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 surges to $65 billion ahead of a fall IPO
Anthropic told investors this week that its annualized revenue run rate topped $65 billion by the end of July, according to Bloomberg, a more than sevenfold jump from roughly $9 billion at the close of 2025. The figure climbed from $30 billion in April to $47 billion in May and then to $65 billion just two months later, one of the fastest revenue accelerations any company has ever reported. Preliminary second-quarter revenue exceeded $11.5 billion, up from $787 million a year earlier, and the company posted positive adjusted operating income for the quarter. Anthropic has confidentially filed for an initial public offering and is working with Morgan Stanley, Goldman Sachs, and JPMorgan Chase on the listing, with trading potentially beginning as early as this fall, ahead of rival OpenAI. Investors reportedly expect Anthropic to finish 2026 between $100 billion and $120 billion in annualized revenue, and internal projections shared with potential backers point toward $190 billion to $200 billion by 2028. For comparison, OpenAI’s own run rate recently crossed $40 billion. Anthropic was valued at $965 billion following a $65 billion Series H round in May, more than double its valuation from just three months earlier. The scale of the jump has made Anthropic’s growth trajectory a bigger draw for investors than OpenAI’s, even though OpenAI’s consumer product still reaches far more people day to day. The IPO filing also comes against the backdrop of Anthropic’s fraught relationship with the Trump administration, including a brief, mandated suspension of its Fable 5 and Mythos 5 models in June over export control concerns that was resolved after roughly two weeks.
2. OpenAI launches a dedicated ChatGPT experience for teens
OpenAI rolled out ChatGPT for Teens this week, a separate experience for users ages 13 to 17 that combines tighter content restrictions, learning-focused tools, and optional parental controls. Anyone who identifies as a teenager, or whom OpenAI’s age-prediction system estimates to be under 18, is automatically routed into the teen version. The system places stricter limits around sexual or romantic roleplay, graphic violence, and self-harm content, and adds safeguards intended to discourage emotional dependency on the chatbot. On the education side, the teen product can steer students toward OpenAI’s existing Study Mode, which walks users through a problem rather than simply producing a finished answer, and parents who link their account to a teen’s can set quiet hours and receive alerts under certain conditions. The rollout lands amid a wave of scrutiny over how AI chatbots affect younger users, including a landmark 29-state trial Meta is heading into over alleged harm to young users of its own AI and social products. OpenAI’s move is widely read as an attempt to get ahead of similar regulatory and legal pressure before it lands on ChatGPT specifically, at a moment when tens of millions of teenagers already use the product informally, often without any of the extra guardrails now being built in. The company has faced its own wrongful-death lawsuits this year alleging that ChatGPT contributed to a teenager’s suicide, which have kept pressure on OpenAI to visibly harden its protections for minors rather than simply adjust them quietly in the background.
3. OpenAI says enterprise revenue has overtaken its consumer business
OpenAI disclosed this week that its enterprise business now generates more revenue than its consumer operations, a crossover the company says arrived earlier than it had expected. The shift reflects how quickly businesses have adopted ChatGPT Enterprise, the API, and related tools for everyday work, even as the consumer ChatGPT app continues to add users at a rapid clip. It also underscores a broader pattern across the AI industry this year: the most durable revenue increasingly comes from companies paying for reliable, high-volume access to models embedded in their own workflows, rather than from individual consumers paying for a monthly subscription. The news arrived in the same week that OpenAI’s ChatGPT advertising business, which launched in the U.S. roughly six months ago, expanded to 31 European markets, with European advertisers initially buying through major agency groups ahead of a planned self-serve rollout. OpenAI says roughly 20% of ChatGPT queries already show direct commercial intent, with additional upper-funnel queries signaling future purchases, a dynamic advertisers are watching closely as they try to understand how AI search and chat are reshaping how people discover products. The enterprise-consumer crossover also matters for how OpenAI positions itself heading toward a public listing: a revenue base anchored in business subscriptions and API usage is generally viewed as more predictable, and therefore more valuable, than one built primarily on consumer app downloads.
4. A 27-billion-parameter open model beats Claude Opus 4.6 Max on coding benchmarks
Alibaba’s Qwen3.8-27B, an open-weight model small enough to run on a single 24GB consumer graphics card, outperformed Anthropic’s much larger Claude Opus 4.6 Max on SWE-Bench Pro, a benchmark that tests how well AI models handle real-world software engineering tasks. The result is notable less for the specific benchmark win, which can shift from month to month as labs release updates, and more for what it signals about the pace of open-weight progress: a model cheap enough to run locally on a gaming PC is now matching or beating a flagship closed model that typically runs in a data center and costs meaningfully more per query. It continues a pattern that has defined much of 2026, in which Chinese labs including Alibaba, DeepSeek, and Moonshot have repeatedly closed the gap with, and in some benchmarks surpassed, their better-funded American counterparts, while charging a fraction of the price. For developers and businesses choosing which model to build on, the practical implication is that good-enough coding capability is becoming available at a much lower price point than it was even a few months ago, intensifying pressure on closed-model providers like Anthropic, OpenAI, and Google to justify their premium pricing through reliability, tool integration, and enterprise features rather than raw benchmark scores alone. Independent replication of any single-company or single-benchmark claim is always worth some skepticism, but the broader trend of shrinking capability gaps between open and closed models has held up consistently across the year.
5. AI meeting notetaker left 181,874 recordings exposed for six months
A single missing security rule in the database behind tl;dv, an AI notetaker used by roughly 2 million people, left 181,874 meeting records from 84,312 users across more than 35,000 domains queryable by any signed-up account. The exposed records reportedly included conference IDs that could have let someone join live calls, not just review past transcripts. The underlying flaw, a missing Firestore rule, was first reported in January and remained unfixed for roughly six months despite the vendor holding SOC 2 compliance certification, a common bar that enterprise buyers use to vet software vendors’ security practices but one that clearly did not catch this particular gap. The episode is part of a wider pattern: as AI notetakers and meeting assistants have become standard fixtures in corporate video calls over the past two years, they’ve also become a concentrated, high-value target, since a single vendor’s database can hold sensitive internal conversations from thousands of unrelated companies at once. Security researchers covering the incident published a checklist of steps organizations should take to audit their own AI notetaker’s data handling, including verifying who can access historical recordings, whether transcripts are shared with any AI training pipeline, and whether access controls are enforced on the backend rather than only in the user interface. Businesses that have adopted AI meeting tools broadly this year are being urged to treat vendor security claims like SOC 2 as a starting point for due diligence, not a substitute for it.
6. Google open-sources a compiler that runs AI inference on encrypted data
Google open-sourced HEIR, a compiler that allows AI models to run inference on encrypted data without the server ever decrypting it, a capability built on a technique called homomorphic encryption. In practical terms, the tool is aimed at situations where an organization wants to use a cloud AI service on sensitive data, such as medical records or financial information, without ever exposing the underlying, unencrypted data to the company running the model. Homomorphic encryption has existed as a research concept for years but has historically been too slow and computationally expensive for most real-world AI workloads; Google’s release is part of a broader industry push to make the technique practical enough for production use as demand grows for AI tools that can work with regulated or highly sensitive data. The release arrived in the same week as several other security-focused stories, including a data leak at an AI meeting notetaker and a critical, actively exploited vulnerability in the Ray framework used to scale machine learning workloads, underscoring how much of this week’s AI news centered on the tension between rapidly expanding AI capability and the security practices needed to keep pace with it. For enterprises in regulated industries like healthcare and finance, tools like HEIR represent a meaningful, if still early-stage, path toward using AI more broadly without expanding the amount of raw sensitive data that has to be exposed to third-party AI providers along the way.
7. Amazon is buying and dismantling rare books to train AI models
Citing a 404 Media investigation, TechCrunch reported this week that Amazon has been purchasing rare and out-of-print books, physically dismantling them, and scanning the pages to build training data for its AI models. The practice raises a set of questions that go beyond the usual copyright debates surrounding AI training data: rare books are often the last surviving physical copies of a given edition, and destroying them to create a digital scan trades an irreplaceable physical artifact for a reproducible one, a tradeoff that historians, archivists, and rare book librarians have flagged as a preservation concern distinct from questions of consent or compensation. The report adds Amazon to a growing list of major AI developers whose data-sourcing practices have drawn scrutiny this year, as the industry’s appetite for high-quality training text has pushed labs toward increasingly aggressive methods of acquiring it, from scraping the open web to licensing deals with publishers to, in this case, buying and destroying physical books outright. Universities, libraries, and other institutions that hold or sell rare materials are being encouraged to think through provenance and preservation questions before entering into any licensing or sale arrangements involving AI training, since a book sold for scanning cannot later be un-destroyed if concerns emerge about how the resulting data was used. Amazon has not detailed the full scope of the program or which AI systems the resulting data is intended to train.
8. Nvidia invests $1.5 billion in the data center developer behind a major OpenAI project
Nvidia announced a $1.5 billion investment in a SoftBank-backed data center developer that is building infrastructure tied to one of OpenAI’s large compute projects, deepening the increasingly circular web of financial relationships among the companies racing to build out AI infrastructure. The deal values the developer at roughly $3.5 billion as it scales Nvidia-based data centers, and comes as chipmakers, cloud providers, and AI labs continue to invest in and lend to one another through overlapping arrangements that make it difficult to fully separate genuine end-user demand for AI compute from investment flows designed to secure supply chains and preferred customer relationships. For builders and buyers of AI infrastructure, the practical takeaway highlighted by industry commentators this week is a blunt one: model capability increasingly depends on power, physical data center capacity, and the cost of keeping GPUs fed, not just on algorithmic breakthroughs. That dynamic has pushed chipmakers like Nvidia further into the business of financing the infrastructure their customers need, rather than simply selling hardware and letting customers arrange their own capital. Critics of the broader AI investment boom continue to point to this kind of circular, multi-party financing as a reason for caution about how much of the AI infrastructure buildout represents durable demand versus vendor-financed capacity that could look very different if growth expectations don’t hold up.
9. Dario Amodei says AI faces a crisis of trust
Anthropic CEO Dario Amodei used a post on X this week to address growing public skepticism toward AI, calling negative public perception of the technology a big problem rooted in what he described as a fundamental crisis of trust. Amodei said people don’t trust AI companies, governments, or the tech industry more broadly, because they suspect these institutions are cooking up some new way to take advantage of them. Pushing back on the idea that his own public messaging has been disproportionately negative, Amodei argued it has actually been roughly balanced between AI’s risks and its benefits. His proposed solution was straightforward, if difficult to deliver on quickly: AI companies, including Anthropic, need to actually follow through on their biggest promises, such as using AI to help cure diseases, rather than continuing to talk about that potential in the abstract. He called the criticism that AI companies have not yet delivered on their big promises to benefit the world the most accurate critique currently being leveled at the industry. The comments landed the same week Anthropic disclosed a dramatic revenue surge and moved toward a fall IPO, a juxtaposition that underscores the tension AI labs are navigating this year: rapid commercial success and enterprise adoption on one hand, and a public that remains broadly unconvinced the technology is being built with their interests in mind on the other.
10. French regulator says AI summaries have cut publisher traffic by up to 38%
France’s communications regulator estimated this week that AI-generated search summaries have reduced traffic to publisher websites by 33% to 38%, reviving a fight over how AI companies use news content that has already produced payment requirements and a €500 million fine against Google tied to earlier copyright negotiations in the country. The dispute centers on AI-generated answers that summarize news articles directly in search results or chat interfaces, letting readers get the gist of a story without ever clicking through to the publisher’s site, and therefore without generating the ad revenue or subscription conversions that click-throughs have traditionally provided. For publishers, the finding adds hard numbers to a concern that has circulated informally for much of the year: that AI search features are quietly cannibalizing the traffic their businesses depend on, even as AI companies continue to rely on that same news content to train and ground their models’ answers. For marketers, the implication is that measuring the value of earned media and organic search visibility is getting more complicated, since a story can perform well in an AI-mediated answer while generating little or no measurable traffic in traditional web analytics. If French or European regulators require compensation or place new constraints on AI-generated news summaries, similar to how earlier disputes with Google were eventually resolved, it could reshape referral traffic patterns, publisher licensing deals, and the broader economics of news distribution well beyond France’s borders.
That’s the week in AI. We’ll be back next Friday with another roundup of the stories shaping how AI is changing work, technology, and everyday life.
PMV Consulting, LLC 
Leave a comment