Rodrigo Liang, CEO of AI infrastructure firm SambaNova, said open-source AI models are becoming a more attractive option for enterprises, citing performance gains that now let companies run, customize and fine-tune models on their own private data.
Enterprise AI · SambaNova Systems
Open-Source Models Are Winning Over the Enterprise
CEO Rodrigo Liang argues that increasingly capable open models — which companies can own, fine-tune and control with their own private data — now beat closed APIs for business. SambaNova's custom RDU chips race to serve them at high speed.
10×+
Faster than the GPU average on DeepSeek R1 671B (250 vs 19 tok/s/user)
5×
More energy efficient than GPUs on comparable workloads
95%
Fewer chips — 16 RDUs match ~320 GPUs on DeepSeek R1
Inference speed — tokens / sec / user
DeepSeek R1 671B: SambaNova RDU vs GPU average
A single 8-RDU node holds large MoE models without sharding across chips. Llama 3.1 405B runs at 132 tok/s in full precision.
Why enterprises want open models
Own & control
Fine-tune privately
Data sovereignty
Samba-1 composes 50+ expert models at trillion-parameter scale — well suited to agentic AI in finance and research. Rack power: ~10kW vs 40–50kW for GPUs.
The pitch
Developers praise the raw speed; holding multiple models at once fits production agentic AI with multi-step reasoning and tool use.
The caveats
Limited to ported open models; less universally useful on complex multi-file coding than benchmarks suggest. Fundraising stalled — a sale is being explored.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…