What To Know
- OpenAI has fired a significant new shot in the global AI hardware race, revealing benchmark results for Jalapeño, its first custom inference chip developed with Broadcom.
- OpenAI presented its first measured Jalapeño results at the Hot Chips conference on August 25, testing the processor against Nvidia GB200 and GB300 systems across three open-weight AI models.
OpenAI has fired a significant new shot in the global AI hardware race, revealing benchmark results for Jalapeño, its first custom inference chip developed with Broadcom. The semiconductor is designed not to train massive artificial intelligence models, but to run them efficiently once deployed, potentially giving OpenAI greater control over one of the most expensive parts of operating AI services at global scale.
Image Credit: Thailand AI News
The early numbers are striking. OpenAI says its 700-watt Jalapeño processor delivered substantially better speed and performance per watt than Nvidia Blackwell systems rated at considerably higher power levels. As demand for ChatGPT-style services and AI agents accelerates, this AI Hardware news report highlights why those efficiency gains could become strategically important: electricity, cooling capacity, and available data center power are increasingly determining how much AI computing companies can actually deploy.
Jalapeño Delivers Impressive Benchmark Numbers
OpenAI presented its first measured Jalapeño results at the Hot Chips conference on August 25, testing the processor against Nvidia GB200 and GB300 systems across three open-weight AI models.
According to the reported results, Jalapeño produced approximately 1.5 to 1.9 times more AI work per watt at peak throughput while delivering between 1.7 and 3.6 times lower end-to-end latency. OpenAI argues that performance per unit of power is increasingly more meaningful than simply comparing performance per chip.
That distinction matters because Jalapeño carries a 700-watt rating, compared with 1,200 watts for Nvidia’s GB200 and 1,400 watts for the GB300.
At data center scale, lower processor power requirements can translate into savings across electricity, cooling systems, and power distribution infrastructure. Alexander Harrowell, senior principal analyst at Omdia, described Jalapeño as particularly impressive in terms of efficiency and said large deployments could materially improve OpenAI’s unit economics.
Power Is Becoming AI’s Critical Bottleneck
The importance of those numbers extends beyond electricity bills.
SemiAnalysis, which conducted benchmarking alongside OpenAI engineers, reported that OpenAI is constrained more by available data center power than by capital or physical floor space. That reflects an industry-wide problem as AI infrastructure consumes rapidly increasing amounts of electricity.
International Energy Agency figures cited in the supplied material show data center electricity demand rising 17% during 2025, while demand from AI-focused facilities surged 50%.
In that environment, every watt becomes valuable computing capacity.
Nvidia CEO Jensen Huang has similarly emphasized the economics of throughput per watt. For an operator working within a fixed gigawatt power envelope, generating more AI output from the same electricity can effectively mean generating more revenue without waiting for additional grid capacity.
Jalapeño therefore represents more than another processor benchmark. It is OpenAI’s attempt to optimize hardware specifically around its own inference requirements and infrastructure constraints.
But the Nvidia Comparison Has Important Caveats
The benchmark results do not mean Nvidia has suddenly been overtaken.
SemiAnalysis described comparisons between Jalapeño and Blackwell as somewhat incomplete because OpenAI’s processor uses newer HBM4 memory, while the tested GB200 and GB300 systems use HBM3E.
Nvidia’s newer Vera Rubin architecture also uses HBM4 and consequently provides a more comparable technological baseline. SemiAnalysis reported that Jalapeño and Rubin produce nearly equivalent output tokens per dollar under its comparison, although differences in techniques such as speculative decoding complicate direct comparisons.
There is also a major difference in availability. Nvidia’s Rubin systems are beginning to ship, while Jalapeño remains at the engineering-sample stage. OpenAI hardware chief Richard Ho has indicated deployment should begin near the end of 2026 in very small volumes before production expands through 2027.
OpenAI also continues to depend heavily on Nvidia hardware for model training and other compute-intensive frontier workloads.
Nvidia Faces a Growing Custom-Silicon Challenge
Even with those qualifications, Jalapeño demonstrates how rapidly the competitive environment surrounding Nvidia is changing.
OpenAI is hardly alone. Google has developed generations of Tensor Processing Units, Amazon offers Trainium processors, and Meta is pursuing custom AI hardware with Broadcom technology. Anthropic has also made major commitments involving Amazon’s custom chips.
Omdia expects custom application-specific integrated circuits, or ASICs, such as Jalapeño to exceed GPUs in unit volume by 2028, although GPUs could retain substantially greater revenue because of their higher prices and broad capabilities.
For Nvidia, the biggest danger may not be one competitor replacing its GPUs. Instead, major customers could gradually move high-volume inference workloads onto processors designed internally for their own needs.
OpenAI has historically been one of Nvidia’s largest GPU customers. Moving even part of its enormous inference demand onto proprietary hardware could reduce costs for OpenAI while applying pressure to Nvidia’s inference margins.
A Nine-Month Development Cycle Changes the Equation
Perhaps Jalapeño’s most consequential achievement is how quickly it was designed.
OpenAI says its models helped engineers move from initial design to manufacturing-ready tapeout in nine months. The broader hardware program took longer, but the compressed design phase illustrates how AI itself could accelerate semiconductor development.
OpenAI reportedly used its upcoming Astra model and Codex during development. That creates an intriguing feedback loop: AI assists engineers in designing specialized chips, those chips run future AI systems more efficiently, and knowledge from operating them informs subsequent hardware generations.
OpenAI is already developing second- and third-generation Jalapeño processors.
The company also designed Jalapeño as a relatively flexible inference ASIC rather than restricting it to one model or a rigid traffic pattern. According to the supplied research, engineers were able to bring three open-weight models outside the original production plan to high performance within approximately two months.
The AI Chip War Is Entering a New Phase
Jalapeño does not eliminate OpenAI’s need for Nvidia, nor do early engineering benchmarks guarantee the same advantages after deployment at enormous scale. Nvidia retains a formidable position through its GPU performance, CUDA software ecosystem, programmability, installed base, and ability to support both training and inference workloads.
Yet the strategic direction is increasingly clear. The world’s biggest AI companies no longer want to depend entirely on general-purpose accelerators supplied by someone else. As power availability becomes a fundamental constraint on AI expansion, owning optimized silicon can deliver economic and operational advantages that extend far beyond headline benchmark scores.
If Jalapeño performs at production scale as its early results suggest, OpenAI could gain faster inference, lower power requirements, more responsive AI agents, and tighter integration between future models and the hardware running them.
Nvidia remains extraordinarily difficult to displace, but its largest customers are increasingly becoming chip designers themselves—and that could reshape the economics of AI infrastructure for years to come.
For more details, visit:
https://openai.com/index/jalapeno-first-results
