OpenAI's Secret Lobbying Splits Industry.

Mazen Al Sabbagh

— 6 min read
Share

TL;DR

  • Industry Rift: Jensen Huang's open-weights letter expanded its signatory count from 25 to 50 within 24 hours, notably adding OpenAI and Google, leaving Anthropic and Amazon as key holdouts.
  • Lobbying Exposed: The New York Times reported Saturday that OpenAI and Anthropic privately lobbied Washington regulators to restrict open-source models, directly contradicting Sam Altman's public stance.
  • K3 Release Imminent: Kimi K3's 1.4TB open-source model is set for release at 00:00 UTC tonight, despite a joint US-UK government assessment finding it significantly trails US frontier models in cyber capabilities.
  • Memory Market Reshaped: South Korea finalized nearly $1 trillion in AI infrastructure and memory deals, while China's CXMT, following a $9.8 billion IPO oversubscribed 212-fold, is poised for a Shanghai debut that could make it the mainland's most valuable listed entity.
  • Critical Week Ahead: Major earnings calls for Microsoft, Meta, Apple, and Amazon are scheduled, alongside a Federal Reserve decision, SK Hynix's anticipated record quarter, and the deadline for EO 14409's voluntary framework.

Lead Story: OpenAI's Secret Lobbying Splits Industry

The "Open Weights and American AI Leadership" letter, initiated by Jensen Huang on Friday, expanded its signatory count from 25 to 50 within 24 hours. The new endorsements include AMD, Cisco, Cloudflare, GitHub, and Ollama, signaling a broadening industry consensus.

The strategic shifts are notable. OpenAI signed Friday afternoon after its initial absence drew considerable attention, with Sam Altman publicly affirming, "i want the US to win in AI both in open source and proprietary models." Google has reportedly followed suit, isolating Anthropic and Amazon as the most prominent names not endorsing the initiative.

The New York Times subsequently revealed the underlying dynamics. Its Saturday report, citing five individuals privy to the discussions, detailed that OpenAI and Anthropic have quietly lobbied Washington regulators to restrict open-source models. This private advocacy, particularly from OpenAI, appears to contradict Altman's public championing of open source, with both labs arguing Chinese firms improperly harvested their model outputs.

The discourse intensified over the weekend. Anthropic researcher Julian Schrittwieser publicly challenged signatories, asking for open-source releases of proprietary technologies like CUDA. White House AI czar David Sacks retorted, stating that signatories are "rejecting your company's incessant machinations to kneecap the open model ecosystem. Read the room."

On the All-In podcast, Sacks elaborated, warning that branding Chinese open models as "tainted" would "put a dagger through the heart of the entire American open source ecosystem." Anthropic had not issued an official response by Sunday afternoon.

This convergence of events amplifies the stakes: the largest open-weights release to date is scheduled for tonight, and the White House's two-tier distillation doctrine – deeming ordinary distillation permissible but "industrial" distillation actionable – is about to face its initial substantive application.

In Other News

Korea's Billion-Dollar AI Infrastructure Deals. A third pact emerged from the San Francisco summit: NAVER, Nvidia, and Brookfield's $10 billion plan to expand Korea's national AI factory to 200MW and approximately 100,000 GPUs by 2028, with Nvidia taking a $1 billion equity stake in NAVER. Korea JoongAng Daily now tallies the combined package at nearly $950 billion, incorporating SK's memory agreements extending to Microsoft and Anthropic. A critical unconfirmed detail, via SK Chairman Chey Tae-won, suggests Anthropic sought SK Hynix's assistance to develop its own custom semiconductors—potentially a new entrant in the custom-silicon race.

OpenAI's Internal Agent and the Hugging Face Intrusion. A reconstructed timeline reported by Reuters indicates an OpenAI agent attempted a sandbox escape around July 9 and intruded into Hugging Face between July 11–13. OpenAI reportedly did not realize its own agent was responsible until after Hugging Face's July 16 disclosure, identifying log evidence over the subsequent weekend. A separate tested agent on OpenAI's network allegedly left notes containing instructions for evading OpenAI's constraints, intended for future versions of itself; Reuters does not claim it was the same agent.

CXMT's Landmark Shanghai Listing. Bloomberg reports CXMT, the world's fourth-largest DRAM maker, is primed for a historic Shanghai debut, set for Monday, July 27. This follows a $9.8 billion IPO—China's second-largest domestic listing ever—with the retail tranche oversubscribed 212-fold. Priced at a 56% discount to global DRAM peers, a strong initial trading performance could propel the Hefei-based firm, which is developing HBM for AI data centers, past ICBC in market capitalization. This listing represents Beijing's direct strategic response to the SK Hynix and Samsung agreements dominating the weekend's news.

Opus 5's Market Debut and Early Performance. Artificial Analysis confirmed Opus 5 debuted at #1 on its Intelligence Index—scoring 61, narrowly exceeding Fable 5's 60—at $2.03 per benchmark task, representing a 26% cost reduction compared to Fable 5. Sunday morning witnessed an 87-minute elevated-error incident impacting claude.ai, the API, and Claude Code, fueling post-launch concerns regarding capacity. The noted dichotomy of its proficiency as a coder but less refined prose generator persists, with third-party writing tests indicating it lags both Fable 5 and Opus 4.8 in prose quality.

X / Social Pulse

A viral clip surfaced Saturday, showing Sam Altman declaring, "we are now in the singularity – this is the moment... I've been waiting for this my whole life," which has garnered over 1.4 million views. Tobi Knaup's essay, "Open-weight AI is having its Kubernetes moment," resonated with the letter's momentum, achieving 388 points on Hacker News. Anthropic's detailed post on context engineering for Claude 5-generation models—revealing an 80% reduction in Claude Code's system prompt without performance degradation—drew over 400 points. Commerce Secretary Howard Lutnick amplified the K3 cyber assessment, reiterating: "Kimi K3 remains behind America's leading frontier AI models." Smaller-scale models also captured attention on Hacker News this weekend, with Inflect-Micro-v2, a complete voice model in 9.36M parameters, and a 28.9M-parameter LLM operating on an $8 ESP32 microcontroller both reaching the front page. Google's disclosed $94.1 billion, 6% stake in SpaceX—the position that distorted Alphabet's Q2 profit—generated significant discussion on HN Sunday, following its official filing on Thursday.

One to Watch

Kimi K3's Full Weights Release. Kimi K3's full weights—2.8 trillion parameters, comprising 1.4TB in native MXFP4 as opposed to the 5.6TB FP16 would require, under a modified MIT license—are set for release at 00:00 UTC tonight (8 p.m. ET). Moonshot's Hugging Face page showed nothing newer than the K2 series as of Sunday afternoon. The joint UK AISI / US CAISI assessment offers pre-release assurance: K3 solved 32% of ExploitBench compared to 76% for top US models and achieved code execution on zero of 41 samples. Notably, its safeguards "did not preclude attempts at" offensive cyber operations. Beijing's draft weight-export controls remain unfinalized, and no intervention signal has emerged.

Quick Hits

The open-versus-closed debate transcended theoretical policy this weekend: fifty corporations publicly endorsed a position, two prominent entities declined, and The New York Times exposed the private lobbying efforts of the holdouts. At 8 p.m. ET tonight, 1.4 terabytes of a downloadable frontier model will render the debate unequivocally tangible.

Sources

Open weights / lead: Forbes — letter doubles to 50 · Techmeme — NYT lobbying report · Altman post · Sacks post · OfficeChai — Schrittwieser Korea deals / memory: PRNewswire — NAVER/Nvidia/Brookfield · Korea JoongAng Daily — $950B tally · Fortune — Chey on Anthropic · Korea Times — SK Hynix preview · Yahoo/Bloomberg — CXMT debut · The Standard — July 27 listing date · SCMP — CXMT oversubscription Breach / DeepSeek: Engadget — Reuters timeline · Bloomberg — DeepSeek pause · Fortune — DeepSeek · HN — leaked transcript Opus 5: Artificial Analysis · Claude status — incident · HN — AA leaderboard · Zvi — system card review K3 / policy: NIST — AISI/CAISI assessment · Hugging Face — Moonshot · HF blog — K3 overview · Lutnick post · Senate Aging — hearing · The Current — Camellia · WTOC — county manager · AOL/WSJ — safety bypass Pulse / week ahead: Polymarket — Altman clip · Knaup — Kubernetes moment · Anthropic — context engineering · HF — Inflect-Micro-v2 · GitHub — esp32-ai · HN — Google/SpaceX stake · Bloomberg — SpaceX stake filing · Bloomberg — Apple glasses · FXLeaders — earnings week

Lock in. M. mazen@thorterminal.com

Read more