Thursday, July 23, 2026

OpenAI had a busy day too, shipping Presence (a governed enterprise voice/chat agent platform) and unveiling a $30B, 3.2GW self-built Georgia data center. And there's a hard deadline for engineers today: OpenAI retires ~15 legacy model snapshots (Codex, deep-research, chat, realtime, computer-use).

TLDR

White House accuses Moonshot of distilling Anthropic's Fable for Kimi K3; Treasury threatens sanctions — OSTP's Kratsios calls it covert, industrial-scale IP theft (Jul 22) Policy Trending
OpenAI launches Presence — a managed enterprise platform for governed voice/chat agents with guardrails, escalation, and Codex-driven tuning (Jul 22) Agents
~15 legacy OpenAI model snapshots retire today — Codex, deep-research, chat-latest, realtime, computer-use aliases stop resolving (Jul 23) Pricing
OpenAI plans a $30B, 3.2GW Georgia data center ("Project Camellia") — its first self-designed/built campus, online from 2028 (Jul 22) Infra
2026 tech layoffs cross ~205K workers; 54% of events cite AI as a driver, per running trackers (Jul 23 YTD) Hiring

Action Items

Things your engineering team should look at or do this week.

Check whether any code pins the OpenAI model snapshots retiring today (Jul 23). Five pre-5.3 Codex variants (gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.2-codex), gpt-5-chat-latest, gpt-5.1-chat-latest, both deep-research variants (o3-deep-research-2025-06-26, o4-mini-deep-research-2025-06-26), plus assorted realtime/search/computer-use aliases stop resolving. Grep your codebase and CI for these strings and repin to current snapshots. (OpenAI deprecations)
Reassess reliance on open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ Chinese models given the Moonshot/Treasury escalation. With Kimi K3 full weights still slated for Jul 27 and sanctions now on the table, teams building on GLM/Kimi/Qwen should have a licensing and continuity contingency in case distribution or support gets disrupted. (TechCrunch)

Models & Tools

OpenAI launches Presence, a governed enterprise agent platform (Jul 22) Agents Customer AI — A managed system for building and running production voice/chat agents, pairing model reasoning with company-defined policies, guardrails, testing, and human-escalation rules. Each deployment is scoped to a specific job (billing, claims, IT service) with least-privilege data/system access, and uses Codex to analyze signals and propose updates that humans approve before rollout. OpenAI says Presence already runs its own English phone support and resolves 75% of inbound calls autonomously; it's in limited rollout to selected enterprise customers. (OpenAI | VentureBeat)

APIs & Pricing

OpenAI retires ~15 legacy model snapshots today (Jul 23) Pricing — A batch shutdown covering five pre-5.3 Codex variants, gpt-5-chat-latest/gpt-5.1-chat-latest, the two deep-research snapshots (o3/o4-mini), plus search, audio, realtime, and computer-use aliases. OpenAI posted the "Legacy GPT model snapshots" notice Apr 22, honoring its ~3-month runway for specialized variants. Calls to retired names will fail — repin now. See Action Items. (OpenAI deprecations)

Industry & Business

OpenAI plans a $30B, 3.2GW Georgia data center — its first self-built campus (Jul 22) Infra — "Project Camellia" at the Savannah Gateway Industrial Hub in Effingham County is OpenAI's first facility it will design and build itself (vs. leasing from Oracle/AWS), with a Georgia Power contract for 3.2GW. Initial cost ~$20B, potentially $30B+ at full scale; several hundred MW expected online from 2028, buildout through 2032. OpenAI is seeking financing/build partners and pledged ~$80M in local benefits plus Codex credits for Georgia college students. (Bloomberg | Data Center Dynamics)
VC roundup, Jul 22 (Jul 22) — Defense-tech startup Cathedral (ex-DOGE founders) closed $160M at a $1.4B valuation, co-led by a16z and Sequoia; healthcare-infrastructure firm Candid Health raised a $120M Series D (Sixth Street Growth, Oak HC/FT, 8VC), reaching ~$220M total. Signal: mega-rounds keep concentrating in AI-adjacent infra, defense, and health. (Tech Startups)
AI-cited tech layoffs keep mounting in 2026 (Jul 23 YTD) Hiring — Running trackers put 2026 at ~322 layoff events / ~205K workers, with ~54% of events explicitly naming AI/automation. Roles with high AI overlap (programming, support, content, data entry) are most exposed; ML infra, AI safety, and applied research stay in demand. Useful context for headcount and reskilling planning. (TechCrunch)

Policy & Safety

White House accuses Moonshot of distilling Anthropic's Fable to build Kimi K3; Treasury threatens sanctions (Jul 22) Policy Trending — OSTP Director Michael Kratsios alleged, in a post on X, that Moonshot ran "covert, industrial-scale" distillationTraining a smaller model to copy a larger model's outputs, so it is cheaper to run.Hinton et al. 2015 ↗ of Anthropic's Fable — using a purpose-built internal system and rotating access routes to stay hidden — to train Kimi K3, the 2.8T-parameter open-weight model that topped several coding arenas. Treasury signaled possible sanctions; Anthropic earlier claimed it traced 3.4M Claude exchanges to the startup. Moonshot hasn't responded, and full K3 weights are still due to go public Jul 27. A major escalation in the US–China model-IP fight worth watching for anyone deploying open-weight Chinese models. (TechCrunch | CyberScoop)

The US–China AI fight went nuclear on IP: the White House publicly accused Moonshot AI of "industrial-scale" distillation of Anthropic's Fable to build Kimi K3, and Treasury threatened sanctions — the sharpest escalation yet over open-weight Chinese models.