Sruti Kopakkar

AI Summary

→ WHAT IT COVERS NVIDIA's Sruti Kopakkar breaks down tokenomics — the framework for valuing, supplying, and monetizing AI tokens — into four pillars: token utility, token supply, token demand, and token monetization, giving business leaders a structured approach to deploying AI infrastructure profitably and measuring true return on investment. → KEY INSIGHTS - **Token Value Framework:** Token value depends on two variables: the intelligence embedded (determined by model complexity and context length) and interactivity (tokens per second per user). Map each use case to the appropriate point on this spectrum — agentic workflows require high interactivity, while enterprise search or chat interfaces do not, avoiding costly over-provisioning. - **Demand Forecasting Multipliers:** Base token demand (users × requests × tokens per session) understates actual requirements. Apply three multipliers: reasoning models generate hidden "thinking tokens" that never reach end users; agentic workflows multiply LLM calls significantly; and KV cache hit rate reduces recomputation. Factor in daily, seasonal, and user-growth variability for accurate forecasting. - **Cost Per Token vs. Input Metrics:** Evaluating AI infrastructure on GPU hourly cost or FLOPS per dollar misrepresents true ROI. Cost per token — GPU cost divided by tokens produced — captures both expenditure and delivered output. NVIDIA Blackwell delivers 50x more tokens per watt than Hopper, versus only 2x on raw FLOPS-per-dollar comparisons. - **Jevons Paradox in AI Scaling:** Lowering cost per token does not reduce GPU demand — it unlocks new use cases that consume the freed capacity. Each efficiency gain historically triggered a new scaling wave: generative AI led to reasoning models, which led to agentic AI. Organizations should plan infrastructure for expanding token consumption, not static or shrinking demand. - **Four Token Monetization Models:** Businesses convert tokens into revenue through four paths: selling tokens directly (Fireworks, Together AI, DeepInfra); building AI-native products (Perplexity, Cursor); infusing AI into existing products (Adobe Firefly inside Photoshop, Shopify, Airbnb); or improving internal operations and employee productivity. Start from the customer use case and work backward to infrastructure decisions. → NOTABLE MOMENT Kopakkar reveals that NVIDIA Blackwell's advantage over Hopper looks modest on paper — just 2x on hourly GPU cost and FLOPS per dollar — but when measured by actual delivered output, Blackwell produces 50 times more tokens per watt, demonstrating how conventional spec-sheet metrics can dramatically obscure real-world infrastructure value. 💼 SPONSORS None detected 🏷️ AI Infrastructure, Tokenomics, Inference Optimization, Agentic AI, Total Cost of Ownership