When Does Investment in AI Data Centers Stop Paying Off?

However relentless the demand for more and bigger data centers that use ever more power may seem, they are not exempt from basic economics. To be sure, competition among hyperscalers to build the biggest and most impressive models is ferocious. That does not mean that adding the next data center will always make economic sense.

The simple economic question is whether the next additional graphics processing unit (GPU) can generate enough revenue to cover its full cost. If not, the investment is uneconomic, even in a growing market.

Rather than forecasting AI demand or cloud pricing, we should assess the cost structure of a GPU-based data center step by step. That will give us the Required Revenue per Hour of productive GPU use to break even.

Our yard stick

“Required Revenue per GPU-Hour” is defined as the minimum revenue that must be generated by a single GPU for every hour it is productively used in order to recover its full annual cost. Full annual cost includes the annualized capital cost of the accelerator and its attached memory, as well as the electricity required to operate it.

This metric is not a cloud pricing rate, or a billing figure charged to customers. It is an internal economic threshold implied by costs and utilization. If realized revenue per productive GPU-Hour exceeds this threshold, the investment is profitable. If it does not, the investment destroys value.

Capital cost

GPUs and their attached memory are large upfront investments that must be amortized over their useful life. This cost exists regardless of utilization.

In the baseline case used here, the combined cost of the accelerator and its attached memory is assumed to be $24,000 per unit, based on research from www.theinformationnet.com. We assume an economic life of five years. Dividing the upfront cost by the assumed life yields an annual capital recovery cost of $4,800 per accelerator. That’s the minimum annual revenue contribution required simply to recover the hardware investment.

Power is money

Electricity is a recurring operating cost that scales with time rather than utilization. Even if a GPU is lightly used, power must be supplied whenever the system is active.

I calculate electricity with three inputs: average power draw per accelerator—about 1.2 kilowatts—total hours in a year—8,760—and the price of electricity estimated at $0.10 per kilowatt-hour. Multiplying these yields an annual electricity cost of approximately $1,050 per accelerator.

Thus, the total annual cost per accelerator is the sum of annualized capital recovery ($4,800) plus annual electricity expense ($1,050) for a total annual cost of $5,850 per accelerator. This is the numerator in the Required Revenue per GPU-Hour calculation.

The annual cost alone does not determine profitability. We need to know how many hours per year the accelerator is productively used. Remember the unit draws power as long as it is active, even if it is not running a model. A year contains 8,760 hours, but only a fraction of those hours generate revenue.

At 50% utilization, an accelerator is productively used for 4,380 hours per year. At lower utilization rates, productive hours decline proportionally, forcing the same annual cost to be recovered over fewer revenue-generating hours.

We can now calculate Required Revenue per GPU-Hour by dividing total annual cost per accelerator by productive hours. At 50% utilization, dividing $5,850 by 4,380 productive hours yields a required revenue of approximately $1.34 per GPU-Hour.

Table 2 shows how sensitive the result is to utilization alone. At 40% utilization, the required revenue rises to $1.67 per GPU-Hour. At 60% utilization, it falls to $1.11 per GPU-Hour. Utilization is one of the most powerful drivers of data center economics.

In reality, none of our costs are static. In Table 3, we allow the prices of multiple inputs to change at the same time. DRAM prices increase, memory cost per accelerator rises, electricity prices drift higher, and utilization declines as infrastructure scales.

Each row in the table represents one internally consistent scenario. The quarter labels are used only to order the scenarios. They are not forecasts. The Required Revenue per GPU-Hour in the final column is calculated from the inputs in the same row using the same equation applied throughout the article.

For example, the 2026 Q1 row assumes DRAM pricing of $3.10 per GB, memory cost per accelerator of $1,587, electricity pricing of $0.13 per kilowatt-hour, and utilization of 42%. Under those assumptions, total annual cost per accelerator rises to $6,850. Dividing that cost by the reduced number of productive hours yields a required revenue of approximately $1.95 per GPU-Hour.

If any of those inputs change, the required revenue changes.

Investor Takeaway

As memory prices rise, electricity prices increase, and utilization becomes harder to sustain at scale, the revenue required per productive GPU-Hour rises. The values shown here are conditional results tied directly to explicit assumptions, not forecasts. However, they demonstrate how easily the data center economics can tighten, even when AI demand remains strong.

For investors, the key insight is that AI infrastructure spending is constrained by the rising revenue required to justify each additional unit of compute. On the flip side, cutting any of the costs cited here, especially the price of memory or the price of electricity, would lead to more demand and larger buildouts.