Anthropic's investors are floating a number that would have sounded absurd even in the 2021 software market. According to Financial Times reporting picked up by Quartz and others, some backers expect an October IPO at a valuation of $2 trillion or more. The claim is not that Anthropic has set a target. The claim is that investors can get there from the company's revenue trajectory, with annualized revenue reportedly expected to reach $100 billion to $120 billion by year-end after a $47 billion run rate in May.
That is the clean bull case. If revenue is compounding that quickly, a huge sales multiple looks less deranged on a spreadsheet than it does in a headline. At $2 trillion against $100 billion to $120 billion of annualized revenue, the market would be paying roughly 17 to 20 times forward run-rate sales. Against the May figure, the number is closer to 43 times. Neither is small. But public AI beneficiaries have already taught investors to tolerate strange-looking multiples when the growth story is attached to frontier models, inference demand, and enterprise adoption.
The better question is not whether the multiple can be defended for one roadshow. A multiple is a price on a distribution. The question is what has to be true in the rest of the capital stack for that distribution to pay.
Anthropic is not selling software in the old sense. It is selling access to a production function that consumes chips, data-center capacity, power, networking, engineering labor, and partner balance sheets. Every extra dollar of model revenue arrives with a claim somewhere else: on Amazon or Google capacity, on Broadcom chips, on cloud contracts, on energy infrastructure, or on customers who must decide whether the model bill replaces labor or simply joins the SaaS pile.
That makes the IPO math inseparable from the financing math.
Apollo's latest credit work puts a number on the other side of the trade. Its midyear outlook estimates that the AI ecosystem could support more than $2 trillion of additional investment-grade debt, especially when chipmakers and highly rated counterparties provide guarantees or contractual backstops. Apollo also argues that public investment-grade markets may absorb less than $1 trillion of that through 2030 because of concentration and rating constraints. Forbes summarized the implied gap at more than $1 trillion, with private credit, asset-backed structures, project finance, and infrastructure debt filling part of it.
That is a useful frame because it moves the AI-boom debate away from vibes. The constraint may not be whether Microsoft, Amazon, Alphabet, or Meta can borrow in the ordinary corporate sense. The hyperscalers entered the cycle with excellent balance sheets. The constraint is capacity, concentration, and correlation. Bond investors can own only so much exposure to the same handful of buyers, vendors, data-center projects, and power assumptions before diversification becomes mostly a word in a slide deck.
Equity investors can ignore that for longer. Credit investors cannot. Debt asks a less romantic question. If the asset fails to earn the expected cash flow, what is the recovery path?
For AI infrastructure, that question is still awkward. A data center has hard assets, but the economics depend on location, power contracts, utilization, tenant quality, and how quickly the useful life of the installed compute decays. GPUs can be pledged, financed, or leased. They are not risk-free collateral. A chip bought for one generation of frontier inference may not have the same value after the next architecture shift, export-control change, model-efficiency jump, or price cut.
This is where the pricing war matters. The Financial Times reported that leading U.S. labs' token prices have fallen by almost a quarter since mid-July, using Silicon Data's token price index, as open Chinese models pressure the closed labs. That is good for customers. It is less comfortable for a valuation story built on the idea that frontier access keeps premium margins while usage explodes.
There is a bullish version of falling token prices. Cheaper inference expands the market. More tasks become economical. Developers and enterprises route more work to models. Elasticity saves the day. If demand grows faster than unit prices fall, revenue keeps compounding and the infrastructure gets used.
I buy part of that argument. The model bill is not a fixed market. Lower prices can create workflows that did not make sense at higher prices. Coding assistants, customer support, document review, research, tutoring, and internal operations all have demand curves. The mistake is to treat every price cut as margin destruction without asking how much latent demand appears below the new price.
But the same argument has a denominator. Revenue may scale with tokens while profit depends on the spread between what customers pay and what it costs to serve them. If competition pushes token prices down faster than model efficiency, hardware utilization, and power procurement improve, the revenue multiple is capitalizing a thinner claim than the headline suggests. A company can have extraordinary revenue growth and still pass much of the surplus to customers, chip suppliers, cloud partners, and financiers.
That is why the Anthropic number and the Apollo number belong in the same essay. One is the equity market saying the frontier model layer may deserve a sovereign-scale valuation. The other is the credit market being asked to fund the physical substrate that makes that valuation plausible. They are not identical bets, but they rhyme.
The bargaining map is roughly this. The labs own the customer relationship and the model roadmap. The hyperscalers own distribution, data-center operations, and much of the balance sheet. The chipmakers own the scarce input when demand outruns supply. Power providers and landowners own the local bottlenecks. Private lenders may own the structured claim when public debt markets hit concentration limits. Customers own the option to route across providers, delay adoption, or use cheaper open models when quality is good enough.
The lab's dream is to sit at the choke point. The risk is becoming an expensive tenant in everyone else's choke point.
A $2 trillion IPO would need investors to believe several things at once. Revenue has to keep scaling after the early adopter wave. Gross margins have to settle at software-like or at least very attractive infrastructure-software levels. Capex obligations and cloud commitments must not turn the model layer into a pass-through for compute vendors. Open models must pressure prices without destroying differentiation. Enterprise customers must find enough productivity to keep paying after experimentation budgets become operating budgets. Public and private credit markets must keep financing the buildout without demanding terms that move too much of the upside away from equity.
None of those assumptions is impossible. None is free.
The hard part for outside investors is that the most important variables are not visible in a clean way. Annualized revenue can be disclosed. Token prices can be indexed. Data-center debt can be counted. But the distribution of inference margin by product, customer cohort, model family, and capacity contract is harder to see. So is the true duration of demand. A corporate customer using Claude heavily in August may be a sticky enterprise account, a temporary migration, or a price-sensitive workload waiting for a cheaper model.
That uncertainty does not make the IPO irrational. It makes the underwriting unusually path-dependent. If frontier models remain meaningfully differentiated and inference demand is highly elastic, the labs can grow into valuations that look silly on trailing numbers. If model quality converges and buyers treat intelligence as a routing problem, the pricing power moves away from the lab and toward whoever controls distribution, procurement, and compute cost.
The credit market will probably reveal the answer earlier than the equity market. Watch the terms on data-center debt, GPU-backed facilities, cloud-capacity commitments, and private-credit deals tied to AI projects. Watch whether guarantees get stronger, maturities shorten, spreads widen, or structures become more asset-specific. Equity narratives can run on TAM. Credit documents have to say who gets paid if the TAM arrives late.
My prior is that the AI buildout is real and still easier to finance in PowerPoint than in maturity schedules. Anthropic may be a remarkable business. It may even deserve a valuation that makes conventional software comps look timid. The $2 trillion figure prices more than Claude. It prices cash flows from intelligence arriving fast enough to pay the chip suppliers, landlords, utilities, clouds, lenders, and public shareholders without leaving the model lab as the squeezed middle.
That is the distribution I would want to see. Not the TAM slide. The cash waterfall.
Sources: Financial Times reporting on Anthropic IPO expectations, as summarized by Quartz and Fortune; Apollo 2026 Midyear Credit Outlook; Forbes coverage of Apollo's AI financing-gap estimate; Financial Times reporting on AI model price cuts and Chinese open-model pressure.
Originally published at https://deanlee.info/essays/anthropic-ipo-credit-market/.
Top comments (0)