Anthropic Before the IPO: What Is Intelligent Capital Pricing In?
From AI to SI: Can the Returns Match the Ambition?
Cypher Capital | Kim Wong | November 2026
AI infrastructure is real. The return on the investment is a separate question. Conceptual illustration.
Anthropic’s reported 2025 net loss was $42 billion. More than $34 billion of that came from an accounting charge tied to financing instruments that may convert into shares. That charge is not cash burn, but it does not make the loss disappear. When a number comes with a footnote that large, read the footnote.
I do not think the useful question is whether AI is “a bubble.” The technology works, customers are using it and companies are spending heavily to build capacity. The harder question is whether that spending will earn a good return for the model companies, the data-centre operators, their lenders and their shareholders. Those are four different bets.
At an IPO roadshow, the revenue chart usually gets the attention. The power bill, the money still needed to build, and the shares that may be issued tend to come later in the presentation. They should be on the first slide too. Would you rather underwrite a growth chart or a power schedule? That choice matters when the bill arrives before the revenue.
Anthropic is growing fast. Is the price too high?
Anthropic told prospective investors that Q3 revenue was $16.8 billion, up 45% from $11.6 billion in Q2. Q2 revenue had more than doubled from Q1’s $4.7 billion. The company added $5.2 billion of revenue in one quarter. That is impressive. But growth slowed from about 147% to 45%. Investors will look at both the extra dollars and the slower rate. [1]
Reuters reported the figures from people who attended Anthropic’s investor meeting. The company also reportedly said it made an adjusted profit in Q3. Until the filing explains what “adjusted” means and shows audited results, treat that as encouraging-but not proof that the business is generating cash. [1]
At the meeting, the conversation is easy to imagine: management points to $5.2 billion of new revenue; an investor asks why the growth rate slowed. Both questions belong in the room. Both should be reflected in the price.
Anthropic’s 2025 prospectus reports $4.6 billion of revenue, $12.65 billion of operating expenses and an $8.06 billion operating loss. Compute and infrastructure accounted for $7.33 billion. It held $20.28 billion in cash and short-term investments; planned cloud, compute and infrastructure obligations total at least $518 billion over the next decade. [2]
That $34 billion charge affects the reported loss and may lead to future dilution; it is not cash the company spent. Even without it, Anthropic had an $8 billion operating loss. I would focus on what the company spends to secure compute, what that spend brings in, and how much more capital it will need.
Investors have lowered the number they discuss
Before the latest revenue disclosure, IPO talk was around $2 trillion. On 15 October, Reuters said some investors were discussing closer to $1.5 trillion. That is not an offer range. Reuters expects the listing after the U.S. midterms, but Anthropic has not announced a date. [1,3]
At $1.5 trillion, Anthropic would be worth about 326 times its 2025 revenue. At $2 trillion, about 435 times. If we multiply Q3 revenue by four, the valuation is about 22 and 30 times that simple annualised figure. Reuters also cited analysts who expect a $110 billion revenue run rate by year-end. That is an estimate, not company guidance or reported revenue. On that estimate, the multiples fall to about 14 and 18 times. The lower multiple looks better, but it does not tell us the margin, dilution or cash required to get there. [1]
I would not call $1.5 trillion cheap. It would still be an extraordinary valuation for a company founded in 2021. But if the price comes down from $2 trillion, that does not mean AI has failed. It may mean investors are finally charging for the cost of keeping the business ahead.
An IPO will not make capital cheap
An Anthropic listing would give investors a public price for one of the leading AI companies. That is useful. It would also give early investors and employees a chance to sell shares, while putting the company under quarterly scrutiny as it signs long-term compute commitments. Public investors can be patient, but only if they understand what they are funding.
The same applies to neoclouds. Shares can be a better source of money than debt for a data centre that will take years to build and fill. But an IPO does not remove the risk of late power, construction overruns or customers that delay. It transfers some of that risk to public shareholders. If the deal mostly lets existing investors sell, the company gets less benefit than the headline deal size suggests.
I expect investors to be more selective from here. Look at how much new capital the company receives, how much it still needs to spend, and whether its customers can pay for the capacity they have booked. A group of listed companies will give us more information about the AI economy. It can also spread a sell-off quickly from one balance sheet to another.
Pre-IPO perps show what traders think—not what Anthropic is worth
Kaiko’s 1 October snapshot showed Anthropic perpetual-futures open interest at about $57 million, up from roughly $11 million in June and July. A separate 18 October snapshot across Binance, Bitget, Gate, OKX and HTX showed about $62 million, with Binance at roughly 67%. On that snapshot Binance implied about $1.88 trillion, OKX about $1.92 trillion, and Hyperliquid’s Entropy io:ANTH market about $1.91 trillion. The OKX raw quote is about $192.37 because its contract is on a different scale; the valuation comparison is normalized. These marks sit roughly 25% to 28% above the $1.5 trillion valuation some investors discussed on 15 October. They are different instruments and dates. The gap is evidence of disagreement, not a trade with a guaranteed convergence point. Hyperliquid open interest of about $40 million is separate from the centralized-exchange figures and should not be added as a consolidated total. [4,12]
There is no listed Anthropic share price, so each exchange sets its own reference price and contract rules. Funding may be switched off or fixed. The share count used to turn the price into a company valuation is an estimate and can change at listing. Open interest tells us the value of positions still open. It does not show whether traders are bullish or bearish, who they are, or how much could be traded at the screen price. Binance holds most of the centralized-exchange interest, so its book has an outsized effect on the headline.
A few large orders can move these prices more than the company’s news would justify. That creates room for manipulation, but a sharp move alone is not proof that anyone manipulated the market. I would check the depth, liquidations, index and exchange notices before drawing a conclusion. A price shown to two decimal places is still only as good as the market behind it. When a perp price jumps, ask: did the company’s outlook change, or did a thin order book move?
These markets make one thing more visible: the price traders are quoting and the size of positions left open. They do not tell us who holds those positions or whether they are hedges. That is useful information, but it is not the same as an IPO book, where the offer, allocation and cash shares are visible.
The exchange prices still differ after adjusting for contract scale. There is no share price to bring them together. The screen can look precise even when there is not much money behind the quote.
Can a perp hedge an IPO allocation?
If you have a confirmed IPO allocation, you can short perps against it to reduce the risk of a weak opening. The hedge is more reliable after the contract converts and the stock has a live price. Before that, you are relying on the exchange’s share count, conversion and settlement rules. You can reduce one risk and add funding, margin and price-gap risk.
SpaceX shows how this could work. The reported offer was $135. At the cash-market open, Binance perps were quoted at $154.52. The shares closed the day at $160.95. If an investor received 10,000 shares at the offer, sold at the close and shorted the same exposure at the Binance quote, the gross gain would have been about $195,200 before fees, funding, slippage and tax. The shares gained $259,500; the perp short lost about $64,300. [5]
That would have been a useful first-day hedge if the investor had the allocation and could execute both trades at those prices. It does not prove you can safely short before knowing whether you will receive shares. If the IPO is delayed, the short can lose money while you have no shares to sell. If the perp rises before the allocation is released, you may have to post more cash even though the shares would later offset the loss. And if the share count changes at conversion, the hedge may no longer match. Would you still call it a hedge if it forces you to post cash before your IPO shares are released?
Shorting before you have an IPO allocation is like selling a house before your bid has been accepted. If you do not get the shares, you are left with the short position—and may have to add cash before the IPO is settled.
Other IPOs show how different the outcomes can be
SpaceX open interest rose from $24 million on 21 May to $271 million on 11 June—about eleven times. More than half the increase came in the final week, after reports gave an offer price and date. Binance held about 82% of open interest on the eve of listing. Trading surged after the contracts converted and the stock began trading. That shows how quickly interest can build when the date is real. It does not mean Anthropic will follow the same path. [5]
Cerebras’ final-hour perp price was reported at $354.54; the IPO offer was $185 and the stock opened at $350. Unitree’s perp implied about $38 billion, against an IPO value near $9 billion; the cash-market opening was reported around $66 billion. In other words, one perp price came close to the opening trade, while the other sat well below it. Neither makes a reliable rule for the next IPO. [6,7]
Compare prices at the same point in time. The offer is for IPO investors; the perp quote is from one exchange; the opening trade comes from the stock-market auction; and the close reflects a full day of trading. Choose the comparison after the event and you can make almost any price look right. But a perp price only tells us what traders may pay for expected growth. It does not tell us what it costs to deliver that growth. If AI demand is real, who funds the buildings, chips and power before customers start paying?
The other side of the AI IPO: who funds the capacity?
Anthropic’s valuation debate is about what investors will pay for AI growth. CoreWeave shows the other side: what it costs to build that growth. Its June bonds were issued at a 9.675% yield. In the week of 12 October, Reuters said they were trading around a 15% yield. The company’s debt was reported at roughly $50 billion, and its shares fell 15.1% from Monday’s open to Thursday’s close. That is a sharp rise in borrowing costs—not proof that demand for compute has vanished. But a company that needs to refinance will feel the difference. [8]
Nscale shows the same issue in its contracts. Its SEC filing reported $103.4 billion of active and contracted total contract value at 31 August. Only $2.6 billion was active—about 2.5% of the total. In the first half of 2026, revenue was $140.6 million and the net loss was $1.02 billion. Its largest customer accounted for 52% of revenue. The important question is how much of that future contract value can be powered, delivered and billed on time. [9]
Think of TCV as a booking, not a sale. The customer may have reserved capacity, but the site still needs power and the customer still has to start paying. Take-or-pay terms help, but they do not switch on the site.
On 18 October, Reuters reported that Microsoft had ended its agreement for Nscale’s delayed Loughton site in Essex. Reuters did not give a reason; another report said grid power might not arrive until the 2030s. Nscale said its wider relationship with Microsoft was strong, other contracts were unaffected and it was looking at alternative power. I see this as a serious warning about one project, not evidence that the whole company has failed. [10]
Follow the money from model to data centre
A customer pays for a model or an application. The model company pays for computing. The cloud or neocloud pays for servers, buildings and power. Equipment suppliers are paid while the data centre is still being built. The same dollar can show up as revenue at several points. That is why headline revenue does not tell us who is actually making money.
The question I care about is who gets paid—and who is left holding the asset if demand arrives late. A long contract helps only if it can be enforced, the customer can pay and the site is ready. A hotel reservation tells you someone plans to stay. It does not prove they have checked in or paid the bill.
These businesses run on different clocks. A software feature can launch in weeks. A data centre can take years to get power and finish construction. If demand arrives first, capacity is valuable. If the buildings and equipment arrive before customers are ready to pay, fixed costs become a problem. The best order book in the world does not earn a return until customers use it and pay.
Three ways this could play out
I do not think we need to choose between “AI works” and “AI is a bubble.” The real question is whether customers will pay enough to cover the cost of the infrastructure being built. These are three outcomes I would watch—not predictions.
I would become more positive if customers renewed, usage spread beyond a few big names and cash came in faster. I would become more cautious if the gap between contracted capacity and powered, billable capacity kept growing. A contract is one thing. A site sending invoices is another.
Public listings can help because companies have to show their debts, spending and customer concentration. But the market will judge them every quarter while the data centres take years to build. More disclosure is useful. It does not make a construction schedule shorter.
Questions I would ask before buying a neocloud IPO
I would want the prospectus to connect customer commitments to revenue, cash and the cost of building the next site. A large contract figure is not enough on its own.
Definitions matter. Backlog, reserved capacity, active capacity and revenue are different. Follow each contract through equipment delivery, power-on, customer use and cash collection.
For Anthropic, prove customers keep paying for models. For neoclouds, prove the capacity gets built, powered and paid for.
My November rule: follow the cash to the balance sheet that must repay the debt. Who carries the loss when capacity arrives late—the customer, operator, lender or shareholder?
Five questions I am asking in November
I believe AI usage will keep growing. I am less certain that every company building for it will earn a good return. Cheaper compute could bring in more users, while also pushing prices down. A bigger AI market does not guarantee that every model company or neocloud gets its share of the profit.
In November, I will watch what is already being used and paid for, how much capacity has power, what cash is coming in, and whether margins cover depreciation and interest. I will also watch what Nscale says about Loughton and whether its other sites can be powered and activated. If those signs improve, the market may be too pessimistic. If they do not, the AI revolution can be real while investors still lose money. That is not a contradiction; it is investing. The question for investors is simple: if AI demand keeps growing, which companies will actually collect the cash?
Sources and notes
[1] Reuters, Anthropic Q3 investor disclosure and expected post-midterm IPO timing, 14 October 2026.
[2] Reuters, Anthropic IPO prospectus and costs, 28 September 2026.
[3] Reuters, investors temper Anthropic valuation expectations, 15 October 2026.
[4] Kaiko, Anthropic pre-IPO perpetual futures snapshot, 7 October 2026.
[5] Reuters, SpaceX listing prices and perps; Binance market review, June 2026.
[6] Coin Metrics / Talos, Cerebras pre-IPO perp price discovery, 9 June 2026.
[7] CoinDesk, Unitree IPO and pre-IPO perp comparison, 19 August 2026.
[8] Reuters, CoreWeave bond repricing, 16 October 2026.
[9] Nscale plc, Form F-1 registration statement, 18 September 2026.
[10] Reuters, Nscale Loughton project and company response, 18 October 2026.
[11] Reuters, OpenAI will not go public in 2026, 12 September 2026.
[12] DeFiLlama, Anthropic pre-IPO perpetual market snapshot, 18 October 2026, 11:56 UTC; OAK Research, Hyperliquid Entropy io:ANTH market snapshot, accessed 18 October 2026.
Company forecasts, private-market valuations and investor discussions are attributed to the reporting source; they are not confirmed offer terms. Historical market prices are dated. Illustrations are not trade recommendations
Disclaimer: This report is published by Cypher Capital (BVI) Limited, a Business Company incorporated in the British Virgin Islands. Cypher Capital (BVI) Limited is not licensed or regulated by the Central Bank of the UAE, the Securities and Commodities Authority of the UAE, or the Virtual Assets Regulatory Authority of Dubai.
This report is provided for informational and educational purposes only and does not constitute investment advice, a recommendation to buy or sell any asset, or an offer or solicitation to invest in any fund, product, or strategy.
This report contains forward-looking statements and third-party price forecasts subject to significant uncertainty. Third-party forecasts cited reflect the views of those institutions, not of Cypher Capital. Cypher Capital, its affiliates, and employees may hold positions in the assets discussed herein.
No representation or warranty is made as to the accuracy or completeness of the information contained herein. Recipients should conduct their own independent analysis and consult qualified advisors before making investment decisions.
Copyright 2026 Cypher Capital (BVI) Limited. All rights reserved.*