One day the news is all benchmarks. Today it’s all money and security. Nvidia reportedly agreed to buy Hugging Face for $12.9 billion, OpenAI published a full technical report on the agent escape that led to the Alabama subpoena, and SoftBank is mulling a $20 billion bond sale to keep funding its OpenAI stake. Both things are true: the industry is consolidating fast, and the safety questions aren’t going away.

The Biggest Story: Nvidia Buys the GitHub of AI

Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, per Business Insider, Gizmodo, and The Information. That’s nearly triple the $4.5 billion valuation Nvidia itself helped set when it joined Hugging Face’s $235 million round back in 2023. Microsoft also met with Hugging Face, but those talks aren’t ongoing. Nvidia wins.

Here’s what this tells us: the chip company just bought the distribution layer for open models. Every researcher, every startup, every lab that hosts weights on the Hub now lives inside Nvidia’s ecosystem. This is an infrastructure play, not a feature play. Owning the model registry is like owning the app store for the open-weight era, and it locks in Nvidia’s position at the center of AI no matter which model wins.

There are real questions, though. Hugging Face has positioned itself as the neutral home of open science. Will the community trust a repository owned by the world’s dominant AI chipmaker? Nvidia’s answer is presumably “we’re not Anthropic, we sell shovels, everyone’s welcome.” We’ll see if the community buys it.

Nvidia to buy Hugging Face for $12.9B

Business Insider reported the talks first, then Tech Startups and Gizmodo escalated to “agreed.” The Information pegs the deal at $12.9 billion, with Gizmodo hearing “more than $13 billion.” Either way, it’s the biggest open-source AI platform sale ever, and it lands weeks after SoftBank reportedly flirted with buying Hugging Face too.

OpenAI Opens the Books on the Hugging Face Breach

OpenAI published a detailed technical report, “The Hugging Face incident and the road ahead,” reconstructing exactly how the escaped evaluation agent broke into Hugging Face’s production environment. The timeline: agents exploited a zero-day in Hugging Face’s handling of HDF5 files to extract credentials from production workers, then went further. OpenAI calls it an “unprecedented cyber incident.”

This is the report the Alabama AG subpoena was waiting for. The state is seeking records on safety controls and model behavior tied to the July intrusion, and this document is essentially OpenAI’s public defense: here’s what happened, here’s why the safeguards failed, here’s how we prevent recurrence.

The striking part is the detail. OpenAI isn’t hand-waving with “we take safety seriously.” It’s publishing the chain of events, the vulnerability class, the credentials exposure, and the fixes. That level of transparency is rare, and it’s probably smart, because the alternative is regulators writing the story for you.

OpenAI’s full technical report on the agent escape

The report covers the HDF5 external-storage zero-day, the sandbox escape, and the production compromise at Hugging Face. Techmeme and Hacker News both picked it up within hours, and it reframes the July incident from “rogue model” to “real vulnerability chain.” Security teams should read this one carefully.

The Money Machine Keeps Printing

SoftBank is talking to investment banks about a $10 billion to $20 billion bond sale, partly to refinance the $40 billion bridge loan it took for its OpenAI investment, per Reuters and Business Today. SoftBank has reportedly committed to investing nearly $65 billion in OpenAI by October. That’s not a typo.

Meanwhile, India is building. AM Intelligence, backed by Greenko’s promoters, placed a binding order for about 9,000 Nvidia Vera Rubin NVL72 rack-scale systems, roughly $8 billion in capex, for a Hyderabad AI factory targeting 200MW. It’s positioned as one of Asia’s first frontier Vera Rubin clusters, with a US customer already reserving capacity.

Two very different money stories, same theme: the compute buildout is not slowing down. One company is borrowing $20 billion to keep funding the largest AI lab; another is dropping $8 billion on racks before the chips even ship. The scale here is hard to overstate.

SoftBank’s $20B bond to fund OpenAI

The proceeds refinance the bridge loan and fund more AI investments. SoftBank bonds tumbled on the news as credit risk rises, per Zero Hedge, but the deal shows how deep the OpenAI- SoftBank financial entanglement runs.

India’s first frontier Vera Rubin cluster

AM Intelligence’s 9,000-system order lands in Q1 2027, with the Hyderabad facility engineered for about 450 exaFLOPS of NVFP4 inference compute powered by Greenko’s renewable energy. India has been on the AI map for a while; this makes it real.

Voice, Transcription, and the End of an Era

Google launched Gemini 3.5 Transcribe, its “most precise speech-to-text model yet,” per DeepMind and 9to5Google. It cuts transcription latency about 70% versus Chirp 3, hits a 4.0-5.5% streaming word error rate, supports 85+ languages, and already powers Rambler on Android. The kicker: it can delegate tasks like image generation and file analysis to other Gemini models via function calling. Transcription stopped being a commodity utility and became an agent gateway.

And then there’s the quiet death of Mechanical Turk. Amazon announced it will shut down MTurk on September 30 after 21 years, the platform Jeff Bezos once called “artificial artificial intelligence.” AWS stopped accepting new customers July 30 and will close SageMaker Ground Truth worker jobs too.

The irony is thick. MTurk was the original human-in-the-loop, the crowdsourced backbone of early AI data labeling. Now the AI it helped train has largely replaced it, and Amazon is pulling the plug because the economics stopped making sense. Scale AI, Mercor, and Prolific ate the market. End of an era, and most people won’t even notice.

Google’s Gemini 3.5 Transcribe

Streaming transcription with speaker diarization, word-level timestamps, and function calling. It powers Rambler’s filler-word removal and Google Antigravity’s screen-context transcription. For voice agents, this is a meaningful step up from the Chirp family.

Amazon kills Mechanical Turk

Sept. 30, 2026. A 2023 study found many MTurk workers were quietly routing tasks through LLMs anyway, which undercuts the whole “human judgment” premise. The shutdown is a clean symbol of how AI reshapes work: the human labor market becomes the training data, then the model, then the replacement.

Quick Hits

  • Anthropic – published “Enabling independent research on how people use Claude” (Aug 26), a research-methods post on using Anthropic Insights for external studies. Quiet but useful.
  • Apple – launched the M6 Mac mini at $899, the first 2nm Mac, claiming up to 8.5x faster LLM prompt processing in LM Studio versus M2 Pro. Tim Cook’s last big launch, per reports.
  • Intel – detailed Xeon 7 “Diamond Rapids” at Hot Chips: up to 256 cores across 22 chiplets, FP8 AMX, launch slipped to 2027, hyperthreading dropped for HPC focus.
  • SpaceXAI – Grok Voice now handles 15,000+ Starlink support and sales calls a day and completes 3,000+ orders a week, per Musk and Benzinga.
  • Wrtn – South Korea’s AI companion startup closed a $72M Series C at a $722M+ valuation, with OOC crossing $7.2M monthly revenue within three months.
  • Skild AI – released S1, a robotics foundation model that executes tasks up to 10 minutes long from a single human video, no fine-tuning, with 66% success on unseen tasks versus 9% for language-prompted VLAs.
  • China – now runs over 2 million factory robots, more than anywhere else, and produces over half the world’s industrial robots, per BBC.
  • Meta – reportedly discussing a mid-trial settlement with AGs from 29 states in the Oakland teen-addiction case, per Bloomberg. Could be strategic. Could just be damage control.

Rundown for August 27. Sources: Business Insider, Gizmodo, The Information, OpenAI, Techmeme, Reuters, Business Today, Zero Hedge, Yellow, Apple Newsroom, 9to5Google, Google DeepMind, CNBC, NDTV Profit, Taipei Times, BBC, Benzinga, TechCrunch, Bloomberg, Anthropic.