Inference will account for roughly two-thirds of all AI compute in 2026, up from half this year and a third in 2023, according to Deloitte's Technology, Media and Telecommunications Predictions report published Nov. 18. The shift does not mean smaller data centers or cheaper chips at the edge. Deloitte expects most of that compute to stay inside large facilities running power-hungry accelerators.
The common assumption has been that once models are trained, running them is cheap and can happen on phones, laptops, and small edge boxes. Deloitte argues the opposite. The market for inference-optimized chips is forecast to pass $50 billion in 2026, up from over $20 billion this year. But chips inside large data centers and on-prem enterprise systems still represent the bulk of the spending, with data center capital expenditure alone estimated at $400 billion to $450 billion for 2026.
The stakes land in budgets and power grids, not just in research labs. An AI data center in 2026 will cost as much as some national infrastructure programs, and the compute load is rising four to five times per year through 2030 even as individual chips get more efficient. Anyone planning capacity for 2027 needs to read the inference shift as a reason to build more, not less.
Inference grows, but the racks get heavier
Inference is the act of asking a trained model a question and getting an answer. It sounds lighter than training, and a single request is. A large language model summarizing a document uses a tiny fraction of the compute needed to train that model. But billions of users and enterprise workers are making those requests, and doing so daily, so the total adds up.
The bigger surprise is what happens inside the model before it answers. Post-training techniques, such as fine-tuning, pruning, quantization, distillation, and reinforcement learning, use an estimated 30 times the compute of the original foundational training run. Test-time scaling, the reasoning step that lets a model think through a question before responding, uses more than 100 times the compute of a simple inference like summarizing an email. Both techniques are now standard practice across AI companies.
The result is that inference-optimized does not mean inference-cheap. One recent product built for inference pre-fill avoids high-bandwidth memory and uses GDDR7 instead, yet each rack draws 370 kW, nearly three times the power density of the training version from the same supplier. Deloitte's read is that optimization here targets throughput and cost per token, not a lighter footprint.
Where the money goes
Deloitte puts AI data center capital expenditure at $400 billion to $450 billion globally in 2026, with $250 billion to $300 billion of that going to chips and the rest to land, construction, power, and permitting. The figure rises to a projected $1 trillion by 2028, with AI chips alone exceeding $400 billion that year.
The chip breakdown tells the same story. The market for inference-optimized chips, including application-specific integrated circuits from Meta, Google, Amazon, Intel, AMD, Qualcomm, Groq, SambaNova, Cerebras, and Graphcore, is expected to reach $50 billion or more in 2026. Those chips sit in data centers and on-prem enterprise systems, not in phones. Deloitte notes that hundreds of millions of PCs and smartphones with on-device AI accelerators were sold in 2025, and that edge processing will keep growing. The edge share of total compute, on this forecast, stays small.
Geography shifts too. Pre-training a 100-trillion-parameter model once required a single campus, because a failed component or a slow handoff between processors could wipe out weeks of work. Inference and post-training are more interruptible, so the load can move between data centers across a country or across the world. Deloitte expects a range from gigawatt-scale campuses to smaller inferencing sites closer to metro areas, plus sovereign AI capacity in each region and enterprise edge deployments as part of hybrid cloud setups.
Power and the grid
More compute spread across more sites does not lower total demand. It changes where the load lands and who manages it. Because post-training and test-time scaling can be paused and resumed, AI operators can take part in demand response programs, shifting tasks between locations or slowing processor clocks during peak hours. Deloitte estimates that this flexibility could let large new data centers support grid reliability rather than strain it, and that distributing workloads helps spread electrical demand more evenly across regions.
The forecast is not a guarantee. Deloitte itself frames the $300 billion to $400 billion ramp to $1 trillion by 2028 as directionally realistic, not certain, and notes that compute demand growth for initial training has already slowed. The open question is whether the new scaling techniques keep consuming power at the current rate or whether efficiency gains catch up. Deloitte's answer is that demand is rising faster than efficiency, at four to five times per year through 2030, so the gap widens.
What this means for capacity planning
For cloud and edge operators, the takeaway is blunt. Inference is the growth story, but it is a growth story for large facilities, metro inferencing sites, and on-prem AI factories, not for a wave of cheap edge deployments replacing data centers. Planning for 2027 means securing power, land, and permits on the assumption that the load keeps climbing, and treating edge AI as a complement rather than a substitute.
For enterprise buyers, the practical point is procurement. Inference-optimized chips cost less per token but still arrive in racks that need 370 kW and liquid cooling. Budgets that assume a lighter footprint will be short. The Reuters coverage of AI infrastructure spending has tracked the same capex curve, and the numbers are not moving down.
The report closes on a note that is hard to argue with. Optimized for inference does not mean less power. It means more compute, in more places, at higher density. The data center is not going away. It is getting bigger.
One more data point worth keeping in view: Deloitte notes that compute demand is rising four to five times per year through 2030, even as individual chips grow more efficient under Moore's Law. That is the gap that makes the capex curve plausible. If efficiency ever outpaced demand, the $1 trillion figure would not hold, and edge deployments would look more attractive than they do today. For now, the edge is a place to run a model, not a place to replace the facility behind it. Industry coverage from Reuters on AI infrastructure spending has tracked the same curve, and the numbers are not moving down.
For more on cloud and edge coverage, see our cloud and edge computing category.