How Much More AI Can We Get From the Power We Have?

I think that businesses will want more work from AI for the money they already spend. As it becomes part of daily operations, paying more every time they expand its use will become harder to justify. A small company that starts using AI to draft documents will eventually want help reviewing them, finding information, and carrying out the work those documents describe. Its budget has limits, even if its ideas for using AI keep growing.

That creates a problem for the providers. They need to serve more demanding work at prices customers will accept, while the infrastructure supporting that work takes time to expand. My expectation is that investment will bring additional data centers and electricity supply. The question is how much useful work providers can deliver while those projects are being built, and what it will cost afterward.

The subscription plans make that tension visible. A fixed monthly payment gives a customer access to a variable amount of computing. Heavy users can receive token allowances that would cost thousands of dollars at the same provider’s API rates, where usage is billed separately. SemiAnalysis’s October 2026 subscription research describes those plans as subsidized and measures how their value changes with the model, workload, and usage limits. The API-equivalent amount describes what a customer would otherwise pay; it differs from the provider’s cost of supplying the service.

A customer building daily work around that access has a practical concern: will the same payment continue to support that work? A plan can become less useful without its monthly price changing. Smaller allowances, restrictions on particular models, or longer waits all affect what the customer gets. Anthropic’s published plan documentation, for example, specifies both five-hour session limits and weekly limits, with paid usage available after those allowances are exhausted.

Building capacity takes time

The International Energy Agency’s 2026 assessment describes a mismatch between the speed of AI growth and the physical systems supporting it. Grid connections, electrical equipment, and chip manufacturing all face constraints. Its central projection puts data center electricity consumption at about 950 terawatt-hours in 2030, roughly double the 2025 level. That projection covers data centers overall; electricity use at AI-focused facilities is expected to grow faster.

Investment can fund a new facility, but the project still needs equipment, permission to operate, and a reliable electricity supply. Generating power onsite introduces another set of construction and equipment requirements. The IEA warns that gas turbine supply constraints can limit that route too. Those delays leave providers with demand they must serve using the capacity available to them now.

Usage limits are one way to manage that demand. Charging more is another. Both protect a provider from having to supply unlimited computing at a fixed price, but both also give customers a reason to look elsewhere. I would expect that pressure to increase as businesses become more dependent on these tools. A weekly allowance is harder to work around when employees need the service to finish ordinary work.

Getting more output from existing infrastructure gives providers another way to respond. That requires progress in the equipment and in the software deciding how the equipment is used. Buying faster chips alone leaves part of the opportunity untouched.

Software changes how much the hardware can do

Running a trained model to answer requests is called inference. The software serving those requests decides how to group them, allocate memory, and reuse work already performed. Poor memory management limits how much work a server can handle, even when the model itself has not changed.

The researchers behind vLLM demonstrated this in their 2023 PagedAttention paper. Their system managed the memory used during text generation more efficiently and delivered two to four times the throughput of the systems they compared it with, at similar latency. That result came from the serving software. It illustrates how additional capacity can come from changing what runs on the hardware already installed.

Other improvements reduce the resources needed for a particular request. Reusing calculations for repeated prompt content avoids processing the same material from scratch. Lower-precision storage of model weights, known as quantization, reduces memory requirements, although accuracy needs to be checked. A smaller model that handles a task reliably can also avoid the expense of sending every request to the largest available model.

New chips matter alongside that software work. NVIDIA already presents inference economics using tokens per second per megawatt and cost per million tokens. Those measures are closer to the question than tokens per square foot: how much output does a facility deliver within its power budget, and what does each unit cost? A denser rack needs enough electricity and cooling to operate. Fitting more equipment into the building only helps if those requirements can be met.

The comparison still needs a defined workload and an acceptable response time. A server answering short requests cannot be compared fairly with one handling long documents and extended reasoning. Research published in Joule in 2026 estimates efficiency gains across models, serving systems, and hardware, while also showing how longer reasoning increases energy use. Lower resource use per token can coexist with higher consumption per task.

The improvement has to reach the customer

There is no single multiplier that tells us how much more AI every existing data center could supply. The gains depend on what is installed, how it is being used, and what customers ask it to do. Some improvements arrive through software updates. Others require replacing equipment or changing cooling and power systems. The evidence does show that construction is only one source of additional usable capacity.

Better efficiency also gives a provider choices. It can serve more customers, expand allowances, reduce prices, or retain the savings as margin. Competition creates pressure to share the benefit, but a technical improvement does not guarantee a particular subscription price or allowance. Whether providers continue with subscriptions or change their billing methods, customers will judge how much work they can afford to complete.

Open-weight models and alternative providers belong in that discussion. They give a business other ways to run a workload, but the comparison needs to include hosting, support, reliability, and staff time. Moving away from a subscription can exchange a predictable bill for responsibilities the company previously left with its provider. The alternative earns its place when it performs the required work at an acceptable total cost.

That is where I think business owners need to keep their attention. More tokens are useful when they let the company finish more work reliably. A lower token price is less helpful if the system takes more attempts, produces answers employees must repair, or becomes unavailable when they need it. The provider’s efficiency matters to the owner through the cost and dependability of the result.

I expect continued pressure to make AI do more within the same customer budget. New facilities will add supply, and better chips will increase what that supply can deliver. Software improvements and better choices about which model handles the work can help in the meantime. A business deciding how much to depend on AI needs to know how much dependable work the same dollar will buy.