A new analysis from Epoch is challenging the dominant narrative that OpenAI is cashing in big on generative models. While ChatGPT and GPT‑5 dominate headlines, the reality is subtler — OpenAI’s high revenue masks persistent losses, and the real profits are flowing to the companies powering the technology behind the scenes.
Revenue Looks Big — But Profit Is Another Story
OpenAI’s reported revenue surged to an annualized $20 billion by late 2025, yet the company still loses billions each year. GPT‑5, for example, achieved roughly a 50% gross margin — respectable, but far below the 80%+ margins typical in SaaS. Its primary lifespan of just four months meant it couldn’t recoup the $5 billion R&D cost of training the model.
Epoch frames the economics as a treadmill:
“Profitability isn’t a straight line from launch to revenue — it’s a treadmill where you must run faster just to stay in place.”
The metaphor captures the escalating compute costs, competitive pressures, and short model half-life that dominate today’s landscape.
The Real Winners: Picks, Shovels, and Power Plants
While product creators burn cash to stay at the frontier, the underlying infrastructure layer is thriving. Epoch highlights the key players:
| Company Type | Key Player | 2026 Status | Net Profit Margin |
|---|---|---|---|
| Compute / Silicon | NVIDIA | Revenue hit $57B in Q3 2026 | 53% |
| Cloud Leasing | CoreWeave | Contracted $22.4B in capacity | Volatile but high growth |
| Memory / Storage | Micron | High demand for HBM3 memory | 22.9% |
| Energy / Nuclear | Constellation | Restarting Three Mile Island to power Microsoft | Surging utility valuation |
In other words, the flashy products capture attention, but the real economic power is in chips, cloud, memory, and energy.
The 2026 Treadmill Metrics
The demands on product creators have scaled exponentially:
-
Training compute is increasing fivefold per year.
-
OpenAI’s power consumption jumped from 0.2 GW in 2023 to 1.9 GW in 2025, with projections of 6 GW in 2026.
To offset high costs, OpenAI shifted to outcome-based pricing for reasoning tasks and launched an $8 “Lite” subscription with ads in January 2026. These measures reflect a structural reality: creating models alone no longer guarantees profitability.
Intelligence as a Utility
In 2026, it’s increasingly clear that raw computational power is becoming a utility. OpenAI provides the “energy,” but the companies building specialized applications for legal, medical, or coding tasks are the ones generating sustainable value.
A practical analogy: OpenAI is like a high-end restaurant with a line out the door, yet it can’t pay its rent because the cost of ingredients rises faster than the price of the steak.
Market Shifts in 2026
Even brand dominance doesn’t guarantee a moat. As of January 2026:
-
Anthropic leads in enterprise LLM API usage with 32% market share.
-
OpenAI has dropped to 25%, showing that compute constraints, pricing pressure, and rapid obsolescence are redefining the competitive landscape.
The Bottom Line
Epoch’s report reveals a clear economic truth: the market rewards control over critical infrastructure, not visibility or hype. OpenAI may be the most recognizable name, but the lasting profits are flowing to the companies that supply the hardware, cloud capacity, memory, and energy that make the services possible.
As the technology boom continues, the lesson is simple: look below the surface to see where the real value is concentrated. The flashy labs capture headlines, but the true winners are quietly building the foundations that allow the entire system to function.
Related: OpenAI’s Cash Crunch Exposes the Real Cost of Building AI
