As AI workloads drive hovering cloud payments, extra firms are weighing whether or not to maneuver computing out of public clouds and into their very own knowledge facilities. However constructing and working AI infrastructure is way extra difficult than merely shopping for servers — networking has turn into one of many greatest technical hurdles.
That’s the chance Seattle startup Hedgehog is chasing.
Based in 2022 by CEO Marc Austin, a Cisco networking veteran, Hedgehog develops open-source software program designed to make personal AI knowledge facilities function extra like hyperscale clouds. It has raised $11 million in seed funding, with plans to boost a collection A financing spherical.
We caught up with Austin for the return of GeekWire’s Startup Highlight to be taught extra concerning the 20-person firm, the AI networking growth and what shocked him most about constructing a startup in considered one of tech’s fastest-moving markets.
In 50 phrases or much less, give us your elevator pitch?
Hedgehog is open-source software program that makes AI networking easy. AI clouds and enterprises use it to run GPU networks the way in which hyperscalers do — deployed in hours as a substitute of months, operated by DevOps groups as a substitute of armies of community engineers, on open {hardware} with no vendor lock-in.
What downside are you obsessive about fixing?
Time to GPU worth. A GPU cluster is the most costly asset most firms will ever purchase, and each day it sits idle ready on the community is cash burning. That wait is never the {hardware} — it’s the material: weeks or months of scarce community engineers hand-designing, cabling, tuning, and validating it throughout proprietary CLIs and locked-in vendor gear.
In the meantime the folks informed to “personal the community” normally aren’t community engineers in any respect — they’re platform and DevOps groups. We’re obsessive about collapsing that timeline: declare your community like intent in Kubernetes and go from racked GPUs to inference in hours as a substitute of months — on open {hardware}, no lock-in, no room stuffed with specialists. Cloud-grade networking with out hyperscaler headcount.
What shocked you after speaking to clients?
How not often the customer is a community engineer. It’s platform and DevOps groups, typically at AI clouds who simply took supply of hundreds of GPUs who’re informed “you personal the community now.” They don’t need to be taught BGP; they need a community that behaves like the remainder of their cloud-native stack. The opposite shock: they don’t simply need to run the community, they need to promote it by carving up capability for their very own clients, like a cloud supplier does.
How has AI modified the way in which you construct your organization?
Twice over.
Our product exists as a result of AI broke conventional networking. Coaching and inference site visitors melts networks designed for net apps.
And AI modified how we construct: we use it closely throughout engineering, testing, and go-to-market, which lets a small staff repeatedly take a look at each supported system and configuration in our lab and ship with hyperscaler-grade rigor. AI raised the bar for what a startup-sized staff can ship.
What’s one factor folks misunderstand about your startup?
That “open supply” means hobbyist. The alternative is true: openness is the enterprise function. Our clients can audit each line of code that runs their cloth, prolong it, and by no means get locked in. Practically each competitor markets “open networking” whereas transport a proprietary controller. Hedgehog is the one one that truly publishes the repo.
What’s the hardest resolution you’ve made previously yr?
Betting completely on Ethernet. We determined open, standards-based Ethernet would win AI networking and put all the pieces behind it. Watching the trade’s largest AI operators now standardize on that very same method makes us be ok with the decision — however saying no was arduous.
What’s the one piece of recommendation you give to different entrepreneurs?
Choose the wave, not simply the surfboard.
Product selections are recoverable; betting in opposition to a structural trade shift isn’t. Discover the usual, the structure, or the customer habits that’s inevitable, align all the pieces to it early, and be affected person whereas the market catches as much as your guess.
We’ll know our firm has made it when…
Networking is boring once more. When a platform engineer stands up a multi-tenant GPU cloud and the community is just some traces of declared intent that no person thinks twice about. When “community like a hyperscaler” describes each AI cloud, not simply the giants working on Hedgehog, then we can have made it!
