OpenAI loses its data-center chief amid a broader exodus
OpenAI's infrastructure bench just took a hit, and forensic researchers meanwhile put hard numbers on the mystery model everyone has spent the week probing. Two stories about who — and what — is really behind the AI frontier's biggest names.
Chris Malone, who joined OpenAI in March 2025 as head of data centers, quietly left the company last week, according to a Wall Street Journal report citing people familiar with the matter — the latest departure in what the paper describes as a broader executive exodus. The role is less glamorous than a model release and arguably more load-bearing: Malone oversaw the physical backbone of OpenAI's ambitions, the gigawatt-scale campus program (Stargate) that converts the lab's roadmap into land, power contracts, and racks of accelerators. When your competitive position depends on locking up electricity and construction capacity years ahead of revenue, the person running that portfolio matters as much as any researcher. The timing is awkward, with the buildout entering its most capital-hungry phase just as rivals like Meta, Anthropic, and xAI bid for the same tiny pool of grid-scale infrastructure talent. OpenAI has not publicly commented on the departure or named a successor.
The mystery Ox Alpha model censors by blacklist, not by instinct — and its fingerprints point squarely to Zhipu's GLM line. That's the upshot of new research from AI firm CTGT, which ran the anonymously listed model through LineageEval, its matched-pair censorship instrument, after it surfaced on OpenRouter on August 20. On most topics censorship audits probe in Chinese models — Xinjiang, Taiwan — Ox Alpha answers identically to American open models, but seven topics tied to domestic Chinese political risk (including Xi Jinping personally) account for virtually all of its measured censorship, a sharply bimodal "switch" profile unlike DeepSeek V4 Flash's censor-everywhere tilt. CTGT's independent fingerprinting also landed where community sleuths already had: an exact 11-of-11 tokenizer match to the GLM-5.x vocabulary plus matching API behavior, alongside a hidden system prompt ordering the model never to reveal its origins. We covered the launch-day tests and the GLM attribution — Mystery model Ox Alpha tops GPT-5.6 — sleuths say it's Zhipu's GLM — and the practical lesson here is for auditors: a benchmark built only on foreign-interest topics would score this model as fully uncensored.
What to watch: whether OpenAI's exodus slows once its restructuring settles — and whether whoever inherits the data-center portfolio keeps Stargate's timelines intact.
Should labs be allowed to ship models whose provenance and censorship rules are both hidden? Tell us in the comments.
