Episode 04 Nikola Borisov

Nikola Borisov

Co-founder & CEO of DeepInfra. Serving 5 trillion AI tokens a week across 8 US data centers. The man building the inference layer that enterprise AI runs on.

Episode 04
About the Guest

Nikola Borisov.

Co-founder and CEO of DeepInfra, the AI inference infrastructure company that processes nearly 5 trillion tokens per week. DeepInfra runs 190+ open-source models on bare-metal GPU infrastructure across 8 US data centers. In 2026, the company closed a $107M Series B led by 500 Global, with NVIDIA and Samsung Next participating, bringing total funding to $133M. Nikola is building the inference layer that enterprise AI runs on, from the silicon up.

Visit deepinfra.com →

2.7K views & listens

Download transcript

Key moments

0:58 The Origin Story of DeepInfraFounder explains how they started DeepInfra in 2022 after noticing AI models need hundreds of billions of computations per token, drawing on experience scaling a messaging app to 200M+ users. 5:56 B200 GPUs at $275 vs $6-10 on HyperscalersDeepInfra offers Nvidia B200 GPUs at $275 compared to $6-10+ on major public clouds like AWS and Azure. 11:27 Pricing a Model at 27¢ When Rivals Charged 60¢DeepInfra aggressively priced Mistral's mixture-of-experts model at 27 cents per million tokens while competitors charged 60 cents, based on correctly costing only the active experts used at runtime. 21:16 Why Inference Isn't Like High-Frequency TradingUnlike HFT which needs proximity to exchanges, inference is so power-hungry that GPUs are best placed in rural areas near cheap electricity rather than near users for latency. 29:56 Open Source Models Are Catching Up FastThe founder says with models like Kimi 2.6 and DeepSeek 4, the gap between open and closed frontier models has nearly disappeared compared to the huge gap seen with Llama 2 vs GPT-4. 47:02 The Server Depreciation Myth in AI InfrastructureDrawing from a decade running data centers, he argues old GPUs like A100s and H100s remain fully utilized and even command higher prices due to overwhelming demand, defying assumptions about rapid obs
Key Moments

What he said.

5 trillion tokens a week is a number that surprises even us sometimes. The demand for inference keeps accelerating and there is no sign of it slowing down. Every week we set a new record.

Nikola Borisov, on the scale of DeepInfra's inference infrastructure

We run 190+ open-source models on our own bare-metal GPU infrastructure. No cloud abstraction layer between us and the silicon. That is how you get the price-performance that makes AI inference accessible to everyone.

Nikola Borisov, on owning the full stack from silicon to API

When NVIDIA and Samsung Next invest in your round, it is not just capital. It is signal to the entire market that inference infrastructure is strategic. This is not a commodity layer. This is where the real compute war is fought.

Nikola Borisov, on closing a $107M Series B with strategic investors

Full Transcript

Every word.

Transcript coming soon ↓

Transcript is auto-generated by YouTube. For a cleaned version, contact [email protected].