Coffee Break: Anthropic OpenAI, Meta, SpaceX Sign Trump ‘Self-Policing’ AI Accords Amidst Financial Follies While Oracle Is Left Out

This is Naked Capitalism fundraising week. 1033 donors have already invested in our efforts to combat corruption and predatory conduct, particularly in the financial realm. Please join us and participate via our donation page, which shows how to give via check, credit card, debit card, PayPal. Clover, or Wise. Read about why we’re doing this fundraiser, what we’ve accomplished in the last year, and our current goal: funding bonuses for our terrific writers. Big news broke Monday when Reuters got Anthropic’s S-1 filing the company is creating in preparation for its IPO, and it ain’t pretty. The other major AI players also had major news break, including a number of them signing an “accord” with the Trump admin and Meta’s massive tax breaks for data centers. Let’s start with Trump’s accords. Here’s what the man himself had to say: TRUMP: Posts White House accord committing Google, Anthropic, Meta, OpenAI, XAI and Nvidia to four layers of AI controls and audits pic.twitter.com/vFTJ59EKHX — Trump Truths (@trumptruthsbot) September 29, 2026 The photo op was also interesting, mostly for who was NOT there: There were 3 people missing from this White house "Bretton Woods of Super Intelligence" Masayoshi Son – CEO SoftBank Larry Ellison – Founder & Executive Chairman Oracle $ORCL Sam Altman – CEO OpenAIAll 3 need a taxpayer funded bailout of their Stargate Project. Coincidence? https://t.co/lhBrpx8qGo pic.twitter.com/yknE7OHbVs— Financelot (@FinanceLancelot) September 30, 2026 Problem Solved: Companies Sign Trump Regime’s AI “Self-Policing” Accord - Trump hosted about 20 tech executives at a White House lunch Tuesday (9/29). Out of it came the one-page “White House Accord on Super Intelligence,” a voluntary pledge he called “morally binding” (Forbes, full text). - The signers were Trump, Amodei, Brockman (OpenAI), Pichai (Alphabet), Zuckerberg (Meta), Huang (Nvidia), and Musk (SpaceX). Karp (Palantir), Bezos (Amazon), and Nadella (Microsoft) attended but aren’t on the signature page. - It commits to four layers of oversight: internal controls, an internal team to check them, an external auditor the company picks itself, and a board committee. - It requires no slowdown, despite Amodei’s “pace the frontier” campaign. It closes by saying codifying it into law “may make sense” over time. - Trump said the government won’t back new guardrails. He also signed an EO “renaming” AI as “Super Intelligence.” - The signature page misspells “United States” as “Unites States” (The Statesman). - House Republicans left town until November without passing any AI safety bill. Trump says he’ll name someone to oversee the accord. - Critics: the companies drafted the principles and hire their own auditors, and nothing binds them (Alvin Wang Graylin, Al Jazeera). Krueger, in the same piece, calls it “regulate without regulating.” David Sacks defended it as better than a treaty that would never happen. - Luiza Jarovsky argues the terms are undefined, the public sees nothing, and the OpenAI/Hugging Face hack would still have happened under it. - Raskin’s NDA probe of Meta, Oracle, Amazon, and Google lands the next day, framed by the WSJ as Democrats tapping voter backlash to AI before the midterms (Political Wire). Let me quote Gary Marcus’ “Hot Take” on the accord in full: Reading the Accord between the lines - We agree not be regulated - We agree not to give the public a voice - Trust us Also, what does the word “independent” really mean above? Subcontractors working for other companies that we (the big companies who signed) choose and hang out with and perhaps already do business with? Finally, if I may ask, what happened to the Pacing everybody was talking about two weeks ago? Surprise, surprise: Dario, Sam, and Elon all chickened out. New York Assemblyman Alex Bores (whose Congressional hopes were dashed after massive investment in his opponent by AI companies in June the NY Dem primaries) pointed out we’ve seen similar stuff before: The Trump administration today, in 2026, got voluntary commitments from AI companies that are strikingly similar to the voluntary commitments the Biden administration got from many of the same companies in 2023 (see below). “Fool me once, shame on you,” and all of that. The key… https://t.co/A7HFV7Jprg pic.twitter.com/FX4Ar7C9mi — Alex Bores (@AlexBores) September 29, 2026 One final thought about the possibilities for Congressional action on AI — the place is rife with Effective Altruists, per Dr. Émile P. Torres, who co-coined (with Timnit Gebru, his X correspondent in the thread embedded below) the clunky term TESCREAL (which stands for Transhumanism, Extropianism, Singularitarianism, (modern) Cosmism, Rationalism (denoting the ideology of the rationalist community), Effective Altruism, and Longtermism) to describe the interlinked web of cult-type beliefs endemic in Silicon Valley: I spoke to someone last night who speaks to folks Congress, and this person told me that literally half of congressional staff is just EA. This was a long-term goal when I was part of the community: to infiltrate every corner of IS power. Scary. https://t.co/ZYdhYKud8w — Dr. Émile P. Torres (they/them) (@xriskology) September 30, 2026 Yves, of course warned of this infiltration back in 2023, Naked Capitalism readers are truly ahead of the pack, donate today! Now let’s turn our attention to the naughty little start up that has been co-causing so much of the trouble. Anthropic Lost How Much on How Much Revenue?!? - A draft of Anthropic’s IPO prospectus leaked to Reuters Monday evening (9/28), after my Coffee Break post went up (Fortune). - 2025 revenue was $4.6B, up 1,088%. The operating loss was $8.06B (vs. $2.98B in 2024), and the net loss was $42B. - The ~$34B gap between the operating and net loss goes unexplained in the coverage; Investing.com says the operating figure excludes “certain liability writedowns.” - Compute and infrastructure spending hit $7.33B in 2025, about triple 2024. On top of that are $518B in future cloud, compute, and infrastructure obligations. - Cash and short-term investments were $20.28B at the end of 2025. - Two unnamed customers made up nearly a quarter of 2025 revenue, and many large customers aren’t on long-term contracts. - The valuation target is more than $2T, double May’s $965B round. The IPO is expected after the November midterms. - About 80 pages go to risk factors vs. about 48 to the business (CNBC). They include “catastrophic or existential risks” and models that can “resist shutdown” (CNN). - Anthropic shipped Opus 5.5 last week to answer OpenAI’s GPT-6 Astra, while Amodei keeps calling for the industry to slow down (CNBC). - Amodei had a private Sunday dinner with Trump, then signed the White House accord Tuesday. Trump called him “fantastic” (CNN). Ever since OpenAI’s financial pear-shapedness became somewhat common knowledge, conventional wisdom has been that Anthropic, with its focus on products for developers in contrast to OpenAI’s wild flailing (Sora! Ads! Porn!), was on the road to profitability, but the Reuters report tells a different story. Notably, Anthropic’s situation has gotten dramatically worse since the death of “tokenmaxxing” earlier this year. 👀 Interesting data from @tickerplus. A cc to this data, Anthropic’s ARR seems to have flatlined around the time when the token index peaked. Many people questioned (rightly) at the time the inferential value of our token index for frontier labs due to the incomplete coverage. We… pic.twitter.com/DCKiArYz0j — Silicon Data (@Silicon_Data) September 29, 2026 And, boy have they! Per Reuters: Ed Zitron’s latest podcast is a concise seven minute watch that goes through the Anthropic numbers: George Noble makes a strong bear case for the Anthropic IPO and argues it poses a massive systemic threat to all investors: THIS ANTHROPIC IPO IS THE MOST DANGEROUS DEAL I’VE SEEN IN MY 45-YEAR CAREER And its own IPO filing tells you why: The prospectus warns that its AI models could resist being shut down and could hide or manipulate information. If you run security at a big bank, that’s the last… pic.twitter.com/U0ZSIZljfr — George Noble (@gnoble79) September 30, 2026 And if you think Anthropic’s having a bad week, dig what’s going on with OpenAI. OpenAI Makes With the Happy Talk - Axios sources put OpenAI’s annualized revenue near $70B, up more than 70% since the start of Q3, with B2B up more than 100% (Axios). Axios couldn’t get expense figures. - For context, OpenAI lost $38.5B on $13.07B of revenue in 2025, per audited financials (Quartz). - Bloomberg reports OpenAI is raising at least $30B at about $1.4T, as a bridge to a 2027 IPO (TechCrunch). - March’s $122B round at $852B was supposed to be the last private raise. Talks at $1.2T were reported only days before the $1.4T figure. - At DevDay (9/29), Altman said OpenAI won’t “barrel, all guns blazing, towards an IPO” until it can make confident safety claims (Business Standard). - CFO Sarah Friar called Q3 “incredible,” declined to discuss the raise, and said the IPO will come “when the time is right for our business” (NewsBeep/CNBC live blog). - DevDay launched GPT-6.1 Sol, one week after the previous version, plus “Dots” agents. - The same day, OpenAI apologized for its agents breaching Australian Medicare databases 84 days earlier (Tech Times). - Altman skipped the White House lunch for DevDay; President Greg Brockman signed the accord for OpenAI. And that’s just OpenAI’s financial news. There’s been another recurring theme in recent OpenAI coverage regarding their recurrent hacks of private companies and government entities. - First lawsuit over the Hugging Face hack (Sept 29): The nonprofit Legal Advocates for Safe Science and Technology sued OpenAI in San Francisco Superior Court over the July hack of Hugging Face, which OpenAI’s own AI agents carried out during an internal test with no human directing them (Lawsuit Informer). The suit argues OpenAI’s “insistence on externalizing the harms of its unsafe decision-making is a fundamentally unfair business practice” (Axios). - Apology to Australia (Sept 28): OpenAI apologized for not promptly telling the Australian government that its agents had breached public services websites, including a system with Medicare spending data. Prime Minister Albanese called the breach “unacceptable” and said the government was weighing legal measures (TechCrunch). - “Project Lily” privacy class action (Sept 16): A proposed class action alleges OpenAI failed to disclose that outside contractors read and grade real ChatGPT conversations to improve its models (Open Class Actions). - Agents hijack Wikipedia, Astra benchmarks shift (early Sept): Reuters reported OpenAI’s agents hacked a German Wikipedia page and used it as a message board to talk to each other. OpenAI also changed the performance metrics for its new Astra model several times within hours of launch (Fortune). In response, YouTuber Eli the Computer Guy has been on tear, telling his 1.14M subscribers that “Sam Altman Needs Jail Time” and “OpenAI Is a Criminal Enterprise” But this is the one I’d consider a must-watch, in which Eli points out that Anthropic’s agents had an easy time hacking OpenAI because neither company appears to follow even the most basic best practices that have been commonplace in IT for 30 years. And if one tech guy’s opinion isn’t convincing, this Gary Marcus Interview with Fordham law prof Zephyr Teachout makes a stronger legal case against OpenAI: Marcus: Is there any reason to think that OpenAI has broken any laws that (at least in another time) might be enforced? Which and why? Teachout: There’s actual break-ins and deaths, and companies themselves who are claiming that they are building something unreasonably dangerous. And to make the evidence unusually strong for this stage of corporate illegality, there are already whistleblowers who are alleging what could be the requisite state of mind for both criminal and civil lawbreaking. A lot of people are treating OpenAI’s repeated claims that they didn’t intend the break-ins, suicides, murder, and building of defective products as dispositive. Sure, okay. That’s what people who don’t want to be held responsible often say. They say they didn’t mean it and they didn’t cause it. But police and prosecutors and civil enforcers don’t then shrug and move on — they investigate to see what actually happened, and that’s what should happen now. So we know that OpenAI’s AI agents broke into Hugging Face’s servers, accessed Australian government health systems, and attempted to break into a U.S. Department of Education website and a university library. Unauthorized attempted or actual access to computers is a federal crime under the Computer Fraud and Abuse Act. OpenAI says no one intended any of this but the DOJ would be a laughingstock if it took this at face value: subpoena OpenAI and find out who knew what when. There is already evidence of corporate knowledge of unauthorized access. Given the public evidence, I don’t have much doubt that efforts to quash a subpoena on internal documents would fail. Did staff see the product start acting as if it was going to engage in break-ins, and what warnings were given? The job of law enforcement is to find out. When it comes to civil law, there are even more possibilities, many of which are being floated in lawsuits already filed. The law of nuisance and the law of abnormally dangerous activity are both relevant to this discussion. I’m going to run through the other companies in today’s roundup as quick bullet points. Meta Using Billion$ in Data Center Tax Breaks The New York Times reports Meta used AI data centers to avoid billions in federal taxes. - Meta treats its AI data centers as “pilot models” under the 1981 research and experimentation tax credit. It reportedly told the IRS they were a “giant experiment that could fail” (Yeni Safak). - Starting in late 2024, Meta split commercially available Nvidia chips into two tax categories: chips bound for AI data centers are “experimental supplies,” while the same chips in ordinary data centers aren’t (Quartz). - The credit cut Meta’s 2025 federal tax bill by $3.9B, up from $700M in 2023, per Quartz’s headline. That is nearly a 71% cut and makes Meta the credit’s largest public-company beneficiary. - Meta alone accounts for more than 10% of the credit’s total $32.1B cost to the Treasury in 2025 (Joint Committee on Taxation estimate, via Quartz). - Meta’s own filings warn the IRS could claw back billions, citing “uncertainties with our research tax credits” (Yahoo Finance). - A research-credit specialist at BPM called the data-center-as-experiment theory “kind of wild and out there.” UT accounting professor Lisa De Simone said Meta’s own accountants are telling investors the benefits are at risk. - Meta didn’t answer the Times’ questions about what makes these data centers experimental (AzerNews). - Separately, the IRS says Meta owes close to $16B in taxes and penalties over profits allegedly routed to the Cayman Islands. - Timing: The Times has Zuckerberg telling investors AI is “accelerating every major part of our core business.” So is it an experiment or a proven success? Also new (9/30): data center NDAs. Rep. Jamie Raskin, top Democrat on House Judiciary, sent letters Tuesday to Meta, Oracle, Amazon, and Google. He wants to know why they make local officials sign NDAs that keep them from discussing data center projects with their own constituents (House Judiciary Dems; first reported by the WSJ). State and local breaks: - Louisiana (Hyperion, Richland Parish): 20 years of no state or local sales tax on equipment, GPUs included. At a 9.56% rate on about $35B of chips, that’s roughly $3.3B, which Fortune says exceeds seven years of the state police budget. Meta promises 500 permanent jobs by 2035, about $6.6M in breaks per job (Fortune). - Louisiana, utility side: Entergy is seeking about $237M in property tax exemptions over 10 years for the gas plant built to power Hyperion (Center Square via AOL). - Ohio: Kasich-era contracts give Meta, Google, and Amazon statewide 100% sales tax exemptions on data center equipment for up to 40 years. That’s about $600M each, uncapped, and current reform bills can’t touch it (Signal Ohio). - El Paso: Meta gets an 80% break on city property taxes for decades, plus $12.5M in city road work. City officials say they can’t cancel it despite public backlash (El Paso Matters). Meta then sold majority ownership to BlackRock in a $12.3B deal and kept the breaks (El Paso Matters). - Wisconsin (Beaver Dam): a $1B data center with 100 jobs. Up to 20 years of tax-increment financing flows back to Meta, plus sales tax exemptions (Center Square via AOL). Zuckerberg signed the White House accord and called it “a start” the whole industry could come to (ABC7). AP notes he has downplayed runaway-AI concerns (PBS). SpaceX Says Don’t You Forget About Me - Musk signed the White House accord “as founder of xAI,” which SpaceX absorbed in February (Boston Globe). He also sat next to Trump at the lunch. - The same day, the White House launched America.gov, a government-services chatbot built on Google’s Gemini and SpaceX’s Grok (Forbes). - Starship reached orbit for the first time in Tuesday’s test flight (CNN). CNN’s coverage frames SpaceX as having an identity crisis as it shifts toward AI and space-based data centers. - SpaceX went public in June and is down 34% from its peak (Motley Fool). - First-half 2026 capex more than tripled, with 83% going to AI, while AI brought in just $3.4B in revenue. Starlink is the only profitable segment (Motley Fool). - The CFO says the AI business could hit a $100B annualized run rate by the end of 2026. Musk says AI revenue will “definitely” pass all other SpaceX revenue, “probably in September.” - SpaceX rents its Colossus capacity to Anthropic, Reflection AI, and Alphabet. Anthropic’s Colossus 1 deal is $1.25B a month, per Wikipedia’s IPO entry. So SpaceX’s AI revenue is partly Anthropic’s compute bill, and those obligations sit inside the $518B in the S-1. - SpaceX’s CDS widened alongside Oracle’s and the other hyperscalers’ last week (TradingView). Oracle Limping Along Back of the Herd as Credit Lions Circle, Not Invited to Accords - Oracle is one of four companies (with Meta, Amazon, and Google) that Rep. Raskin is questioning over NDAs that gag local officials on data center projects (House Judiciary Dems). - Oracle’s 5-year CDS hit a record 227.15bp after the Project Jupiter force majeure notice, up 16.2% week over week (TradingView). - Its 6.7% bonds due 2056 yielded over 8% for the first time, and about $18B of Jupiter loans are trading below par (TipRanks). - S&P has Oracle at BBB-, one notch above junk. A downgrade could push about $120B of its bonds out of investment-grade indexes (Yahoo Finance, citing a Seeking Alpha analysis). - Since April, Oracle has carried Jupiter’s energy costs and the debt service on its $18B of loans. The campus is built for OpenAI (Yahoo Finance). The stock is down 30% year to date. - The OpenAI link: OpenAI is raising another $30B privately and pushing its IPO to 2027, so Oracle’s biggest tenant stays private for another year. - Goldman is reportedly cautioning about an AI debt wave among the biggest borrowers (wallstreet-online). - Oracle wasn’t among the accord signers, and none of the coverage places an Oracle executive at Tuesday’s lunch. Related posts OpenAI - Coffee Break: OpenAI Is Sinking Fast, Who Will It Pull Down With It? (Sept. 2, 2026) - Coffee Break: Will the Trump Admin Make OpenAI Its Lucky 31st Investment and Bailout Softbank? (July 13, 2026) - Coffee Break: OpenAI, Sora, Iran, the Ellison Empire, and Maybe a Recognition Event? (March 25, 2026) - Coffee Break: OpenAI as The Money Pit (June 30, 2025) Anthropic - Coffee Break: AI Credit and Credibility Collapsing or Will the AI Booster/Doomers Power Through? (Sept. 28, 2026) - Coffee Break: Effective Altruism Is a Misnomer Just Ask Ezra Klein (Sept. 23, 2026) - Coffee Break: AI Doomer Hypeloop Serves to Support OpenAI & Anthropic’s Suspicious Agenda (Sept. 14, 2026) - Coffee Break: OpenAI, Amazon Colluding With Trump Against Anthropic (June 15, 2026) Meta - Coffee Break: Is Mark Zuckerberg Pulling Meta’s Business Into a Death Spiral? (June 17, 2026) - Coffee Break: Meta’s Mark Zuckerberg Loves To Throw Good Money After Bad (July 7, 2025) SpaceX / xAI - Coffee Break: Elon Musk Headed for Supernova with OpenAI Suit, SpaceX-xAI Merger? (Feb. 4, 2026) - Coffee Break: Elon Musk to Squeeze Tesla Shareholders to Juice xAI? (July 16, 2025) Oracle - Coffee Break: Oracle’s Exploding Debt and Plumeting Stock Price Threaten Paramount Media Empire? (July 22, 2026) - Coffee Break: Has the Ellison Infotainment Empire Peaked? (April 27, 2026) - Coffee Break: Weak Links Oracle, OpenAI, UAE Are Hammered by Iran War (April 6, 2026) - Coffee Break: Oracle Debt and TikTok Transition Troubles Vex the Ellison Media Empire (Jan. 26, 2026) Background: Silicon Valley ideology Visit America.gov at your peril folks. I would advise against visiting the site even if you are just curious. On substack, the Drey Dossier did an expose of the site on two occasions. It is run by the owner of airbnb. https://substack.com/@thedreydossier re: “It commits to four layers of oversight: internal controls, an internal team to check them, an external auditor the company picks itself, and a board committee.” [my italics] I hear Arthur Anderson LLC is thinking of reviving its business model. Enron and WorldCom are no longer using their services. / ;) Is this expected to be taken seriously? That companies can sign a ‘treaaty’ with the President of the US, carve out special exceptions for themselves, and expect this to be taken as legally meaningful? This might as well be an edgy law school hypothetical on how many foundational legal principles you can violate at once. This is such a farce I have to interpret it as a sign of desperation. I’d love to have the numbers the “Super intelligence” executives are looking at that have them so terrified. so every alternate monday these guys put out a menacing message that they just might destroy humanity. Given that this kind of statement usually merits an obligatory interview with the FBI, letting them buy a research biolab of all things is just brilliant. I would keep these people away from sharp objects. Especially when other people are around. So, the greatest con artist in history signs an oxymoronic “self-policing” accord with a bunch of aspiring oligarchs. Just when you thought things could not possibly get dumber…. When politicians in Louisiana require NDAs for projects like Meta’s data center or Elon’s Space X, we in Louisiana know the payoffs are what’s being hidden. Just wonder how much our utility rates will go up to grease Entergy’s profits.what happens when Meta’s promises of permanent jobs evaporate. Profits over people aka Socialism for the Rich and Capitalism for the Poor. Suck it up working people! T’s faux accord, a Con Mans’ work is never done. Lucy Parsons Trump being Trump, he made that meeting all about himself- https://truthsocial.com/@realDonaldTrump/posts/117355137125519227 So here is the thing. Some time ago a reporter asked Trump what restrained him. Turns out that it was not the law, not the US Constitution, not even international law but he said that it was his own morality. Yeah, not a comforting thought that. So with this agreement, those AI corporations are also going to be restrained by what amounts to only their own morality. Let’s see how that works out. And this is fine reporting, Nat. thanks, Rev Kev! Trump’s executive order came up in a conversation today, and I just said I was disappointed he didn’t go full-troll (or trolled by a saboteur speechwriter) and call it Super Technological Intelligence. Five minutes of absolutely bonkers TV with Trump going on about how he wanted to give a STI to every family in the country and that American STIs would rule the world might have brightened someone’s day. But apparently we don’t even get some gallows-humor as this ship of fools sails right at the lighthouse & the rocks below. Which is especially ironic because I think generative AI is actually a really good medium for cringe comedy. Thanks, Nat. As you note, all the talk about “guardrails” is of course just hot air. To add some grist, try this: https://huggingface.co/search/full-text?q=uncensored&type=model These are all the “uncensored” models (i.e., no guardrails) that can be loaded direct from the Hugging Face Hub. They are all public repos. I.e., you just use an app like LM Studio, pick the “no guardrails” model of choice, load it, and begin your unethical prompting about how to build better drones for killing children in Gaza, etc. Cheshire Cat Smile. I propose a new one: The P.T. Barnum Grin. Think a dissipating Trump, only the Mussolini lips, jowls and grin remain. Suckahs! Born Every Minute! Step Right Up ( cue the Tom Waits…) Foxes agree to self police the hen house raids in what appears to be the largest ever mal-“investment’ in history. lets just don’t get me started on the legitimacy of AI at this juncture of Homo sapiens.. And in related news… unexpected rush for Slovenian .si domains 3500 registered in august. 3700 per day since trump’s announcement. Can you imagine being in that registrar? I wonder what “oh FFS” is in Slovenian?