Thermal Management: The Hidden Tax on AI Factory Profitability

The report, titled "Chip to Chiller," argues that hyperscalers have historically treated cooling as a facilities afterthought rather than a core component of infrastructure design. Because GPU performance is locked at the point of manufacture, thermal management remains the only variable in a operator's control that can prevent "leakage power." As GPU junction temperatures climb, efficiency erodes exponentially, effectively wasting power and hardware investment on heat rather than token generation.

Frore Systems suggests that treating the "Thermal Stack" as a first-class piece of infrastructure can increase an AI factory’s Tokens/Watt—the industry's key metric for economic efficiency—by more than 30%. The company’s own LiquidJet coldplate technology, which utilizes semiconductor manufacturing techniques, demonstrated a reduction in GPU die temperatures of 6 to 12°C in testing. This improvement translates to a 10 to 25% gain in token throughput. Seshu Madhavapeddy, CEO and founder of Frore Systems, notes that at the gigawatt scale, thermal architecture is no longer a minor line item but a fundamental competitive moat. Operators who integrate advanced cooling early in the capital allocation process stand to extract significantly more value from their hardware than those relying on conventional skived coldplates.