Reading the OpenAI and Anthropic IPOs Through an Investor Lens — Risks to Watch Before the Hype

The IPO speculation around OpenAI and Anthropic is not just about two large companies going public. This is the moment when the two companies at the center of today’s AI shift from the private market — where “AI growth” was enough — to the public market, where “revenue quality,” “capital efficiency,” and “questions about circular transactions” face real scrutiny.

The Information reported that Anthropic is considering an IPO as early as Q4 2026, and OpenAI is also said to be pursuing a listing as soon as possible. Anthropic’s annualized revenue has reportedly reached $19 billion and OpenAI’s $25 billion, but IPO advisors see differences in how cloud resale revenue is recognized as a key issue in the review process.

My blog usually covers technical changes in AI and infrastructure. This time, though, I want to look at this IPO from the investment and financial side rather than going deep on the technology itself.

What I am most interested in is not which company goes public first. It is how much of each company’s revenue is backed by actual user demand, outside of circular transactions with their cloud partners.

AI companies can no longer be evaluated by revenue alone. OpenAI in particular has restructured Stargate from “building its own data centers” to “leasing cloud and facilities from others to secure compute as fast as possible,” while cutting peripheral businesses like Sora and payments. This can be read positively as focus and prioritization, but it can also be read as cleaning up unflattering capital allocation ahead of an IPO. The plan to spend $665 billion on cloud servers by 2030, and the management focus on capital raising, supply chains, and data center buildout, should be seen in the same context. (Reuters, The Information)

The Positive Side

The positives are clear. First, both companies are already businesses generating massive revenue.

Anthropic reached an annualized revenue of $19 billion as of early 2026, and reportedly ramped up significantly from prior-year levels within the first two months. Internally, the company has indicated a path to cash generation by 2028. Among bankers, there is a view that Anthropic, with its stronger enterprise and developer focus, may be more attractive to public market investors than the more consumer-heavy OpenAI. (The Information)

OpenAI still leads in scale. According to Reuters, its annualized revenue exceeded $25 billion as of February 2026. Its U.S. advertising pilot for ChatGPT reached an annualized revenue run rate of over $100 million in just six weeks. However, the path to profitability is longer than Anthropic’s — OpenAI targets free cash flow breakeven by 2030, while reportedly expecting to burn over $200 billion until then.

The revenue scale is already massive, and growth is extremely fast. Moreover, the market is starting to pay attention to the timeline to profitability — 2028 for Anthropic, 2030 for OpenAI. Both companies are no longer seen as “companies that build AI” but as companies that have begun converting AI into real money.

Second, OpenAI is focusing and prioritizing.

Shutting down Sora, shifting to leased infrastructure, expanding advertising, and building enterprise channels — these look separate but point in the same direction. Enterprise recurring revenue and securing compute are being prioritized over flashy consumer-facing projects. For a pre-IPO company, this makes sense.

What Really Matters Is Not Revenue Volume but Revenue Quality

This is where the main argument begins. What matters in an IPO is not “is revenue big” but what that revenue is made of.

I think there are three issues to examine.

A. Is revenue reported on a gross or net basis?
B. Is revenue inflated by circular transactions?
C. Are barter deals and offsets making the true revenue picture harder to see?

These three.

A. Gross vs. Net Revenue Recognition

The most straightforward issue is whether revenue is shown as the full amount or just the company’s take.

Consider a simplified Anthropic example.

Anthropic Case 1: Direct Sales

Anthropic sells the Claude API to an enterprise customer for $100 directly. Revenue is straightforwardly $100. If the underlying AWS cost is $30:

Revenue: $100
Cost: $30
Gross profit: $70

This is relatively clean. Anthropic holds the customer, the contract, and the billing.

Anthropic Case 2: Via AWS Bedrock

The customer pays $100 to AWS. AWS handles sales, billing, the customer relationship, and possibly even pricing. Anthropic receives $70 as its share.

The question investors want answered is simple. Is Anthropic’s revenue $100 or $70?

If Anthropic reports $100, the apparent growth rate looks bigger. But if only $70 actually flows to Anthropic and AWS controls the customer relationship, investors will ask whether that is truly Anthropic’s own revenue.

The issue for Anthropic is not whether the product is selling. It is how much of that revenue counts as “their own business.” (The Information)

The structure is similar for OpenAI, but more complex.

Case 1: Via Azure OpenAI Service (Microsoft as the Channel)

An enterprise customer pays Microsoft $100. Microsoft is the sales channel. The contract, billing, and customer-facing layer all sit on the Microsoft side.

OpenAI receives $20 of that as a license-like payment, which is recorded as OpenAI’s revenue.

Customer pays: $100
Microsoft receives: $100
OpenAI’s share: $20
OpenAI revenue: $20

In this model, the service is delivered as Microsoft’s Azure OpenAI Service, so OpenAI does not directly bear the full infrastructure cost for that $100. OpenAI’s P&L looks roughly like:

$20 (revenue) minus R&D and other expenses = OpenAI’s profit

In other words, Azure-channel revenue is not a $100 business for OpenAI. It is a $20 business. (Reuters Breakingviews)

Case 2: OpenAI Direct Sales (ChatGPT / API)

Next, consider the case where a customer pays OpenAI directly $100. The $100 goes to OpenAI first.

But it does not end there. Even on direct sales, OpenAI pays Microsoft $20 as a partnership-like share.

Customer pays: $100
OpenAI receives: $100
Microsoft’s share: $20
OpenAI retains: $80

But since OpenAI runs its models on Azure, it must also pay massive server costs (inference costs) to Microsoft out of that $80. This is the key difference from Azure-channel sales.

If Azure usage costs $50 and other expenses are $10:

OpenAI retains: $80
Azure costs: $50
Other expenses: $10
Profit: $20

So direct sales look like $100 in revenue on the surface, but in practice:

Microsoft takes its share first,
Azure infrastructure costs come next,
Then OpenAI’s own expenses.

This means direct sales appear large but have many deductions, while Azure-channel sales appear smaller but OpenAI does not directly carry infrastructure costs. This difference significantly changes how revenue quality and margins look across channels.

The numbers shown for OpenAI and Anthropic above are not meant to reproduce actual contract terms. They are schematic illustrations to make the issues investors care about in an IPO visible. What can be confirmed as fact is that OpenAI has a deep revenue-sharing, capital, and infrastructure relationship with Microsoft, and that for Anthropic, the accounting treatment of resale revenue through AWS and Google is flagged as an issue in IPO reviews.

B. Could Circular Transactions Make Revenue Look Inflated?

The next issue is circular transactions. Under SEC related-party transaction rules, transactions and relationships between a company and closely connected parties are subject to strict disclosure requirements.

In an IPO, what matters is not just legality but whether transactions look like arm’s-length dealings with independent third parties.

A schematic example of what could be questioned as circular for Anthropic:

AWS or Google makes a large investment in Anthropic.
Anthropic uses that capital or credits to lease GPUs from AWS or Google.
At the same time, AWS or Google resells Anthropic’s models on their cloud platforms.
That resale revenue is counted in Anthropic’s annualized revenue.

On the surface, “investment,” “cloud costs,” and “resale revenue” appear to be separate items. But investors ask the real question: “Is this truly revenue earned from third parties, or is the investor’s money just coming back in a different form?” That is why circular transactions are viewed negatively.

The schematic for OpenAI:

Microsoft provides capital or credits to OpenAI.
OpenAI consumes Azure heavily.
Microsoft sells OpenAI’s models to customers as Azure OpenAI Service.
A portion becomes OpenAI’s revenue or future share.

If this loop grows too large, the “Microsoft → OpenAI → Microsoft” fund flow becomes so thick that from the outside it becomes unclear whether revenue is driven by external demand or just circulating within a large partner ecosystem.

Of course, there are real external customers and real demand. But what IPO investors dislike is not being able to see how much of that demand is truly independent.

If a significant portion of revenue comes from what is essentially internal circulation with a major partner, the apparent growth rate is less reassuring than it looks. Revenue can be recorded, but how much “independent third-party market” exists behind it is a separate question.

C. Barter Deals and Offsets Make Revenue Harder to Read

In startups and fast-growing companies, it is not unusual for revenue to be recognized before cash is collected, or for credits, prepayments, offsets, and non-cash consideration to be used.

The core problem is that when these practices accumulate, it becomes harder to tell what is genuine operating revenue, what is capital support, and what is accounting offset.

Under ASC 606 (Revenue from Contracts with Customers), non-cash consideration from customers can qualify as measurable consideration under certain conditions. In other words, equity, credits, and offsets can all be recognized as revenue under accounting rules. That is precisely why investors look beyond the revenue number to the economic substance behind it.

An OpenAI example:

Example: Azure-channel sales and offsets Microsoft sells an OpenAI-based service to an external customer for $100. OpenAI’s contractual share is $20. Normally, Microsoft would pay OpenAI $20.

But at the same time, OpenAI owes Azure, say, $10 million in inference costs this month. Instead of moving cash, they settle it as: “This month’s $20 share will be deducted from your Azure bill.” On OpenAI’s books, $20 in revenue is recorded, but no cash enters the bank account.

The problem is that it becomes hard to tell from outside whether this $20 is pure operating revenue or simply the result of an offset within a massive transactional relationship with Microsoft. The boundaries between revenue, infrastructure costs, capital support, and related-party transactions become blurred. This complexity is exactly what IPO investors are wary of.

Something similar can happen with Anthropic.

Example: AWS Bedrock channel and credit offsets A customer pays AWS $100. Anthropic’s share is $70. But in the same month, Anthropic owes $200 in AWS GPU usage fees. If AWS processes it as: “We will not pay the $70 in cash — it will be offset against your cloud bill.” Then Anthropic records $70 in revenue, but no cash comes in.

Again, the core issue is not about whether cash was received. What matters is that investors cannot easily tell whether that $70 is revenue backed by third-party demand, or just a book entry within the broader capital-credit-usage relationship with AWS.

That is why annualized revenue figures of $19 billion or $25 billion are not enough on their own. What needs to be examined is operating cash flow, accounts receivable, deferred revenue, related-party disclosures, credit offset practices, and the ratio of direct sales to resale revenue.

In short, what an IPO tests is whether a company can clearly explain the relationship between revenue, capital support, and offset arrangements to investors.

An Alternative View

From another angle, this can be seen as a reflection of the unique nature of the AI industry. Model companies cannot grow without cloud, power, and chips.

So they are securing physical resources using not just cash but equity, priority supply agreements, long-term contracts, credits, and revenue sharing.

OpenAI’s shift toward cloud leasing and its attempt to secure chips from AMD through equity are examples of this.

Both companies are not “ordinary software companies.” They are companies that use AI as collateral to procure infrastructure.

That is their strength, but it is also what makes them hardest to explain in an IPO.

My View

I think Anthropic will be easier to evaluate in the early IPO stage. Its enterprise focus means it does not carry the massive consumer-side maintenance costs that OpenAI does. I understand why bankers see Anthropic going first.

But I do not think that makes it safe. My concern with Anthropic is that it looks like it has fewer risks on the surface.

Once revenue quality, cloud resale structures, and offset practices are disclosed, there is a real possibility that the current perception of Anthropic as a “lean, efficient AI company” could shift.

My concern with OpenAI is that it is hard to see whether it can truly generate profit as an independent public company within a structure where Microsoft holds so much control.

OpenAI has overwhelming visibility and growth. But as an investor, what concerns me is not the size of revenue itself. What matters more is who ultimately captures the profit from that revenue.

Microsoft holds a very strong position over OpenAI — as a shareholder, as a sales channel, and as an infrastructure provider. If the structure is such that most of the profit flows to Microsoft regardless of how much OpenAI grows, then the return available to OpenAI’s public shareholders may be smaller than expected.

OpenAI’s risk is not just business complexity. It is that even as revenue grows, profit distribution may be skewed toward the partner rather than toward OpenAI’s own shareholders. To be truly valued as a public company, OpenAI must demonstrate not just that it is a high-growth company, but that it is a company that can generate profit independently.

Summary

What really matters in the OpenAI and Anthropic IPOs is simple.

A. Is revenue reported gross or net?
B. Is revenue truly from external customers, or does it have a strong circular transaction element?
C. Are offsets and credit arrangements making revenue and cash flow harder to read?

These three questions.

Because these two companies are at the center of today’s AI, their revenue numbers look large. But what an IPO tests is not growth rate itself — it is how transparent that revenue is and how much independent economic substance it has.

Anthropic’s issue is how much resale and related-party structure is embedded within revenue that looks clean on the surface, given its enterprise focus. OpenAI’s issue is whether, within a capital, contract, and infrastructure structure heavily controlled by Microsoft, enough profit remains for OpenAI’s own shareholders even as the company grows.

I do not see this IPO as simply “a chance to invest in companies at the center of AI.” I see it as the first time that revenue quality, profit attribution, and accounting transparency of AI companies will be seriously tested by the market.

If the numbers hold up under scrutiny, both companies are the real thing. If, compared to the surface appearance of revenue, the reality turns out to depend heavily on resale, offsets, and related-party transactions, then valuations will shift significantly — even in the middle of AI enthusiasm.

What investors should really watch is not the listing itself. It is what the prospectus footnotes say about who generates the revenue, who captures the profit, and who holds control.