The biggest stock market debut in history may be an AI company, and the paperwork could land within weeks. Elsewhere today: a survey of thousands of executives found almost none can measure a productivity gain from AI, California is trying again to ban the algorithmic firing, a State Farm defense team admitted its filings cited seven cases that do not exist, private equity started embedding engineers instead of consultants, and a Chinese robot broke Usain Bolt's 100-metre record and then crashed into a barrier.
Anthropic's bankers are floating the largest IPO ever attempted
The banks preparing Anthropic's initial public offering have told potential investors the company could raise more than $100 billion at a valuation approaching $2 trillion, according to reporting first published by the New York Times on August 23. If it prices anywhere near that range, it breaks two records at once: the largest amount ever raised in a listing and the highest valuation a company has ever carried into one.
The comparison is SpaceX, which listed in June 2026, raised $85.7 billion and was valued at $1.77 trillion — the current record on both counts. Anthropic's last private mark was roughly $900 billion. Doubling that in months is being justified by revenue that has moved faster than almost anything in corporate history: $9 billion for full-year 2025, $11.6 billion in the second quarter of 2026 alone, and an annualized run rate above $65 billion as of July, with Claude Code named as the primary engine. A prospectus is expected within weeks and a listing within months, putting Anthropic on the public markets well ahead of OpenAI, which is not expected to list before 2027.
There are asterisks that will be visible in the filing itself. Anthropic has been supply-constrained on chips and servers for most of the year — it hired former Google custom-silicon executive Amir Salek onto its compute team this week as groundwork for an in-house semiconductor effort, which is not a move a company makes when it is comfortable with its supply.
For anyone running a business on these tools, the relevant part is not the number. It is what a public listing does to a vendor. Quarterly earnings pressure changes pricing behavior, changes which customer segments get attention, and changes how quickly unprofitable tiers get repriced. The generous developer pricing of 2026 was funded by private capital that did not have to answer to a quarterly call. That is about to change for at least one of your suppliers.
Thousands of executives cannot find the productivity gain
The most useful AI story this week is not about a model. A National Bureau of Economic Research working paper surveying nearly 6,000 CEOs, CFOs and senior executives across the US, UK, Germany and Australia found that while 69% of firms actively use AI, 89% of executives reported no measurable impact on labor productivity — sales per employee — over the past three years, and more than 90% reported no impact on employment. A separate Atlanta Federal Reserve study landed in the same place. The same executives forecast a 1.4% productivity increase over the next three years: real, but modest.
Alongside that is work from Mark Ma, a professor of business administration at the University of Pittsburgh, who analyzed millions of employee reviews, thousands of financial reports, hundreds of AI-linked layoff announcements and roughly 10,000 earnings-call transcripts. The finding is uncomfortable for anyone planning headcount reductions on an AI thesis: employee sentiment toward AI showed a strong association with firm productivity, AI-related comments in reviews ran markedly more negative than employees' overall tone, and stock market reactions to layoff announcements averaged near zero across more than half the events studied. Ma's conclusion is that AI-driven layoffs and the job insecurity they create are "actively destroying the very conditions needed" for the productivity gain to materialize.
Read together, the picture is not that AI does not work. It is that most organizations deployed it in a shape that cannot show up in sales per employee — scattered individual usage, no process redesign, no measurement.
California takes a second run at the algorithmic firing
SB 947, the No Robo Bosses Act of 2026, would bar California employers from relying solely on an automated decision system to fire or discipline a worker. It requires human review and independent verification of any such decision, prohibits systems that use personal worker data to predict future employee behavior, and requires employers to tell a worker when an automated system was involved. It was introduced by state senator Jerry McNerney on February 2, 2026, co-authored by senator Eloise Gómez Reyes, and sponsored by the California Federation of Labor Unions, AFL-CIO.
This is the second attempt. Its predecessor, SB 7, passed both chambers in 2025 and was vetoed; the current version has been modified to address the governor's objections. "We need stronger guardrails to make sure there is human review and oversight of any decision made by a machine that impacts a worker's job and paycheck," said AFL-CIO president Lorena Gonzalez.
Whether or not it passes, the operational requirement it describes is one small employers should already be able to satisfy. If you use software that scores applicants, flags attendance patterns, ranks performance or generates a disciplinary recommendation — and most modern scheduling and HR platforms now do at least one — you should be able to say who reviewed the output, what they checked it against, and whether the employee was told.
Seven cases that did not exist
Defense lawyers for State Farm in a Los Angeles Superior Court fire-damage case admitted that filings submitted on the insurer's behalf contained seven case citations that do not exist, spread across eight filings, along with incorrect case titles and fabricated quotations. The firm is Musick, Peeler & Garrett; the case is *Meni-Siliga v. State Farm*, brought by a postal carrier whose Carson home burned in 2020 and who sued in July 2024. The problem surfaced at a conference on August 7, with trial scheduled for October. An attorney at the firm, Jacquelene Robinson, had used an AI tool called Irys. Lead trial counsel Kenneth Katel said: "I was not aware that AI had been used, but as lead trial counsel I accept full responsibility."
That last sentence is the whole lesson, and it is not a legal-industry lesson. A senior person was accountable for output produced by a tool he did not know was in use, in a workflow nobody told him had changed. Every business now has some version of that exposure.
Consolidation arrives at the AI infrastructure layer
Hugging Face has engaged a bank to gauge buyer interest at a valuation of $13 billion or more, per reports published August 23 — nearly triple the $4.5 billion the company was worth after its $235 million Series C in August 2023. What is being priced is distribution: more than 3 million public models and over 1 million datasets, plus the developer ecosystem around them. No agreement has been reached and talks are described as early.
It is not an isolated move. Stripe acquired the model router OpenRouter for $7.5 billion earlier this month, and Nvidia is reportedly in discussions to take a multibillion-dollar stake in Perplexity at a valuation above $30 billion, with Perplexity's annualized revenue past $750 million. The pattern is the same across all three: the money is moving toward whoever sits between the models and the people using them. Model weights are commoditizing; routing, hosting and default placement are not. The demand side agrees — open-weight models went from 28% to 62% of token share on Vercel in two months, and Anthropic's Opus 5 overtook Fable 5 in corporate spending within a month of launch, at roughly half the price.
Private equity stops hiring consultants and starts embedding engineers
Blackstone and Hellman & Friedman have formed a joint venture called Ode with roughly 160 AI engineers, backed by approximately $300 million each from Blackstone, Hellman & Friedman and Anthropic — about $1.5 billion in total. The engineers are placed directly inside portfolio companies to build new product lines and grow revenue, explicitly not to cut costs. Blackstone plans to deploy Ode to 25 of its 270-plus portfolio companies, with engineers already working at six as of last week. The showcase example is deliberately unglamorous: at Chamberlain Group, a garage-door-opener manufacturer, engineers extended the software so homeowners get notified when a package arrives. "What we liked was this was a senior engineering team. It's not like a bunch of kids who just learned AI," said Rodney Zemmel, who runs Blackstone's operating team.
Set this against the NBER survey and a resolution appears. The firms that cannot measure a gain bought tools. The firms betting real money are buying engineering time and pointing it at a product line. For a smaller business the affordable version is one competent person given one revenue-facing process and a quarter to rebuild it.
Physical AI
The largest embodied-AI funding round of the year came out of an automaker. XPENG announced on August 24 that its humanoid robotics unit raised more than $900 million at a post-money valuation above $6.3 billion, led by IDG Capital with strategic backing from Tencent and Alibaba — the largest single private round in China's embodied intelligence sector, by the company's account. Its flagship, IRON, carries 76 degrees of freedom across the body with 21 per hand, and runs three in-house Turing chips for a claimed 2,250 TOPS of onboard compute. The plan is to apply automotive-grade manufacturing to robot production, which is the most credible path anyone has described to getting unit costs down, because it is the one problem the company has already solved once.
Drone delivery took its most consequential commercial step to date. Zipline and Uber will offer drone delivery through the Uber Eats app, launching in Dallas and Houston where Zipline already operates, with Uber also making a strategic investment. The stated ambition is 1 million deliveries per day at 5 to 10 minutes each. Zipline's operating record is why that deserves a hearing: more than 20 million items delivered and over 135 million autonomous miles flown, largely in medical logistics serving 5,000-plus hospitals. For restaurants and local retailers in those two metros, this is the first time drone delivery arrives as a channel inside an app customers already have rather than as a pilot.
Waymo revealed its first custom chip — a 5-nanometer ASIC fabricated by TSMC, delivering more than 1,000 TOPS dedicated to front-end sensor processing. It handles the raw feed from 13 high-resolution cameras, four lidars and radar before anything reaches the driving software, and is in production in the Zeekr-built vehicle. The significance is economic rather than technical: purpose-built silicon is how a per-mile cost curve bends, and it signals Waymo expects volume large enough to amortize a chip program.
At the second World Humanoid Robot Games in Beijing, Tiangong Ultra — built by the Beijing Humanoid Robot Innovation Centre — ran a 100-metre heat in 9.39 seconds, beating the 9.58-second human record Usain Bolt has held since 2009; Lightning finished in 9.47. Tiangong followed with a 400 metres in 38.15 seconds, then stumbled off course into a padded barrier; Lightning fell and was carried off on a stretcher. That is the honest summary of the field: extraordinary peak capability, unreliable recovery from the unexpected. The organizers appear to agree — this year's programme added 14 scenario-based events across nine sectors including households, hotels and logistics, plus a dexterous-hand competition, because sprint times do not sell machines to warehouses.
Quick Takes
Apple laid off more than 200 employees across its Vision Pro, Siri and software engineering teams, refocusing engineering on AI and new hardware form factors.
OpenAI expanded ChatGPT Ads to 31 European countries on August 18. Free and low-cost tiers will see ads; paid subscribers remain ad-free.
Anthropic expanded Claude Mythos 5 to defenders, letting Claude Security scan enterprise codebases and suggest patches for human review without giving users the ability to prompt the model directly. It also committed $35 million in credits to help open-source projects find and fix vulnerabilities.
Two unreleased Anthropic models, codenamed Marshmallow and Melon, were spotted on developer platforms as claude-marshmallow-eap and claude-melon-eap. Neither has been confirmed by the company.
Meta hired former OpenAI researcher Luke Metz into Meta Superintelligence Labs, and OpenAI acquired the team behind Instant to build real-time agent memory infrastructure.
Nvidia's AVO coding agent scored 100% on ARC-AGI-3, completing all 183 levels across 25 public environments without instructions, stated rules or explicit goals.
Alibaba and Tencent together spent $18 billion on AI infrastructure last quarter. Alibaba is raising HK$80 billion (about $10.2 billion) in new shares for AI, after Q2 capex of 67.7 billion yuan, up 75% year over year; its cloud and AI revenue grew 45%. CEO Eddie Wu Yongming says the compute investment breaks even within three years.
More than 1 million people have used LinkedIn's AI slop button, the platform's control for flagging machine-generated filler, and business schools are now charging around $28,000 to train executives as "Chief AI Officers."
What This Means for Your Business
Measure one process, not your whole company. The NBER survey is the most important number in this briefing: 69% of firms use AI, 89% of executives cannot detect a productivity effect. That gap is almost never a model problem. It is that AI got distributed as individual licenses rather than applied to a specific process with a before-and-after number attached. Pick one repeatable workflow this month — quote preparation, intake, invoice reconciliation, first-draft proposals — and record two things before you change anything: how long it takes and how often it has to be redone. Then apply the tool to that one process and measure again in thirty days. A single verified result gives you a basis for the next decision. A company-wide rollout gives you a line item and a shrug.
If you are considering AI-attributed layoffs, read the Pittsburgh finding first. Mark Ma's analysis found that the market barely rewards these announcements — reactions averaged near zero across more than half the events — while employee sentiment toward AI, which turns sharply negative when jobs are cut in its name, correlates strongly with actual firm productivity. The mechanism is not mysterious: the people who know where the manual work is buried are the people who stop volunteering that information when they think the tool is aimed at them. If AI genuinely lets you do more with the team you have, the more valuable move is to say so explicitly and redeploy people onto revenue work. That is precisely what Blackstone is paying $1.5 billion to do inside its portfolio companies.
Write down where automated decisions touch your employees, before someone asks. California's SB 947 would require human review, independent verification and employee notification whenever an automated system contributes to a termination or disciplinary action. Even outside California, that is the standard your own HR platform's audit trail should be able to meet. Inventory the software that scores, ranks, flags or predicts anything about your staff — scheduling tools, applicant tracking, performance dashboards, attendance monitors — and for each one, write down who reviews its output and what they compare it against. This is a one-afternoon exercise that becomes very expensive to do retroactively during a dispute.
Separate drafting from verification, and make AI use visible inside your team. State Farm's outside counsel filed eight documents containing seven imaginary cases because the person accountable for the work did not know the tool was in the workflow. Establish two norms: anyone using AI on client-facing or filed work says so, and every factual claim, number, quote and named source gets checked against a document a human actually opened. The check has to be a separate step performed with the source in hand — asking the same tool to confirm its own citations is how these errors survive review.
Do not lock yourself to a vendor that is about to have shareholders. An IPO at anything near the scale Anthropic's bankers are describing changes the vendor relationship: pricing decisions start answering to quarterly earnings, and the tiers that look generous today are the ones with the least defensible margins. Meanwhile open-weight models went from 28% to 62% of token share on Vercel in two months, and corporate spending is already rotating toward cheaper options. The defensive posture is not to pick the winner. It is to keep your prompts, evaluation sets and business logic in your own repository rather than inside a vendor's console, so that switching is a configuration change rather than a rebuild — and to compare providers on cost per successfully completed task, not on the per-token rate.