内容来源：https://www.geekwire.com/2026/startup-spotlight-hedgehog-bets-that-open-source-networking-will-power-the-next-generation-of-ai-clouds/

内容总结：

AI算力成本飙升，企业加速“云回迁”：网络成最大技术关卡

随着人工智能工作负载推动云服务账单持续攀升，越来越多的企业开始重新评估是否要将计算任务从公有云迁移回自建数据中心。然而，构建和运营AI基础设施远非采购服务器那么简单——网络技术已成为最大的技术瓶颈。

这正是西雅图初创公司Hedgehog瞄准的市场机遇。该公司由思科网络资深人士Marc Austin于2022年创立，专注于开发开源软件，旨在让私有AI数据中心像超大规模云平台一样高效运作。目前，Hedgehog已获得1100万美元的种子轮融资，并计划启动A轮融资。

Hedgehog的核心产品是一款开源软件，能大幅简化AI网络部署。其“电梯演讲”直击痛点：AI云和企业用户可像超大规模云服务商一样运行GPU网络——部署时间从数月缩短至数小时，由DevOps团队而非大量网络工程师运维，且基于开放硬件，无供应商锁定。

团队痴迷于解决“GPU价值变现时间”问题。GPU集群是企业最昂贵的资产之一，因网络问题闲置的每一天都在烧钱。痛点往往不在硬件，而在于网络架构：需要稀缺的网络工程师花费数周甚至数月，在专有命令行界面和锁定供应商的设备上进行手工设计、布线和调优。更关键的是，被要求“负责网络”的往往不是网络工程师，而是平台和DevOps团队。Hedgehog的目标是将这一过程压缩：开发者只需在Kubernetes中声明网络意图，即可在数小时内从机架GPU到完成推理，无需大量专家驻场。

在与客户交流中，创始人惊讶地发现：买家很少是网络工程师，而是平台和DevOps团队——他们刚接收了数千块GPU，却被要求“现在你负责网络”。这些人不想学习BGP协议，只希望网络能像其他云原生组件一样工作。另一大发现是：客户不仅想运行网络，还希望像云服务商那样，通过分割容量向自己的客户出售网络服务。

AI浪潮也深刻改变了Hedgehog自身的构建方式。一方面，传统网络为网页应用设计，无法承载AI训练和推理流量，这催生了Hedgehog的产品。另一方面，AI被大量应用于工程、测试和市场推广，使小团队能以超大规模云服务商的标准持续测试所有设备与配置，提高了初创团队的交付门槛。

关于外界误解，创始人强调：开源不等于业余。恰恰相反，开放性是真正的企业级特性。客户可以审计每一行运行网络架构的代码、进行扩展，永不锁定。几乎所有竞争对手都在推销“开放网络”的同时，交付专有控制器，而Hedgehog是唯一真正公开代码仓库的公司。

过去一年最艰难的决定是：完全押注以太网。团队决定，基于标准的开放以太网将赢得AI网络市场，并孤注一掷。如今看到行业最大的AI运营商正采用相同路线，证明了决策正确，但当初说“不”并不容易。

给其他创业者的建议是：选对浪潮，而非仅挑冲浪板。产品决策可以调整，但逆行业结构性转变而行则无法挽回。要找到必然发生的标准、架构或买家行为，尽早对齐，耐心等待市场追上你的判断。

Hedgehog认为公司成功的标志是：让网络重新变得“无聊”。当平台工程师搭建多租户GPU云时，网络只需几行声明性代码，无人再费心考量；当“像超大规模云服务商一样的网络”成为每一家AI云的常态，而不仅仅是运行在Hedgehog上的巨头时，就真正实现了目标。

中文翻译：

随着AI工作负载导致云服务账单飙升，越来越多的企业正在权衡是否将计算任务从公有云迁移至自有数据中心。但构建和运营AI基础设施远比单纯购买服务器复杂——网络已成为最大的技术瓶颈之一。

这正是西雅图初创公司Hedgehog瞄准的机遇。

Hedgehog由思科网络领域资深人士Marc Austin于2022年创立并担任CEO，公司开发的开源软件旨在让私有AI数据中心的运营更接近超大规模云。该公司已获得1100万美元种子轮融资，并计划启动A轮融资。

值此GeekWire"创业聚焦"栏目回归之际，我们采访了Austin，深入了解这家20人团队的公司、AI网络热潮，以及在科技行业最快速演进的市场中创业最令他惊讶的方面。

请用50字以内概括你的电梯演讲？
Hedgehog是简化AI网络的开源软件。AI云和企业用它像超大规模云商那样运行GPU网络——数小时而非数月部署，由DevOps团队而非成群的网络工程师运维，基于开放硬件，无厂商锁定。

你执着于解决什么问题？
缩短GPU价值实现周期。GPU集群是多数企业将购买的最昂贵资产，每闲置等待网络一天就是烧钱。瓶颈很少出在硬件——而是网络架构：稀缺的网络工程师花数周数月，通过专有CLI和锁定厂商设备手动设计、布线、调优和验证。

而被告知"负责网络"的人往往根本不是网络工程师——而是平台和DevOps团队。我们执着于压缩这个时间线：像在Kubernetes中声明意图那样声明网络，从机架安装GPU到推理运行只需数小时而非数月——基于开放硬件，无锁定，无需满屋子专家。实现超大规模云级别的网络，却不需要超大规模的团队。

与客户交流后什么让你惊讶？
购买者极少是网络工程师。通常是平台和DevOps团队——往往是刚接收数千块GPU的AI云公司，被告知"现在网络归你们管"。他们不想学BGP，只想要一个像云原生技术栈其余部分那样运行的网络。另一个惊讶：他们不仅想运行网络，还想像云提供商那样划分容量出售给自家客户。

AI如何改变了你构建公司的方式？
双重改变。
我们的产品之所以存在，是因为AI打破了传统网络。训练和推理流量会压垮为Web应用设计的网络。
AI也改变了我们的构建方式：我们在工程、测试和市场推广中大量运用AI，让小型团队能持续测试实验室里每台支持的设备和配置，以超大规模云商级别的严谨性交付产品。AI提升了创业规模团队的能力上限。

人们对你的初创公司有哪些误解？
认为"开源"意味着业余玩家。事实正相反：开放性才是企业级特性。客户可以审计运行其网络的每一行代码，进行扩展，永远不会被锁定。几乎每个竞品都在推销"开放网络"的同时搭载专有控制器。Hedgehog是唯一真正公开代码仓库的公司。

过去一年你做过最艰难的决定是什么？
全力押注以太网。我们认定开放、标准的以太网将赢得AI网络，并将所有资源投入其中。看到行业最大的AI运营商现在纷纷标准化同一方案，让我们对当初的选择感到欣慰——但当初说"不"确实艰难。

你会给其他创业者什么建议？
选择浪潮，而非冲浪板。
产品决策可以修正；但逆行业结构变革而行则不可逆。找到必然发生的标准、架构或买家行为，尽早让一切与其对齐，耐心等待市场印证你的判断。

当出现什么信号时，我们才算真正成功？
当网络重新变得无趣。当平台工程师搭建多租户GPU云时，网络只是几行无需多想的声明式意图。当"像超大规模云商的网络"成为每朵AI云的标配，而不仅仅是运行在Hedgehog上的巨头时——那时我们就真正成功了！

英文来源：

As AI workloads drive soaring cloud bills, more companies are weighing whether to move computing out of public clouds and into their own data centers. But building and operating AI infrastructure is far more complicated than simply buying servers — networking has become one of the biggest technical hurdles.
That’s the opportunity Seattle startup Hedgehog is chasing.
Founded in 2022 by CEO Marc Austin, a Cisco networking veteran, Hedgehog develops open-source software designed to make private AI data centers operate more like hyperscale clouds. It has raised $11 million in seed funding, with plans to raise a series A financing round.
We caught up with Austin for the return of GeekWire’s Startup Spotlight to learn more about the 20-person company, the AI networking boom and what surprised him most about building a startup in one of tech’s fastest-moving markets.
In 50 words or less, give us your elevator pitch?
Hedgehog is open-source software that makes AI networking simple. AI clouds and enterprises use it to run GPU networks the way hyperscalers do — deployed in hours instead of months, operated by DevOps teams instead of armies of network engineers, on open hardware with no vendor lock-in.
What problem are you obsessed with solving?
Time to GPU value. A GPU cluster is the most expensive asset most companies will ever buy, and every day it sits idle waiting on the network is money burning. That wait is rarely the hardware — it’s the fabric: weeks or months of scarce network engineers hand-designing, cabling, tuning, and validating it across proprietary CLIs and locked-in vendor gear.
Meanwhile the people told to “own the network” usually aren’t network engineers at all — they’re platform and DevOps teams. We’re obsessed with collapsing that timeline: declare your network like intent in Kubernetes and go from racked GPUs to inference in hours instead of months — on open hardware, no lock-in, no room full of specialists. Cloud-grade networking without hyperscaler headcount.
What surprised you after talking to customers?
How rarely the buyer is a network engineer. It’s platform and DevOps teams, often at AI clouds who just took delivery of thousands of GPUs who are told “you own the network now.” They don’t want to learn BGP; they want a network that behaves like the rest of their cloud-native stack. The other surprise: they don’t just want to run the network, they want to sell it by carving up capacity for their own customers, like a cloud provider does.
How has AI changed the way you build your company?
Twice over.
Our product exists because AI broke traditional networking. Training and inference traffic melts networks designed for web apps.
And AI changed how we build: we use it heavily across engineering, testing, and go-to-market, which lets a small team continuously test every supported device and configuration in our lab and ship with hyperscaler-grade rigor. AI raised the bar for what a startup-sized team can deliver.
What’s one thing people misunderstand about your startup?
That “open source” means hobbyist. The opposite is true: openness is the enterprise feature. Our customers can audit every line of code that runs their fabric, extend it, and never get locked in. Nearly every competitor markets “open networking” while shipping a proprietary controller. Hedgehog is the only one that actually publishes the repo.
What’s the toughest decision you’ve made in the past year?
Betting entirely on Ethernet. We decided open, standards-based Ethernet would win AI networking and put everything behind it. Watching the industry’s largest AI operators now standardize on that same approach makes us feel good about the call — but saying no was hard.
What’s the one piece of advice you give to other entrepreneurs?
Pick the wave, not just the surfboard.
Product decisions are recoverable; betting against a structural industry shift isn’t. Find the standard, the architecture, or the buyer behavior that’s inevitable, align everything to it early, and be patient while the market catches up to your bet.
We’ll know our company has made it when…
Networking is boring again. When a platform engineer stands up a multi-tenant GPU cloud and the network is just a few lines of declared intent that nobody thinks twice about. When “network like a hyperscaler” describes every AI cloud, not just the giants running on Hedgehog, then we will have made it!