NVIDIA AI Chips Lose Ground With Engineers as Cooling Costs Push Hyperscalers Toward Custom ASICs, Evercore Warns

NVIDIA faces growing skepticism from AI engineers regarding its data center chips, according to analyst reports. Hyperscalers are increasingly evaluating custom ASICs and alternative accelerators as they shift toward inference-led workloads.

Analysts report engineers prioritize cooling and power costs over raw performance per watt when evaluating AI chips.

Morgan Stanley notes that Blackwell GPUs deliver up to eight times higher performance per watt than custom AI chips. This efficiency advantage exists despite the fact that building a data center with NVIDIA hardware costs twice as much.

Evercore ISI reports that average AI engineers are unimpressed by NVIDIA's claimed 35x performance improvements. Analysts believe the company's 70 percent gross margins appear excessive to practitioners who prioritize cooling and power costs over raw speed.

The firm projects NVIDIA's inference market share will decline to 50 percent by 2028 as alternatives like AMD TPUs, Trainium chips, and Groq hardware improve. Engineers are reportedly willing to adopt good enough options from competitors to improve their economics.