A look at OpenAI’s economics
Or why OpenAI must pivot from being a consumer app to an enterprise leader — before competitors and market dynamics erode this opportunity.
OpenAI is one of the world’s most influential businesses — and one of the most confounding when analyzing its business model.
Founded in 2015 as a non-profit, OpenAI has dominated the global conversation around generative AI (genAI) since the first public release of ChatGPT two years ago. While still technically governed by a non-profit entity, OpenAI has a for-profit business (which limits returns that investors can earn) that is reportedly on track for $3.7 billion in revenue this year with forecasts of $11 billion for 2025.
OpenAI just raised another $6.5 billion round in September at a $157 billion valuation, bringing its total raised to $17.9 billion.
Meanwhile, the company is innovating at a furious pace, releasing its Strawberry — or OpenAI o1 — model in September, which taps into a powerful new genAI model that uses a reasoning technique called “chain-of-thought reasoning.” Bringing reasoning capabilities to LLMs is a technological necessity in a future of agents (i.e., autonomous systems) and a significant step towards OpenAI’s ultimate goal of AGI. These models are designed to process information more accurately than the first wave of LLMs, making them ideal for complex tasks like scientific research, coding, and strategic decision-making. Per Noam Brown at the TedAI conference I attended last month in SF, “Now we have a new parameter, one where we can scale up system two thinking as well — and we are just at the very beginning of scaling up in this direction.”
This release comes as rival Anthropic has released its Claude 3.5 Sonnet, the first frontier AI model to offer computer use in public beta, or in other words, an AI agent. Though OpenAI should not be far behind, this reflects the intense competition in the space.
Still, on the innovation front, OpenAI is now beginning to roll out its search product — potentially one of the biggest challenges Google has faced in decades to this core business.
This has all the trappings of an unstoppable juggernaut. And yet…
The company has some significant vulnerabilities. Chief among them are expenses. As The New York Times wrote in a recent headline, “OpenAI Is Growing Fast and Burning Through Piles of Money.” Citing the documents that OpenAI circulated to investors as part of its most recent funding round, the Times reported that OpenAI “expects to lose roughly $5 billion this year after paying for costs related to running its services and other expenses like employee salaries and office rent, according to an analysis by a financial professional who has also reviewed the documents. Those numbers do not include paying out equity-based compensation to employees, among several large expenses not fully explained in the documents.”
In a business model in which the ability to continuously innovate is critical, though a breakthrough, the Strawberry model is just at its beginnings and far more costly than others. Aside from its o1 model, OpenAI still faces plenty of competition from fellow genAI LLM startups in a market that is evolving and innovating at a breathless pace.
Given OpenAI’s stature, it’s critical to understand just how resilient its business model is. So is OpenAI unstoppable…or in trouble?
Amid the company’s scrutiny during the latest fundraising, I applied our firm’s Advanced Growth Intelligence (AGI) methodology to OpenAI’s business using the publicly available data to test the durability of its growth and thus assess its “durable growth” moat. In this case, AGI provides a comprehensive assessment focusing on four critical pillars: Quality of Revenue, Quality of Growth, Quality of Margins, and Balance Sheet Strength. We develop the first three in this note.
Quality of Revenue
OpenAI’s revenue model is primarily derived from consumer subscriptions to ChatGPT (c.75% of total revenue per CFO, Sarah Friar), bolstered by a recently introduced enterprise-grade version, ChatGPT Pro. While these are valuable sources of recurring revenue, certain aspects compromise the overall revenue quality as they command lower margins.
Consumer-Driven Revenue Base: ChatGPT consumer subscriptions are OpenAI’s primary annual recurring revenue (ARR) source, shaping its valuation. The company boasts 250 million weekly active users (!) with a 5–6% conversion rate.
However, consumer-driven revenue can be less reliable and more sensitive to market fluctuations, especially for high-tech products, as user interests shift quickly.
The revenue stream’s dependency on consumer subscriptions also raises questions about retention, where sustained growth is highly contingent on OpenAI’s ability to retain and continually engage a large user base.
From a revenue model perspective, there are emerging questions about potential advertising models. Rumors suggest that OpenAI could adopt an ad-driven model similar to Meta.
But even with that, we consider this revenue stream to be of lower quality, given its lower gross margins due to the high cost of running ChatGPT and scale limitations. This is a crucial difference vs. traditional consumer tech models: the cost to serve a $ of ChatGPT revenue is much higher than the cost to serve a $ of any other online model. I often hear that it is a winner-takes-all race, as in other consumer tech plays. I argue that precisely because the technological aspect of revenue (i.e., the gross margin per revenue stream) is significantly lower, it is not.
Enterprise Expansion with Room for Optimization: On the enterprise side, OpenAI generates revenue through API tokens and custom fine-tuning solutions for professional users. This segment offers a more stable revenue base — though ARR has to be defined adequately given the usage-based nature — and potentially a higher lifetime value (LTV) per client.
The problem is that inference costs are falling as models become less differentiated in customers’ eyes, leading to steep price competition. Per Nathan Benaich, in his latest “State of AI” report, the cost of outputs from OpenAI’s GPT-4o today is 100 times less per token (which is about equivalent to 1.5 words) than it was for GPT-4 when that model debuted in March 2023. Google’s Gemini 1.5 Pro now costs 76% less per output token than when that model was launched in February 2024.
This is why Strawberry could provide headroom for OpenAI: though many customers will not choose a significantly more expensive model than ChatGPT 4.0, a fraction of enterprise customers with complex use cases could see the associated benefits provided the solution delivers its promises technically. This is a welcome move to avoid commoditization and thus preserve margins in the LLM space.
Quality of Growth
In evaluating OpenAI’s growth, the AGI framework emphasizes the speed of development (expressed in Net Revenue and ARR growth) and its quality. In other words, what are the key levers of the company’s growth equation, and are they each sustainable over the long term?
Consumer as a Gateway to Enterprise: The rapid, viral adoption of ChatGPT among consumers has undeniably been a success. ChatGPT is already the first genAI web product (by unique monthly visits) and the first genAI mobile app (by monthly active users). Competition is heating up, and virality is likely to taper in the near term.
However, this consumer market could serve as a stepping-stone for enterprise growth by positioning OpenAI as the go-to solution for AI-driven applications, thereby reducing CAC on the enterprise side.
High-Quality Enterprise Growth Potential: OpenAI’s enterprise market presents fertile ground for higher-quality growth. By positioning its consumer segment as an acquisition channel, OpenAI can unlock greater value in the enterprise space, where high retention (evidencing the stickiness of a company’s product) and recurring usage create more stable and predictable revenue.
Competition is intense among LLM providers vying for enterprise use cases. Most companies have already integrated generative AI capabilities and commonly utilize multiple LLMs to address various needs. While one LLM might be well-suited for a customer chatbot, another may be more appropriate for complex tasks involving mathematical or scientific reasoning. To effectively navigate this landscape, adopting a fundamental approach to growth dynamics, with the proper level of detail, is essential for assessing the quality of growth. This will involve breaking down growth dynamics on a quarterly basis by product and end-use case. By doing so, we can better understand both the effective usage and monetization of each product (such as Da Vinci, Strawberry, ChatGPT 4.0, etc.) and use case, as indicated by a higher number of tokens consumed for API calls and fine-tuning. This analysis will help identify whether truly successful products or use cases are emerging; or if companies are merely transitioning from one proof of concept to another while still testing the solution.
In an environment where compute infrastructure is becoming a critical bottleneck for AI model development — with each new generation requiring exponentially more resources — it’s essential to validate existing models’ commercial traction and ROI before committing resources to next-generation development. In an extreme case, there might be no new model or it could be significantly delayed.
High Customer Acquisition Costs (CAC): Customer acquisition costs (CAC) in the enterprise market remain steep, necessitating strategic investments in customer education and value demonstration, especially with more complex offerings like Strawberry and fine-tuned chatbot capabilities. OpenAI’s enterprise revenue and growth may be of higher quality. Still, the financial burden of onboarding and scaling these clients is substantial, potentially tempering the net gains from the enterprise segment.
Enterprise acquisition, while providing higher returns, involves heavy upfront investments in client onboarding and support. These costs eat into potential profits, particularly as OpenAI continues to scale up operations in enterprise markets.
Quality of Margins
The sustainability of OpenAI’s model also depends heavily on its ability to improve margins — a significant challenge given the high computing costs and customer acquisition, the latter having been developed above.
Compute and Infrastructure Costs: A critical component of OpenAI’s cost of goods sold (COGS) is compute, which relies heavily on costly infrastructure investments with Nvidia for GPUs, significant energy consumption, and the need for scalable data centers. With high computing costs and geopolitical tensions potentially affecting supply chains, OpenAI’s margins are at constant risk. Any disruption in these inputs would quickly impact the bottom line, rendering the current model potentially unsustainable in the long term without alternative strategies for optimization or partnerships that reduce infrastructure costs.
Note that GPU costs for developing and training can potentially be treated as either opex or capex. The criteria may well depend on the actual use of the GPU.
Path To Sustainability
The AGI framework allows us to draw two clear conclusions:
First, while the consumer segment has successfully driven early growth, it requires a clear pathway to profitability. With increasing price pressure, intensifying competition, and high serving costs (particularly compute), the unit economics are challenging. This segment should be viewed as a customer acquisition channel for enterprise-grade offerings rather than a standalone profit center.
Next, the enterprise segment, with its potential for high-quality growth and robust Annual Recurring Revenue (ARR), represents OpenAI’s true source of value. This aligns with the AGI methodology’s emphasis on sustainable, scalable assets. However, the quality of its revenue, growth, and margins is yet to be demonstrated at this stage.
Despite this potential, OpenAI’s margins remain a concern, with high costs of goods sold (COGS), customer acquisition costs (CAC), and capital expenditures (capex) posing significant obstacles to profitability and cash flow generation. To address these challenges, OpenAI may need to explore strategic pivots or partnerships that can lower infrastructure costs, reduce dependency on compute resources, and optimize customer acquisition expenses. Ultimately, the path to profitability for OpenAI will be steep.
While investors are currently willing to subsidize a transformative company like OpenAI, this generosity will not last indefinitely. OpenAI faces a critical window in the next 12 to 24 months to demonstrate traction in the enterprise sector and a sustainable unit economics model. This is crucial before competition intensifies, investor patience diminishes, infrastructure costs increase, and market dynamics shift. Until then, the company and its founder’s reputation and first-mover advantage should enable it to maintain its dominance in the space, provided there are no new significant internal or external challenges.
