OpenAI Jalapeño Chip Crushes Nvidia Blackwell: 1.9x Throughput & 3.6x Latency Drop in First Benchmarks

OpenAI Jalapeño Chip Crushes Nvidia Blackwell in First Published Benchmarks

On August 25, 2026, OpenAI released the first benchmark results for Jalapeño, its custom inference accelerator co-developed with Broadcom. The numbers are striking: across three public open-weight models, Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency compared to Nvidia's flagship Blackwell GB300 GPU.

This isn't a marketing claim from a chip company's slide deck. OpenAI published results on SemiAnalysis' InferenceX, a public benchmark that measures the full process of serving an AI request — from prompt ingestion to final token delivery. The comparison is apples-to-apples: normalized using each accelerator's published chip power rating.

What Jalapeño Actually Is

Jalapeño is OpenAI's first Intelligence Processor — a purpose-built ASIC designed from scratch for LLM inference. It's not a general-purpose GPU adapted for AI. It's a clean-sheet design built around the specific demands of serving modern language models at scale.

Key specifications:

• 700W rated power, but sustained at 550W or below during testing
• Designed for both prefill (prompt processing) and decode (token generation) phases
• Co-developed with Broadcom (silicon, networking, Tomahawk switch integration) and Celestica (board, rack, production systems)
• Taped out in nine months — with AI accelerating the design process
• Multi-generation platform; deployment at gigawatt scale planned with data center partners starting in 2026

The Benchmark Numbers: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T

OpenAI tested Jalapeño against leading commercially available AI systems across three open-weight models that represent different architectures and scales:

The advantage widened further on frontier OpenAI models in internal testing, suggesting the architecture becomes more valuable as workloads grow larger and more demanding.

For highly interactive workloads — the kind agents need for real-time tool calling and multi-step reasoning — Jalapeño delivered 2.1 to 4.1 times higher performance.

How It Achieves This: The Full-Stack Advantage

Jalapeño's gains come from co-designing the chip, memory, network, software, and rack-scale system around real language-model workloads. Here's what's different:

1. Explicit KV Cache Placement

Language-model inference moves through distinct phases with different bottlenecks. Prefill is compute-intensive; decode is memory-bandwidth constrained. Communication adds latency when data moves between cores and chips. Jalapeño is designed to minimize data movement by explicitly placing and keeping model state — including the KV cache — local to the compute that needs it.

2. Integrated Network Architecture

The network is integral to the architecture, not an afterthought. Its large domain allows entire workloads to remain within one connected system, minimizing data movement and keeping the complete request fast from beginning to end.

3. Balanced Fungible Accelerator

Jalapeño supports changing model architectures and excels at both prefill and decode phases, adapting as the balance between them changes — a defining feature of agentic workloads where models alternate between reasoning and generating.

AI Designed the Chip, and the Chip Was Designed for AI

OpenAI used AI to design Jalapeño, enabling the team to go from initial design to tapeout in nine months. AI explored implementations, shortened design and verification loops, and optimized arithmetic circuits to fit more compute performance into the chip on schedule.

The chip was also designed as a clear, predictable programming target for both humans and AI. Engineers describe work through local tensors, explicit communication, and predictable synchronization. AI then optimizes how that work is mapped, placed, scheduled, and coordinated across the system.

The results speak for themselves: Using Codex with GPT-Astra, the team brought three open-weight models to high performance within two months. For selected attention and mixture-of-experts blocks, AI-generated implementations ran 1.5 to 1.8 times faster than existing human-expert-written implementations.

What This Means for the AI Industry

The Nvidia Moat Narrows

Nvidia's dominance has been built on the combination of CUDA software ecosystem and hardware performance. Jalapeño demonstrates that a vertically integrated company can design silicon that outperforms Nvidia on the specific workloads that matter most for AI inference. The 700W Jalapeño beating the 1,400W Blackwell is a power-efficiency story that changes the economics of deploying AI at scale.

Agent Economics Shift

For AI agents that need to complete many sequential steps, latency compounds across an entire task. Jalapeño's 2.1-4.1x advantage on interactive workloads directly translates to faster agent completion times and lower cost-per-task. This is the kind of infrastructure advantage that could determine which companies can afford to run always-on agent fleets.

The Custom Silicon Race Accelerates

OpenAI is now a chip company. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. The trend toward custom AI silicon is accelerating, and Jalapeño's results prove the approach works for inference — the workload that accounts for the majority of AI compute spending.

Production Ramp Timeline

OpenAI plans to begin deploying Jalapeño in production before the end of 2026, with a multi-year ramp as chips are integrated into more of OpenAI's infrastructure. The deployment will be at gigawatt scale with Microsoft and other data center partners.

The first Jalapeño chips have already run ML workloads in the lab at production target frequency and power, including GPT-5.3-Codex-Spark. A detailed technical report on performance will be presented in the coming months.

The Bottom Line

Jalapeño's first benchmarks represent a watershed moment in AI infrastructure. OpenAI has demonstrated that a custom inference chip, designed from the ground up for LLM workloads and co-optimized with the models it serves, can significantly outperform the best commercially available hardware. For the AI industry, this means faster inference, lower costs, and a more competitive hardware landscape. For enterprises running agent fleets, the power-efficiency gains could be the difference between profitable and unprofitable AI deployment.

Frequently Asked Questions

What is OpenAI's Jalapeño chip?

Jalapeño is OpenAI's first custom AI inference accelerator, co-developed with Broadcom. It's a 700W ASIC designed specifically for LLM inference — not a general-purpose GPU. It was taped out in nine months with AI assistance and delivers industry-leading performance per watt for serving language models.

How does Jalapeño compare to Nvidia Blackwell?

In benchmarks on SemiAnalysis' InferenceX, Jalapeño delivered 1.5-1.9x more AI work per watt and 1.7-3.6x lower end-to-end latency than Nvidia's Blackwell GB300. The advantage is especially pronounced on interactive workloads, where Jalapeño achieved 2.1-4.1x higher performance.

What models was Jalapeño tested on?

OpenAI tested Jalapeño on three public open-weight models: GPT-OSS 120B (transformer), DeepSeek R1 670B (MoE), and Kimi K2.5 1T (massive MoE). The architecture's advantage widened further on frontier OpenAI models in internal testing.

When will Jalapeño be available in production?

OpenAI plans to deploy Jalapeño in production before the end of 2026, with a multi-year ramp at gigawatt scale with Microsoft and other data center partners.

Did AI help design the Jalapeño chip?

Yes. OpenAI used its own AI models to accelerate the chip design process, from initial design to tapeout in nine months. AI explored implementations, optimized arithmetic circuits, and generated kernel implementations that ran 1.5-1.8x faster than human-expert code.