Is the AI Economic Equation No Longer Adding Up?

A Trillion-Dollar Gamble With an Increasingly Ugly Balance Sheet

In 2025, global tech giants' investments in AI infrastructure have reached staggering proportions. Microsoft, Google, Amazon, and Meta alone spent over $200 billion on AI-related capital expenditures in 2024, and that figure continues to climb in 2025. Nvidia is raking in enormous profits from GPU sales, data centers are springing up everywhere, and surging power demands are straining national electrical grids.

Yet an increasingly unavoidable question is surfacing: Does the AI economic equation actually add up?

Sequoia Capital partner David Cahn once offered a pointed calculation — for the current level of AI infrastructure investment to generate a reasonable return, the entire AI industry would need to produce at least $600 billion in annual revenue. The reality is that even if you combine all AI-related revenue from OpenAI, Anthropic, Google Cloud AI services, Microsoft Copilot, and every other major player, the total still falls enormously short of that figure.

This begs the question: Are we witnessing an epic misallocation of resources?

The 'Valley of Death' Between Input and Output

The Cost Side: A Bottomless Pit

The cost structure of large AI models — from training to inference — devours astonishing amounts of capital at every stage.

Training costs are escalating rapidly. GPT-4's training cost was estimated at over $100 million, and the next generation of flagship models is expected to break the $1 billion barrier. The compute required for each model iteration is growing exponentially. Anthropic CEO Dario Amodei has publicly stated that within the next few years, training a single model could cost $5 billion or even $10 billion.

Inference costs are equally significant. Training is a one-time investment, but inference is an ongoing expense. Every time a user asks ChatGPT a question, substantial GPU compute is consumed. Analysts estimate that OpenAI's inference computing costs reached billions of dollars in 2024 — and that was after continuous optimization of inference efficiency.

Infrastructure costs are truly astronomical. A modern AI data center costs billions of dollars to build and faces a series of bottlenecks including power supply, cooling systems, and land approvals. Microsoft has even begun exploring nuclear power solutions, while Google and Amazon are signing long-term clean energy procurement agreements. The payback period for these investments often stretches beyond ten years.

The Revenue Side: Rapid Growth but Insufficient Scale

AI is indeed generating revenue, and at an impressive growth rate. OpenAI's annualized revenue has reportedly surpassed $5 billion, and Microsoft's AI-related cloud service revenue is also growing rapidly. But the problem is that relative to the scale of investment, these revenues remain a drop in the bucket.

More critically, users' willingness to pay has a clear ceiling. ChatGPT Plus at $20 per month is already not cheap for individual users, but that price may be far from sufficient to cover the underlying compute costs. Enterprise customers are willing to pay more, but enterprise AI adoption is moving much more slowly than on the consumer side, with many companies still in the "trial" and "wait-and-see" phase.

A harsh reality: Most AI startups have gross margins far below traditional SaaS companies. Traditional cloud software companies typically enjoy gross margins of 70%-80% or higher, while AI companies, burdened by steep inference costs, often see margins of only 50%-60%, with some even lower. This means that even as revenue grows, the room for profit growth is severely compressed.

Echoes of History: Is This Time Really Different?

Optimists frequently draw parallels to the dot-com bubble — people questioned the internet's economic model back then, too, but ultimately Amazon, Google, Facebook, and others proved that the internet could create trillions of dollars in commercial value. Won't AI repeat the same story?

The analogy has merit, but it also has critical blind spots.

First, the internet's marginal cost approaches zero; AI's does not. Once a website is built, an additional visitor adds virtually no cost. But every AI inference requires real GPU compute power — every additional user and every additional query means more electricity consumption and hardware wear. While inference efficiency continues to improve, the pace of demand growth may be even faster.

Second, the internet created entirely new business models; AI is primarily replacing and augmenting existing ones. Search engines created pay-per-click advertising, social networks created news feed advertising, and e-commerce redefined retail — all representing value creation from scratch. Most current AI applications — whether intelligent customer service, coding assistants, or document summarization — are fundamentally improving the efficiency of existing workflows rather than opening entirely new revenue streams. Efficiency gains certainly have value, but that kind of value struggles to support a trillion-dollar market.

Third, the competitive landscape may drive 'value to zero.' As open-source models (such as Llama, Qwen, DeepSeek, and others) increasingly close the gap with closed-source models, AI inference is rapidly becoming commoditized. When multiple models can deliver similar quality output, price wars are virtually inevitable. We have already seen mainstream API call prices drop by over 90% in the past year. This is great for users but a nightmare for AI companies' business models.

Who Is Footing the Bill for This Feast?

The AI industry's current economic cycle has a structure worth pondering: Nvidia is the biggest winner, cloud providers are the biggest buyers, AI application companies are the biggest "hope," and venture capital and public market investors are the ultimate bill payers.

Nvidia's data center revenue exceeded $47 billion in fiscal year 2024, with strikingly high profit margins. But this revenue is essentially downstream customers (cloud providers and enterprises) "pre-paying" for anticipated future AI demand. If downstream AI applications ultimately fail to generate enough revenue to justify these hardware investments, GPU procurement volumes will inevitably decline.

Meanwhile, a large number of AI startups survive on venture capital lifelines. Anthropic has raised over $10 billion cumulatively and is valued at tens of billions of dollars, yet it remains unprofitable. This model can be sustained in a low-interest-rate, capital-abundant environment, but once capital market sentiment shifts, many companies could face severe existential challenges.

Where Are the Possible Ways Out?

Although AI's economic equation looks unfavorable at present, this does not mean AI is destined to be a bubble. Several potential paths could change the picture:

1. Continued Breakthroughs in Inference Efficiency

If inference costs can drop fast enough, the gross margins of AI services will improve significantly. In fact, the pace of inference efficiency gains over the past two years has been genuinely impressive — through techniques such as model distillation, quantization, and speculative decoding, the compute cost for equivalent-quality output has already fallen by an order of magnitude. If this trend continues, AI's economic model could fundamentally improve within the next two to three years.

2. The Emergence of a Killer Application

The internet's economic model only truly worked after search engine advertising and e-commerce emerged. The AI space may also need its own "killer application" — a use case that creates an entirely new business model rather than merely replacing existing workflows. AI Agents represent one such promising direction.