AI Newsletter: Top 10 AI Stories — Week of August 15–21, 2026

Welcome back to the PMV Consulting AI Newsletter. Each week we round up the ten AI stories that mattered most, in plain language, with links to the original reporting so you can dig deeper on anything that catches your eye. Here’s what happened in AI during the week of August 15–21, 2026.

1. Anthropic’s annualized revenue tops $65 billion ahead of expected IPO

Anthropic told investors this week that its annualized revenue run rate climbed above $65 billion at the end of July, a sevenfold jump from a year earlier and an increase of roughly $18 billion in just two months. The company is preparing for a highly anticipated initial public offering, and reports this week indicated its pre-IPO revolving credit facility is set to rise past its roughly $10 billion target as it lines up financing ahead of a Wall Street debut that could come as soon as this fall. Investors reportedly expect Anthropic to keep growing at a similar pace through the rest of the year, putting it on track to finish 2026 somewhere between $100 billion and $120 billion in annualized revenue, and the company is said to be targeting a public valuation of $2 trillion or more, which would make it the largest market debut on record. Separately, reports emerged that Anthropic plans to grant CEO Dario Amodei and other co-founders shares carrying extra voting power ahead of the listing, a structure increasingly common among tech companies going public that lets founders retain control even as outside shareholders take a larger economic stake. Rival OpenAI has also been growing quickly, reportedly doubling its revenue to $40 billion over the past year, though the two companies calculate their metrics somewhat differently. Anthropic’s growth rate has nonetheless captured particular attention on Wall Street, and the company is expected to go public ahead of OpenAI. The numbers underscore how quickly enterprise and developer demand for Claude has scaled this year, even as questions persist across the industry about whether the enormous capital being poured into AI infrastructure will ultimately pay off.

2. OpenAI launches ChatGPT for Teens with built-in safety guardrails

OpenAI introduced ChatGPT for Teens on August 18, a version of the chatbot tailored for users between the ages of 13 and 17 that OpenAI’s systems will apply automatically whenever it estimates a user falls in that age range. The teen experience restricts content related to self-harm, eating disorders, and romantic or sexual conversations, and OpenAI says it is taking additional steps to reduce the risk of teens forming unhealthy emotional attachments to the chatbot. On the academic side, the release introduces a Study Mode built to guide students toward working through problems themselves rather than simply handing over finished answers, alongside homework reminders and quiz tools, and OpenAI announced a partnership aimed at helping teens understand how AI systems work rather than just how to prompt them. Parents retain access to family controls introduced earlier this year, including safety notifications and the ability to set quiet hours. The launch follows sustained scrutiny of how teenagers use ChatGPT for schoolwork, everyday questions, and companionship, and comes amid ongoing lawsuits alleging the chatbot provided harmful information related to self-harm to minors in the past. OpenAI framed the release as an effort to meet what it called the first generation to grow up with AI on its own developmental terms, rather than treating teenage users identically to adults or younger children. The company already offers a teacher-facing version of ChatGPT, and observers noted that a purpose-built teen product could also give OpenAI a stronger foothold in K-12 and secondary education, an area every major AI lab is racing to establish itself in as schools work out their own policies on AI-assisted schoolwork.

3. Nvidia backs an 8-gigawatt Ohio AI campus built for OpenAI

Nvidia announced on August 17 that it will invest $1.5 billion in SB Energy and serve as the exclusive AI compute provider for a massive new data center campus in Pike County, Ohio, where OpenAI will be the customer under a 20-year lease. The first phase of the PORTS-Pike Technology Campus is designed to deliver 4.25 gigawatts of AI computing capacity, with Nvidia holding an option on an additional 3.75 gigawatts that would bring the site to a full 8 gigawatts, a scale that could ultimately require at least 10 gigawatts of new power generation. SB Energy, backed by SoftBank and OpenAI, will build, own, and operate the facility using Nvidia’s DSX AI factory platform, and Nvidia says the site alone could represent roughly $600 billion in computing systems purchased through 2030. To address local concerns about who bears the cost of powering these facilities, SB Energy and SoftBank committed to building at least 10 gigawatts of new energy generation and investing at least $4.2 billion in regional grid infrastructure through a partnership with the local utility, alongside an $80 million community benefits fund. The deal is the latest example of Nvidia financing the very companies that buy its chips, a strategy that has fueled explosive demand for its hardware but also drawn scrutiny over the circular nature of money flowing between chipmakers, cloud providers, and AI labs. It also reflects a broader shift in how frontier AI companies are securing capacity: rather than simply buying chips, they are increasingly locking up land, power, and multi-decade leases years in advance to guarantee the infrastructure their models will need.

4. Alibaba’s Qwen3.8-27B open model closes in on Claude Opus on agentic tasks

Alibaba’s Qwen team released Qwen3.8-27B on August 14, a 27-billion-parameter open-weight model under the permissive Apache 2.0 license, and the model became one of the most discussed AI releases of the week as independent reviewers dug into its benchmark claims. Small enough to run on a single high-end consumer graphics card, Qwen3.8-27B accepts text, images, and video, and ships with a native context window of more than 260,000 tokens. On Alibaba’s own published model card, it outperforms Claude Opus 4.6 Max on a majority of benchmark categories, with its largest advantages in agentic software engineering, instruction following, and computer-use tasks like navigating a desktop or browser. It trails Opus 4.6 Max on pure knowledge-reasoning benchmarks such as Humanity’s Last Exam and GPQA Diamond. Independent commentators were quick to note an important caveat: every comparison figure comes from Alibaba’s own model card, no outside lab has reproduced the results, and the “beats Opus” framing circulating online overstates what a mixed scorecard actually shows. Even with that caveat, reviewers who tested the model described it as the strongest self-hostable option in its size class, capable of trading blows with much larger, closed frontier systems on real coding work despite costing nothing to download. The release adds to a wave of increasingly capable open-weight models arriving from Chinese labs this year, intensifying pressure on U.S. AI companies to justify the premium pricing of their closed, proprietary systems for agentic and coding-heavy use cases.

5. Google open-sources HEIR, a compiler for running AI on encrypted data

Google released HEIR, short for Homomorphic Encryption Intermediate Representation, as an open-source compiler toolchain that lets developers run AI models on encrypted data without ever decrypting the underlying information. The technology, known as homomorphic encryption, allows a server to process sensitive information, generating a useful result such as a fraud score or an intrusion alert, while only ever seeing scrambled, unreadable data. The approach has existed in theory for years but has historically been too slow and required too much cryptography expertise for most companies to deploy in production. HEIR is part of Google’s broader Private Computing Toolkit and is designed to convert models built to run on ordinary, unencrypted data into versions that operate on encrypted inputs, with Google saying its long-term goal is to make the process close to a one-click experience for developers who aren’t cryptography specialists. Google has already demonstrated the toolkit on private recommendation systems, credit card fraud detection, network intrusion detection, and hotword recognition for voice assistants, applications where companies want AI-driven insight without ever exposing raw customer data to the systems doing the analysis. The release quickly drew hundreds of comments on Hacker News, with technologists debating whether the approach is now genuinely practical for production workloads or still too computationally expensive to matter outside of narrow use cases. For industries like healthcare and finance, where strict regulations limit how sensitive data can be shared or processed, tools like HEIR point toward a future where AI assistance and airtight data privacy don’t have to be a trade-off.

6. Pennsylvania signs what its governor calls the nation’s strictest AI data center rules

Pennsylvania Governor Josh Shapiro signed Executive Order 2026-05 on August 18, establishing what he described as the strictest state-level guardrails on AI data centers in the country as more than 100 proposed data center projects sit in various stages of planning across the state. The order directs the state’s Department of Environmental Protection to review a data center permit application only if the developer has made a legally binding commitment to meet the state’s new Responsible Infrastructure Development standards and has already secured approval from the local community where the project would be built. Under those standards, developers must pay the full cost of any new electricity generation, transmission, and distribution their project requires, so the expense doesn’t land on ordinary ratepayers, and an increasingly significant share of that power must come from clean sources like solar, advanced nuclear, or battery storage. In the event of a grid emergency, the order specifies that data centers would be the first facilities cut off, not residential customers. The order also removes AI data center proposals from Pennsylvania’s fast-track permitting process and bans the use of nondisclosure agreements on these projects, aiming to increase public transparency around deals that have often been negotiated quietly. Shapiro said he acted after the state Senate failed to pass related legislation that had already cleared the House, framing the order as a response to a wave of speculative proposals rather than a rejection of data center investment outright, and pointing to earlier, well-financed projects he had championed as the kind of development the state still welcomes. The move reflects growing political pressure nationwide as rising electricity bills get tied, fairly or not, to the AI industry’s soaring power demand.

7. Claude autonomously designs protein binders that hold up in the wet lab

Anthropic published research this week showing that Claude Opus 4.8 and its Mythos Preview model autonomously designed functional protein binders that were independently synthesized and tested by outside laboratories Adaptyv Bio and Twist Bioscience, with promising real-world results. Working against 15 biologically significant targets, including proteins tied to cancer immunotherapy, Alzheimer’s disease, and inflammatory conditions, Claude produced 1,320 candidate protein designs, of which 354 were confirmed to successfully bind their targets, a hit rate between roughly 23% and 35% depending on how the sessions were configured. That compares favorably to the 10% to 15% success rate Anthropic says is typical for this kind of protein design campaign industry-wide, and Claude successfully designed at least one working binder against 14 of the 15 targets attempted. In one standout case, against a target called RBX1, Claude’s Mythos Preview model achieved a 40% hit rate in a single-target session, compared with a 3.7% success rate among human participants in a prior public design competition on the same target. Rather than relying on a specialized biology model trained in-house, Anthropic gave Claude access to established open-source protein design tools and let it act as an autonomous coordinator, selecting binding sites and choosing among dozens of workflow combinations without step-by-step human guidance. Anthropic and outside researchers were careful to frame the results as an early and genuinely notable research finding rather than a drug-development breakthrough, since a working protein binder is only one early step in the years-long process of developing an actual medicine, but the results mark a real shift from AI simply assisting scientists toward AI directly running consequential parts of the experimental process.

8. AI companies keep buying, and destroying, millions of secondhand books

Reporting continued this week on the unusual and controversial practice of AI companies buying enormous quantities of secondhand physical books, scanning them, and then destroying the originals to build higher-quality training data for their models. The trend traces back to unsealed court filings from Anthropic’s Project Panama initiative, in which the company hired contractors to slice the bindings off millions of purchased books, run the pages through industrial scanners, and shred what remained, an effort an internal document described as digitizing “all the books in the world.” The legal basis for the practice rests on the first-sale doctrine, the longstanding copyright principle that once someone legally buys a physical book, they can do essentially whatever they want with that specific copy, including destroy it, even though the underlying text remains protected. A federal judge ruled last year that scanning legally purchased books and training AI on the resulting text counts as fair use, though Anthropic separately agreed to a $1.5 billion settlement over its earlier use of pirated digital copies obtained outside that legitimate purchase-and-scan process. Booksellers, including small independent shops in Europe, have reported unusually large, indiscriminate bulk orders they suspect are destined for this pipeline, often facilitated anonymously through book-data intermediaries. Pre-2022 books are considered especially valuable for this purpose because they predate the flood of AI-generated text now circulating online, meaning they offer AI labs cleaner, purely human-authored writing at a moment when high-quality training data has become an increasingly scarce and contested resource across the industry.

9. Grok starts spitting out gibberish, and xAI struggles to explain why

Some users of xAI’s Grok chatbot began receiving bizarre, nonsensical responses to ordinary requests starting around August 19, in an incident that spread across social media and drew coverage from multiple outlets before the week was out. In one widely shared example, a user who asked Grok to generate a PDF instead received several paragraphs of unrelated words strung together in broken grammar, and another user found that the chatbot’s cited source links pointed to unrelated reinforcement-learning research sites rather than anything resembling a real answer. The issue was reported primarily among users of Grok Lite, and some described the gibberish persisting even after starting a fresh conversation or refreshing the page, though xAI said its official status page continued to show all services as fully operational throughout. The company eventually acknowledged the problem in a brief social media statement, calling it a rare, temporary generation glitch and suggesting affected users start a new chat or regenerate the response, but it offered no technical explanation for what caused it or confirmation that a fix had been deployed. Commentators noted that the vague response looked especially thin given that xAI is currently navigating a formal European Commission investigation under the Digital Services Act along with several ongoing lawsuits, and that rivals including OpenAI, Anthropic, and Google have each published more detailed public postmortems after past incidents degraded their own chatbots’ response quality. The episode, while apparently limited in scope, is a reminder that even frontier AI systems remain vulnerable to unexplained failures as they become more deeply embedded in everyday work and research.

10. Anthropic’s Dario Amodei says the AI backlash is “fundamentally a crisis of trust”

Anthropic CEO Dario Amodei pushed back this week against the argument that his own public warnings about AI risk are what turned public opinion against the industry, in an unusually candid exchange that played out on social media. The comments came in response to investor Gavin Baker, who had argued on a podcast and on social media that Amodei’s repeated warnings about AI’s dangers have helped fuel a broader backlash in the United States, including growing local resistance to new data center projects. Amodei agreed that the public holds a genuinely negative view of AI and called that a serious problem, but argued the real cause runs much deeper than any one executive’s messaging, describing it instead as a broad crisis of trust in companies, governments, and the technology industry generally, rooted in a widespread suspicion that powerful institutions are working against ordinary people’s interests. He pointed to what he called the most accurate criticism of AI companies, including his own: that the industry has made big promises about benefiting humanity, from curing diseases to accelerating scientific discovery, without yet consistently delivering on them at scale. Amodei said Anthropic is working to close that gap through expanded research in biology and medicine, and pointed to what he described as early positive signs without offering further specifics. On the question of regulation, he rejected the idea that AI rules inevitably concentrate power among the largest incumbents, arguing that Anthropic has specifically pushed for policies designed to slow down frontier labs while leaving room for smaller competitors to catch up. The exchange offered a rare glimpse into how one of the industry’s most prominent and most publicly candid leaders is thinking about the widening gap between what AI companies promise and what the public currently believes about them.