Open models skip US safety tests + Texas audits data centers
The White House told AI labs its safety framework will not test open-weight models, including China's. Texas separately ordered an audit of every data center in its grid queue. Plus 6 more stories.
Listen to this brief
Audio & video are paid features
Plus unlocks audio streaming and PDF downloads. Pro adds offline MP3 downloads, video, certificates, and more.
- Audio streaming
- Downloadable PDFs
- All AI Playbooks
- Personalized content
- Certificates of completion
- Audio MP3 downloads
- Video lessonssoon
- & More…soon
Watch this brief
The Trump administration has settled which artificial intelligence models the United States will ask to test before release, and open-weight models are not among them — not America's, and not China's. Texas moved the opposite way on infrastructure, ordering a full audit of every data center waiting to plug into its grid. Compute deals, a moderation model and a memory standard fill out the rest.
- 1
The White House safety framework will not test open-weight models, including China's
Officials walked executives from Meta, Nvidia, Microsoft, OpenAI and Anthropic through the completed frontier-model framework, and its shape is now clear. Only closed, proprietary American models that hit top scores on cybersecurity benchmarks are asked to submit for government testing — voluntarily, up to thirty days before release. Open-weight models are exempt, and Bloomberg reports that officials told American firms the open-weight models coming out of China will not be tested either. The framework itself will not be published; Chris McGuire of the Council on Foreign Relations called that baffling, arguing the world cannot run on secret, voluntary rules for its most important technology.
- 2
Texas will audit every data center waiting to connect to its power grid
Governor Greg Abbott directed the state's utility commission and its grid operator to verify and audit each data center moving through the interconnection queue, and any project that fails the requirements will be denied a connection. That queue now holds roughly 474 gigawatts of requests, more than five times the state's record peak demand, and about ninety percent of it is data centers. Applicants must disclose projected power draw, on-site generation, water sourcing and cooling method, noise and lighting mitigation, tax incentives received, and who actually owns the project. An earlier voluntary survey went mostly unanswered.
- 3
Anthropic signs a $10 billion, six-year compute deal with a startup founded this year
Volta, founded in January by former Brookfield Asset Management executives, leases artificial intelligence capacity and helps customers finance the chips behind it. It will supply Anthropic under a six-year agreement worth $10 billion, starting with a 133 megawatt site in Norway built with the crypto-mining firm Bitdeer and running Nvidia's Vera Rubin systems. Volta raised $300 million recently, valued at $2.4 billion, so the contract is worth more than four times the company selling it — the same lease-the-compute structure that now sits underneath much of Anthropic's capacity.
- 4
INTERPOL links artificial intelligence to more than half of Africa's reported cybercrime
The African Cyberthreat Assessment Report 2026, drawn from 36 member countries, finds that 55 percent of reported cybercrime on the continent is now AI-enabled, and that losses climbed from $192 million in 2024 to $484 million in 2025. Neal Jetton, who directs INTERPOL's cybercrime unit, said AI is automating every stage of an attack, from reconnaissance and phishing through extortion and evasion. Business email compromise aimed at victims in Europe and North America is the clearest case, and 72 percent of the countries surveyed reported scam centres operating on their soil.
- 5
Mistral ships Shieldstral, a 3 billion parameter open-weights moderation model
Shieldstral judges text, images, and text-image pairs against moderation policies written in plain language at inference time, rather than against a fixed set of harm categories frozen in during training. A platform can therefore change what counts as unsafe by rewriting a sentence instead of retraining a classifier. Mistral says the model matches or beats open guard models up to seven times its size, runs on a single 16 gigabyte Nvidia GPU, and ships on Hugging Face under the Apache 2.0 license — which lands the same week that Washington decided open weights are the category it will not be testing.
- 6
Spotify licenses Merlin's independent labels for AI covers and remixes
Merlin, which represents independent labels and distributors across more than 70 countries and roughly 15 percent of the global recorded-music market, signed on to Spotify's forthcoming fan-made covers and remixing product. Member labels choose whether their artists take part, and the tool will launch as a paid add-on so participating artists earn from it. "Every creation drives listeners back to the original work," said Spotify senior vice president Charlie Hellman. Universal Music Group had already backed the product separately, which puts both the major-label and independent halves of the market inside a format that did not exist a year ago.
- 7
Waymo opens its Dallas robotaxi service to anyone with the app
The waitlist is gone. Dallas riders no longer need an invitation to hail a fully autonomous car, six months after a February commercial launch that carried nearly 150,000 people drawn from the interest list. Waymo says fully autonomous testing is already running at the Love Field Airport terminals, and that freeway testing begins shortly — the last step before highway routes open to public riders.
- 8
SK hynix and SanDisk publish the first High Bandwidth Flash specification
High Bandwidth Flash sits between high-bandwidth memory and solid-state storage, using NAND flash to hold far more data than HBM does at speeds much closer to it. The first specification, published through the Open Compute Project six months after the consortium formed, defines stacks up to 512 gigabytes with three performance grades running from roughly 0.4 to 3 terabytes a second. The target is inference, where model weights and context windows have outgrown the memory attached to each accelerator. Google and Tenstorrent are also in the consortium.
Get Top AI Stories by email
The day's most important AI news — free, daily, unsubscribe anytime.
Sources
- 1.Spotify and Merlin Announce Licensing Agreement for Fan-Made Covers and Remixes — Spotify · August 4, 2026
- 2.Dallas, open to all — Waymo · August 4, 2026
- 3.Governor Abbott Directs Comprehensive Data Center Audit — Office of the Texas Governor · August 3, 2026
- 4.China's Open-Weight Models to Be Spared US Tests, US Firms Told — Bloomberg · August 5, 2026
- 5.INTERPOL report finds AI linked to more than half of cybercrime in Africa — INTERPOL · August 3, 2026
- 6.White House won't publicly release AI model evaluation framework it reviewed today with Meta, Nvidia, Microsoft, OpenAI, Anthropic — Fortune · August 4, 2026
- 7.Introducing Shieldstral — Mistral AI · August 4, 2026
- 8.SK hynix Unveils First HBF Standard Specifications with Sandisk, Presenting AI Memory Solutions at FMS 2026 — SK hynix · August 4, 2026
- 9.Anthropic signs $10B deal with AI cloud startup Volta — TechCrunch · August 4, 2026
- 10.Texas halts new data centers as governor calls for audits — TechCrunch · August 4, 2026
This brief was published on August 5, 2026. Cited URLs above point to third-party publishers and may move, paywall, or be retired over time. If a link no longer resolves, original article titles are preserved so you can recover them via search; the canonical web edition at aiproplaybook.com/top-ai-stories/2026-08-05 may carry updated source URLs.