Anthropic OpenAI Revenue Looks Comparable Until Cloud Accounting Changes the Score

Anthropic OpenAI revenue comparisons just acquired a major caveat despite the enormous annualized figures presented to investors. Anthropic reportedly counts the full value of certain cloud-partner sales. OpenAI records only the portion it retains from comparable partner transactions.

That difference can move billions of dollars onto or off the top line without changing what customers purchased. It also complicates the apparent contest between the two largest independent AI labs. A larger reported run rate does not necessarily mean one company captured more cash or built a healthier business.

The issue surfaced as investors tried to reconcile several striking numbers. Anthropic reportedly reached a $65 billion annualized pace in July. OpenAI approached $50 billion by late September, while forecasting at least $70 billion by year-end.

Those figures sound comparable because they share the same label. Their underlying calculations are not necessarily comparable.

The distinction matters beyond an accounting debate. Anthropic and OpenAI depend on Amazon, Google, and Microsoft for infrastructure, distribution, financing, or some combination of all three. How each company records those relationships can shape perceived growth, margins, market leadership, and eventual public valuations.

The Same Cloud Sale Can Produce Two Revenue Numbers

The central problem is simple: Anthropic and OpenAI reportedly count some partner-generated sales at different points in the transaction.

Anthropic makes Claude available through cloud marketplaces operated by Amazon Web Services and Google Cloud. Enterprise customers can purchase model access through those existing cloud relationships rather than contracting directly with Anthropic.

According to Bloomberg’s reporting, Anthropic records the entire customer payment from certain partner channels as revenue. It then recognizes the cloud provider’s share as an expense. This is commonly called gross presentation.

OpenAI reportedly takes another approach for sales involving Microsoft. It records only the portion of the customer payment that belongs to OpenAI after Microsoft receives its share. This is net presentation.

Consider a simplified transaction involving a customer payment of 100 units. Under gross presentation, the model developer might record all 100 as revenue and 20 as a partner expense. Under net presentation, it might record only the remaining 80 as revenue.

The developer can retain the same 80 units in both examples. Yet the gross method produces 25 percent more reported revenue than the net method.

This example does not describe the companies’ actual contractual percentages. Those terms can vary by platform, product, and customer agreement. It shows why the accounting choice can materially alter headline comparisons.

The distinction reflects the principal-versus-agent framework in revenue accounting. A principal controls the promised service before transferring it to the customer. An agent arranges for another party to provide that service.

Companies acting as principals usually record the transaction’s gross value. Agents generally record only their fee or retained share.

An SEC accounting example explains that the assessment turns on control, responsibility, pricing authority, and transaction structure. It is not a voluntary switch selected only to improve appearances.

Bloomberg’s account indicates that both Anthropic and OpenAI consider their methods compliant with generally accepted accounting principles. The mismatch can therefore exist without either company violating the standard.

The contracts can assign different obligations to each model developer and cloud provider. One partner might merely process a sale. Another might control the customer relationship, bundle services, or assume additional performance responsibilities.

That contractual detail determines the accounting result. It also means investors cannot safely normalize the companies by subtracting one universal partner percentage.

The problem grows when media reports use annualized revenue as a scoreboard. Annualized revenue projects a short period of sales across a full year. It is useful for illustrating current momentum, but it is not the same as completed annual revenue.

A run rate already amplifies short-term conditions. Combining it with inconsistent gross and net presentation adds a second layer of uncertainty.

This does not make the figures meaningless. It means they require definitions, channel disclosures, and reconciliation before they can support a credible ranking.

Anthropic OpenAI Revenue Comparisons Need a Common Baseline

A direct ranking becomes unreliable when one company includes cloud partners’ shares and the other excludes them.

Anthropic’s annualized revenue reportedly exceeded $65 billion at the end of July. That represented more than seven times its pace at the end of 2025, according to annualized revenue reporting.

OpenAI’s annualized revenue was roughly $50 billion at the end of September. People familiar with the figures said the company expected to reach or exceed $70 billion by year-end, driven partly by enterprise growth.

The year-end projection implies another significant increase during the final quarter. However, the forecast and the September run rate answer different questions.

The $50 billion figure reportedly reflects OpenAI’s normal net accounting. A previously circulated figure near $70 billion was designed to approximate OpenAI’s revenue under Anthropic’s gross methodology.

That adjustment explains why apparently conflicting OpenAI figures appeared within a short period. One represented the company’s reported accounting basis. The other tried to create an investor comparison.

Axios subsequently reported that OpenAI’s run rate was near $50 billion, roughly $20 billion below the earlier grossed-up estimate. Its revenue reconciliation said investors had requested an apples-to-apples view.

That episode shows how quickly a normalized estimate can escape its original context. Once repeated without a definition, it begins to look like ordinary reported revenue.

The gap does not automatically establish that Anthropic overstated its business. It also does not prove that OpenAI adopted a more conservative company-wide philosophy.

The companies have different contracts, partner networks, customer relationships, and delivery obligations. Those differences can lead to different accounting conclusions under the same standard.

Still, the result is unsatisfactory for investors. A top-line comparison should distinguish at least three measures.

The first is recognized revenue under each company’s normal accounting. The second is gross customer spending across direct and partner channels. The third is the amount retained after partner distributions.

Recognized revenue answers the formal accounting question. Gross customer spending shows the economic activity flowing through the company’s products. Retained revenue comes closer to the funds available before compute and other operating costs.

None of these measures alone describes business quality. Each reveals a different part of the commercial model.

Gross customer spending can show product reach, especially when a cloud marketplace handles procurement. However, it can exaggerate the economic value retained by the model developer.

Net revenue better reflects the developer’s direct share of a transaction. Yet it can understate customer demand when compared with a rival using gross presentation.

The choice also affects margin interpretation. A company recording gross revenue usually records the partner payment in cost of revenue or another expense category.

Its revenue appears larger, but its reported margin reflects that added partner cost. A company using net presentation starts with a smaller top line and does not repeat the same partner amount as an expense.

Investors therefore cannot compare revenue growth and gross margins independently. Both numbers must be adjusted together.

Changing only the top line would distort the analysis again. A valid normalization must rebuild revenue, partner costs, and margins on the same basis.

Cloud Distribution Is Now Part of the Competitive Contest

Cloud marketplaces are not passive checkout pages; they influence adoption, customer ownership, economics, and reported scale.

Large companies already buy infrastructure, databases, security software, and applications through cloud agreements. Purchasing Claude through AWS or Google Cloud can fit existing procurement systems and committed-spending arrangements.

That convenience can shorten a sales process. It can also place a model inside the tools that developers and information-technology teams already use.

Anthropic’s dependence on this route is substantial. A confidential filing reviewed by Reuters reportedly showed that 47 percent of its 2025 sales flowed through Amazon and Google marketplaces.

Those marketplace sales totaled about $2.16 billion. Anthropic reportedly paid approximately $351 million in distribution fees, or about 16 cents per marketplace revenue dollar.

The cloud channel figures illustrate both the value and the cost of partner distribution. Cloud companies deliver customers and procurement access, but they retain part of the transaction.

Those historical figures should not be applied mechanically to Anthropic’s current run rate. Channel mix, contract terms, and product demand can change quickly.

They still demonstrate why gross-versus-net treatment matters. When almost half of sales pass through partners, presentation can meaningfully affect the total reported top line.

OpenAI has a similarly consequential relationship with Microsoft. Azure has supplied computing infrastructure and distributes OpenAI models through enterprise products and cloud services.

Microsoft also has its own customer contracts, pricing structures, and bundled offerings. Those features can change which party controls a promised service under accounting rules.

The commercial contest is therefore broader than Claude against ChatGPT. It includes the distribution systems surrounding each model.

Anthropic benefits from access to both AWS and Google Cloud customers. OpenAI benefits from Microsoft’s enterprise reach and its integrations across productivity and development products.

These relationships create leverage and dependence at the same time. A cloud partner can make a model easier to buy, but it can also control billing, packaging, and customer access.

The model developer might gain rapid distribution while surrendering part of each sale. It might also depend on a partner that sells competing models or develops its own.

Amazon supports Anthropic while offering other models through Bedrock. Google invests in Anthropic while building Gemini. Microsoft supports OpenAI while expanding its own model and cloud strategy.

The United States Federal Trade Commission has previously examined these relationships. Its cloud partnership study described how investment, cloud capacity, distribution, and technical dependencies can become intertwined.

That structure makes revenue attribution more than bookkeeping. It reveals how value and control move across the AI supply chain.

For enterprise buyers, cloud distribution can simplify governance and billing. It can also make model spending harder to separate from broader infrastructure commitments.

A company might select Claude because it already has unused AWS commitments. Another might deploy OpenAI services through an established Microsoft agreement.

Those decisions represent genuine demand, but they do not always resemble a direct software subscription. The commercial channel influences both the purchase and its measurement.

That is why the primary contest is not simply Anthropic against OpenAI. It is gross channel accounting against net channel accounting as two views of commercial scale.

What the Headline Numbers Still Do Not Show

Revenue presentation cannot answer the harder questions about cash retention, compute costs, customer concentration, or durable demand.

Annualized revenue begins with a recent sales period and extends that pace across twelve months. It works best when demand and pricing remain relatively stable.

AI demand does not meet that condition consistently. New model launches, agent adoption, enterprise contracts, and temporary usage surges can change monthly consumption sharply.

A company can produce a large run rate from one exceptional month. That does not guarantee the same activity will continue for the following eleven months.

Usage-based API revenue creates another complication. Customers can rapidly increase workloads when a new model performs well. They can also redirect traffic when a rival improves quality, latency, or cost.

The run rate therefore measures present velocity more effectively than future certainty. It should not be treated as contracted recurring revenue unless disclosures support that interpretation.

Customer concentration creates a related risk. A few large buyers can produce substantial usage while making future revenue sensitive to individual renewal or routing decisions.

Cloud marketplaces can obscure that concentration from outside observers. Investors see the aggregate channel total, but not necessarily which end customers generated it.

Gross accounting can further enlarge the visible number without changing the model developer’s retained proceeds. Net accounting can hide the full scale of customer spending while presenting a closer view of the company’s share.

Neither method reveals compute economics by itself. Serving advanced models requires accelerators, networking, power, storage, and engineering support.

Those costs can vary by model, workload, hardware provider, and contract. They also change as companies optimize inference or introduce more compute-intensive systems.

Partner fees are only one deduction from revenue. Model-serving costs, training expenses, stock compensation, sales costs, and research spending remain central to profitability.

A company could lead in gross customer spending while retaining less cash from each partner sale. Another could report less revenue but keep a larger proportion of what it records.

The current figures do not resolve that possibility.

Profit claims require similar caution. Anthropic has reportedly told investors that it achieved quarterly operating profitability under a measure excluding some expenses.

That result would mark a meaningful operational milestone. It would not necessarily equal net profitability under the disclosures expected from a public company.

The same standard should apply to OpenAI. Its growth forecasts need expense, cash-flow, and infrastructure context before they support conclusions about economic strength.

Another uncertainty concerns consistency across channels. A company can use gross presentation for one partner and net presentation for another when the contracts allocate control differently.

Even a company-level label can therefore oversimplify the accounts. Investors need channel-level explanations rather than a single statement that one company reports gross.

Both businesses remain private, which limits public access to audited statements and detailed footnotes. Reported figures often come from investor documents or people familiar with internal results.

That sourcing does not invalidate the information. It does demand precise attribution and restraint.

The strongest conclusion is narrower than the headline contest. Anthropic’s and OpenAI’s disclosed run rates cannot be ranked cleanly without adjustments.

It would be equally misleading to assume the entire difference comes from accounting. Product demand, channel mix, timing, and enterprise growth also affect the totals.

A rigorous comparison must separate reported revenue, grossed-up transaction value, retained proceeds, and completed-period sales. Until then, the apparent leader depends partly on the chosen column.

Investors Are Comparing Growth Before Comparing Definitions

The accounting mismatch pressures investors and reporters to define the metric before declaring a winner.

AI companies use annualized revenue because their businesses are expanding faster than conventional annual reporting can capture. A trailing twelve-month total can look outdated before it is published.

That makes a run rate useful during rapid expansion. It also makes the metric vulnerable to selective timing and inconsistent construction.

The latest Anthropic OpenAI revenue debate demonstrates that vulnerability. One company’s number can include partner economics that another company leaves outside its top line.

Investors can address this with a bridge between reported and normalized results. The bridge should show direct sales, partner sales, partner deductions, and the final recognized amount.

It should also state the measurement period. A July monthly pace, September quarterly pace, and year-end forecast should not appear in one ranking without clear labels.

Completed quarterly revenue offers a firmer starting point. It captures an actual reporting period instead of extending a short interval across a year.

However, quarterly figures still need accounting normalization. A full quarter measured gross is not directly comparable with a full quarter measured net.

Cash receipts provide another useful check, though timing differences can make them noisy. Deferred revenue, customer prepayments, and partner settlement schedules can move cash across periods.

Gross profit can help when revenue and cost classifications are understood. Yet it becomes misleading if one company includes partner fees in cost of revenue and another never records them.

Customer activity can provide an independent signal. Enterprise seats, active developers, API consumption, and workload retention can show whether reported growth reflects durable use.

These metrics also need careful definitions. Registered accounts, paying customers, and active workloads describe different things.

The pressure extends to public cloud providers. Amazon, Microsoft, and Google benefit when model demand drives customers toward their infrastructure.

However, cloud revenue linked to model developers can involve reciprocal investment, infrastructure commitments, and marketplace fees. The same economic activity can appear in several corporate narratives.

That creates a risk of double counting across the broader AI market. A model company reports customer activity, while a cloud provider reports infrastructure or marketplace revenue related to that activity.

Each accounting treatment can be valid within its own company. Adding every headline number together can still exaggerate the sector’s independent end demand.

Public investors will ask harder questions as the model developers approach possible listings. Audited filings require clearer revenue policies, channel concentration, related-party disclosures, and expense classifications.

Those documents should reduce uncertainty, but they will not eliminate judgment. Principal-versus-agent analysis depends on contract details and control assessments.

An auditor can determine that both approaches follow accounting standards. Investors must still decide which economic measure best fits their valuation model.

For enterprise buyers, the dispute offers a practical reminder. Vendor scale should not be inferred from one annualized figure.

Buyers should examine availability, model performance, governance, switching options, service commitments, and total deployment costs. Accounting presentation does not determine which model fits a workload.

Developers face a similar choice. Distribution through a familiar cloud can simplify credentials, billing, and deployment. Direct access might provide different features or economics.

The partner channel shapes the product experience as well as the financial statement. That makes channel strategy a competitive variable worth tracking.

Three Signals Will Clarify the Real Revenue Race

The next stage of the Anthropic OpenAI revenue contest will depend on audited disclosures, normalized channel data, and sustained customer usage.

The first signal is a detailed revenue-recognition policy in an audited filing. Investors need each company to explain when it acts as principal and when it acts as agent.

That disclosure should identify material partner channels and describe how fees enter the income statement. It should also explain whether the policy changed between reporting periods.

Clear filings would strengthen confidence that the companies’ figures can be normalized. Vague language would preserve the current uncertainty and weaken simple rankings.

The second signal is a reconciliation between annualized revenue and actual quarterly sales. That bridge should identify the period used for the run rate and any gross-up adjustment.

For OpenAI, investors will watch whether the company reaches its forecast of at least $70 billion by year-end. They will also ask whether that figure uses the normal net basis.

For Anthropic, the key question is whether growth after July supports the reported $65 billion pace. A completed quarter would provide stronger evidence than another point-in-time extrapolation.

Sustained quarterly performance would strengthen the claim that both companies are experiencing durable enterprise demand. A sharp gap would weaken conclusions drawn from annualized snapshots.

The third signal is the amount of revenue retained after cloud distributions and model-serving costs. This measure gets closer to the economic value each company keeps.

Investors should watch partner fees, inference costs, gross profit, cash use, and customer concentration together. No single figure can substitute for that combination.

Better retention would show that scale is translating into operating leverage. Deteriorating retention would suggest that headline growth depends heavily on costly infrastructure or distribution.

These signals matter because the market is funding more than two software vendors. It is funding an interconnected system of model labs, cloud providers, chips, data centers, and enterprise applications.

Revenue from one part of that system often becomes an expense or investment elsewhere. Understanding who ultimately retains the customer’s dollar is essential.

The current accounting difference does not erase Anthropic’s growth or OpenAI’s reach. It changes what their headline figures can prove.

The reported run rates establish that customers are spending heavily on both companies. They do not establish a definitive revenue leader on a common basis.

Readers should resist the next simple leaderboard unless it includes definitions beside the numbers. Ask whether each figure is gross, net, annualized, completed, or adjusted.

Then ask which company keeps more of the transaction after partners and compute providers are paid. That answer will reveal more about the AI business race than the largest headline alone.