Tag
inference
3 articles
OpenAI's Jalapeño Chip Cuts Inference Costs 50%—If It Ships on Time
Custom silicon outperforms Nvidia Blackwell on efficiency metrics, but volume production delayed to 2027 as Nvidia's next architecture looms.
OpenAI's Custom Chip Cuts Inference Costs 50 Percent on First Try
Jalapeño, built with Broadcom on TSMC's 3nm process, matches Nvidia Blackwell performance while addressing the economics of running models at scale
Cerebras Doubles CS-4 Inference Throughput Without New Chip
The chip maker extracts twice the token-per-second performance from its existing WSE-3 wafer through clock speed increases and architectural improvements, letting customers double AI inference revenue on flat hardware budgets.