SambaNova

You can serve a billion tokens a second and lose money on every one of them. Speed is expected. Payback is the benchmark. Ours is 6 months. Here’s a fun experiment for you: Ask an AI infrastructure vendor about payback and watch how quickly the conversation gets awkward. 😬 Every buyer cares about one thing: When does this hardware earn back what I paid for it? For most infrastructure, the answer is a year or more. We’ve somehow decided that’s normal. Well, we don’t. 6-month payback. 5-year lifespan. You can do the math. Economics is the real difference-maker.

"The benefit of SambaNova, uniquely, which is a result of our architecture, is our systems are fully air cooled. They're 30 kilowatts and they can go into most datacenters in the world today." — Anton McGonnell Many AI racks need 600 kW of power. No datacenter in the world can handle that today. SambaNova runs on 30kW and fits in existing infrastructure. 🦾 🎥: Forbes

Training gets the headlines, but inference is where AI comes to life. Every token, every query, every agent interaction. We're focused on making that inference faster and more efficient by tackling one of the biggest challenges at scale: data movement. ⚡

"As the cost of tokens continue to rise because the infrastructure is so expensive, people are looking for alternatives. So at SambaNova, we come in with a much lower-cost infrastructure, much higher performance. We drive it at 10 kilowatts, an order of magnitude lower power than the traditional GPUs." — Rodrigo Liang During his chat with Matthew Miller on Bloomberg Television, Rodrigo explained how we're able to make AI more affordable: disaggregated inference. Instead of relying on a single architecture, it combines GPUs for prefill, RDUs for high-speed decode, and CPUs for orchestration, allowing each processor to do what it does best. The result is faster inference, lower latency, and significantly lower-cost tokens for enterprises and service providers deploying AI at scale. This architecture powers the world's first commercially available disaggregated inference cloud, including VC2, helping organizations bring AI into production with better performance and better economics. Watch their conversation below ⬇️

What happens when you match the right chip to the right stage of inference? At #Computex, we demonstrated premium inference for AI agents running live inside the new VC2 data center. Using NVIDIA B200 GPUs for prefill, SambaNova RDUs for decode, and CPUs for orchestration, the system delivered up to 2X the inference speed of B200-only configurations. The demo highlights where AI infrastructure is heading next: heterogeneous systems optimized for real-world agentic workloads 🦾

"Where dataflow comes in is... you can actually just flush these things like a river flushes through their own environment with a very fast pace." - Rodrigo Liang One of the biggest challenges in AI today isn't the model, it's getting data to the right place at the right time. That's where dataflow architecture comes in, helping eliminate memory bottlenecks and unlock faster, more efficient inference at scale. 🦾 🎥: Forbes