OpenAI, Intel and Meta Unveil Lower-Power AI Chips as OpenAI Plans $750 Billion Data Center Buildout
Updated
Updated · Computerworld · Aug 31
OpenAI, Intel and Meta Unveil Lower-Power AI Chips as OpenAI Plans $750 Billion Data Center Buildout
2 articles · Updated · Computerworld · Aug 31
Summary
Hot Chips 2026 showcased new AI processors from OpenAI, Intel, Meta and others aimed at cutting token-generation costs and power use as companies push for more productive AI spending.
OpenAI used the event to detail its in-house Jalapeño inference chip, which it said should help serve more demand at lower cost; SemiAnalysis said the first-generation design outperformed tested Nvidia, AMD and Google chips on several open-source models.
Intel highlighted Diamond Rapids server CPUs with AI extensions for on-chip inferencing and Wildcat Lake PC chips with NPUs, reflecting a broader shift of some AI workloads from GPUs toward CPUs, PCs and edge devices.
Analysts said the show underscored a more heterogeneous AI market beneath Nvidia, with future deployments spreading across cloud, sovereign data centers, local servers and AI PCs rather than relying on one GPU vendor.
That diversification is being driven by economics: enterprises still expect AI budgets to rise, but they increasingly want architectures flexible enough to switch among GPUs, CPUs and specialized chips as performance-per-watt and cost improve.
As OpenAI's custom Jalapeño chip threatens Nvidia's dominance, will the quest for cheaper AI compute finally shatter the GPU monopoly in 2027?
Will the breakthrough efficiency of these new AI architectures genuinely save power, or just trigger a massive explosion in total energy consumption?
Could the complex orchestration required for split inference and heterogeneous chips actually create worse bottlenecks than the traditional GPU systems they replace?