构混训）、模型压缩（Llama-3-70B可压缩至16GB显存运行）与行业适配层（已预集成电力、交通、医疗三类垂域知识图谱）。值得注意的是，该轮融资由国家级产业基金联合多家银行系资本主导，非纯市场化VC，表明政策性资金正加速下沉至AI底层工具链。对早期AI应用团队而言，这意味着：第一，不必自建全栈Infra——可直接调用无问芯穹提供的‘开箱即用’推理API（含动态批处理与KV缓存优化），将Agent上线周期从平均6周压缩至72小时内；第二，融资节奏需匹配基础设施成熟度，若团队仍依赖手动微调+本地部署，将面临客户对SLA（如99.95%可用性）和合规审计（等保三级日志留存）的刚性要求，而这些能力需至少18个月工程沉淀。

Unisound raises $230M in single round, $430M total — new record for Chinese AI Infra

This $230M round — the largest single raise in China’s AI Infra space — brings Unisound’s cumulative funding to $430M, surpassing all peers. Its stack delivers production-grade capabilities: heterogeneous cluster training (10k+ GPUs), model compression (e.g., Llama-3-70B runs in 16GB VRAM), and pre-integrated domain knowledge graphs for power, transport, and healthcare. Critically, state-backed industrial funds and policy banks led this round — signaling capital is shifting from speculative model bets to foundational tooling. For early AI application teams, this means two immediate implications: First, avoid reinventing the wheel — Unisound’s managed inference APIs (with dynamic batching and KV cache optimization) let teams deploy production agents in under 72 hours, cutting typical time-to-value from 6 weeks. Second, align fundraising strategy with infra maturity: teams still relying on manual fine-tuning and local deployment will struggle to meet enterprise SLAs (e.g., 99.95% uptime) or compliance audits (e.g., China’s Level-3 Cybersecurity Protection requirements), which demand 12–18 months of dedicated engineering effort.

来源：量子位 — 单轮23亿，累计43亿，无问芯穹刷新国产AI Infra赛道融资双重纪录