We are already in an era where intelligence comes with a price tag. Every answer, inference, agent, and AI workflow carries a real infrastructure cost, and that cost isn't determined by the model alone. It rises or falls with compute efficiency, utilization, power, cooling, networking, latency, and workload architecture. The same intelligence can have radically different economics depending on where, when, and how it is delivered.

Back in March 2023, running a GPT-4-class model cost roughly $20 per million tokens. Today, that same intelligence costs less than $0.50, a fall of nearly 98 percent in three years. Yet the average enterprise AI budget has grown nearly six times over the same period, from $1.2 million in 2024 to approximately $7 million in 2026.

Intelligence has never been cheaper. AI has never cost more.

Unit price is falling while total spend rises. That is the paradox defining AI infrastructure in 2026, and it is why the industry’s most important conversation is shifting from how much capacity we can build to how much usable intelligence that capacity can deliver, per dollar, per watt, and per second.

What follows is the case for that shift: the economics behind it, the chips accelerating it, and the power constraint quietly capping it.

From capacity to productivity

The first phase of the AI infrastructure race was necessarily about capacity. Compute was scarce, demand was accelerating, and hyperscalers, enterprises, and governments moved quickly to secure GPUs, power, and data center capacity.

That capacity remains essential, but it doesn't by itself determine AI economics. A GPU cluster running at high utilization has very different unit economics from the same cluster sitting partially idle. The question is evolving from how much compute is available to how productively that compute converts into usable intelligence.

Earlier this year, a widely circulated figure claimed DeepSeek had trained a frontier model for just $5.6 million. That number actually referred to pre-training compute for an earlier model, not the total cost of the frontier model itself, a distinction lost in the headlines. Weeks later, Microsoft, Meta, Amazon, and Google all reaffirmed or increased their AI capex guidance rather than cutting it.

This is Jevons’ paradox in action: when efficiency gains lower the price of a resource, total consumption tends to rise, not fall. Every dollar saved per token becomes the reason to run ten more tokens in parallel, not the reason to build less.

The price of a GPU is not the cost of intelligence

Nvidia’s own Blackwell-vs-Hopper analysis illustrates the distinction. While newer-generation infrastructure can carry a higher GPU-hour cost, its delivered cost per million tokens can be substantially lower because of gains in throughput, memory, interconnect, software optimization, and energy efficiency.

Buyers evaluating AI infrastructure on hardware price or theoretical compute performance alone are therefore optimizing for an incomplete metric.

The true cost of intelligence is determined by the entire stack. Model efficiency determines how much computation an output requires. Compute architecture determines how efficiently accelerators, memory, and interconnects execute the workload. Utilization determines how much installed infrastructure is doing productive work rather than sitting idle. Infrastructure efficiency determines how effectively power and cooling support high-density compute. And workload placement determines the latency, network, and data-movement economics of where intelligence is produced and consumed.

Optimizing any one layer in isolation can produce misleading economics. A highly efficient model running on poorly utilized infrastructure can still be expensive. The fastest accelerator can lose its advantage if constrained by memory, networking, or power. And low-cost compute may not represent the lowest-cost outcome if latency and data movement undermine the economics of the workload.

Cost of intelligence is therefore a stack-wide metric, not a hardware metric.

Why power now matters more, not less

Each new generation of accelerators delivers more AI output from the underlying compute. In theory, those efficiency gains should ease infrastructure pressure. Nvidia's Blackwell Ultra already delivers 1.5x the throughput of the original Blackwell generation; its successor, the Vera Rubin platform confirmed for late 2026, promises up to a 10x reduction in inference cost at scale.

In practice, the opposite dynamic is emerging. Average data center rack density has gone from roughly 7kW in 2021 to about 27kW today, with megawatt-scale rack designs already on chipmakers' roadmaps for 2027. Every efficiency gain at the chip level has been reinvested into denser racks, not lower power draw.

In many markets, power availability is consequently emerging as a more consequential constraint than access to compute itself. The AI infrastructure race is increasingly becoming a race for power.

Global data center electricity demand is projected to reach approximately 132GW in 2026, climbing toward 290GW by 2030, according to Gartner. In several major markets, securing new grid interconnection capacity now takes three to four years longer than it takes to physically construct a facility. Time-to-power has become as critical a planning variable as time-to-deployment.

Where intelligence runs is an economic decision

Cost of intelligence is also changing how the industry thinks about the geography of compute. Large-scale training still benefits from concentrated compute in hyperscale environments. But latency-sensitive enterprise and real-time applications create a different equation, where moving data back and forth to distant infrastructure introduces latency, bandwidth costs, and unnecessary data movement.

The infrastructure question is no longer simply hyperscale versus Edge. It is: where should each workload run to deliver the right combination of cost, latency, resilience, and governance? The emerging answer is architecture that is increasingly distributed, with large-scale compute at the core, complemented by regional and edge capacity closer to where intelligence is actually consumed.

Can India compete on the cost of intelligence?

This is where India has an opportunity to compete differently.

The country’s AI ambition should not be measured only by how many GPUs it deploys or how many megawatts of data center capacity it builds. A more consequential measure may be how economically it can convert those resources into usable intelligence.

India enters this phase with structural advantages: expanding renewable capacity, competitive power economics, strong engineering talent, a rapidly growing digital economy, and AI demand that will increasingly be distributed across enterprises, industries, and geographies.

Much of India’s next generation of AI infrastructure is also being built now. That creates an opportunity to design for high-density power, advanced cooling, efficient utilization, and distributed workload placement from the outset rather than retrofitting infrastructure designed for an earlier generation of compute.

The industry has spent the past few years celebrating GPUs deployed and megawatts commissioned. Those milestones still matter, but they measure what goes into AI infrastructure rather than what comes out of it.

GPUs, FLOPS, and megawatts will remain important. But ultimately, they are inputs. The next measure of AI infrastructure leadership will be how much trusted, usable intelligence those inputs can deliver per unit of capital, power, and time.

That is the shift from capacity to the cost of intelligence.