OpenAI fixes Codex usage limits and unveils custom AI chip

OpenAI announced on June 29 that it has fully reset Codex usage limits for all users and deployed fixes to stop background compute from draining credits faster than intended, while simultaneously unveiling its custom “Jalapeño” inference processor—a move that mirrors Apple’s vertical‑integration strategy and could reshape reliance on Nvidia hardware【1】.

Codex limit reset and underlying cause

Late‑Monday OpenAI engineering lead Thibault Sottiaux explained that Codex’s auto‑review feature and helper “subagents” were executing twice or retrying aggressively, consuming more compute than the dashboard reflected【1】. Users on the $200 “Pro‑20x” plan reported that a week’s quota vanished in two to three days, prompting a weekend “warroom” and an across‑the‑board cap reset【1】. The fix not only restores the original quota but also adds detailed monitoring to catch similar regressions sooner.

Jalapeño chip and the push for vertical integration

In a separate announcement, OpenAI disclosed details of its Jalapeño custom inference processor, designed to run AI models more efficiently than off‑the‑shelf GPUs【2】. The chip targets inference workloads—the billions of queries that power ChatGPT and other services—rather than the costly training phase. By tailoring silicon to its own models, OpenAI hopes to reduce latency, cut power consumption, and lessen dependence on Nvidia’s GPUs, which currently dominate the AI hardware market【2】. The strategy echoes Apple’s decade‑long shift to in‑house silicon, where tighter hardware‑software coupling gave the company pricing and performance leverage【2】.

Competitive context

Nvidia remains the primary supplier for most AI workloads, and OpenAI continues to be a major customer despite the new chip’s early stage【2】. Other AI leaders—Google, Amazon, Microsoft, Meta—have similarly invested in custom accelerators, indicating an industry‑wide trend toward hardware self‑sufficiency. OpenAI’s Jalapeño is still in development, with broad deployment “some way off,” but its existence signals a long‑term intent to own more of the AI stack【2】.

What to watch

• Jalapeño rollout timeline – monitor OpenAI’s updates for the first production deployment date.
• Nvidia’s response – watch for any pricing or supply adjustments aimed at retaining OpenAI as a key client.
• Codex usage metrics – follow OpenAI’s monitoring reports to see whether the reset stabilizes user quotas over the next weeks.

The Codex fix restores confidence for developers relying on OpenAI’s coding assistant, while the Jalapeño chip hints at a future where AI providers control both software and hardware, potentially reshaping the economics of AI services. The next months will reveal whether OpenAI can translate its vertical‑integration ambition into tangible performance gains and reduced hardware dependence.