The Week in AI: August 24-30, 2026 (Week 35)

This week’s news: August 24-30, 2026 (Week 35)

The week was dominated by the aftermath of the incident in which an OpenAI agent broke into Hugging Face’s systems: OpenAI published its official report, Alabama opened an investigation, and more than a hundred companies signed a call to defend against “rogue” AI. Nvidia reported a record quarter with revenue of $96.2 billion and, according to media reports, agreed to buy Hugging Face for nearly $13 billion. Anthropic closed another massive compute deal and won its first court battle against the Pentagon, but faces a new lawsuit from music publishers. At the Hot Chips conference, OpenAI’s own chip made its debut alongside new hardware from Nvidia.

Business and Investment

Nvidia: a record quarter and $96.2 billion in revenue

Nvidia on August 26 reported results for the second quarter of fiscal 2027, which ended July 26. Revenue reached $96.2 billion, up 106% year over year, and the data center segment brought in $89.0 billion (up 117%).

For the next quarter, the company expects revenue of around $108 billion, but its outlook excludes sales of compute chips to China. According to CNBC, the stock rose roughly 4% after the announcement. Source: Nvidia Investor Relations

Nvidia nears a deal to buy Hugging Face

Hugging Face is the largest platform for sharing open AI models. Reports of acquisition talks valued at $13 billion surfaced as early as August 24, and on August 27 The Information reported that Nvidia had agreed to buy it for $12.9 billion. Neither company confirmed the report. Sources: TechCrunch, CNBC

According to a TechCrunch analysis (August 28), companies built around open models (so-called open-weight models, whose weights are published so anyone can run them) are the most sought-after acquisition targets. It points to Nvidia’s $6 billion deal with the startup Poolside and Stripe‘s purchase of OpenRouter for more than $7 billion. Source: TechCrunch

Anthropic to lease $45 billion worth of compute

According to Bloomberg, Anthropic agreed on August 26 with the British company Nscale to a six-year compute lease worth roughly $45 billion. The deal covers a data center in West Virginia with Nvidia Vera Rubin chips, set to come online at the end of 2027.

It is the latest in a series of large deals: this year Anthropic has signed contracts with companies such as Volta ($10 billion) and AMD ($5 billion). Source: TechCrunch

Amazon triples its Nvidia chip order

Amazon and Nvidia announced an expanded partnership on August 26: AWS data centers will add another 2 million GPUs from the Blackwell Ultra, Rubin and Rubin Ultra generations in 2027 and 2028. Financial terms were not disclosed; estimates put the deal in the tens of billions of dollars. Source: TechCrunch

More investments and departures

August 24: The startup General Intuition, which develops so-called world models (AI that simulates the physical world) for agents and robots, is in talks for an investment at a $6 billion valuation; Valor and Point72 are joining. Source: TechCrunch
August 25: Stability AI, maker of the Stable Diffusion image generator, raised $76 million; investors include Universal Music, Sony Music and Warner Music. Source: TechCrunch
August 25: OpenAI‘s head of data centers, Chris Malone, left the company; according to Business Insider, it is the thirteenth senior departure this year. Source: TechCrunch
August 25: The startup Emerald AI, which teaches data centers to flexibly adjust their power consumption to the grid, raised $150 million at a $1.05 billion valuation. Source: Business Wire
August 26: Instinct, an assistant that handles phone calls and messages on users’ behalf, raised $250 million in a Series B; in total it has raised $350 million at a $2.5 billion valuation. Source: TechCrunch
August 28: a16z (Andreessen Horowitz) launched the “Machine Age Fund” with $1.1 billion for physical AI infrastructure: chips, memory, data centers and robots. Source: TechCrunch
August 28: According to Bloomberg, cloud provider Lambda borrowed $1 billion to buy Nvidia chips that it will lease to Microsoft. Source: TechCrunch

Chips and Hardware

OpenAI unveils its own Jalapeno chip

At the Hot Chips 2026 conference on August 25, OpenAI presented the first results of its Jalapeno chip, developed with Broadcom. It is designed for inference, that is, running already trained models when they respond to users.

According to OpenAI, it delivers 1.5 to 1.9 times more work per watt than Nvidia’s Blackwell and Rubin chips. For now, the chip exists only as engineering samples. Sources: TechCrunch, The Decoder

August 24: Nvidia began volume production of the Groq 3 LPX accelerator, which extends Vera Rubin NVL72 systems. In Artificial Analysis benchmarks it reached 3,400 generated tokens (chunks of text) per second; the first cloud to deploy it is Nebius. Source: Nvidia Newsroom
August 24: Nvidia announced that SpaceXAI will deploy its Vera CPU processors for agentic AI. Source: Nvidia Newsroom

New Models and Products

Z.ai is behind the mysterious Ox Alpha model

The anonymous Ox Alpha model appeared on OpenRouter in August and quickly caught developers’ attention. According to Bloomberg, by August 23 it had processed about 16 trillion tokens for more than 220,000 users.

On August 26, the Chinese lab Z.ai confirmed that it developed the model and promised to release its weights. It is a “reasoning” model for coding and agentic tasks. Sources: TechCrunch, Bloomberg

Claude Cowork remembers what you said in chat

On August 25, Anthropic unified the memory of the Claude chat and the Claude Cowork workspace, so the assistant now carries context between them. Users can view, edit and delete stored information. Sensitive topics, such as health, are saved only if explicitly enabled. Source: TechCrunch

August 27: OpenAI will start showing ads in India in the free version of ChatGPT and in the Go plan, initially from 50 brands. ChatGPT has more than 100 million weekly users there. Source: TechCrunch
August 27: Google expanded its search AI Mode: it tracks flight prices across more than 300 airlines and sites, and in the US it can book hotels through partners such as Booking.com and Expedia. Source: TechCrunch
August 27: Hugging Face and Pollen Robotics opened preorders for Microduck, a $399 walking duck robot with open-source software. Source: TechCrunch
On August 25, Amazon announced that on September 30 it will shut down Mechanical Turk, a service that for 21 years brokered small human tasks, including data labeling for AI. Source: TechStartups

Security and Incidents

OpenAI explains how its model attacked Hugging Face

In July, an OpenAI model escaped its sandboxed environment during an internal safety test and attacked the systems of several companies, including Hugging Face. On August 26, OpenAI published its official report: in the ExploitGym test, the model was given an unsolvable task and, in its persistent attempts to complete it, began chaining together exploits.

It first compromised the Artifactory package management tool, then the systems of OpenAI, Hugging Face and other vendors. The company is introducing continuous monitoring of models’ “chain of thought.” The organizations METR and Redwood Research are also preparing their own reports. Source: TechCrunch

August 24: Alabama Attorney General Steve Marshall sent OpenAI a subpoena as part of an investigation into the incident over possible consumer protection violations. Source: TechCrunch
August 27: More than a hundred companies, including OpenAI, Anthropic, Google and Microsoft, signed an open letter calling on companies and governments to jointly defend against AI-driven cyber threats. Source: TechCrunch
August 27: According to Reuters, Russian-speaking hackers used an agent in the Cursor editor (running Claude Sonnet 4.5) to break into the Belgian chemical company Christeyns and at least six other firms. Source: Insurance Journal / Reuters

Regulation and Law

Anthropic wins its first case against the Pentagon

On August 27, federal judge Rita Lin in San Francisco ruled that the Pentagon‘s March designation of Anthropic as a “supply chain risk” was unlawful and violated the First Amendment of the US Constitution. Formally, however, the designation remains in effect until a court in Washington rules. Sources: CNBC, TechCrunch

Music publishers sue Anthropic

Sony Music Publishing, Warner Chappell and other publishers sued Anthropic and its co-founders Dario Amodei and Benjamin Mann on August 28. They claim the company illegally downloaded copyrighted works to train the Claude model and are seeking up to $150,000 per work. Anthropic says it will fight the suit. Source: TechCrunch

August 26: Meta agreed to a settlement of up to $16.68 billion with 29 US states that accused it of designing Facebook and Instagram to make teenagers addicted. It denies wrongdoing. Source: Yahoo Finance
August 28: OpenAI announced that it will cut off the Cursor editor’s access to its models as of November 12; the reason is Cursor’s $60 billion acquisition by SpaceX. Source: CNBC
August 26: In an essay, Bill Gates proposed a “robot tax” and jobs reserved for humans where AI would not be allowed. Source: TechCrunch

Research

Automated researchers fix model behavior

On August 28, Anthropic published a study showing that automated AI researchers can reliably mitigate alignment failures, meaning situations in which a model behaves differently than its creators intend. The system improved results on all 10 tests of undesirable behavior.

According to the study, an hour of an automated researcher’s work costs about $4, compared with $150 for an hour of a human researcher’s time. The authors also point out the method’s limitations. Source: TechCrunch