Short Answer: OpenAI's Jalapeño chip, co-developed with Broadcom on TSMC 3nm in a record nine months, promises to cut AI inference costs by roughly 50% per token versus Nvidia GPUs. If delivered, this would be the most significant shift in AI economics since ChatGPT launched — directly reducing API bills for every developer running LLM workloads and breaking Nvidia's decade-long pricing dominance in inference hardware.
Bloomberg Television's Ed Ludlow reports on the OpenAI and Broadcom Jalapeño chip announcement (June 24, 2026).
What Is OpenAI Jalapeño?
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom "Intelligence Processor" — an inference-only ASIC designed from scratch for LLM workloads. Built on TSMC's 3nm node, the chip measures ~840mm² (near the EUV reticle limit) and packs six HBM3 memory stacks, with HBM4 already sampling from Micron.
This isn't a research project. Engineering samples are currently running GPT-5.3-Codex-Spark at target frequency and power in OpenAI's labs. The chip emerged from concept to tape-out in just nine months — believed to be the fastest ASIC development cycle in the semiconductor industry's history.
We covered the initial announcement in our earlier piece, OpenAI Jalapeño: How OpenAI's First Custom Chip with Broadcom Is Rewriting the AI Infrastructure Playbook.
The 50% Cost Target: How Jalapeño Rewrites Inference Economics
The headline number — roughly 50% lower cost per inference token versus Nvidia GPUs — comes directly from Broadcom CEO Hock Tan in a CNBC interview. OpenAI's blog uses more reserved language, describing "substantially better performance-per-watt." As Tom's Hardware notes, this is a design target — independent benchmarks have not yet been published.
OpenAI spent an estimated $8.4 billion just serving ChatGPT inference in 2025, a figure that contributed to a $3.7 billion Q1 2026 burn rate. With 900 million weekly ChatGPT users pushing costs toward $14 billion in 2026, the incentive to build cheaper inference silicon is existential.
A visual explainer of how Jalapeño's ASIC architecture delivers better performance-per-watt for LLM inference.
The Full-Stack Strategy: Chip to API
What makes Jalapeño different is the depth of vertical integration. As TechCrunch reported, OpenAI is designing not just the chip architecture but the kernels, memory systems, networking, deployment, and product experience — the entire stack from silicon to API endpoint.
Richard Ho, OpenAI's Head of Hardware Program (former Google TPU lead): "Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers."
What This Means for Developers
Cheaper API Bills Are Coming
If Jalapeño delivers even half its 50% cost target, OpenAI's API pricing drops significantly — every developer running GPT-4-scale workloads sees their inference bill shrink.
Nvidia Pricing Pressure Benefits Everyone
OpenAI committed $30 billion in February 2026 for Nvidia hardware plus a 10GW Vera Rubin agreement. By building its own inference silicon, OpenAI breaks Nvidia's pricing floor. DeepSeek's DSpark took a software approach; Jalapeño is the hardware equivalent.
Custom Silicon Is Now Table Stakes
Google has TPU v7 "Ironwood," Amazon has Trainium and Inferentia, Microsoft has Maia 200, Meta has MTIA, Anthropic is working with Broadcom on custom silicon. OpenAI's Jalapeño completes the set — every major AI company now designs its own silicon.
Timeline Reality Check
Prototype deployment begins late 2026, volume ramps through 2027, full production in H1 2028. The next 12-18 months are still Nvidia's game. Nvidia's RTX Spark superchip shows the competitive response is already underway.
The Bottom Line
OpenAI Jalapeño is more than a chip announcement — it's a declaration that inference economics must change for AI to scale. The proof will be in benchmarks and API price sheets. If Jalapeño delivers on its cost target, 2027 will be the year AI inference pricing broke open.
Featured Image: OpenAI CEO Sam Altman and Broadcom CEO Hock Tan with the Jalapeño Intelligence Processor wafer at the announcement event, June 24, 2026. Image via Tom's Hardware / Future plc.
Frequently Asked Questions
When will OpenAI Jalapeño chips be available at scale?
Initial prototype deployment begins in late 2026, volume ramping through 2027, full production in H1 2028.
How much cheaper is Jalapeño compared to Nvidia GPUs?
Broadcom CEO Hock Tan claims roughly 50% lower cost per inference token. OpenAI's blog uses "substantially better performance-per-watt." Independent benchmarks not yet published.
Did AI really design this chip?
OpenAI models accelerated testbench generation, verification, and timing closure. Physical design tasks like floorplanning and routing remained human-led.
Originally published on TekMag.
Top comments (0)