Upscale AI raised $500 million to develop networking silicon purpose-built for AI training clusters. The round targets GPU cluster interconnect bottlenecks that standard Ethernet and InfiniBand cannot solve at scale.
Key facts
- Upscale AI raised $500 million for AI networking silicon.
- Google Cloud is a strategic partner in the round.
- Round targets GPU cluster interconnect bottlenecks.
- Valuation and revenue figures were not disclosed.
Upscale AI has closed a $500 million funding round to build AI-native networking silicon, according to Fierce Network. The company aims to replace traditional Ethernet and InfiniBand with custom interconnect hardware designed for the unique traffic patterns of large-scale GPU clusters.
Google Cloud is listed as a strategic partner, signaling potential integration with Google's TPU and GPU fleets. The investment comes as hyperscalers grapple with network latency and bandwidth limits when training models on clusters exceeding 100,000 accelerators.
Upscale AI did not disclose valuation or revenue figures. The company competes in a space that includes established players like Nvidia's NVLink and emerging startups such as Enfabrica and Celestial AI.
Why this matters
Standard networking protocols were designed for general-purpose data center traffic, not the all-to-all communication patterns of distributed training. As model parameters grow into the trillions, the network fabric becomes the bottleneck — a problem traditional Ethernet standards can't solve without sacrificing utilization. Upscale's custom silicon approach mirrors the broader trend of hyperscalers building their own interconnects, from Google's TPU pods to Amazon's Trainium clusters.
The competitive landscape
Nvidia's NVLink dominates GPU-to-GPU communication within nodes, but inter-node networking remains fragmented. InfiniBand, while fast, is proprietary and expensive. Ethernet-based solutions sacrifice throughput for compatibility. Upscale's pitch is a third path: purpose-built silicon that matches InfiniBand latency while maintaining Ethernet's flexibility. Whether they can deliver at scale — and at what cost — remains an open question.
What to watch
Watch for Upscale's first customer deployments and whether the $500M round includes production orders from Google Cloud. If Google integrates Upscale's silicon into its next TPU generation, it could signal a broader shift away from standard networking for AI workloads.
Source: news.google.com
Key Takeaways
- Upscale AI raised $500M for AI networking silicon, with Google Cloud as a strategic partner.
- The deal targets GPU cluster interconnect bottlenecks.
