The Truth About Cloud Costs
Who Is Paying the Rent in the AI Era?
*On the quiet rise of cloud infrastructure costs*
As the AI industry has accelerated, the thing that has grown fastest is not model performance. It's the price of compute.
In 2023, the average hourly cloud rental cost for an NVIDIA H100 GPU sat around $2 to $3. By mid-2024, that number had climbed to $4 to $6. By 2025, some regions were reporting hourly rates as high as $8 to $10. Same hardware, same usage hours, but the price has doubled or tripled. Meanwhile, AWS, Azure, and GCP have grown their AI workload revenue by more than 50% annually.
These numbers tell a simple story. **Someone is paying more rent every year.**
Who absorbs the rent
Look at the P&L of nearly any AI startup, and the largest line item is, with rare exception, cloud infrastructure. OpenAI runs on a structure that would be impossible without Microsoft's Azure credits. Anthropic depends heavily on AWS. The companies smaller than them have it worse — 30 to 50 percent of their revenue flows directly to cloud providers.
That cost ends up somewhere. Some of it is passed to users through price hikes. Some is absorbed through margin compression and pushed onto investors. And some is paid through closure — companies that simply couldn't sustain the cost and disappeared from the market.
The bigger problem is that the pricing power sits in the hands of three companies. AWS, Azure, Google Cloud. Together they hold more than 65 percent of the global cloud market. Pricing in this space is not the result of competition. It is closer to a cartel.
The economics of distributed infrastructure
This is where distributed infrastructure becomes meaningful. Not because "decentralization is good" as a slogan, but because without distributing pricing power, the cost structure of the AI era is fundamentally unsustainable.
A distributed compute network can run on a different cost structure. A node operator only needs to recover their electricity bill and hardware depreciation. The real estate, marketing, and shareholder dividends that big cloud providers build into their pricing simply don't enter the equation. That's where the math for delivering the same unit of compute at half the price comes from.
Distributed infrastructure isn't a fit for every workload. Ultra-low-latency training and single-shot training of the largest models will continue to live in centralized data centers. But inference, fine-tuning, distributed training, and data storage — these are workloads that can move to distributed infrastructure today. And these workloads, in aggregate, are where most AI spending actually happens.
To summarize
Infrastructure costs in the AI era will keep rising. Models will grow larger. Inference demand will explode. Data center power costs are already approaching their limits. In this trajectory, someone has to pay more rent each year.
The question is whether that rent continues to flow to three companies, or whether it gets distributed to the participants of an actual network. That single choice will define the next decade of AI-era infrastructure.