Anthropic is about to attempt the largest IPO in history. Nothing can stop it — except the thing, it acknowledges, everyone is afraid of — and now everyone in the industry seems terrified too.
Anthropic confidentially filed its S-1 on the first of June. The public version is expected this month, the listing reportedly pushed back to mid-November at the earliest so the offering can include third-quarter results, and investors are talking about $2 trillion — which would make it the largest public offering ever completed, past SpaceX. That record is nearly four months old. SpaceX listed on the twelfth of June at $135 a share, raised $75 billion — nearer $86 billion once the underwriters maxed their allotment — and closed its first day worth about $2.1 trillion. The price bought factories, launch pads, a satellite constellation already in orbit and a working fleet of reusable rockets, and Morningstar still valued the lot at $780 billion — less than half the target — calling the company significantly overvalued. That is the bar Anthropic means to clear: a price the appraisers couldn't defend for a company that at least owns rockets. Reuters has read the confidential prospectus: 261 pages, about 80 of them risk factors against some 48 describing the business, including a warning that advanced AI could bring "catastrophic or existential risks to humanity." The same section, per reporting from August, lists public backlash against AI — the objection and the extinction risk, filed side by side. This disclosure warrants more attention than the genre is built to attract — risk factors are written to be skimmed, filed somewhere between currency fluctuation and the loss of key personnel. People objecting to what is being done to their livelihoods appear in the document as a threat to the share price. The objection is filed as a risk to investors, not a reason to reconsider the plan. That is a moral inversion executed as a legal document. The tempting defence is that no villain drafted it — just a template, correctly filled in. But templates don't file themselves, and this one carries signatures.
The valuation maths runs like this: $47 billion annualised revenue in May, investor expectations of $100–120 billion by December. Two trillion is north of forty times the first number and 17–20 times the second — a multiple stranded between a denominator already obsolete and one not yet earned. Bulls prefer the end-of-July run-rate of $65 billion, which brings the multiple down to about 31. Forty still holds on revenue actually booked: $11.5 billion in the second quarter, about $46 billion annualised. The bridge is a formula, offered to the Financial Times by an actual investor with actual money: a company growing 800% a year commands at least 30 times revenue. Run the formula on the December projection and it prices the company past $3 trillion, which tells you the formula is not where the number came from — the price came first, and the arithmetic was fitted afterwards. Growth justifies the multiple, the multiple attracts the capital, the capital funds the growth. If that loop sounds familiar, it's because you've seen it before, in 1999, wearing a different lanyard.
So the question worth a week of anyone's time: what, precisely, is the $2 trillion buying?
The moat inspection
Not infrastructure. Anthropic rents its physical layer — the tangible assets sit on other people's balance sheets, including SpaceX's. Not exclusivity of distribution: Amazon holds somewhere between 15 and 20 per cent of Anthropic and sells OpenAI's newest models on Bedrock, through the same pipe, to the same customers. The distributor has no loyalty because the distributor owns the shelf, not the products on it. Amazon wins whichever model wins, which is the correct way to hold a lottery and a poor thing to be a ticket in.
What Anthropic owns is weights, talent, and enterprise relationships. Weights are a strange asset class: they depreciate in months, which is why the revenue has to grow 800% a year just to hold the story upright. The closest precedent had more underneath it: Sun Microsystems sold real servers from real factories, traded at ten times revenue at the peak — a multiple its own chief executive later said no assumptions could defend — and still lost 96%. The multiple here, on annualised revenue, is four times Sun's fatal one — on far faster growth, the bulls will say, which is the argument every fatal multiple has ever made for itself — with less floor beneath it.
The standard rebuttal is switching costs, and in almost any other enterprise software market it lands: ERP migrations run for years, databases have gravity, cloud egress is priced like a hostage situation. With AI the rebuttal dissolves, and the reason is structural. Switching costs are accumulated state — the data, schema, and integrations that pile up on the vendor's side. An inference API accumulates nothing: you re-send the context with every call, so there is nothing on their side to hold you. Thirty years of enterprise moat-building has nowhere to sit. What's left is friction — procurement, security review, evals, the annual commitments Anthropic sells to manufacture more — and friction isn't gravity. An ERP deal is a multi-year term with penalty clauses wrapped around a system the business cannot run without: the customer is bound, so the revenue is safe, and the multiple prices the safety. The paper Anthropic sells runs the lock the other way — committed spend, mostly annual, a discount instrument that binds the customer's money without binding the workload; even the multi-year option merely holds a rate in a falling market, a contract protecting the buyer from the vendor's prices rather than the vendor from the buyer's exit. Gravity versus a deposit. The S-1 puts figures on the asymmetry: $518 billion of compute and infrastructure commitments to six partners over roughly a decade, about 80% of it non-cancellable or payable regardless of usage, against a customer book where two unnamed clients each generated 12% of 2025 revenue and where, the filing concedes, many of the largest customers have no long-term contracts and could cut or stop spending. Fixed obligations it cannot cancel, against customers who can walk. And in May the deposit became compulsory: Anthropic restructured enterprise billing around mandatory spend commitments sized to its own estimate of your usage, payable whether you consume it or not, with the old volume discounts eliminated — a company that must show $100 billion of committed-looking revenue by December, requiring customers to guarantee spend it estimated itself, months before the filing. Anyone building for production builds redundancy first anyway; uptime demands it. At the serious end a second vendor is already wired in and load-balanced, so "switching" isn't a decision anyone convenes a meeting for. It's a weight in a router, drifting toward whoever answers — cheaper, faster, at all. The multiple is borrowed from SaaS, where revenue renews by signature and survives a mediocre quarter on contractual inertia. Metered API revenue has to re-win the workload continuously — on price, capability, latency, reliability — request by request. SaaS defends a contract. Inference defends a title.
The market leader, meanwhile, is sitting this one out. OpenAI filed a week after Anthropic, then pushed its listing to 2027 rather than price below its trillion-dollar anchor. The company confident in its trajectory chose to wait. The one going first is the one selling the snapshot — cashing in a relative position weeks before the next reshuffle of a leaderboard nobody, including the labs, can predict two quarters out. Going first isn't only the scarcity premium of being the lone public ticker for AI-lab exposure. It's also what you do with a lead you suspect is perishable.
The transformer business
Press on what the company actually does, and the electricity metaphor the industry loves — intelligence too cheap to meter, tokens as kilowatt-hours — turns out to be load-bearing, just not where the marketing points it. Anthropic doesn't generate. The compute is rented from Amazon, Google, and lately SpaceX. The fuel — the accumulated corpus of human writing — was acquired at roughly the price of scraping it; the $1.5 billion authors' settlement stands as the only invoice the fuel ever produced. What the company does is transform: current in, current out, stepped down to the voltage of your query, minus losses as heat. In 2025 the current cost more than it sold for: $7.33 billion of compute and infrastructure against revenue of nearly $4.6 billion, an operating loss of $8.06 billion. The industry even named the architecture with perfect accidental honesty. The T in GPT stands for transformer, and a transformer is a passive component — no fuel of its own, nothing created — which in an electricity economy is precisely the layer with no pricing power. Generation has scarce fuel. Distribution is a natural monopoly. The substation in between is anonymous industrial equipment: nobody knows, or has ever cared, whose transformer serves their house. Anthropic sits at exactly that layer, with Amazon upstream selling it generation and Amazon downstream owning the wires to the customer. Bedrock is National Grid; the router is the socket. A transformer between two ends of the same company is not a business with leverage. It's a component with a logo. And strip the metaphor entirely and the position holds on documented facts alone: $110 billion of compute commitments to Amazon, another $111.1 billion to Google, rented capacity on SpaceX-owned hardware, distribution through the cloud marketplaces — Amazon's and Google's, shelves that stock the competition — which together carried 47% of 2025 revenue, a reliance the filing itself says "creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute." The transformer is the illustration. The value-chain placement is the argument.
Open source slots into the metaphor without forcing: rooftop solar. Capacity installed on your own roof, capex once, no meter running — undercutting the utility from below on exactly the workloads it priced too high, while the utility's fixed costs stay fixed. Utility economics has a name for what follows, because defection of the best customers raises unit costs for the remainder, which drives more defection. And generators of a fungible product trade at utility multiples — low single digits of revenue — because nobody pays a premium for premium electrons. The honest limit of the metaphor is also its sharpest edge. A kilowatt-hour is fungible by physical law — no instrument can tell a solar electron from a coal one, so generation commoditised totally and at once. Tokens are fungible only task by task: below a capability threshold every competent model's output is interchangeable and the buyer routes on price; above it, differences are worth paying for. All pricing power in this industry lives in that sliver above the line — and the line moves in one direction. Electricity commoditised once. Inference commoditises continuously, as each capability becomes table stakes, crosses over, and never comes back. The failover architecture every serious customer already runs is the market building that fungibility on purpose.
Table stakes are set by the table, not the vendor. OpenAI, Google and xAI all ship image and video generation inside the same product as the text; Anthropic ships neither, a restraint that read as principle in 2024 and reads as absence now that every rival sells the whole appliance. A customer who has to keep a second vendor for pictures already has the second vendor wired — the redundancy argument from the moat section, arriving from the other direction.
Electricity also comes with a hundred-year-old price history for exactly this manoeuvre. The last time generation was sold as a growth technology rather than a commodity — Samuel Insull's leveraged utility empire — the collapse took the savings of some six hundred thousand small investors with it, Roosevelt campaigned against the Insulls by name, and the wreckage produced the Securities Acts and the disclosure regime under which S-1s are filed. The form Anthropic submitted on the first of June is, in part, a fossil of the previous occasion somebody sold shares in the future of electricity at the top.
The sophisticated bull case accepts every word of this and answers: that's what the money is for — a hundred billion of proceeds as the ladder out of the transformer layer. Buy generation, buy distribution, stop being the component. But look at what the money is already for. The weights depreciate in months, so the raise is spoken for before it lands: the next training cycle, the one after that, the multi-year compute commitments already signed, the operating losses in between. That isn't leverage. It's fuel. A company whose sole asset expires quarterly cannot divert the renewal budget into becoming something else, because stepping off the treadmill to build the ladder means the only thing it owns goes stale mid-build. And the ladder leads into the territory of the companies being paid — a large fraction of the proceeds round-trips to Amazon, Google and Nvidia as compute payments. The fallback is equity as currency — don't spend the proceeds, print the shares, buy the moat by acquisition. But stock is a currency denominated in its own multiple: it spends at forty times revenue only while the market honours forty, and every acquisition big enough to matter is a bet that the honouring outlasts the integration. The moat bought with inflated paper deflates with the paper. You cannot out-build your landlord with money raised by promising to keep renting from him. The raise doesn't buy an exit from the transformer layer. It buys the electricity bill for staying in it.
The invoice
When Artificial Analysis ran its Intelligence Index against Sonnet 5 at launch, at standard pricing, the budget tier — the workhorse, the model sold as the cheap way to run agents — came out at $2.29 per completed task: roughly 15% above Opus 4.8, the model it was meant to undercut, with the full benchmark costing more on Sonnet than on Fable. The mid-tier out-billed the top of the range. The Opus/Sonnet/Haiku architecture is a promise about ordering: pay less, get less, but pay less. At the delivery layer, the ordering inverted. A price list that misleads the customers who follow it isn't a price list. It's décor. And the list has since confessed. Sonnet 5 launched with a 50% step-up written into the rate card — $2/$10 introductory through August, $3/$15 from the first of September, a rise timed to land in the same weeks as the public prospectus. On the tenth of August, Anthropic cancelled it: the introductory price is now the permanent price, the step-up footnote deleted from the page. Read generously, that's a concession to customers. Read at all, it's this essay's argument arriving on the vendor's own pricing page — a company that supposedly needs $100 billion by December looked at collecting a pre-announced 50% rise on its volume model and concluded the market wouldn't pay it. Pricing power is the ability to collect your own listed price; declining to is the commoditisation thesis, signed by the vendor — and signed again on the twenty-second of September, when Opus 5.5 arrived at $4/$20, a fifth below Opus 5's $5/$25, its claimed 40% lower running cost at default settings not yet independently measured. And the cancellation touches none of the arithmetic above, because the problem was never the price of a token.
The mechanism is not mysterious, and it was measured: roughly 40% more output tokens per task than its predecessor, about three times the agentic turns on knowledge-work evaluations, and a new tokenizer inflating counts up to 1.35x on top. Cheaper per token, dearer per job — and hold onto that tripled turn count, because it comes back later as a personality. And note the label on Fable's bar: "with fallback" — the benchmark's cost includes rerouted attempts. Whatever the precise mix of refusals and errors behind that label, the fallback architecture is priced into the chart: part of the premium is the model not answering the first time.
The defence is value per task: enterprises buy outcomes, not tokens, and a model with a lower error rate can be cheaper per unit of realised value even at a higher price per run — a failed task costs a human escalation, and escalations dwarf tokens. It's a real argument, and it defends exactly one bar on the chart: the flagship. It cannot defend the mid-tier, because Sonnet losing to Fable on delivered cost is the same lab's economical option priced above the thing it exists to economise on. There is no quality story that explains that. The premium defence and the tier structure cannot both be true, and the company is selling both. The broader version of the defence — that $2.29 a task embarrasses any loaded salary — is true and aimed at the wrong target. The threat to the price was never the human it replaces; it's the open-weight model that does the same task for ten cents six months from now. Substitution defends the utility. Nothing on the page defends the margin.
Here's the part that should bother the people writing the prospectus. On a per-token business, tokens-per-task is the revenue line. A model that burns twice the tokens to deliver the same work books twice the revenue for the same value — which means some unknowable fraction of May's $47 billion run rate is inefficiency, monetised. Every lab faces the temptation that verbosity is billable. A company that must show $100 billion by December to hold a $2 trillion print is the one least able to resist it, because in the brutal near-term accounting, making the model efficient is revenue destruction.
The temptation has a fresh timestamp. Fable 5.1 shipped at the start of September advertised as roughly 25% cheaper for typical workloads — up to 45% for agentic ones — with the entire saving sourced from one billing line: cache reads, cut 75%. Base rates didn't move. Independent measurement found the new model spending about 1.7 times the output tokens of its predecessor per task, and output is the dear lane: at full effort, $3.69 per completed task against Fable 5's $3.14 — the most expensive model on the index, 57% above the lab's own second tier. The fair reading is effort-dependent — at default settings on cache-heavy workloads the discount can be real — but notice the construction either way: the price cut lives on the billing line that doesn't grow, and the consumption growth lives on the line that does. Marketed as a saving, measured as a rise. That makes two consecutive launches in which the advertised direction of the price and the measured direction of the bill disagree — once is an accident of benchmarking; twice is the revenue model. Sonnet 5.5 followed on the twenty-eighth of September, and at maximum effort the shape repeated: the list price held at $2/$10, the advertised saving of up to 30% per task — "in our testing" — sourced from token efficiency, and the index's measurement finding roughly 60% more output tokens per task than Sonnet 5 or Opus 5.5 — $7.60 per completed task against Opus 5.5's $5.98. The run used a pre-release build with a structured-outputs bug since fixed, and Anthropic itself pitches the model at lower effort settings; the fair reading is as effort-dependent as Fable's — a data point that fits the pattern, not a third proof. And the same divergence is in litigation at the subscription counter: a proposed class action filed in June alleges the $200 "Max 20x" plan delivers six to eight times the base tier rather than twenty, the multiplier applying to a five-hour session window while a quieter weekly cap does the gating. Unproven and contested — but the same shape as every other row: the number advertised and the number delivered, parting at the layer where the customer pays.
The flagship also ships with a change the launch post files under compliance: Fable 5.1 is the first Claude to watermark every text output — an invisible statistical signature, applied worldwide because there's no reliable way to switch it off by region, readable only through a detection API in private preview for eligible organisations, with the older models due to be marked over the coming months. For chat, that's furniture. For professional writing — legal drafts, copy, anything whose author cannot deliver work that a third party with the key can test for machine involvement — it's an exit sign, hung in the S-1 window, pointing at the unwatermarked open-weight models next door. And on the craft itself, the early verdict from this desk runs the wrong way — tested, blind-checked through a rival model, and A/B'd against Fable 5 on identical prompts; the full evaluation is a separate post. Anthropic's own prompting guide gestures at the same conclusion, conceding the new model can produce denser, longer-sentenced prose and advising users to tell it to remove "mannered prose." For the flagship, the roadmap into the listing now reads: dearer per completed task, litigated at the subscription counter, watermarked at the output, and — per the vendor's own guide — in need of being told to write less like itself.
The demand side of the same chart is running on a transient of its own. Tokens are a brand-new line on the corporate P&L, and companies are in the build phase — paying frontier prices to prototype, migrate, and replace software: the Salesforce seat, the Marketo licence, the SaaS stack that can be swapped without severance, morale, or a tribunal. Build-phase spend looks like demand while behaving like capex — construction traffic mistaken for the traffic the road will carry. Klarna ran the full arc in public: ripped out Salesforce and Workday, automated the service desk, then rehired the humans when the quality bill came due. The software stayed gone; the people came back. The sequencing generalises, because any clear-headed company replaces the tools first and keeps the staff to do the replacing — which means a good share of what the prospectus books as the future of work is, today, the past of enterprise software.
Underneath all of it sits a variable the projections treat as free: operator skill. The substitution thesis prices the capability as if it deploys itself, and it doesn't. Extracting labour-grade work from these systems is a skill; the skill is scarce; and the scarcity is visible in every walked-back automation and every bloated token bill for work a competent operator gets at a tenth of the spend. The bull will call that gap headroom — skills diffuse, revenue grows as they do. But it's a race between two clocks: competence diffusing through the buy side, and the model commoditising underneath it. Everything above says the commodity clock runs faster. By the time the median enterprise can actually operate the capability, the router will be buying it at whatever the morning quotes.
Replacement revenue is real, and it recurs — a captured Salesforce workflow becomes a workload that runs indefinitely, and the SaaS pool being drained is worth hundreds of billions a year. But a recurring workload is exactly what gets optimised, because every new line on a P&L gets optimised the moment finance notices it. Caching, prompt compression, distillation onto smaller models, routing to the cheap tier, self-hosting whatever stabilises: every technique cuts tokens per unit of work, and the build-to-run transition cuts the units. The run rate supporting the multiple therefore has two transients stacked inside it — an inefficiency premium, verbosity billed as value, and a construction premium, migration spend billed as recurring demand. What survives both is genuine: run-phase workloads at optimised token counts. But that is a commodity annuity, repriced at whatever the router quotes each morning, and the multiple was borrowed from businesses that sell contracts. The valuation is priced off peak roadworks, on the assumption that this is the traffic the road will carry. The roadworks have now produced the bull's best number: second-quarter revenue of $11.5 billion, fourteen times the $787 million of a year earlier, and the first quarter of positive adjusted operating income — preliminary, and adjusted to exclude costs including stock-based compensation — with a second forecast to investors. Grant the bull its reading, volume outrunning commoditisation and margins arriving: a quarter of adjusted earnings at peak demand shows the business can earn at the peak, which is consistent with both transients still billing and says nothing about how much of the demand is build and how much is run. The quarter the $2 trillion needs — earned after finance has optimised the line and the build phase has ended — hasn't happened yet.
The flywheel
Which brings us to Opus 5, and a confession nobody noticed they were making. The model shipped in July to strong benchmarks and a field report of a machine that argues — instructions treated as openings for negotiation, simple directives costing multiple turns — and the fix, per Claude Code's own lead, was to delete over 80% of the product's system prompt, because instructions written for older models cause the new one to over-behave. Read as engineering, that's a patch. Read as evidence, it's an admission of overtraining: each instruction token now carries more behavioural weight than the environment it was written for, and the remedy chosen was to sterilise the environment rather than harden the model. Every enterprise system prompt and standing instruction anyone has accumulated became, overnight, a compatibility hazard — which also settles the last hope for the prompt moat, since a vendor that breaks its own prompt compatibility every generation makes the re-tuning cost of staying equal to the re-tuning cost of leaving. Meanwhile the internal instruments read the same events as progress: Anthropic's transparency reporting finds Opus 5 better aligned than its predecessor on precisely the axes the field is filing bugs about — ignoring explicit limits, discouraging tone, condescension. Both readings are honest, and that is the problem. Optimise hard on "never ignores an instruction" and you get a model that enforces every instruction with maximum literalism, which is the reported behaviour; the metric improving may be the very mechanism of the degradation. Proving it would take a training ablation only Anthropic can run. The company holds the experiment. The public holds the bug tracker.
Now assemble the machine. The first link is the interpretation — overtraining produces the verbosity, the reading that best reconciles the audit with the bug tracker — and every link after it is bookkeeping: the verbosity produces the token burn, the token burn produces the run rate, the run rate produces the valuation, and the valuation produces the pressure to never test the first link. Theo Browne's cost charts went round X at 736,000 views with a skull emoji, doing the price discovery the price list abdicated — while the company sits in an SEC quiet period where every public word is a securities-law event, constrained from arguing back at the precise moment the argument is being lost. That bar chart is not a pricing embarrassment. It's the flywheel, photographed.
The exit that isn't there
The fashionable reading is that machines like this need no villains — incentive structures produce the outcome with everyone locally blameless, which is what makes it durable. It's a comforting piece of sociology, and it fails at the top of the org chart. The exculpation of the cog is not seeing the machine; the chief executive's job is the assembled view. Dario Amodei said publicly in 2025 that AI could eliminate half of entry-level white-collar jobs within five years and push unemployment to double digits. Then his company built the headcount pitch, priced the prospectus on the substitution, and filed. And unemployment at that scale is not an abstraction with a retraining programme attached — it has a mortality literature. Meta-analysis puts the increase in all-cause mortality risk among the unemployed at roughly 63%; the BMJ attributed more than ten thousand suicides to the 2008 recession across Europe and North America alone. A forecast of double-digit unemployment is, by published epidemiology that predates the forecast, a forecast of deaths. The temptation here is prosecution, and prosecution invites the cheapest defence in the industry: the capability arrives regardless; we only change who profits. Let it be true — hypocrisy has no counterfactual defence. The same man issued the forecast, urged governments to prepare for it, built the headcount pitch that executes it, priced a prospectus on the substitution, and filed the objections of the people inside the forecast as a risk to his investors. A warning issued while behaviour continues unchanged is not mitigation, it's liability management — the forecast living in the communications strategy when it belongs in the company's theory of responsibility. Every sentence of that is public record. No two of them survive being read together. Whether the future was coming anyway is beside the point: the warning and the filing are both his, and they require no court — only a reader willing to hold them side by side.
The instinctive analogy for what breaks is 1999, and the balance sheet says 2006. Dot-com was an equity bubble, and equity losses stop at the shareholder. Housing was a debt bubble, and debt cascades through counterparties to people who never bought the asset. The AI buildout crossed that line over the past eighteen months, and the people paid to measure such things have started saying so in official documents. The Bank of England's July Financial Stability Report records the inflection: in 2025 required investment exceeded what AI companies could fund from their own cashflows, the sector turned to external finance, and the share of AI investment financed by private credit jumped from 9% to 34% in a single year — with the Bank flagging the growing complexity and opacity of the debt structures as a stability risk in its own right. The FSB is warning of "sizeable" losses for private-credit investors, in the same weeks that JPMorgan, Morgan Stanley and SMBC are reported to be offloading AI data-centre construction debt at scale. Senate Banking Democrats have formally pressed FSOC to investigate the trillion-plus in projected AI debt as a threat capable of triggering a broader crisis. Add lending collateralised on GPUs that depreciate faster than the loans mature, and OpenAI carrying a reported $1.4 trillion in infrastructure commitments against an annualised run-rate nearing $70 billion — twenty to one, a ratio that in 2006 was called a stated-income loan and had a ruder name in the trade. The mappings arrive pre-assembled. "House prices never fall nationally" has reissued as "AI demand only grows" — in both cases the comfort is the correlation, the single assumption that fails for everyone simultaneously. Vendor financing — chipmakers investing in the labs that spend the round on chips — is lending people the money to buy the asset that secures the loan; the deal web has now been drawn in the press as a literal circuit diagram, and Anthropic's corner of it is the tightest loop on the page: both of its major investors are also its landlords, and one of them is also the shop. An internal audit certifying the product improved on the exact axes users are filing bugs about is doing the ratings agency's job in this story — the certifier stamping AAA while the underlying defaults — with the 2006 conflict of interest streamlined: the agency is in-house. And the IPO is the distribution step — securitisation's equity-side analogue, a different instrument performing the same function. The banks of 2006 didn't keep the mortgage risk; they packaged it, rated it, and distributed it into pension funds while collecting the fee. Early money exiting at two trillion into index-tracking money — on whatever schedule the lock-ups stagger it — is the same manoeuvre in equity form. The S-1 is the prospectus on the tranche, and the tranche's buyers are pre-enrolled: SpaceX cleared into the total-market indexes within days, under fast-track rules rewritten for exactly this class of arrival, and the default funds — which is to say you — sit downstream of the stamp. The one mercy: the Lehman plumbing of deposit runs and overnight funding is absent, because this debt sits in private credit and insurance balance sheets, which fail slower and quieter than banks. 2008 with a longer fuse — which cuts both ways, since a longer fuse is also more time for the exposure to grow before anyone marks it.
Can the IPO be stopped? Not an idle question — my pension is in the vicinity and so, via the diversification illusion, is yours — a default index fund now carries the concentrated AI names whether you ticked a box or not, and a serious unwind is an event measured in tens of trillions. The answer is no, and the past six months prove it, because the stopping has now been attempted from both ends of American politics and the filing is still on schedule.
The administration went first. In March the Pentagon designated Anthropic a supply-chain risk — the first American company ever given a label built for foreign adversaries — after the company refused to let Claude near fully autonomous weapons or mass surveillance of Americans. Trump ordered every federal agency off the products; defence contractors had to certify non-use. Anthropic's own CFO told a court the actions could cost multiple billions of 2026 revenue. It took until the end of August for a federal judge to void the designation as unconstitutional retaliation and order the directives rescinded — and the parallel case in D.C. has since come back the other way: on the twenty-fifth of September a D.C. Circuit panel upheld the Pentagon's designation under a second statute, 2–1, Judge Henderson dissenting that the law targets foreign sabotage, not an American company's safety guardrails. The courts are now split: the government-wide ban stays blocked, the Pentagon may keep stripping Claude from its systems and barring it from defence work, and Anthropic says it is "considering all options, including further review" — an en banc rehearing or the Supreme Court, neither guaranteed. The executive branch swung, connected, and was eventually made to put down one weapon; the appeals court has left the other in its hand. What it demonstrated in the meantime is the fragility no risk factor quite states: a revenue line that can shed billions on one designation letter is being asked to carry forty times itself.
The other wing swung within a week of the ruling. Sanders and Casar announced the Ban Artificial Superintelligence Act: a permanent ban on superintelligent systems, a pause on all advanced AI development until a new federal regulator writes the rules, a cabinet-level agency to enforce it, and penalties imported from nuclear-weapons law — twenty years for persons, a corporate death penalty for entities. It will not pass this Congress, and that is not what it's for. It's a sample, and not a lone one — a bipartisan House pair filed their own Stop Rogue AI Act the same day, and the same Senate office already has a data-centre moratorium and an AI windfall tax in the drawer. And the bill's own method is the tell: Sanders built the justification out of the industry's public admissions that the technology is escaping its control. The labs spent two years narrating their power to investors, and the narration is now arriving back as legislative findings.
Neither wing can block the offering. The SEC runs a disclosure regime: it checks whether you told the truth, not whether the price is sane; the procurement weapon has been ruled unconstitutional by one court and lawful by another; a bill needs a Congress. If it lists, Anthropic will go to market as a supplier the US government's own defence department has blacklisted. But stopping the IPO and repricing the asset are different operations, and the second needs no statute. Forty times annualised revenue on 800% growth is a bet that the next several years are politically frictionless — purchased in the same eighteen months in which a nationalist Pentagon and a socialist Senate office both moved against the company, from opposite motives, aimed at the same layer. Look where every constraint points: not at tokens, at the frontier. A superintelligence ban is a ceiling on precisely the capability sliver above the commoditisation line where all the pricing power lives. Open source eats the layer below; Washington is drawing a line above; the space between is the asset. Being the best is the whole moat, and being the best is the specific thing both wings of American politics have now, within one year, tried to punish or prohibit.
The company itself calls the filing an option, dependent on market conditions.
Which leaves market conditions. The only force that has ever stopped an IPO priced on belief is belief collapsing during price discovery — WeWork met its investors, the investors met the financials, the offering died in the room. SpaceX's debut has already wobbled hard enough to give OpenAI second thoughts. So the one remaining regulator of the largest IPO in history is the correction everyone is dreading, arriving slightly early. There is no adult left in the room. There is only the room, repricing.
We built a probabilistic system, spent five years making it deterministic, and are now selling shares in the certainty. The prospectus lists your objection as a risk factor. The router lists it as a weight.