"It's a land grab right now": SambaNova's Rodrigo Liang on why inference, not training, is AI's real bottleneck
Key takeaways from the Sourcery interview with SambaNova Systems co-founder and CEO Rodrigo LiangDealroom has a profile for this one. Try Dealroom → (July 2026):
The billion-dollar round. SambaNova has done a first close of a $1bn fundraise at an $11bn valuation, led by General Atlantic with Seligman Ventures, T. Rowe Price and Capital Group among the backers — bringing total capital raised to around $2.5bn. Liang, a 32-year semiconductor veteran, says he has "never seen the interest in semiconductors higher."
The inference thesis. Founded in 2017 to lower the cost of training, SambaNova now targets the harder problem the industry has hit at scale: deploying models for millions of daily users efficiently, where inference chips deployed will vastly outnumber those used for training.
The efficiency edge. Its SN40 rack replaced a 130–140kW rack of Nvidia GPUs with a 10kW, air-cooled rack that outperformed them — running a trillion-parameter model in a single rack where rivals need dozens, with no liquid-cooling retrofit and standard 19-inch racks, Kubernetes, Red Hat Linux and Ethernet.
Six chips in seven years. SambaNova has taped out six chips since founding and will tape out its seventh next year, with generation five shipping later in 2026.
Scale-out, not scale-up. Because inference scales out, SambaNova's minimum quantum is a single rack — versus 10–20 racks for some providers to run a 1.5-trillion-parameter model like DeepSeek — letting operators grow with their user base and deploy in existing downtown data centres for low latency.
Why latency matters. In an agentic world where 20 agents orchestrate a single prompt, a two-second response per agent compounds to 40 seconds; users expect one to two seconds, so per-agent latency must fall to fractions of a second — driving demand for compute placed near users in metropolitan areas.
Premium inference defined. Two dimensions: model size (accuracy, with frontier models heading toward 10 trillion parameters) and speed. SambaNova runs the largest models in full, unquantized precision faster than rivals — comparing the shift to 5G, arguing everyone will eventually want the fastest inference.
Coopetition and the edge. SambaNova offloads the 70–80% of racks running inference in mixed data centres so operators can resell freed-up GPU capacity and lift margins; its 10kW racks also fit into shipping containers for edge deployments (a partnership with Armada) in remote or infrastructure-constrained regions.
Read more: SambaNova
