SambaNova CEO on Raising $1B at $11B: "It's a Land Grab Right Now"

Summary

Rodrigo Liang frames SambaNova’s latest financing as validation of a broader AI infrastructure land grab: the company announced a first close on a $1B round at an $11B valuation while he says semiconductor demand is the strongest he has seen in 32 years. The conversation centers on why inference, not training, is now the bigger systems problem, especially as large frontier models and agentic workflows make latency and rack efficiency critical. Liang highlights SambaNova’s chip-and-rack strategy, including the SN40 and the new SM50, and argues that a 10 kW air-cooled rack can displace far larger GPU racks for certain workloads. He also emphasizes sovereign, on-prem, and partner-led deployment models, saying enterprises and governments want to keep private data out of shared foundation models and are increasingly buying infrastructure to run locally. His view of the market is that a few chip vendors and deployment platforms will win by scaling fastest, with customers judged on per-rack economics, speed, and real business ROI rather than raw token volume.

Notes

• SambaNova announced a first close on a $1B financing round at an $11B valuation.
• Liang says he has 32 years of experience building high-performance chips.
• He says current semiconductor demand is the highest he has ever seen.
• AI usage at Anthropic, OpenAI, and Gemini has pushed the industry from model creation toward scale inference.
• SambaNova originally focused on lowering training cost.
• The company shifted from training to inference as the larger infrastructure problem.
• SN40 became popular because it outperformed a 130–140 kW Nvidia GPU rack using a 10 kW air-cooled rack.
• SambaNova says it can run a trillion-parameter model in a single rack.
• The rack-level efficiency claim is central to SambaNova’s pitch on inference deployment.
• Training requires thousands of GPUs synchronized in one cluster.
• Training failures force checkpointing because one node failure can break the whole cluster.
• Inference scales by adding racks as demand grows, rather than by synchronizing thousands of devices.
• SambaNova says the minimum deployment quantum is one rack.
• Some providers may need 10 to 20 racks to serve a 1.5T-parameter DeepSeek model.
• SambaNova standardizes on 19-inch racks, air cooling, Kubernetes, Red Hat Linux, and Ethernet.
• The company is designed to avoid liquid-cooling retrofits in customer data centers.
• Building a gigawatt data center can take nine to 18 months, according to Liang.
• Power procurement and even nuclear generation can become part of the AI infrastructure stack.
• Liang argues agentic systems make latency additive across many agents.
• A 2-second response time becomes 40 seconds when 20 agents are orchestrating.
• Premium inference is defined by both model size and speed.
• Liang says SambaNova runs the biggest models in original, full precision rather than quantized form.
• He says models are moving toward 10T parameters, with open source already at 1T to 2T.
• The inference market is a land grab across hyperscalers, frontier labs, neo-clouds, and sovereign clouds.
• Liang says scale is the main competitive advantage in this market.
• SambaNova has raised about $2.5B over the company’s history.
• Liang says only a few chip vendors are likely to dominate AI infrastructure over time.
• SambaNova chooses not to build its own cloud to compete with AWS.
• Instead, SambaNova focuses on shippable infrastructure such as racks, chips, and software.
• Partner companies run the NEO cloud services while SambaNova ships the infrastructure.
• The partner model shifts data-center, energy, and facilities costs to service providers.
• Sovereignty is framed as protecting private data from entering shared model pipelines.
• Countries are investing heavily to train homegrown national models from scratch.
• Liang expects heterogeneous AI stacks that combine Anthropic, OpenAI, open source, and local models.
• He says on-prem repatriation is already happening after years of cloud-first migration.
• He argues enterprise differentiation erodes when everyone uses the same commodity model.
• He says that realization arrives on a roughly two-year horizon, not a decade away.
• Liang says the aim is to reduce current costs and unlock new services, not simply spend more on AI.
• He says token usage is not the goal; service creation is.
• Liang says companies should ask how they will differentiate and scale in a constrained AI market.
• He argues AI will help companies create more services and compete harder, not just cut jobs.
• He says business-scale AI is constrained by energy, data-center, chip, and cost bottlenecks.
• He says resilience matters because chip businesses are never a straight line.
• SambaNova’s latest chip is the SM50.
• Liang says SM50 is showing strong results in aggregated inference.