The Shift Is Already Measured in Dollars
IDC says global spending on edge computing will hit 61 billion in 2025. That isn't a forecast from a vendor pitch deck; it's a hard number from a research firm that tracks purchase orders. The money isn't going to bigger warehouses in northern Virginia or Frankfurt. It's going to factories, ports, and cell towers where latency kills a deal faster than price.
Why This Matters Right Now
The stakes are simple and immediate. Cloud providers built their margins on centralization. Now, AI inference is splitting apart. A vision model running quality inspection doesn't need to leave the plant floor. The cost of sending every frame to a distant GPU cluster is now larger than the cost of running a small rack locally. Companies don't want to move data; they want answers.
Liquid cooling is helping, but it isn't the only change. Rack density is climbing because new AI chips draw 700 watts or more. That heat has to go somewhere. Direct-to-chip cooling isn't a luxury anymore; it's a requirement for keeping high-density edge nodes online.
What the Numbers Look Like on the Ground
According to STL Partners, liquid cooling is quickly becoming a design standard rather than an experiment. The reason isn't environmental branding. It's physics. At 700 watts per chip, air can't carry heat away fast enough. Direct-to-chip loops do. They also shrink the physical footprint, which matters when you're installing at a factory that wasn't built for a data center.
McKinsey's 2026 report on colocation notes that hyperscaler capacity expanded sharply in 2025, but the growth is uneven. Some regions have power constraints, not space constraints. Edge nodes sidestep both by using smaller, distributed footprints.
The Hard Part Remains Maintenance
There's a catch. Edge nodes don't have the 24-hour crews that a hyperscaler can afford. A failure at a remote site can mean hours of downtime while a technician drives out. Companies are addressing this with predictive maintenance and redundant links, but it's not solved. The technology works; the operations don't always.
It's also not clear this will hold. IDC's 61 billion assumes sustained enterprise investment. If AI inference costs drop faster than expected at the central cloud, some edge deployments will be abandoned. The economics are tight. They work when latency is the bottleneck, not when it's merely convenient.
What's Next at the Edge
Edge AI isn't replacing cloud; it's splitting the workload. Training stays central. Inference moves out. The result is a hybrid mesh that doesn't look like the old hub-and-spoke model. For operators in Asia, that mesh is already being built.
McKinsey colocation research shows the infrastructure race isn't finished; it's just changing lanes.
Read more desk coverage.
Cooling, Capacity, and the Cost Curve
The economics only work if the cost curve keeps falling. Nvidia and its rivals have shipped chips that deliver more inference per dollar than the parts they replace. That means the argument for local inference gets stronger every quarter. A plant running 10,000 inspections a day no longer needs to negotiate a contract with a cloud provider or accept queue times. It can buy a rack and know the price per inference in advance.
What This Means for Builders
The teams most affected are platform engineers at retailers, logistics firms, and telecom operators. They spent 2024 building private clouds. They're now being asked to build inference fleets. The tools are converging, but the training on orchestration still lags behind the hardware rollout. Expect the next twelve months to sort the adopters from the dabblers.
The View From Asia
The shift to edge AI favours markets with dense manufacturing bases. Indonesia, Vietnam, and India are the clearest beneficiaries. Cheap power, skilled labour, and policy support point the same direction. The money quoted by IDC is global, but the near-term pull is regional.
For the industry in Southeast Asia, the race isn't to the biggest data centre. It's to the farthest factory gate.
Power, Politics, and Permission
Edge deployments need permits. Grid connections in Indonesia's manufacturing belt move slowly. Each site needs an environmental review that can drag for months. The fastest operators pre-file permits and negotiate power contracts before ordering hardware.
The Tooling Gap
Kubernetes works at the edge, but it's heavy. Engineers are testing lighter frameworks like KubeEdge and open-source projects from LF Edge. Hardware vendors sell bundles that include the software stack pre-installed. The promise is simpler operations, though the reality varies by vendor.
What Stays Central
Training a model still costs more than a car. Those runs stay in hyperscaler campuses with thousands of GPUs. The edge handles inference only. That split will define architecture decisions for years.
Looking Ahead
The next wave of deployments lands in early 2026. Watch factory gates, not server rooms.
A Note on Measurement
IDC's number includes software licenses, consulting fees, and hardware. It counts only direct enterprise spending. It misses the indirect costs operators face when they add backup links or hire technicians who can drive to remote sites. The real cost is higher, though the exact premium varies by country.
What the Vendors Don't Advertise
Liquid cooling works well in dense environments. In a factory with high humidity or dusty air, the same loops can fail faster than air-cooled systems. Maintenance contracts often exclude environmental damage. Operators must read the exclusions carefully before signing multi-year agreements.
The Regional Difference
Not every market moves at the same pace. Southeast Asia has manufacturing density but slower grid buildouts. The gap between hardware delivery and grid connection is already stretching some timelines. The companies that plan ahead are the ones whose nodes go live first.
The Supply Chain Factor
Semiconductor shortages still slow deliveries. Lead times for edge GPUs stretch to eight weeks or more. Operators are stocking spare units and extending warranties. The companies that buy ahead and install in batches avoid the worst delays.
Latency Budgets Tighten
Real-time vision systems allow roughly five milliseconds of compute time. Cloud round-trips add dozens of milliseconds. The gap grows with distance. Edge racks sit inside the five-millisecond budget. That is the practical line that defines every deployment decision.