OpenAI's first custom chip "Jalapeño" reportedly beats Nvidia's Blackwell and Rubin in inference benchmarks
Source: The Decoder (opens in a new tab) · Matthias Bastian
Intel Summary
At the Hot Chips conference, OpenAI unveiled 'Jalapeño,' its first custom in-house inference chip. According to benchmark assessments reported by research firm SemiAnalysis, the first-generation silicon reportedly outperforms Nvidia's Blackwell and Rubin architectures in both operational throughput and energy efficiency. While first-generation custom accelerators historically lag incumbent merchant silicon, initial disclosures indicate OpenAI's specialized design targets major bottlenecks in running production large language models at scale.
Why It Matters
If independent production yields match conference benchmarks, OpenAI could significantly decrease its reliance on Nvidia hardware, improving its operating margins and compute availability. The development represents a major strategic shift in AI datacenter hardware, intensifying competition for semiconductor market share and pushing hyperscalers toward proprietary application-specific integrated circuits for model inference.
Part of an ongoing development
Independent reportingOpenAI unveiled custom inference chip Jalapeño
At the Hot Chips conference, OpenAI unveiled 'Jalapeño,' its first custom in-house inference chip. According to benchmark assessments reported by research firm SemiAnalysis, the first-generation silicon reportedly outperforms Nvidia's Blackwell and Rubin architectures in both operational throughput and energy efficiency. Claims are as reported; this summary makes no determination about accuracy or significance.
- Confidence
- High confidence
- Corroboration
- Strongly corroborated
More coverage of this development
- Jalapeño’s first results show industry-leading speed and efficiency in AI inferenceOpenAIPrimary source
- OpenAI’s Jalapeño AI chip brings new 'threat' to Nvidia margins as custom silicon gains groundCNBC TechIndependent reporting
Organizations & Entities
Related Intelligence
- ReportAlso involving Nvidia
Nvidia’s Powerful B300 Chips Spotted on Chinese State-Owned Network
The Information reports that Israeli security firm Lava identified 32 Nvidia Blackwell Ultra B300 data-center chips operating on systems linked to the network of China Telecom, a Chinese state-owned telecommunications enterprise, despite export restrictions barring their sale to China.
The Information - ReportAlso involving Nvidia
Nvidia turns the AI factory into a product, and agents into the workload
SiliconANGLE reports that at GTC Berlin, Nvidia Chief Executive Jensen Huang shifted focus away from standalone GPU hardware narratives, framing the 'AI factory' as a packaged product and positioning AI agents as the primary operational workload.
SiliconANGLE - ReportAlso involving Nvidia
The neocloud shakeout is coming, and Nvidia has the most to lose
Industry executives project that only seven to ten specialized AI cloud providers will survive an impending market shakeout, following stock drops, Firmus scrapping its IPO, and CoreWeave losing a quarter of its market value. The contraction is highlighted by Microsoft terminating a deal with Nscale for 23,000 Nvidia chips, exposing Nvidia because non-hyperscaler buyers generate 45 percent of its data center revenue.
The Decoder - ReportAlso involving Nvidia
Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers
TechCrunch reports that Google has released a developer tool designed to help PyTorch users migrate their applications away from Nvidia GPUs, mirroring earlier efforts by Amazon to challenge Nvidia's dominance in AI developer workflows.
TechCrunch