OpenAI’s run rate tops $40 billion. Anthropic’s IPO valuation hinges on a 2028 revenue forecast of $190–200 billion. Nvidia is guaranteeing OpenAI’s data centre debt while mobilising $500 billion in outside capital for its own customers. The AI buildout is now financing itself — and the public markets are about to price it.
In six trading days, every layer of the AI stack moved toward the public capital markets at once. The events are separate in form — an IPO filing here, a bond offering there, a guarantee restructured on a Friday — but they describe one structure: a buildout whose funding loop closes back onto itself.
| Date | Event | Amount |
|---|---|---|
| Aug 10 | Nvidia signs MoUs with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to establish “compute financing platforms” for its customers[1][2] | $500B+ third-party capital target |
| Aug 12 | CoreWeave Q2: revenue $2.575B (+112% YoY), backlog $104B; raises FY capex guidance to $35–39B after raising ~$18B in the quarter across debt, convertibles and equity[3][4] | $18B raised in Q2 |
| Aug 13 | AMD launches four-tranche senior unsecured bond offering (maturities 2029–2036), settling Aug 17[6][7] | $4–5B |
| Aug 14 | Bloomberg: OpenAI revenue run rate tops $40B, up from ~$20B in April; Anthropic Q2 preliminary revenue above $11.5B[8][9] | $40B+ run rate |
| Aug 14 | Nvidia scales back its guarantee for OpenAI’s proposed Ohio data centre to less than $120B, down from the $250B previously discussed (WSJ)[10] | $250B → <$120B |
| Aug 15 | Reuters: Anthropic’s IPO valuation hinges on a $190–200B 2028 revenue forecast; DeepSeek quadruples pricing of its flagship models ahead of its own listing (Bloomberg)[11][12] | $190–200B by 2028 |
A July memo on the capex reckoning covered the debt side of this buildout — who is borrowing and what happens when credit markets turn. This one covers what happens next: the equity market starts pricing it.
Three of the largest private AI companies are moving toward public markets simultaneously, and each presents a different pricing problem to investors.
OpenAI is priced on run rate. At over $40 billion annualised — up from roughly $20 billion in April — its IPO will be anchored to revenue that already exists.[8] The question for public investors is not whether the number is real but what it costs: OpenAI’s compute commitments are among the largest single obligations in corporate history. A meaningful share of them is being financed by the very suppliers who sell it the hardware.
Anthropic is priced on a forecast. Per Reuters, its IPO valuation hinges on reaching $190–200 billion in revenue by 2028.[11] The company’s run rate was about $9 billion at the end of 2025 and had risen to more than $47 billion by May; Q2 preliminary revenue came in above $11.5 billion.[9][13] Reaching the low end of that forecast requires roughly a four-fold increase in about thirty months, sustained — not a one-time jump. The IPO will be the first major listing where an audited two-year-out revenue projection is effectively the anchor for valuation.
DeepSeek is priced on monetisation discipline. Quadrupling the price of its flagship models ahead of its own listing (Bloomberg)[12] is a different playbook: raise prices before scale, not after. It signals that even in China’s open-weight segment, the era of subsidising inference to win users is ending — which matters for every company whose cost assumptions were built on cheap frontier compute.
Look at the Aug 10 and Aug 14 events together, because they are two halves of one structure.
Nvidia signed memorandums with six of the largest asset managers on Wall Street — Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR — to mobilise over $500 billion in third-party capital for its customers. Those customers are hyperscalers, frontier labs and enterprises building data centres and buying Nvidia hardware.[1][2] Jensen Huang’s framing was explicit: the chips are an “investable asset,” financed like commercial real estate or toll roads.[2]
Four days later, Nvidia scaled back its guarantee of OpenAI’s proposed 10-gigawatt Ohio campus from $250 billion to less than $120 billion — initially covering only half the project — after investors pressed on the chipmaker’s risk exposure (WSJ).[10] The deal is reportedly close to signing.
Source: CNA — “Nvidia scales back $250 billion OpenAI data center guarantee, WSJ reports” (15 Aug 2026). channelnewsasia.com
The loop: Nvidia’s revenue depends on AI labs’ capex. The labs’ valuations depend on compute availability, which depends on Nvidia. And now the capital markets finance both legs — banks and asset managers lend against the data centres that buy Nvidia chips, while Nvidia itself backstops part of the debt for its largest customer. AMD is doing the same at smaller scale with a $4–5 billion bond offering to fund AI-related investment.[6][7] CoreWeave raised roughly $18 billion in a single quarter — debt, convertibles and equity combined — and guided capex up to $35–39 billion.[3][4]
Vendor financing is not new; it has run through semiconductors and telecom for decades. What is new is the scale at which the same dollars are counted on both sides of the loop. In a normal vendor-financing cycle, a demand miss hits participants sequentially — first the buyer, then the supplier. In this structure, a slowdown in AI revenue flexes the chipmaker’s sales, the neocloud’s backlog and the lab’s valuation at the same time.
Anthropic’s IPO is the cleanest test case, because its valuation explicitly depends on a number that does not exist yet.
From a run rate above $47 billion in May to $190–200 billion by end-2028 implies sustained growth of roughly 35–40% per quarter for the next two and a half years. That is not impossible — Anthropic went from ~$9 billion at end-2025 to over $47 billion in about six months, and CoreWeave’s backlog grew 246% year-over-year to $104 billion.[5][13] But note what that evidence actually is: contracted demand concentrated among a handful of frontier labs whose own funding depends on this same cycle.
The neocloud numbers are the strongest demand signal in the stack — CoreWeave’s Q2 revenue grew 112% year-over-year, adjusted operating income reached $128 million, and management described near-term capacity as effectively sold out.[4][5] But a backlog is only as good as the credit of the companies writing it. If one major lab slows its buildout or reprices itself in an IPO, the neocloud’s contracted revenue and the chipmaker’s forward orders flex together.
This is why the Anthropic filing matters beyond one company: how auditors and underwriters treat a 2028 revenue forecast as an IPO anchor will set the standard for every AI lab listing that follows. If it passes scrutiny at $190–200 billion, the next two filings will be priced against it.
Singapore’s answer to the listing wave is connectivity rather than competition. The Equities Market Review Group’s final report, published 19 November 2025, set out a dual listing bridge between SGX and Nasdaq aimed at Asian companies with market capitalisation of S$2 billion and above. The new board envisaged to go live around mid-2026[14]. It came alongside a S$30 million Value Unlock package and a second batch of asset managers that took Equity Market Development Programme allocations to S$2.85 billion[14].
The improvement is real but sits below the AI mega-caps. Small- and mid-cap turnover grew 88% quarter-on-quarter, with net institutional inflows for nine consecutive months after the first measures in February 2025[14]. A Singapore investor’s exposure to the AI listing cycle therefore continues to run through US venues; the domestic change is the plumbing that lets an Asian company reach both markets at once.
The bear case rests on circularity. The bull case has four answers, and they are not weak.
1. The revenue is real and growing fast. CoreWeave grew 112% year-over-year with expanding operating margins. Anthropic roughly quintupled its run rate in six months. OpenAI doubled it in four. These are contracted, invoiced numbers — not projections. A bubble requires the underlying demand to be fictitious; so far the evidence points the other way.
2. The debt markets are pricing this at investment-grade spreads. AMD’s notes priced around 70–115 basis points over Treasuries across maturities to 2036.[6] That is a statement about credit quality made by institutional buyers with their own capital, not by the issuers.
3. The structure is familiar; only the scale is new. Vendor financing, take-or-pay contracts and supplier backstops have run through every major infrastructure buildout in history — telecom fiber in the 1990s, data centres in the 2000s. Markets have digested these structures before at smaller sizes.
4. The guarantee cut shows risk management working. Nvidia reducing its OpenAI backstop from $250 billion to under $120 billion — before signing, in response to investor concern — is the system correcting itself rather than discovering the problem after the fact. That is a healthier outcome than the alternative.
Four near-term events will tell you how the market prices this structure:
1. The Nvidia–OpenAI deal terms. Reportedly close to signing, with an initial guarantee under $120 billion covering half of the Ohio project.[10] Watch whether the remaining half is financed through the new asset-manager platforms — if so, the circularity becomes explicit in the capital structure rather than implicit in the supply chain.
2. OpenAI’s IPO terms and timing. Whether it prices on run rate or forecast, and what share of its compute commitments is vendor-financed versus equity-funded, will set the template for the entire wave.
3. AMD’s bond settlement on Aug 17. Order-book depth on a $4–5 billion offering tells you where institutional appetite for AI-related debt actually sits[6] — and whether spreads hold as more issuers enter the market.
4. Anthropic’s S-1. The first public document to anchor an IPO valuation on a 2028 revenue forecast. How auditors treat it, and what covenants or disclosures attach to it, will define the standard for every AI lab listing that follows.
The buildout is real, the demand is contracted, and the financing is circular. The question the public markets are about to answer is not whether AI infrastructure gets built — it is who holds the bag if growth slows. How much of the risk has already been transferred from balance sheets into bond tranches.
This analysis is based on publicly available data as of 2026-08-16. For related coverage, see The AI Safety Crisis and Algorithmic Harm as Legal Liability.