OpenAI and Broadcom unveiled Jalapeño on June 24, a custom chip designed from scratch for LLM inference and targeting roughly 50% lower cost per token than current NVIDIA-based systems. When inference gets that much cheaper, the economics of automating knowledge work change significantly for every team running AI workflows at scale.
The marketing team running a 10-million-output-token-per-month research workflow on OpenAI's API spends roughly $150 to $200 a month in inference costs today. If the efficiency numbers behind Jalapeño hold and get passed through to API pricing, that same workflow could cost half as much within 18 months without any change to how it operates.
OpenAI and Broadcom unveiled Jalapeño on June 24, OpenAI's first custom-built intelligence processor, designed from scratch specifically for large language model inference rather than adapted from general-purpose GPU hardware. Early lab tests show performance per watt "substantially better than current state-of-the-art," and independent reporting points to a cost-per-token target of roughly 50% lower than current NVIDIA-based systems. The chip is already running GPT-5.3-Codex-Spark in lab conditions and is scheduled to deploy at gigawatt scale with Microsoft and other data center partners beginning by end of 2026.
The reason this matters to a business leader is not the chip itself. It is what happens to the economics of knowledge work automation when inference gets substantially cheaper.
The cost floor on every AI workflow
Every automated task that runs through an OpenAI model, whether a content pipeline, a research synthesis agent, a customer support flow, or a code review process, has a cost floor set by inference pricing. Right now that floor is determined largely by what OpenAI pays NVIDIA for GPU access, plus the premium that comes with booking scarce compute in a market where demand outpaces supply. When OpenAI controls its own silicon, that relationship changes.
The strategic move here is straightforward. NVIDIA's H100 and B200 series are general-purpose AI accelerators designed to handle the full range of possible AI workloads, from image generation to video to scientific simulations, not just transformer-based text inference. Jalapeño is a blank-slate design built entirely around the memory movement patterns, kernel operations, and networking requirements that matter specifically for LLMs. You do not pay for capabilities you are not using. That focused architecture is where the efficiency gains come from.
Greg Brockman, OpenAI's President and co-founder, said in the announcement: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access." That framing describes a real structural shift in how the cost of AI inference gets set over time, not just a hardware upgrade.
What changes for business teams
The immediate implication is that the break-even threshold for automating AI-assisted work is about to move. Right now, teams do the math on a workflow and decide whether the output value justifies the token cost. A lot of moderately useful automations, the ones that are helpful but not quite worth $300 a month in API spend, sit below that line. A 50% reduction in per-token cost does not just make existing automations cheaper. It pulls some of those marginal workflows across the line from "not worth it" to "obviously yes."
Longer agentic runs are the most affected category. A 30-step research and synthesis workflow that calls a model at each step, checks its output, and branches based on the result is expensive today at frontier model rates. When inference costs drop substantially, the economics of long autonomous work sessions improve directly. Codex-based workflows and multi-tool agents that interact with external systems, pull data, and write and review copy across a full production cycle are the workflows that get meaningfully cheaper as the per-token cost falls.
The API pricing timeline is worth watching separately from the chip announcement. OpenAI has not announced specific price changes tied to Jalapeño. The chip is scheduled for initial deployment by end of 2026, with the full production ramp in 2027 and 2028. When and how much of the efficiency gain gets passed through to API pricing is an open question. The previous pattern at OpenAI, where price reductions followed infrastructure improvements by several months, suggests the business impact will start appearing in 2027, not tomorrow.
The honest caveat
Jalapeño will not be sold to external parties. It is OpenAI's proprietary infrastructure, and the efficiency gains flow to you through lower API pricing and faster, more reliable service, not through any direct hardware access. That also means the 50% cost reduction figure is a target efficiency metric under lab conditions, not a confirmed API price reduction. The gap between chip-level performance-per-watt improvement and actual model pricing reduction depends on factors OpenAI has not disclosed, including what they choose to retain as margin improvement versus pass to customers.
There is also a dependency question. OpenAI has committed to deploying Jalapeño alongside NVIDIA hardware, not in place of it. For the next two years, the company will be running a mixed infrastructure, which limits how quickly the new chip's efficiency can shift the cost structure of the entire platform. Businesses whose AI spend decisions depend on a specific price reduction should treat Jalapeño as directional signal, not a near-term budget line item.
The closing observation
For the last three years, the business case for AI automation has been built partly on the assumption that inference prices would eventually fall the way storage, bandwidth, and compute prices always fall. Jalapeño is the first piece of evidence that OpenAI intends to drive that decline deliberately, on its own timeline, using its own hardware. The companies whose workflows are designed to get more valuable as inference gets cheaper are already positioned. Everyone else is still running costs based on a price floor that is now officially in motion.