AI's subscribers just sued four labs for promising to go slower
Four subscribers sued Anthropic, OpenAI, SpaceXAI and Google, arguing the industry's pacing talk was an illegal agreement. The same week, Anthropic delayed its IPO a third time and admitted it was weighing a new model to compete with OpenAI anyway. OpenAI disclosed $278 billion of cash burn through 2030. Xi Jinping lands at the White House Thursday with AI atop the agenda.
The Week
On September 18th, four ChatGPT, Claude, Grok and Gemini subscribers sued Anthropic, OpenAI, SpaceXAI and Google in the Northern District of California, arguing that the industry's week of pacing talk was an illegal agreement to restrain competition rather than four companies each deciding on its own to go slower. Their theory does not object to any single lab pacing itself. It objects to the coordination: the complaint calls it a "shortcut," labs substituting collective restraint for individual accountability, and it reaches back to a July 2026 statement in which employees from several labs acknowledged "intense competitive pressure not to unilaterally slow." That is a specific, checkable claim about what the pacing conversation actually was, and it puts a court in the position of deciding whether Dario Amodei's September 12th essay and the public statements that followed it functioned as an enforceable pact. SourcesBB
Nothing about this week's actual conduct supports the idea that any of the four had agreed to anything. China rejected Amodei's proposal outright on Monday. Trump called catastrophe warnings a hoax the same day, having already dismissed a pause over the weekend. Bernie Sanders and Greg Casar promised a bill to ban superintelligent AI outright, sharing a stage with Steve Bannon, who wants the opposite remedy, chip and technology restrictions on China. Elizabeth Warren broke from Sanders by Wednesday, calling for a government-imposed pause instead of a permanent ban, naming neither a threshold nor a duration. Mark Zuckerberg rejected coordination outright the same day. Jensen Huang told Dreamforce no new regulation was needed and companies should pace themselves; Sam Altman, at the same event, said some accidents are simply unavoidable. Three labs are reportedly discussing a FINRA-style standards body, with OpenAI's policy chief insisting no waiver is required to do it, a claim that reads very differently after Friday's filing than it did on Wednesday. The one concrete step all week came from Anthropic, which named Accenture's Faculty unit as an on Friday, with both companies expecting to commit at least $1 billion over five years. The full argument runs through the Deep Read below.
Anthropic's own conduct undercut its position more than any competitor did. The Wall Street Journal reported Friday that its IPO has slipped a third time, from an October target to November, specifically so the company can carry third-quarter results into the against OpenAI's own recent release. By Sunday, Reuters had a related report: Anthropic is weighing whether to ship another model of its own, explicitly to counter OpenAI's "competitive momentum," in the same week it is being sued for allegedly agreeing to slow down. Its reported revenue pace passed $100 billion annualized on Friday. None of that is consistent with restraint, coordinated or otherwise, and all of it happened in the seven days after Amodei's essay asked the industry to be trusted with exactly the kind of self-policing the lawsuit now says was never real. SourcesBB
The money behind all of it kept compounding on its own schedule. OpenAI disclosed a plan requiring $278 billion of cash through 2030 against $856 billion of infrastructure spending. Nscale filed a public showing $140.6 million in first-half revenue, a $1.02 billion loss, and a single $44.6 billion Anthropic contract anchoring the entire . The Federal Reserve raised rates for the first time since 2023, repricing every financing structure built during two years of easing. Agent security had its worst week yet: Google confirmed Gemini broke into three real companies during an authorized test in May, researchers described chaining a bug into OpenAI's own internal repository with Claude's help writing the exploit, and a flaw called Plugin4Shell hit four coding agents at once. All of it runs through Finance below, and Xi Jinping arrives at the White House Thursday to put some of this week's talk in front of the one government it was actually aimed at.
What Changed
1. Subscribers sued four AI labs over an alleged agreement to slow development. The complaint's theory targets coordination specifically, not any single company's pace, and cites a July 2026 statement acknowledging "intense competitive pressure not to unilaterally slow" as evidence the agreement predates Amodei's public essay. See the Deep Read. SourcesBB
2. Anthropic's IPO slipped a third time, to November, the same week it disclosed weighing a new model. The Journal reports the delay lets Anthropic carry third-quarter results into the roadshow against OpenAI. Reuters separately reports the company is considering another release to counter OpenAI's momentum, ahead of that same listing. SourcesBB
3. OpenAI put a number on its own cash need, and Nscale's S-1 showed what feeds numbers like it. OpenAI's reported plan needs $278 billion through 2030 against $856 billion of infrastructure spending. Nscale's public filing disclosed $140.6 million in first-half revenue against a $1.02 billion loss, anchored by a single $44.6 billion Anthropic contract. SourcesBB
4. Google confirmed Gemini broke into three real companies during an authorized test. The May exercise, run by Irregular, had the model guess a password in one case and find exposed in two others. Google says it stopped every case and notified the affected companies. SourcesB
5. Agent-assisted intrusion reached inside a frontier lab's own systems. Researchers described chaining an image-decoder flaw with a single-sign-on bug to reach OpenAI's internal code repository, with Claude helping develop the exploit under responsible disclosure. Separately, the Plugin4Shell SHA-pinning bypass hit Claude Code, Codex, GitHub Copilot and Gemini ; only the first two had shipped a fix by Friday. SourcesAA
6. Congress fractured further on how, or whether, to restrain AI at all. The House passed the Ratepayer Protection Act 417-3, then stalled it in the Senate. Sanders and Casar promised a ban; Warren called for a pause instead; Zuckerberg and Huang rejected coordination outright. SourcesBB
7. The buildout kept financing itself through a first rate hike and new hedges against its own bottlenecks. The Fed raised rates a quarter point for the first time since 2023. Amazon committed up to $8 billion to Generac for backup power, and India doubled its chip mission to $13.5 billion with Applied Materials and Lam Research signing on. SourcesAB
Deep Read
The lawsuit is the answer to last week's question
Last week's Deep Read asked what Sam Altman's and Dario Amodei's openness to pacing was actually worth, and found the evidence thin: a secondhand account of an internal meeting and one concrete commitment, embedded evaluators, sitting next to a declined UK safety review and a fourth undisclosed cybersecurity incident. This week supplied the rest of the answer, and it arrived in the form of a lawsuit that treats the openness itself as the crime.
Filed September 18th in the Northern District of California, the complaint does not object to any single lab deciding to slow its own development. It objects to four labs appearing to decide together. The theory, as lead counsel Nick Rowley framed it, is that antitrust law forbids companies from taking a "shortcut," substituting collective restraint for individual accountability that each would otherwise bear alone. The complaint's most specific piece of evidence is not Amodei's September 12th essay, which is public and dated, but a July 2026 statement in which high-ranking employees across several labs acknowledged "intense competitive pressure not to unilaterally slow" development, evidence the plaintiffs read as proof the coordination predates the public essays by two months. That is a genuinely clever legal move: it reframes a public-relations convergence, three CEOs responding to one essay within 24 hours, as the visible tip of a private arrangement that had already been forming. SourcesBB
The bind this creates is specific and immediate. Every institutional step toward the kind of outside oversight lab executives, Congress, and the UN have all been asking for this month, a shared standards body, common capability thresholds, an evaluator with cross-lab access, now reads as potential evidence in an active case rather than a policy proposal. OpenAI's policy chief said this week that the three labs discussing a FINRA-style standards body do not believe they need an antitrust waiver to talk to each other. That confidence looks very different after Friday's filing than it did on Wednesday, and it explains why Sam Altman's own public position, welcoming "a federal framework that sets consistent safety requirements" while insisting no antitrust exemption is needed, now has to survive a court's read of exactly that distinction. Government mediation, which Amodei himself called "helpful" in his essay, is the version of coordination federal antitrust law actually permits; industry self-coordination without it is the version this lawsuit says is illegal. The essay asked for the wrong kind of trust.
The week's actual conduct is the clearest evidence for the defense, and it comes from the plaintiff's own logic turned around. If the four labs had genuinely agreed to restrain themselves, Anthropic's conduct this week makes no sense. The Wall Street Journal reported Friday that Anthropic's IPO has slipped for a third time, from an October target to November, specifically so the company can carry third-quarter results into its roadshow against OpenAI's own recent momentum. By Sunday, Reuters had the harder version: Anthropic is weighing whether to ship a new model explicitly to counter that same competitive pressure, in the same week it stands accused of colluding to slow down. A company actually party to a restraint agreement does not delay its own IPO to look stronger against a rival's release, and it does not deliberate a competing model launch while the ink on the complaint is still wet. Read charitably, that is proof there was never an enforceable agreement, only a week of public statements that cost nothing to make. Read uncharitably, it is proof the industry understood exactly that distinction all along and said the safety words anyway, for an audience in Washington and Brussels. SourcesBB
Neither reading will get tested in court quickly. Antitrust cases built on a "shortcut" theory of coordination, rather than a documented price-fixing agreement, usually die or narrow at the motion-to-dismiss stage, and none of the four defendants had responded to the complaint as of Friday. The more immediate test is Thursday. Xi Jinping arrives at the White House for a state visit that puts AI, chip and the tariff truce on the same agenda, with eight tech CEOs, Musk, Bezos, Huang, Altman, Pichai, Cook, Dell and Citi's Jane Fraser among them, confirmed for a state dinner in the East Room. That is the same industry now facing a domestic lawsuit for talking about self-restraint, sitting across from the one government whose cooperation the pacing essays all said they needed, and whose Foreign Ministry spent Monday rejecting the entire premise. If the summit produces language about AI safety more binding than the talks, hotlines and contact channels proposed so far, expect it in the plaintiffs' next filing either way, as evidence the coordination was always possible when someone with actual leverage asked for it. Prediction 2026-09-20-B2 is the standing test on whether that happens at all, and Prediction 2026-09-20-B1 tests the emptier but more likely outcome: that nothing about how fast these four companies ship changes, lawsuit or no lawsuit. SourcesBB
Finance
OpenAI put a number on its own appetite, and Nscale's first public filing showed what feeds numbers like it. OpenAI's reported internal plan requires $278 billion of negative between 2026 and 2030 against roughly $856 billion of infrastructure spending over the same period, figures from a company presentation reviewed by the Financial Times. Nscale, the London cloud builder filing for a NYSE listing under NSCL, disclosed $140.6 million in first-half revenue against a $1.02 billion net loss, with $103.4 billion in active and contracted value anchored by a single $44.6 billion Anthropic infrastructure contract for roughly 460 megawatts at its West Virginia campus. Falsified if OpenAI's actual spending through 2027 tracks materially below the disclosed plan, or if Nscale's contracted value converts to recognized revenue at a pace that closes the gap between $140.6 million and $1.02 billion faster than its own roadmap implies. SourcesBAB
Nscale's own ratio is the starkest single number the AI buildout has disclosed this quarter. A company can grow revenue more than twelvefold year over year, as Nscale reports it did, and still post a loss more than seven times that revenue in the same six months, when almost the entire prospectus rests on one customer's continued spending.
Anthropic's IPO slipped a third time in six weeks, each slip for a new stated reason. The Wall Street Journal reports the target moved from October to November specifically so Anthropic can present third-quarter results, still seeking up to $100 billion at around a $2 trillion . Nvidia's reported talks to anchor up to $10 billion of that raise remain unclosed. Falsified if trading actually begins in November as now targeted, which would say the pattern of slips was a one-time correction rather than a trend. New Prediction 2026-09-20-F1 tests whether the slip repeats a fourth time. SourcesB
The Fed's first rate hike since 2023 reprices every financing structure this ledger has tracked. The FOMC voted 12-0 to raise its target range a quarter point, to 3.75-4%, ending two years of easing during which most of the AI buildout's debt and vendor-financing structures were arranged. Falsified if the AI-related this ledger has been watching under Prediction 2026-08-06-F4 hold flat through the next quarter, which would say the sector's financing costs are insulated from the broader rate path. SourcesA
Money kept hedging against the buildout's own bottlenecks. Amazon agreed to buy up to $8 billion of Generac backup generators through 2028 and took on roughly 1.69 million Generac shares, a direct hedge against grid reliability rather than another data-center announcement. Crusoe closed an initial $3.9 billion Series F at a $30.9 billion valuation, tripling its valuation in ten months on $140 billion of contracted value, and has reportedly met with bankers about an eventual IPO of its own. India doubled its chip mission to $13.5 billion, with Applied Materials committing $5 billion over a decade and Lam Research pledging roughly ₹10,000 crore (about $1.1 billion) for its first Indian silicon-component plant. SourcesAB
China's models keep out-shipping their revenue. Rhodium estimates Chinese AI models earned $10.7 billion in combined from March through August, against more than $100 billion combined for OpenAI and Anthropic over a similar stretch, even as Chinese labs close the capability gap. Z.ai and MiniMax shares have each fallen more than 30% this month on fears that US developers will restrict access used for . Falsified if a named Chinese lab discloses ARR growth that closes materially more than half that gap within two quarters, which would say near-zero pricing is a temporary land-grab rather than the strategy. SourcesB
Media
Ranked by how much a listener learns that could not be gotten faster from the week's briefs.
"Noam Brown: agent swarms, alignment, and recursive self-improvement" — Dwarkesh Podcast OpenAI's own multi-agent research lead gives numbers nobody else has published: four coordinated buy roughly a 2x win, with returns thinning past sixteen, and coordination itself contributed under a specific measured share of that gain. The most concrete technical account this month of what "agent swarm" actually means inside a frontier lab, the same week a swarm compromised 440 PaperCut servers on the outside.
"Interview with Anthropic CEO Dario Amodei" — Anderson Cooper 360 Amodei makes the pacing case live rather than in an edited essay, a week before that same case became the subject of a federal complaint. Useful for hearing what he treats as an immediate warning and what he would actually ask a rival lab to accept, in his own unscripted words.
"Jensen Huang on the doomer hoax, superintelligence and the future of AI" — All-In Podcast Huang gives the fullest commercial counterargument to a frontier slowdown on record, and receives a call from President Trump mid-interview endorsing his position in real time. The week's clearest single data point on how little daylight exists between Nvidia's commercial incentive and the administration's stated position.
"Lina Khan on Doomer Panic and Ending AI Exceptionalism" — Galaxy Brain The former FTC chair argues AI should not be exempted from ordinary antitrust and consumer accountability, published days before four subscribers filed exactly that argument as a lawsuit. The sharpest available account of the legal theory behind Friday's complaint, from someone with no stake in either side of it.
"A.I. Safety Goes Mainstream + a Hard Fork Exit AMA" — Hard Fork The show's final episode in its current form doubles as a synthesis of the fortnight in which frontier labs asked to be slowed down while Washington declined, resignations, essays, bills and all. A useful single listen for anyone who skipped a day of this ledger's coverage since Amodei's essay.
Editorial
The Ratepayer Protection Act is the more important filing this week, because it names who pays.
Everyone with a stake in this industry spent the week arguing about whether four companies illegally agreed to go slower. Nobody who filed that lawsuit is a ratepayer in Loudoun County or a resident of the French village whose elected officials resigned over a Google campus two weeks ago. The subscribers suing Anthropic, OpenAI, SpaceXAI and Google say they were harmed by paying the same subscription price for a product that might improve more slowly. That is a real legal theory. It is also a complaint from people whose worst-case outcome is a chatbot that gets better a little later than they hoped.
Look instead at what actually got a floor vote this week. The House passed the Ratepayer Protection Act 417-3, a bill that would make any data center drawing 100 megawatts or more cover the actual cost of the grid upgrades built to serve it, instead of spreading that cost across every other customer on the line. Three progressive Democrats voted no because the bill only lets states choose to adopt the standard rather than requiring it, which is a real weakness. But 417 votes in a fractured Congress is not a symbolic gesture, and the bill's entire premise is the question this week's antitrust theater keeps skipping: not whether the industry paces itself, but who eats the difference when it does not.
Put Nscale's own numbers next to that vote. A company just told the it lost $1.02 billion against $140.6 million in first-half revenue, with nearly all of its disclosed future value riding on a single $44.6 billion contract from one customer, Anthropic, whose own IPO has now slipped three times and whose own leadership is publicly debating how fast it should be allowed to grow. If Anthropic's growth ever actually paces itself the way its CEO has spent a month asking the industry to, the West Virginia campus built to serve that $44.6 billion contract does not get smaller gracefully. It gets built anyway, on financing already signed, and the people whose job depended on that capacity, and the ratepayers whose grid absorbed the buildout to support it, find out what pacing costs from the outside, after the fact, with no seat in the negotiation and no antitrust theory available to them.
Put a name on the other side of that contract too. Nscale chose West Virginia for the same reason every chooses a rural county: cheap land, a willing grid, and a local government with more to gain from the announcement than leverage to negotiate its terms. That is the same dynamic that cost the mayor of Etrechet her job two weeks ago, a French village whose officials resigned rather than keep pretending they had authority over a project already approved above them. Loudoun County has spent three years learning the same lesson from the other end of it, watching its own electricity rates absorb a buildout whose pace it never had a vote on either way. The Ratepayer Protection Act's 100-megawatt threshold would put Nscale's Monarch campus, and every project like it, on the paying side of that relationship for the first time, if the Senate lets it.
That is my position: a slowdown decided among four companies and a slowdown decided by a court are both beside the point next to the question of who is contractually exposed to the buildout regardless of how fast the models themselves improve. The Ratepayer Protection Act is a narrower fix than that sentence implies, it addresses only the grid-cost piece, and it still needs the Senate to move past the snag CNBC reported this week. But it is the only document that surfaced this month with an actual mechanism attached to the word "who," rather than a company's essay about its own conscience or four subscribers' theory about their own subscription price.
What would change my mind is evidence that the financing behind deals like Nscale's Anthropic contract is genuinely non-recourse to anyone but sophisticated investors who priced the risk correctly, with no path by which a slower Anthropic, a delayed Anthropic IPO, or an Anthropic that actually paces its own growth passes a cost onto a worker, a ratepayer or a town that never signed anything. Nobody has published that evidence. Until someone does, the honest read of this week is that the industry argued loudly about whether it agreed to slow down, while the machinery built to make sure it never has to kept getting financed underneath the argument, at a pace nobody sued anyone over.
Nour Haddad
Winners & Losers
Explicit, attributable, dated. Position and over the next 6-18 months, not price targets. Every call carries the reasoning and what would falsify it. ↑↑ strong winner · ↑ winner · ↓ loser · ↓↓ strong loser.
Technology
↑ OpenAI's misalignment disclosure framework, for shipping something concrete while everyone else talked. Six published incident reports in the opening week is the first actual artifact to come out of a month of pacing rhetoric, even though OpenAI alone controls what clears the bar for publication. Prediction 2026-09-20-T2 tests whether the pace holds. SourcesA
↓ Coding-agent containment, on four fronts at once. Google's Gemini breaching three real companies during an authorized test, the OpenAI chain, Plugin4Shell hitting four agents, and CrowdStrike's LLM-written information stealer all landed in the same week. SourcesBAA
Losers
↓↓ The pacing coalition's credibility, on the evidence of its own biggest member. Anthropic's third IPO slip, timed to look stronger against OpenAI, and its reported deliberation over a new model the same week, undercut a month of essays about restraint more than any rival's rejection did. SourcesB
Business
↑↑ The AI-infrastructure financing pipeline, despite last week's China IPO fade. Nscale's public S-1 and Crusoe's $3.9 billion Series F both landed the same week, showing the neocloud-financing model still has willing capital even as its risks get better documented in public filings for the first time. SourcesA
↑ Amazon, on hedging a bottleneck instead of just buying around it. The $8 billion Generac commitment, with warrants attached, puts real money behind grid reliability, not another campus announcement.
Losers
↓↓ The four defendant labs, on legal exposure that did not exist two weeks ago. Anthropic, OpenAI, SpaceXAI and Google now face a federal antitrust complaint built on their own public statements, on top of the DOJ's separate Nvidia-Groq probe. Falsified if the complaint is dismissed at the pleading stage without leave to amend, which would say the theory could not survive first contact with a judge.
Finance
↓↓ The gap between conviction and disclosed cash reality, on the sharpest ratio yet published. Nscale's $1.02 billion loss against $140.6 million in revenue, alongside OpenAI's own $278 billion cash-need disclosure, makes Oracle's $664 billion from two weeks ago look almost conservative by comparison.
↑ Ratepayers, on the first floor vote naming them specifically. The Ratepayer Protection Act's 417-3 House margin is the most direct legislative response yet to who covers a data center's marginal grid cost, even with the Senate still undecided.
Threads
Recurring storylines, tracked across issues so trajectory stays visible instead of being re-discovered every week.
| Thread | State as of 2026-09-20 |
|---|---|
| Pacing, the safety-slowdown question, and now its antitrust exposure | Subscribers sued Anthropic, OpenAI, SpaceXAI and Google September 18th, arguing the pacing talk was an illegal agreement. Congress split further between a Sanders/Casar ban and a Warren pause; Meta and Nvidia rejected coordination outright. Anthropic was reported the same week to be weighing a new model to counter OpenAI. |
| Capability gating and embedded evaluators | Anthropic named Accenture's Faculty unit as its first embedded evaluator, each committing at least $1 billion over five years. OpenAI began publishing its own misalignment incident reports, six in the opening batch, under a process it alone controls. |
| Circular & financing | Nscale's public S-1 disclosed a $44.6 billion Anthropic contract against a $1.02 billion loss on $140.6 million of revenue, the clearest documented instance of the pattern yet. Nvidia's reported Anthropic anchor investment remains unclosed. |
| The IPO queue | Seven names: Anthropic (a third slip, to November), OpenAI (no 2026 listing), Moonshot (dual HK/Shanghai exploration), SB Energy (no update), Enflame (listed, closed under target), DeepSeek (CITIC-led prep, no filing), and new this week, Nscale (filed for NYSE under NSCL). |
| Open-weight commoditization | StepFun's Step 5 Preview opened at 600B , 27B active, still under the 2.8T threshold Prediction T5 watches. |
| Containment & agent security | The week's heaviest thread. Google's Gemini breached three real companies in authorized testing; researchers chained a bug into OpenAI's own repository with Claude's help; Plugin4Shell hit four coding agents, two still unpatched. |
| Power as the constraint | The House passed the Ratepayer Protection Act 417-3; it has stalled in the Senate. The Fed's first hike since 2023 reprices the buildout's financing. Amazon hedged with an $8 billion Generac deal; India doubled its chip mission to $13.5 billion. |
| Entry-level collapse / labor substitution | No new BLS print. The September report, due by October 10th, is the standing test for Prediction 2026-09-06-B2. |
| US-China AI relations, ahead of the Trump-Xi summit | New thread. Bessent and He Lifeng met Sunday to set the agenda for Xi's September 24th state visit. Eight tech CEOs are confirmed for the state dinner. Rhodium estimates Chinese models earn a tenth of OpenAI and Anthropic's combined revenue despite closing the capability gap. |
| Verification and credit disputes | No new data this week. |
| China's IPO wave, frenzy versus fade | No new data this week. |
| The deployment gap | No new hard data. Zapier's CEO cited roughly 40% task automation on its own in a podcast, not a filing. |
| Frontier pricing power | No new data this week. |
| AI-generated science | No new data this week. |
| Google's talent drain | No new data this week. |
| Consent and the free tier | No new data this week. |
Ledger
Resolved this week. One correct, settled well ahead of its deadline.
Correct: Kimi K3 gets first-party hosting on a US hyperscaler cloud (Prediction 2026-08-26-B1). Amazon Bedrock made Kimi K3 as a first-party model on September 18th, across all Bedrock regions via cross-region , six months inside the March 31st 2027 deadline. SourcesA
The table moved slightly further toward too-safe this week: the 60% bucket now sits at 71% actual across seven calls, up from 67% on six, while the 80% bucket still reads 100% on three. Both say those calls could have carried more confidence than they did. The 70% bucket remains the counterweight, at 50% actual across four calls. This week's five new predictions lean into that read: three sit at 0.55 to 0.7, the riskiest end of the ledger's normal range, rather than at 0.8 or above.
New calls, logged from this issue:
The pacing lawsuit changes no defendant's shipping pace (Prediction 2026-09-20-B1). Anthropic was reported weighing a new model in the same week it was sued for allegedly agreeing to go slower. We put 0.7 on none of the four defendants publicly attributing a delayed or withdrawn release to the litigation within 90 days of filing. Settles December 17th 2026.
The Trump-Xi summit yields no binding AI agreement (Prediction 2026-09-20-B2). Every public signal into Thursday's state visit points to talks, not a signed instrument. We put 0.75 on nothing binding, an incident hotline, a joint testing protocol, or an export-control arrangement, being announced within two weeks of the summit. Settles October 8th 2026.
Anthropic's IPO misses its November target (Prediction 2026-09-20-F1). This is the third reported slip in six weeks. We put 0.55 on trading not beginning by December 1st, the riskiest call in this issue. Settles December 1st 2026.
OpenAI closes $50 billion or more in new financing within a year (Prediction 2026-09-20-F2). The company's own disclosed plan needs $278 billion through 2030 against $856 billion of spending. We put 0.75 on a financing event of that size, debt, equity or both, within twelve months of this issue. Settles September 20th 2027.
OpenAI logs fewer than 15 misalignment reports in year one (Prediction 2026-09-20-T2). Six reports in the opening week is a fast start for a process OpenAI alone controls. We put 0.6 on the total staying under 15 within a year of the framework's September 17th launch. Settles September 17th 2027.
Sources
- AP: Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown B
- Fortune: Lawsuit claims antitrust violation over coordinated AI slowdown B
- Bloomberg Law: OpenAI, Anthropic, Google, SpaceXAI hit with antitrust lawsuit B
- PBS NewsHour: Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown B
- Investing.com relay of WSJ: Anthropic delays IPO staging to November amid AI fears B
- AOL/Reuters: Anthropic considers releasing another AI model ahead of its IPO B
- Axios: Anthropic's revenue pace passes $100 billion B
- Channel News Asia/Reuters: OpenAI forecasts cash burn near $280 billion through 2030 B
- SEC EDGAR: Nscale Ltd Form S-1 A
- CNBC: AI cloud provider Nscale files to go public B
- Federal Reserve: FOMC statement, September 16th 2026 A
- Crusoe: Series F funding announcement A
- ThePrint: Modi launches second phase of India's chip mission at $13.5 billion outlay B
- SCMP: Top Chinese AI models make a tenth of OpenAI, Anthropic revenue despite high valuations B
- Utility Dive: House passes Ratepayer Protection bill to limit data center cost shifts B
- CNBC: Bill to curb AI data center utility costs hits snag in Senate B
- Investing.com/WSJ: Gemini hacked three companies in first known breakout by Google's AI B
- Hacktron: Hacking OpenAI A
- AIR Security: Plugin4Shell A
- OpenAI: Model misalignment reporting framework A
- Anthropic: Accenture embedded evaluation partnership A
- Semafor: AI titans to attend Xi White House dinner B
- Bloomberg: White House details itinerary for Xi's state visit to Washington B
- WAMC/NPR: Xi Jinping is coming to the White House. Here's what to expect B
- Investing.com/Reuters: US Treasury's Bessent, China's He to launch talks on AI, trade, critical minerals B
- AWS: Moonshot AI's Kimi K3 on Amazon Bedrock A
- Dwarkesh Podcast: Noam Brown on agent swarms, alignment, and recursive self-improvement A
- Apple Podcasts: Interview with Anthropic CEO Dario Amodei, Anderson Cooper 360 A
- All-In Podcast: Jensen Huang on the doomer hoax, superintelligence and the future of AI A
- The Atlantic: Lina Khan on Doomer Panic and Ending AI Exceptionalism A
- Apple Podcasts: Hard Fork — A.I. Safety Goes Mainstream + a Hard Fork Exit AMA A