OpenAI Announces Jalapeño Custom Chip Co-Designed with Broadcom and Locks Up 40% of Global DRAM Wafer Supply Through 2029

OpenAI unveiled Jalapeño, a custom AI chip co-designed with Broadcom, purpose-built for the LLM workloads that power ChatGPT, its Codex coding agent API, and what the company calls “future agentic products.” Simultaneously, OpenAI secured supply agreements covering 40% of global raw undiced DRAM wafer output through 2029, according to Yahoo Finance.

The moves represent a sharp pivot from model lab to vertically integrated infrastructure company. OpenAI now controls or has committed to controlling significant portions of its own silicon design, memory supply, and inference stack.

Agents in the Chip Design Loop

The most notable detail about Jalapeño is the design process itself. According to Yahoo Finance, OpenAI positioned Jalapeño as “the first chip designed with AI agents in the loop,” with engineers leveraging agents to write instruction sets faster during the chip design phase. This is a recursive application of OpenAI’s own technology: agents accelerating the hardware that will eventually run agents.

The chip targets LLM inference workloads specifically, optimizing for the token processing, batching, and latency patterns that characterize production agent deployments. Current AI inference at OpenAI’s scale runs predominantly on Nvidia GPUs, and custom silicon represents an effort to reduce both per-query cost and dependency on a single hardware supplier.

The DRAM Supply Play

The 40% DRAM wafer commitment through 2029 is a supply-chain bet on sustained inference demand. Raw undiced DRAM wafers are the upstream input for the memory chips that feed AI accelerators during inference. By locking supply at the wafer level rather than purchasing finished memory modules, OpenAI is hedging against the kind of supply constraints that have periodically disrupted AI infrastructure buildouts.

The deal makes Micron Technology a direct beneficiary, according to Yahoo Finance, given its position as a leading DRAM manufacturer.

The Vertical Integration Trajectory

OpenAI joins a small group of AI companies pursuing custom silicon. Google has run its TPU program for over a decade. Amazon designs its Trainium and Inferentia chips for AWS. Meta has invested in custom inference hardware. But OpenAI’s approach is distinct in scope: combining custom chip design, agent-assisted development, and raw material supply agreements simultaneously.

For agent infrastructure operators, the cost implications matter most. Custom silicon optimized for LLM inference could reduce the per-token economics that currently constrain how many autonomous agents can run at production scale. If Jalapeño delivers meaningful cost or latency improvements over general-purpose GPUs for inference, it changes the unit economics for every product OpenAI ships, from ChatGPT subscriptions to Codex API calls to whatever “future agentic products” the company has planned.

The DRAM lock-up suggests OpenAI expects inference demand to grow substantially through at least 2029, and is willing to commit capital years in advance to ensure supply keeps pace.