/ Insights

Why AI inference belongs close to users, not in a distant campus

Illustration of a small dark data center on a city street corner at dusk

Gartner projected that inference would overtake training in AI-optimized cloud infrastructure spending in 2026. Inference runs continuously, where the work happens. That is why it belongs at the edge.

Gartner projected in October 2025 that, within AI-optimized infrastructure as a service, spending on inference would overtake spending on training in 2026. It expected worldwide end-user spending on AI-optimized infrastructure as a service to reach $37.5 billion in 2026, up from $18.3 billion in 2025, with inference spending reaching $20.6 billion, up from $9.2 billion in 2025. That would make inference 55 percent of AI-optimized IaaS spending in 2026, and more than 65 percent in 2029. These are forecasts for one category of cloud spending, not a count of all AI spending. Gartner’s reasoning is that training involves intensive, large-scale compute during model development and ongoing updates, while inference happens continuously, powering real-time applications such as chatbots, recommendation engines, and fraud detection.

Training and inference are different jobs. Training is a periodic, centralized, batch job. It happens in one place, at massive scale, and the location barely matters. You can train a model in a desert campus and ship the weights anywhere. Inference is continuous and distributed. It runs every time a user, a sensor, or an agent asks the model a question. It has to run where the work happens, because the work never stops.

The economics flip with that shift. Training cost is dominated by compute: how many GPUs, for how long. Inference cost is dominated by how often the model runs and how fast it answers. A model that is trained once and run a million times a day spends its life in inference, not training. The infrastructure that matters is not the campus that trained the model. It is the capacity that runs it, all day, close to the people and systems using it.

The market is moving that way. IDC forecasts worldwide AI infrastructure spending of $497 billion in 2026, and projects it will surpass $1 trillion in 2029. IDC’s FutureScape 2026 predictions put the edge shift this way: by 2030, half of enterprise AI inference workloads will be processed locally on endpoints or edge nodes. IDC expects a large share of that inference work to move out of the central cloud and toward the edge, because that is where the users and the data are.

Why close matters comes down to four things. First, latency. A round trip to a distant campus adds delay a real-time application cannot afford. A customer-facing agent, a vision system on a line, or a clinical documentation tool does not wait well. Second, data gravity. Moving data to a distant campus costs bandwidth and time, and some data should not leave the metro it was created in. Third, resilience. A local inference site keeps working when the wide-area link is degraded. Fourth, control. The workload runs in a facility you can visit, in a metro you can name, under terms you negotiated.

That is why the buying question is different for inference. Do not ask how big the campus is. Ask how close the capacity is to the users, whether the power is already there, and when the hall can be energized. A 100 MW campus three years away does not help a workload that needs 5 MW this year. A 5 MW hall in the metro, on power that is already delivered, does.

The inference layer is capacity close to users, not 100 MW campuses far away. That is the product. Edge AI is a global provider building 1 to 50 MW build-to-suit data centers for AI inference. We have several hundred sites identified, screened for power, fiber, and access, with 7 GW of power identified across them.

Training built the models. Inference is where the value is created, and it happens close to users. The capacity has to be there too.

Sources

All insights

Talk to a specialist.

Tell us the market, the megawatts and the date. We reply within one business day.