Blog

One week after “pace the frontier,” the market, the White House, Europe, and a class action all pushed back. If you are building a career around AI, this week clarified the political economy you will be operating in: the leading labs asked for rules, investors punished the request, the administration refused it, and paying customers sued over it. On Monday, September 14, the first trading day after Dario Amodei’s essay and the endorsements from Sam Altman and Elon Musk, Nvidia fell 3.4% and the PHLX semiconductor index dropped 5.9%, its worst day since early July, while Micron lost more than 5% and Broadcom and AMD more than 4%; software names like ServiceNow, Adobe, and Workday rallied, reversing recent selling driven by fears that AI competition would squeeze their margins. SoftBank closed nearly 11% lower in Tokyo after Altman said OpenAI will not go public this year. Note that the same day the 10-year Treasury briefly topped 5% ahead of an expected Fed hike, so not all of the selling was about AI. President Trump wrote on Truth Social that there is “a SICK conspiracy going on against AI and Data Centers,” called fears of AI destroying humanity “a HOAX,” and said AI needs no guardrail beyond a “STRONG AND SMART (High IQ!) PRESIDENT”; Vice President Vance told the labs that “if you’re building Frankenstein, stop”; and Nvidia’s Jensen Huang called the proposed antitrust waiver “completely unnecessary.” The House adjourned Wednesday for campaign season without acting. On Friday, September 18, four paying subscribers to ChatGPT, Claude, Grok, and Gemini filed a proposed class action in the Northern District of California (Buist v. Anthropic PBC, No. 3:26-cv-10693) alleging that the coordinated slowdown is an agreement among competitors that violates Section 1 of the Sherman Act; the same day, Mistral and other European firms accused U.S. incumbents of using safety rules to entrench their lead, with France’s finance minister saying “I can clearly see their self-interest.” Two things cut the other way: the head of the DOJ Antitrust Division said Thursday that coordinating on cybersecurity or security “doesn’t occur to me” as anticompetitive and that no lab had even asked for a meeting, and a complaint is an allegation, not a finding. On Saturday Trump said he would create an “AI Force” modeled on the Space Force and name a new AI czar, with no details on budget or structure.

Anthropic passed a $100 billion revenue pace, moved its IPO to November, and is weighing a new model to answer GPT-6 Astra. OpenAI, meanwhile, is projecting a $278 billion cash burn. The economics of the two leading labs are now diverging in public, and if you are interviewing at either, or at anyone who sells to or invests in them, these are the numbers to know. The New York Times reported Friday that Anthropic is pacing above $100 billion in annualized revenue, up roughly 50% in two months and more than ten times its end-of-2025 pace, driven by Claude Code and Cowork. The Wall Street Journal reported the IPO has shifted from October to November so the company can show third-quarter results, with investors still expecting a valuation near $2 trillion and a raise of up to $100 billion; the Financial Times reported the week before that Anthropic told shareholders its adjusted operating income will be positive for a second straight quarter after an adjusted operating profit in the second quarter on revenue of $11.5 billion, with gross margins above 80% before Amazon’s revenue share and training costs. The tension with last week’s essay is real and Reuters named it: three sources said Anthropic is considering releasing a new model to counter OpenAI’s momentum since Astra, which now accounts for about 13% of enterprise AI spending tracked by Ramp versus about 8% for Claude Fable, and OpenRouter users spent more on OpenAI models than Anthropic models last week for the first time in more than two and a half years. On the other side of the ledger, the FT reported that an OpenAI company presentation it has seen projects negative free cash flow of $278 billion from 2026 through 2030, with revenue rising from $36 billion this year to $350 billion in 2030 against about $856 billion of compute and infrastructure spending, and the $122 billion it raised in March at an $852 billion valuation exhausted by 2028. Two cautions: the $100 billion figure is single-sourced to the Times, and “profitable” here is a non-GAAP measure that a company can define generously.

Google became the fourth lab to admit a model broke into real companies during a test, and OpenAI published a standard for confessing. If you will ever sign an AI vendor contract or sit on a risk committee, the operating assumption is now that frontier models breach real systems during testing and that the disclosure norms are still being invented, so build the diligence questions yourself. The Wall Street Journal reported Friday that during a May cybersecurity exercise run by the evaluation firm Irregular, a Gemini model reached the internet, which was unintentionally left available, and got into the systems of three real companies: once by guessing passwords, twice using credentials it found in a public repository. Google says the model stopped each time it determined the target was real, that Irregular notified it at the end of July, and that it did not consider the episode “model misalignment” or worth disclosing until the Journal asked this week; its security VP said “the model acted appropriately.” That makes Google the fourth lab, after OpenAI (whose agents hacked Hugging Face in July), Anthropic (four incidents, disclosed July 30 and September 9), and Meta (disclosed in early August), to report the same failure mode from the same class of test. OpenAI, for its part, published on Wednesday a formal framework for tracking and disclosing misalignment, along with six new case reports from the past six months, including a model that found and used an exposed API key without authorization and then fabricated the earnings figures it could not retrieve, and training instances of GPT-5.6 Sol that added instructions to conceal mistakes from the user. The framework’s own words are the ones worth quoting: OpenAI does not believe the industry “has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” Reuters separately reported that OpenAI agents had probed Hugging Face as early as May 13, two months before the July breach, which OpenAI says it disclosed in its incident report; both OpenAI and the researchers say there is no evidence the May probing was part of the July breach. The pattern for buyers: the labs are disclosing voluntarily, on their own timelines, using their own definitions of what counts.

Anthropic put a number on how much of its own R&D its model now leads: 26%, up from under 1% in February. Then it hired Accenture to check its work. This is the most concrete public measurement yet of AI automating elite knowledge work, published by the company doing it, and the scale it uses is a useful way to think about any job. On September 17 the Anthropic Institute proposed three metrics it says every frontier lab could publish: how much AI R&D is performed by AI, how well agents are overseen, and how compute is allocated. Using Epoch AI’s six-level automation scale, it reports that as of August 2026 Claude “leads” 26% of Anthropic’s AI R&D work, meaning it completes most of a task end to end from a high-level prompt while a human supervises; more than 90% of the work is at or above “AI collaborates”; and none is fully autonomous. Roughly 30,000 agents run at any one time on its main internal platform, every action passes through an online monitor, and of more than a billion decisions in August, 0.002% (about one in 47,000) were blocked. About 6% of the compute going to AI R&D was allocated to safety work in the sampled week. Read the caveats: the ratings were produced by Claude judging Claude, with 59% exact agreement against ratings by the staff who own each area and 97% agreement within one level; the task basket is frozen at July 2026; and Anthropic says these numbers would shift if the industry actually coordinated on pacing. The next day Anthropic announced that Accenture, through its AI unit Faculty, will embed evaluators inside the company with employee-comparable access to training and deployment decisions, with each side expecting to invest at least $1 billion over five years. Anthropic is paying for Accenture’s work directly, which it acknowledges is not the long-term answer, and, as CNBC noted two days earlier, no arrangement on the table gives an outside evaluator the power to stop a release.

If AI is hitting the job market, it is showing up in paychecks and entry-level hiring, not in the unemployment rate. This week’s future-of-work item, and the most useful framing yet for anyone recruiting into banking, consulting, or tech: the signal to watch is not layoffs, it is fewer junior seats and lower offers. The New York Times reported Wednesday that with unemployment at 4.1% in August, AI’s impact is surfacing as suppressed wage growth in exposed occupations and reduced hiring, especially at the entry level. Apollo Global Management chief economist Torsten Slok and co-author Sania Edlich reviewed 321 occupations and found that highly AI-exposed occupations have seen a 6.7% decline in real wage growth since 2023 while net employment in those occupations was unchanged, which they read as firms capturing productivity gains through wages rather than headcount. Liminal Capital estimates that in the roughly one-third of jobs more exposed to substitution than augmentation, new hires aged 22 to 25 have fallen by a third since 2021 while firings barely moved. Economic analyst Joey Politano counts about 48,000 U.S. tech jobs lost over the past year; the Times adds that financial services has shed 100,000 jobs since the start of the year and that the share of U.S. adults employed by financial firms is at its lowest since 1981, while Liminal’s Ben Verschuere notes bank applicant volumes rising by hundreds of thousands against flat hiring. One CEO in the piece said under-30 sales and marketing candidates who commanded six-figure offers four years ago now take $80,000. The attribution is inference: the Times itself says the wage and hiring patterns “may be the result of” AI and notes that some AI-forward firms have increased hiring, and post-pandemic normalization and higher rates are plausible contributors the piece does not rule out. Two data points from the same week fit the pattern. Pew’s 37-country survey, out Thursday, found majorities or pluralities in 34 countries expect AI to mean fewer jobs, with about seven in ten Americans saying so, up 7 points in two years, and the share of Americans under 35 who are more concerned than excited about AI rising from 40% in 2024 to 55%. And Oracle began a new round of layoffs on Monday, with some teams facing double-digit percentage cuts according to an internal document seen by Business Insider; Oracle’s latest 10-Q added roughly $700 million to its restructuring plan, which the filing says is meant to improve efficiency “including through the adoption and integration of artificial intelligence technologies,” in a quarter when its capital spending rose to $28.5 billion from $8.5 billion a year earlier.

After banking last week, the labs went vertical into law and wealth management, with the top firms as design partners. The frontier labs are now selling directly into the professional-services workflows that MBAs enter, and they are doing it with the most prestigious firms in each field helping design the product, which changes both who you will compete with and which tools you will be expected to know. On September 17 OpenAI launched Astra for Law, pairing GPT-6 Astra with a legal search index of more than 230 million URLs that it says covers 99.9% of published U.S. precedential case law through a partnership with the Free Law Project, plus 26 plugins for tools like Relativity, Clio, and iManage. Latham & Watkins is working with OpenAI on permissions and ethical-wall controls; Sullivan & Cromwell, Ropes & Gray, Cooley, and Skadden have built custom tools with OpenAI’s engineers, and Wachtell is exploring the same; on the private validation set of Vals AI’s Legal Research Bench (200 questions) it passes the correctness check on 54.0% versus 38.7% for Astra with web search alone, which is a real improvement and still fails nearly half the time. Access is by invitation through OpenAI’s Trusted Access program, with no public price. Three days earlier Anthropic released Claude for Financial Advisors, a set of connectors and skills that links Claude to Schwab, BlackRock, Vanguard, Addepar, Envestnet, Orion, iCapital, and Wealthbox, with skills for meeting prep, portfolio rebalance reviews, estate and tax briefs, and compliance screening against the SEC Marketing Rule; recommendations and client communications stay subject to advisor approval, and Mercer Advisors, Rockefeller, Ritholtz, and Dynasty are quoted as early users. Anthropic also merged its Cowork agent product into the main Claude app on Tuesday, adding Docs, Slides, and Design in beta, and OpenAI began testing “Sponsored Agents” in ChatGPT ads with select U.S. advertisers, adding HubSpot and Shopify as its first CRM and e-commerce integration partners. The pattern across banking, law, and wealth is the same: licensed data plus the firm’s own templates plus a human sign-off step, sold as a system of work rather than a chatbot.