The Week in AI: August 17–23, 2026 (Week 34)

News of the week: August 17–23, 2026 (Week 34)

The week was dominated by Anthropic’s numbers: a $65 billion annual revenue run rate, its first quarter with higher revenue than OpenAI, and reports of an October IPO at a possible $2 trillion valuation. Money keeps pouring into infrastructure, with Nvidia backing a giant OpenAI data center in Ohio and Google securing a warrant for more than $12 billion worth of Marvell shares. On the product side, AI agents are moving into team tools such as Slack and Cursor, and ChatGPT on the Mac can now read and send messages. Chinese company Unitree had one of the wildest stock market debuts of the year, and a new Pew survey shows Americans are growing increasingly worried about AI.

Business and Investment

Anthropic: a $65 billion revenue run rate and a lead over OpenAI

Anthropic, the maker of the Claude models, told investors that its annualized revenue (monthly revenue extrapolated to a full year) reached $65 billion at the end of July. In May it was $47 billion, and at the end of 2025 roughly $9 billion. The news was reported on August 17 by CNBC.

A day later, The Wall Street Journal reported that Anthropic had for the first time also overtaken OpenAI in actual quarterly revenue. In the second quarter it took in $11.6 billion and posted a small operating profit, while OpenAI brought in $6.7 billion and its operating loss widened to $12.3 billion (August 18, The Wall Street Journal, summary in The Next Web).

Anthropic could go public at a $2 trillion valuation

According to The New York Times, bankers told investors that Anthropic’s October initial public offering (IPO) could raise more than $100 billion at a record valuation of around $2 trillion. These are bank estimates, not confirmed terms of the offering (August 21, The New York Times, PYMNTS).

For comparison, the current record is held by SpaceX, which raised $85.7 billion on the stock market this June. Based on the company’s metrics, Anthropic could beat that record.

Nvidia backs OpenAI’s Ohio data center with up to $105 billion

In a filing with the US regulator, the SEC, Nvidia said it will provide up to $105 billion in credit support so that OpenAI can lease a data center in Pike County, Ohio, for 20 years. The campus will be built and operated by SB Energy, part of the SoftBank group, in which Nvidia is also investing $1.5 billion (August 17, CNBC).

The first phase is set to have a capacity of 4.25 gigawatts, with an option for another 3.75 GW, and the first 800 megawatts are due to come online in 2028. Earlier there had been talk of a guarantee of up to $250 billion, so Nvidia has scaled it back significantly (AI Business).

Google secures a $12.2 billion share warrant from Marvell

Chipmaker Marvell gave Google the right to buy up to 58.97 million of its shares at $206.58 each, worth roughly $12.2 billion. The warrant (an option to buy shares) vests gradually, with each $500 million that Google spends on custom chips tied to its TPU processors (August 19, Bloomberg, The Next Web).

Marvell shares rose as much as 14 percent during the day, while shares of rival Broadcom fell.

Unitree: a robotics company debuts up 629 percent

Chinese humanoid robot maker Unitree listed on Shanghai’s STAR Market on August 19. The shares opened at 1,100 yuan, 629 percent above the offer price of 150.80 yuan, and closed the day at 845 yuan (August 19, Al Jazeera, South China Morning Post).

A record of roughly 9.8 million accounts applied for the shares. Unitree is the first general-purpose humanoid robot company to be listed on a mainland Chinese stock exchange.

More money in the industry

Groq raised $350 million at a valuation of $3.5 billion, roughly half of last year’s peak. Following a licensing deal with Nvidia, the company is shifting from its own chips to renting out compute as a so-called neocloud (August 17, TechCrunch).
British chip developer Fractile is in talks to raise $600 million at a $6.5 billion valuation, according to Bloomberg. It was helped by a deal to supply chips to Anthropic worth roughly $250 million, although the chips will not arrive until 2027 (August 20, The Next Web).
Brazil set aside roughly $444 million for two national AI supercomputers. One will be built with Chinese companies Huawei and iFlytek, the other will go to a tender in which Nvidia is the favorite (August 21, Al Jazeera).
The Ramp platform tracks the spending of more than 70,000 US companies. According to its data, in July Anthropic had just under 44 percent of companies paying for either of the two labs and OpenAI just under 40 percent, but OpenAI is closing the gap (August 20, TechCrunch).
Startup micro1, which supplies human experts for model training, grew from $100 million to $500 million in gross annual run rate in eight months (August 20, TechCrunch).
Amazon raised prices on its devices, for example the Echo Dot from $49.99 to $79.99. It cites more expensive memory, whose prices are being driven up by demand from AI data centers (August 21, Fortune).

Products and Agents

Slack Code: coding agents in team chat

Slack launched Slack Code, a feature in which AI coding agents work alongside people in shared channels. Among the first partners are Claude Code from Anthropic, Devin from Cognition, GitHub Copilot, and an agent from Vercel (August 20, Slack, TechRepublic).

A team summons an agent by mentioning it in a conversation, and the agent creates a temporary channel for the task. Deploying code to production requires human approval. Slack Code itself is free on all plans, but the agents are paid for separately.

ChatGPT on the Mac reads and sends messages

OpenAI added an extension for the Messages app (Apple Messages) to the ChatGPT desktop app for macOS. ChatGPT can search and summarize conversations in iMessage, SMS, and RCS, suggest replies, and send them once approved (August 20, 9to5Mac, MacRumors).

It only works on computers with Apple Silicon chips and requires full disk access. The media are therefore pointing out the privacy risk, since the app also gains access to years-old private conversations.

Smaller product news

The Cursor editor gave its cloud agents a /goal command for long-term objectives (for example, “fix all flaky tests”). Agents can also wake up on an event, such as a new pull request or a Slack thread, and helper agents run in separate virtual machines (August 19, Cursor).
OpenAI cut the price of GPT-5.6 Sol for developers for three months: input from $5 to $4 and output from $30 to $20 per million tokens (tokens are the chunks of text used for billing). The promotion runs at least until November 21 (August 21, AWS, Outlook Business).
Google offered publishers a Preferred Sources button they can place on their websites. Readers use it to mark a source as preferred, which then appears more often for them in Google Search, Discover, and News. It is meant to ease the loss of traffic that publishers attribute to AI Overviews (August 20, TechCrunch).
Legal AI company Harvey introduced Tenet, an open-weights model. It is based on Kimi K3, and according to Harvey’s own tests it handles almost twice as many legal tasks as the base model (Harvey).
Anthropic is preparing a change for enterprise customers, according to Reuters. Data that must be retained for 30 days for safety review of the most powerful models will be allowed to stay in the customers’ own cloud (August 20, Reuters via The Star).

Research

Claude designs proteins that work in the lab

Anthropic published the results of a project in which the Claude Mythos Preview and Claude Opus 4.8 models independently designed protein binders, proteins that attach specifically to another molecule. Working binders were found for 14 of the 15 target proteins (August 18, Anthropic).

The designs were independently produced and tested by Adaptyv Bio and Twist Bioscience. The success rate reached 22.6 to 26.7 percent, compared with the 10 to 15 percent typical in the field. The results have not yet been peer reviewed (The Next Web).

Nvidia AVO and the ARC-AGI-3 benchmark

Nvidia unveiled AVO, an agentic system that completed all 183 levels of the public portion of the ARC-AGI-3 benchmark, scoring 100 percent. The benchmark consists of interactive puzzles meant to measure the general ability to solve novel problems. Inside the system runs Claude Opus 5, which on its own scores around 30 percent (NVIDIA Developer, The New Stack).

The result shows how much difference the “harness” around a model makes, meaning the tools and the orchestration of the process. However, these are Nvidia’s own measurements on the public set, not an independent verification on the hidden task set.

More research

A new paper on arXiv lowered the bound on the matrix multiplication exponent from 2.371339 to 2.371177. The authors were helped by machine learning, including Google DeepMind’s AlphaEvolve system. Matrix multiplication is a fundamental computational operation on which neural networks themselves are built (arXiv).
In a UK trial called Operation Blue Skies, Google is testing how an AI model can reroute aircraft over the Atlantic so they do not leave behind contrails, which contribute to warming (Google).

Regulation, Politics, and Society

OpenAI wants a stricter California SB 53

OpenAI urged California to expand SB 53, its law on the safety of the most powerful AI models. According to the company, oversight should apply already during model training, not only before release. This is a shift from its earlier stance, when the company tended to push back against similar regulation (August 21, Politico).

A Trump-linked firm and Chinese AI models

Reuters found that the crypto company World Liberty Financial, linked to the family of Donald Trump, is working with the Hong Kong platform WorldClaw. The platform sells access to AI models, and 43 of its 90 models come from Chinese companies under US government scrutiny, such as Alibaba, Baidu, and Z.ai (August 17, Reuters).

Reuters does not allege any violation of the law. However, experts contacted by the agency consider the partnership at odds with the administration’s policy toward Chinese technology.

Pew: Americans are increasingly worried about AI

A Pew Research Center survey of 3,488 adults found that 52 percent of Americans are more concerned than excited about the growing use of AI. In 2021 the figure was 37 percent. For the first time, a majority of people under 30 feel the same way, specifically 55 percent (August 18, Pew Research Center).

Security and Incidents

GitHub suffered a worldwide outage that affected the web interface, the API, Actions, and the AI assistant Copilot. Microsoft confirmed the outage (August 17, BleepingComputer).
British supermarket chain Sainsbury’s paused the Facewatch facial recognition system in one store. For the second time this year, it had wrongly flagged a customer as a shoplifter (August 17, The Guardian).
According to The Register, OpenAI acknowledged that safety monitoring of its most powerful models costs roughly 20 percent of the compute it oversees. It paused some training runs for two weeks to tighten security (August 19, The Register).
404 Media described the spread of so-called “subtlefakes” on X. These are real photos slightly altered into intimate images without consent, some bearing the watermark of the Grok model (August 21, 404 Media).