Summary
Speechify founder and CEO Cliff Whitesman joins Harry Stebbings on 20VC to explain why the text-to-speech startup has invested tens of millions of dollars in Nvidia GPUs — including paying a $100,000 premium per card for early delivery — rather than relying on cloud providers like AWS, Azure, or GCP. Whitesman breaks down the economics in detail: an H100 that costs roughly $30,000 to purchase would run $35,000–$50,000 to rent for a single year, making ownership significantly cheaper over the hardware’s useful life. He also argues that owned GPUs change engineer behavior, removing the friction that causes teams to under-utilize rented compute.
The conversation covers Speechify’s latest model, Simba 3.2, which Whitesman claims ranks first globally for text-to-speech quality while being ten times more affordable than competitors including ElevenLabs. He addresses the competitive dynamics directly, comparing Speechify’s position to Anthropic’s rise from second-place challenger to OpenAI to a top-tier player, and signals an aggressive push into B2B voice AI and agent APIs.
Whitesman also discusses the accelerating GPU upgrade cycle — from A100s through H100s to Blackwells and the new Rubin architecture — and how that affects the calculus of hardware ownership versus depreciation. For founders and AI infrastructure decision-makers, this episode offers a rare candid look at the build-versus-buy decision for compute at scale, complete with real cost figures and strategic reasoning from a founder who has lived both approaches.
📺 Source: 20VC with Harry Stebbings · Published September 05, 2026
🏷️ Format: Podcast
