Why top AI labs may struggle financially, even with working tech
The initial puzzle
Open-weight AI models are nearing the capabilities of frontier ones with much lower costs. This is because major labs bear high expenses for research and trying new things. Competitors can then use these advancements without the cost. AI improvements also don’t scale perfectly with more computing power, so a small difference in computing doesn’t create a large performance gap. Open models can also use techniques like distillation to benefit from expensive research without the expense. The real question isn’t if AI tech will improve, but if businesses can be built around it.
The main challenge
Top AI labs train models and then charge based on usage (per token). This is difficult for several reasons:
Labs must recover rapidly increasing costs in a short period, while their service prices fall each year. Since prices can’t rise due to competition, they need to massively increase usage volume, about 30x per year, to stay afloat.
Why volume growth is hard
Increased usage means higher operating costs. Labs need to grow profit \* volume, not just volume. As profit per token shrinks, extreme volume growth is required. Customers want tasks completed, but labs earn money based on tokens used. Better AI models use fewer tokens for the same task, reducing revenue. Conversely, AI performing longer, independent tasks uses more tokens. The unpredictability of volume growth stems from these competing trends. Usage growth often comes from cheaper, distilled versions or older models, which funds past investments rather than new research.
The hardware aspect
GPUs last about 4 years, becoming outdated due to better efficiency rather than wear. While shorter than assets like solar panels, GPUs generate higher annual revenue relative to their cost. The short lifespan isn’t the primary issue. The problem is that solar has long-term contracts, while GPU revenue depends on short-term market prices that are rapidly falling. This means the payback period for GPUs, based on current prices, might exceed how long those prices remain valid.
A fitting comparison
This situation is similar to pharmaceuticals, but without patent protection. There are huge research costs, but very low production costs per unit. Value relies solely on exclusivity. However, in AI, this exclusivity lasts only about six months, lacks legal protection, and is undermined by labs themselves through distillation. Pharmaceutical companies can recoup costs due to twenty years of protection. Without that, the business model falters.
What this suggests
Since both costs and revenues grow exponentially, decline isn’t gradual. Failure looks like this: a new AI run costs three times more, revenue only increases 1.5 times, and the next run cannot be funded, causing the operation to halt. This can happen even as AI models continue to improve. AI advancement doesn’t guarantee business health. User satisfaction can also be misleading. The key indicators are how long AI can perform tasks independently, its success rate, and how quickly open-weight models catch up.
Possible solutions
All potential solutions involve moving beyond selling AI access per token:
Major AI labs are pursuing these strategies, indicating they agree with this assessment.
An important note
It’s possible that extremely high usage could sustain the business. Complex tasks requiring reliable, top-tier AI can consume vast numbers of tokens, and adoption is still low. This isn’t a prediction of failure, but a statement of what’s required for success.
A historical parallel is the fiber optic cable build-out in the late 1990s. The technology was sound, and demand eventually grew, but the companies that laid the cables went bankrupt. The infrastructure remained, but the investments did not.
