Fresh research from Epoch AI suggests that any earnings surplus generated by one model is ‘eclipsed’ by the development costs of its successor.
Recent discoveries from a study conducted by Epoch AI, a non-profit research institution, appear to significantly challenge the belief that AI companies, particularly OpenAI, will eventually achieve profitability.
The research document, authored by Jaime Sevilla, Hannah Petrovic, and Anson Ho, indicates that while operating an AI model might generate sufficient revenue to cover its operational expenses, any generated profit is negated by the substantial cost involved in creating the subsequent advanced model. Consequently, it highlighted, “even if individual models are profitable, companies can still incur annual losses.”
The report aims to address three fundamental inquiries: How financially viable is the operation of AI models? Do models maintain profitability throughout their operational lifespan? And will AI models ultimately achieve profitability?
To tackle the first question, researchers devised a scenario dubbed the GPT-5 bundle. This bundle encompassed all of OpenAI’s offerings available during the period GPT-5 served as the flagship model, including GPT-5 and GPT-5.1, GPT-4o, ChatGPT, and the API. They then estimated the earnings and expenditures associated with running this bundle. All figures were compiled from various information sources, including statements from OpenAI and its personnel, alongside reports from media outlets, primarily The Information, CNBC, and the Wall Street Journal.
The projected revenue, they explained, “is fairly simple to determine.” Since the bundle covered all of OpenAI’s models, it represented the company’s total income during GPT-5’s operational span from August to December of the previous year, amounting to $6.1 billion.
And, they noted, “initially, $6.1 billion sounds impressive, until one considers it against the costs of operating the GPT-5 bundle.” These expenses originate from four primary areas, according to the report. The first is inference compute, costing $3.2 billion. This figure is derived from public projections of OpenAI’s total inference compute expenditure in 2025, assuming that compute allocation during GPT-5’s activity was proportionate to the share of the year’s revenue generated in that timeframe.
The remaining expenditures comprise staff remuneration ($1.2 billion), sales and marketing efforts ($2.2 billion), and finally, legal, office, and administrative overheads, totaling $0.2 billion.
The Core of the Calculation
Regarding methods for computing profit, the paper stated, “one approach involves examining gross profits. This considers only the direct operational expenses of a model, which in this instance is solely the $3.2 billion inference compute cost. Given that revenue stood at $6.1 billion, this yields a profit of $2.9 billion, or a gross profit margin of 48%, aligning with other forecasts. While this figure is lower than those seen in other software enterprises, it is sufficiently high to form a viable business foundation.”
In essence, they affirmed, “the operation of AI models appears to be profitable, characterized by respectable gross margins.”
However, this presents only a partial view of the financial landscape.
The paper highlighted that accepting the premise that only gross margins should be considered when assessing profitability means, “on those terms, operating the GPT-5 bundle was indeed profitable. Yet, was it profitable enough to offset its development expenses? Theoretically, yes—one simply needs to keep running them, and eventually, sufficient revenue will accumulate to recover these costs. However, in reality, models might possess too brief a lifespan to generate adequate revenue. For example, they could face obsolescence due to superior offerings from competing labs, necessitating their replacement.”
The crux, the authors explained, lies in contrasting gross profits with the nearly $3 billion against the firm’s R&D expenditures: “To properly assess AI products, we must examine both inference profit margins and the speed at which users transition to improved alternatives. In the case of the GPT-5 bundle, we determine that it is decisively unprofitable across its entire lifecycle, even when considering only gross margins.”
Regarding the overarching question of whether AI models will achieve profitability, the paper asserted, “the most critical takeaway is that these model lifecycle losses are not necessarily a cause for concern. AI models do not need to be profitable immediately, so long as companies can persuade investors of their future profitability. This is a common pattern for rapidly expanding tech enterprises.”
The ultimate conclusion, according to the three authors, is that profitability is highly attainable because “compute margins are shrinking, enterprise agreements offer more stability, and models can maintain relevance for longer than the GPT-5 cycle implies.”
When questioned about the market’s irrationality persisting long enough for OpenAI to achieve financial stability, Jason Andersen, VP and principal analyst at Moor Insights & Strategy, commented, “it’s a possibility, but not a certainty. I anticipate that in 2026, these companies will refine their strategies. In my view, OpenAI and other general-purpose AI providers have three main levers to enhance their financial standing (or at least decelerate their cash burn).”
The first, he elaborated, is pacing, “which I believe is already underway. We observed a slower cadence of major model releases last year. Thus, by moderating their speed, they can mitigate some costs or at minimum, distribute them more effectively. Frankly, customers also need time to adapt, so they can plausibly slow down, allowing the market to catch up with their existing offerings.”
The second, Andersen noted, is to broaden their product range, and the third involves securing revenue from other software vendors.
As for whether OpenAI and similar entities can endure long enough for AI to truly realize its potential, he stated, “OpenAI and Anthropic possess the best prospects for sustained independence. However, I also wish to approach ‘truly effective’ with caution. If ‘truly effective’ signifies achieving AGI, that remains theoretical, likely requiring significant advancements in hardware and energy. But if ‘effective’ means reaching profitability over a span of years, then yes, those two have a credible chance.”
The key to profitability, he explained, “will involve discovering a strategy to compete and succeed against companies that have inextricably linked their future to AI. Notably, Google, Microsoft, and X have now integrated their models so deeply into their other products and services. So, is there adequate time and sufficient diversification opportunities to rival them? My prediction is that a select few pure-play companies will thrive and perhaps even disrupt the market, but many others will not survive.”
Describing the paper’s conclusions as “quite linear” and based on short-term analysis, Scott Bickley, advisory fellow at Info-Tech Research Group, mentioned that OpenAI has been “rather transparent about its current lack of profitability. Their response hinges on a compelling chart illustrating exponential revenue growth over the next three-plus years, which is why they are currently seeking to raise $200 billion to build the infrastructure necessary to support hundreds of billions of dollars in annual business.”
Numerous Fortunes Linked to OpenAI
He estimated that OpenAI’s total financial commitments, stemming from agreements with Nvidia and hyperscalers, alongside data center constructions, now amount to $1.4 trillion, adding, “They’re aiming to make themselves too significant to fail, to secure the extensive runway they’ll require for these investments to hopefully yield returns over years, or even decades.”
Currently, he observed, the company is “strengthening its financial position. They are striving to construct every possible asset to secure future operational time. But either they achieve unprecedented success, creating applications that I cannot currently imagine as realistic, or they experience a dramatic failure, in which case everyone would be guaranteed to acquire portions of their empire for a fraction of the cost, or something similar. But I see it as either a boom or a bust scenario; there’s no middle ground.”
As things currently stand, Bickley stated, all major vendors have “intertwined their destinies with OpenAI, which is precisely what Sam Altman intended. He’s going to compel the most influential players in the sector to contribute to his success.”
Should the company eventually fail, he predicted the implications for firms utilizing its AI initiatives would be minimal. “Regardless of what becomes of OpenAI as a commercial entity, the intellectual property developed, and the existing models, will persist. They will come under someone’s stewardship and continue to be employed. There is no risk of them becoming unavailable.”
This content was originally featured on Computerworld.
