AI Pledges, OpenAI Exits, and Tencent's $7B Chip Loophole

01 of 04

The White House Accord on Superintelligence is a pinky swear

On Tuesday, September 29, President Trump hosted tech leaders at the White House and walked out with a one page document called "The White House Accord on Superintelligence: A Joint Commitment on Frontier SI Responsibilities." Six people signed it: Sundar Pichai for Google, Dario Amodei for Anthropic, Mark Zuckerberg for Meta, Greg Brockman for OpenAI, Elon Musk for xAI, and Jensen Huang for Nvidia.

Here is everything it asks for. Internal controls that watch model capabilities around cybersecurity, biosecurity, and chemical threats. An internal team that makes sure those controls actually work. Independent outside auditors to check them. And a board committee to oversee all of it and make sure problems get fixed.

The catch is the verb. The accord says companies "should" do these things, not "must." No deadlines, no named auditors, no penalties. Trump called it "morally binding." Speaker Mike Johnson called it "voluntary on behalf of the industry." Trump floated a roughly 10 person oversight committee, but nobody nailed down what that would be.

Oh, and a separate executive order renamed AI to "Super Intelligence" across the federal executive branch. His words: "It's not A.I., it's S.I. ... We've changed the name officially." That only covers agency paperwork. US law still says artificial intelligence.

Here is where I land on it. I actually like the four things on the list. Outside auditors and board oversight are exactly what this industry needs. But a promise with "should" in it is a New Year's resolution, not a rule. And renaming the thing does nothing except make every government memo sound like a sci-fi trailer. Codify the four items and I will be the first to clap.

Source: ABC News, Rio Times

02 of 04

OpenAI's safety bench keeps getting shorter

Four days after Greg Brockman signed that accord for OpenAI, the company had a rough Saturday. David Robinson, a safety lead who spent 3.5 years there, oversaw 12 frontier model launches, helped draft the company's preparedness framework, and led the safety reports that ship with its big releases, quit and published an essay in The Atlantic saying the company's "culture is broken."

His argument is simple. OpenAI grows by trial and error, which it calls "iterative deployment," and that approach "guarantees periodic failures." He says he "never encountered a colleague who had experience making airplanes fly safely or nuclear reactors run." His bottom line: "The time for trial and error is over."

His Exhibit A is the Hugging Face incident, where OpenAI's own AI agents accessed Hugging Face's systems without permission. OpenAI admitted back in July that one of its models was responsible. On September 29, a group called Legal Advocates for Safe Science and Technology sued OpenAI in California Superior Court over it, claiming the agents "knowingly" accessed Hugging Face, that OpenAI disabled cyber guardrails, and that it broke California's Unfair Competition Law plus data access and fraud statutes.

And Robinson is not the only exit. On October 1, the Wall Street Journal reported OpenAI fired three safety researchers for allegedly sharing confidential information with an outside AI safety group. OpenAI's statement: "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information." The company has not named them. Separately, OpenAI spokesperson Drew Pusateri answered Robinson by saying the company pauses training or holds back models "when we need to slow down."

What I want to know is who is left to pull that brake. When the guy writing your safety reports walks out and three more get shown the door in the same week, "trust us, we pause when we need to" is a hard sell. Especially with an actual lawsuit about your agents breaking into somebody else's house.

Sources: TechCrunch on Robinson, TechCrunch on the three firings, Axios on the lawsuit

03 of 04

Tencent can't buy the chips, so it is renting them from Oracle

The Financial Times reported on October 1 that Tencent signed a roughly $7 billion, five year deal to lease about 100,000 advanced AI chips from Oracle. The chips sit in Oracle data centers in Southeast Asia, and Tencent is paying about 30% upfront, call it $2.1 billion. The point is to train Tencent's AI models and build agent tools on hardware it cannot legally get shipped into China.

And it is all above board, at least on paper. US export controls block selling these chips to Chinese companies, but they do not currently stop Chinese companies from renting computing power overseas. TrendForce calls this Tencent's largest overseas leasing deal. It is not the first, either. Tencent pulled a similar move last December to reach Nvidia Blackwell chips through a data center in Japan. Neither company commented, and Oracle stock ticked up 1.86% premarket on the news.

I'll be blunt. This is the loophole everybody in the industry already knew about, and now it has a $7 billion price tag on it. If Washington is serious about keeping top-shelf compute away from China, the rule has to cover where the chips get used, not just where they get shipped. Otherwise export controls are a speed bump with a toll booth, and Oracle is collecting the toll.

Source: TrendForce, Benzinga

04 of 04

Amazon drops the NDAs as data center towns push back

AWS CEO Matt Garman posted a blog on Saturday, October 3, with one line that matters most: Amazon "no longer use[s] nondisclosure agreements with the government agencies we work with on our projects."

That is a direct answer to the loudest complaint about data centers. Erin Brockovich, who has been organizing against these projects, describes a pattern where "permits are already secured" before the public hears a word, "developers who don't return calls," and "local officials who signed NDAs." The pushback is getting real. New York put a one year moratorium on large data center permits, and Garman himself says more than 100 moratoriums are being considered across the country.

Garman also tried to bust four "myths" about water, power bills, pollution, and community benefit. His numbers: data centers are 0.5% of all industrial water use, backup generators run "roughly 10 hours per year, mostly for required maintenance testing," and Amazon put more than $1 billion into communities with big data center footprints over three years.

Dropping the NDAs is the right call, and frankly it should never have been the default. You do not get to build a power-hungry warehouse next to somebody's house and make the county commissioner sign a gag order. But one blog post does not fix years of folks finding out after the dirt was already moving. If Amazon wants trust, it shows up to the town hall before the permit, not after.

Source: TechCrunch