A new analysis reveals OpenAI’s GPT-5 achieved a surprising 50% gross margin – a figure calculated by Exponential View and Epoch AI through a detailed forensic examination of the company’s finances. Published January 28, 2026, the study by Jaime Sevilla, Hannah Petrovic, and Anson Ho tackles the critical question of whether AI companies can actually turn a profit, despite valuations in the hundreds of billions. While seemingly positive, this margin – lower than typical software companies – doesn’t tell the whole story. As Azeem Azhar notes, “running each AI model generates enough revenue to cover its own R&D costs,” but this surplus is often outweighed by the expense of developing the next generation model, potentially masking overall losses. The analysis ultimately finds GPT-5’s four-month lifespan was insufficient to recoup its own research and development costs.
GPT-5 Bundle Revenue and Inference Compute Costs
A detailed economic analysis of OpenAI’s GPT-5 reveals a complex picture of profitability, challenging the notion that current AI models immediately recoup their substantial development costs. The study, a collaboration between Exponential View and Epoch AI, meticulously dissects revenue and expense data to assess the unit economics of the GPT-5 “bundle” – encompassing GPT-5, GPT-5.1, GPT-4o, ChatGPT, and the associated API. Total revenue generated by this bundle between August and December of last year reached $6.1 billion, a figure that initially appears promising. However, a deeper dive into operational costs paints a different scenario. The largest single expense was inference compute, totaling $3.2 billion, based on estimates of OpenAI’s 2025 spend. Staff compensation accounted for a further $1.2 billion, with researchers conservatively estimating that 40% of stock compensation is allocated to running models rather than research and development. Sales and marketing contributed a significant $2.2 billion, alongside $0.2 billion in legal, office, and administrative costs. Calculating gross profit—revenue minus only inference compute—yielded $2.9 billion, resulting in a 48% gross margin, “lower than the norm for software companies (where 60-80% is typical) but high enough to eventually build a business on.”
However, factoring in all operating costs flips the script, revealing a likely operating loss of $0.7 billion and a -11% operating margin. Moreover, the analysis indicates that GPT-5, despite its revenue generation, failed to recover its own research and development costs within its four-month lifespan. “Even using gross profit, GPT-5’s tenure was too short to bring in enough revenue to offset its own R&D costs.” OpenAI also shares approximately 20% of its $6.1 billion revenue with Microsoft, further complicating the financial landscape, though the researchers note, “This doesn’t necessarily mean that the revenue deal is entirely harmful to OpenAI.” The study concludes that while gross margins might be decent, achieving overall profitability remains a significant hurdle for OpenAI and, potentially, the wider AI industry.
OpenAI’s Operating Costs: Staffing and Marketing
Beyond the headline figures surrounding OpenAI’s GPT-5, a detailed economic analysis reveals a complex picture of operational expenditure, particularly concerning staffing and marketing. While the “GPT-5 bundle” – encompassing models like GPT-5, GPT-5.1, GPT-4o, and ChatGPT – generated $6.1 billion in revenue between August and December, a significant portion was immediately absorbed by core operating costs. Staff compensation alone reached $1.2 billion, a figure derived from OpenAI staff counts, stock compensation reports, and H1B filings. Determining precisely how much of this expenditure directly supports model operation versus research and development remains a challenge, with analysts assuming 40% allocation to inference, “matching the fraction of compute that goes to inference.”
This emphasis on personnel highlights the intense competition for skilled AI researchers and engineers, driving up compensation costs. However, even more substantial was the investment in sales and marketing, totaling $2.2 billion, predicated on the assumption that OpenAI’s spending grew in the latter half of the year. This aggressive marketing spend underscores the need to rapidly acquire users and establish market dominance in a fiercely competitive landscape. Combining these costs with legal, administrative expenses of $0.2 billion, and the substantial $3.2 billion spent on inference compute, the analysis suggests a precarious financial position. Indeed, when all operating costs are tallied, the GPT-5 bundle appears to have operated at a loss of $0.7 billion, resulting in an operating margin of -11%. “One option is to look at gross profits,” the analysts note, which, at $2.9 billion, presents a more favourable, though still lower than typical software margins, picture. Furthermore, the 20% revenue share agreement with Microsoft adds another layer of financial complexity, described as “a real drag on OpenAI’s path to profitability,” despite potential revenue sharing from Microsoft itself. This suggests that while running AI models may yield decent gross margins, achieving overall profitability requires careful management of substantial operational overheads.
Last year, OpenAI earned about $13 billion in full-year revenue, compared to $6.1 billion for the GPT-5 bundle.
GPT-5 Gross vs. Operating Profit Margins Analyzed
## GPT-5 Gross vs. However, a breakdown of expenditures paints a different reality. The analysis highlights a 50% gross margin, calculated by subtracting compute costs from revenue, resulting in $2.9 billion in profit. The crucial divergence emerges when broader operational costs are factored in. Inference compute alone reached $3.2 billion, with staff compensation adding another $1.2 billion, and sales and marketing consuming $2.2 billion. Furthermore, the study casts doubt on the ability of GPT-5 to recoup its substantial research and development (R&D) costs within its limited four-month lifespan.
Estimating total 2025 R&D expenditure at $16 billion, analysts determined that even a conservative calculation of GPT-5’s development costs – around $5 billion – exceeded the $3 billion in gross profits generated during its operational period. “In other words, OpenAI spent more on R&D in the four months preceding GPT-5, than it made in gross profits during GPT-5’s four-month tenure,” the report states.
This suggests a model where current profits are often reinvested into future development, a common strategy in rapidly evolving tech sectors. “So if GPT-5 is at all representative, then at least for now, developing and running AI models is loss-making.”
For the investors who are willing to entertain more extreme scenarios, an even stronger effect is when “ intelligence explosion ” dynamics kick in – if OpenAI pulls ahead at the right time, they could use their better AIs to accelerate their own research, amplifying a small edge into a huge lead.
Jevons’ Paradox and AI Model Profitability Debate
The relentless drive towards more powerful AI models is prompting a critical re-evaluation of their underlying economics, with a surprising connection to the 19th-century observation known as Jevons’ Paradox. This principle posits that technological progress increasing efficiency in resource use tends to increase overall resource consumption – a phenomenon increasingly relevant as AI models become cheaper to operate. “Jevons’ paradox suggests that as tokens get cheaper, demand explodes,” explain Azeem Azhar and Hannah Petrovic, authors of a recent analysis of OpenAI’s finances. The question now is whether this surging demand translates into sustainable profitability for AI developers.
A detailed examination of OpenAI’s GPT-5 reveals a complex financial picture, challenging the narrative of immediate returns. The study highlights that even with a seemingly healthy gross margin, OpenAI likely incurred an operating loss. Total operating costs, including staff compensation ($1.2 billion) and sales & marketing ($2.2 billion), pushed expenditures close to $6.8 billion, exceeding the $6.1 billion in revenue generated between August and December. Furthermore, the analysis indicates that GPT-5’s four-month lifespan proved insufficient to recoup its substantial R&D investment, estimated at around $5 billion in the four months prior to release.
Microsoft Revenue Deal Impacts on OpenAI Finances
The financial relationship between OpenAI and Microsoft is proving more complex than simple revenue sharing, with implications for the long-term profitability of AI model development. Analysis of OpenAI’s finances reveals that while gross margins from running models like GPT-5 appear healthy – around 50% – a deeper look at operational costs complicates the picture. Specifically, the agreement to hand over approximately 20% of revenue to Microsoft, totaling a significant portion of the $6.1 billion generated by the “GPT-5 bundle” (including offerings like ChatGPT and the API) between August and December, creates a substantial drag on OpenAI’s bottom line.
This isn’t necessarily detrimental; as stated by researchers, “the deal probably shouldn’t significantly affect how we see model profitability — it seems more to do with OpenAI’s economic structure rather than something fundamental to AI models.” However, ongoing renegotiations of the deal underscore its impact on OpenAI’s financial trajectory. The analysis highlights a substantial difference between gross profit and operating margin, with the latter potentially negative at -11% when all costs are factored in – including $3.2 billion for inference compute, $1.2 billion for staff compensation, and $2.2 billion for sales and marketing.
Beyond immediate operating costs, recouping research and development (R&D) expenditure remains a key challenge. Estimating that OpenAI spent $16 billion on R&D in 2025, analysts found that even with a $3 billion gross profit from the GPT-5 bundle’s four-month lifespan, the model failed to offset its development costs. Ultimately, the study proposes viewing frontier models as “rapidly-depreciating infrastructure,” requiring rapid revenue extraction before obsolescence. While operating margins may be misleading for a fast-growing company, the interplay between Microsoft’s revenue share and the speed of model iteration presents a critical economic puzzle for OpenAI and the wider AI landscape.
In fact, we could be substantially lowballing the R&D costs. GPT-5 has been in the works for a long time – for example, early reasoning models like o1 probably helped develop GPT-5’s reasoning abilities.
R&D Costs and GPT-5 Lifecycle Revenue Recoupment
While initial gross margins appear respectable, a deeper examination of operational expenses suggests a different reality. However, the cost of maintaining this suite of models quickly erodes those gains. Staff compensation alone reached $1.2 billion, with researchers consuming a significant portion of these funds, while sales and marketing accounted for a substantial $2.2 billion expenditure. Legal, office, and administrative costs added another $0.2 billion to the tally. Crucially, the four-month lifespan of GPT-5 proved insufficient to recoup its substantial research and development investment. This finding is stark: “in practice, it seems like model tenures might indeed be too short to recoup R&D costs.” The rapid pace of innovation and competitive pressure, exemplified by Gemini 3 Pro surpassing GPT-5 within three months, necessitates a constant cycle of development, rendering models “like rapidly-depreciating infrastructure.”
Despite these financial challenges, the analysis doesn’t dismiss the value of models like GPT-5. “Even an unprofitable model demonstrates progress, which attracts customers and helps labs raise money to train future models.” The R&D invested in GPT-5 isn’t lost; it informs the development of subsequent iterations, like GPT-6, potentially paving the way for greater profitability in the future.
