Two different things are being called "moving off the cloud," and conflating them produces bad predictions. One is enterprises pulling steady-state AI inference out of public cloud and back into private data centers they own or lease — Broadcom's 2026 survey found public cloud's share of production AI inference fell 15 percentage points in a year, from 56% to 41%. The other is a smaller set of workloads moving to the network edge — onto factory floors, cell towers, and delivery vehicles — because physics won't let a robotic arm wait for a round trip to any data center, public or private. Both get called "edge computing." Only one of them actually is.

"Off the cloud" is doing two jobs at once

Ask an IT leader why their company is moving AI workloads and you'll get one of two unrelated answers. The first is about money: renting GPUs at scale got expensive enough that owning the hardware pays for itself. The second is about distance: a defect-detection camera on an assembly line can't wait even a few hundred milliseconds for a verdict to come back from anywhere off-site. These are different problems with different fixes, and industry coverage routinely blends them into a single "edge computing is having a moment" story. Cost-driven moves tend to land in private cloud or colocated data centers — still centralized, just not rented from a hyperscaler. Latency-driven moves land in genuinely distributed locations, physically close to where the data is created. The rest of this explainer keeps them separate, because the businesses making each move, and the tradeoffs they're weighing, don't overlap much.

The repatriation move: enterprises leaving public cloud, but not for the edge

Broadcom's Private Cloud Outlook 2026, based on a Radius Tech survey of 1,800 senior IT decision-makers conducted in February and March 2026, found that 56% of enterprises now run or plan to run production AI inference on private cloud, up from near-parity with public cloud a year earlier. Sixty-two percent of respondents said they were very or extremely concerned about the infrastructure costs of running generative and agentic AI, and 97% reported some degree of cloud spending waste, with more than half estimating that waste at over a quarter of their budget. That's a cost and governance story, not a distance story. Private cloud in this context usually still means a company's own data center or a colocation facility — a building full of servers, just not Amazon's, Microsoft's, or Google's. Eighty-three percent of respondents told Broadcom they were considering workload repatriation, and half said they'd already moved some workloads back. None of this requires the AI to run any closer to the people using it. It requires the bill to be smaller and the auditors to be happier.

The actual edge move: latency that a private data center can't fix either

A separate shift is happening at the physical edge of networks — inside factories, retail stores, vehicles, and cell towers — and it's driven by something a bigger private data center doesn't solve: distance itself. "As the focus of AI shifts from training to inference, edge computing will be required to address the need for reduced latency and enhanced privacy," IDC research vice president Dave McCarthy said, predicting that this shift would enable business models "previously not possible with centralized infrastructure." ZEDEDA's 2026 Edge AI Survey, fielded by Censuswide among 600 IT and operations leaders across the US and Germany in late February 2026, found that only 24% of enterprises now rely primarily on centralized cloud or data center infrastructure for AI, while 47% run hybrid cloud-edge architectures. Computer vision and customer-experience use cases lead current production deployments at 45% each, followed by real-time monitoring and predictive maintenance. Gartner's own 2026 predictions put a number on where this goes: more than two-thirds of enterprises globally will deploy edge AI by 2029, up from about 10% in 2025. ZEDEDA sells edge orchestration software, so a survey showing edge adoption accelerating serves its own interest — worth weighing against the more neutral fact that Gartner, an analyst firm with no orchestration product to sell, landed on a similarly steep adoption curve independently.

The part that looks contradictory: hyperscaler spending is also exploding

If enterprises are leaving public cloud, the largest cloud providers don't seem to have noticed. Amazon, Microsoft, Alphabet, and Meta guided toward roughly $700 billion in combined 2026 capital expenditure, up from about $410 billion in 2025 — a jump of more than 70% in a single year, based on earnings-call figures reported by CNBC and analyzed by Futurum Group. That's not a market in retreat. Both things are true because the surveys and the capex numbers are measuring different customers. Enterprise inference — the steady, predictable workload a single company runs day after day for its own products — is the specific thing Broadcom's respondents said they're pulling back to private infrastructure. Hyperscaler capex is overwhelmingly aimed elsewhere: training frontier models, and building capacity to rent to AI labs like OpenAI and Anthropic, whose usage patterns are nothing like a bank running its own fraud-detection model around the clock. A company repatriating its own inference workload and a cloud provider building more data centers to sell capacity to a foundation-model company are not competing for the same square footage. One is a customer leaving; the other is a landlord building for different tenants.

What actually decides where a workload runs

For the repatriation decision, the math is mostly about utilization. GPU cloud providers price H100 GPU-hours at roughly $2 to $8 depending on configuration and provider, according to pricing data published by GMI Cloud in April 2026. A workload that runs continuously at high utilization crosses the break-even point for owned hardware fairly quickly; one that's bursty or unpredictable rarely does, which is why cloud remains the default for experimentation and spiky traffic even among companies repatriating their steady workloads. For the edge decision, utilization doesn't enter into it. A camera on a production line either gets an answer fast enough to stop a bad part before it ships, or it doesn't — no amount of cheaper GPU-hours in a private data center changes the speed of light. That's the distinction Broadcom's and ZEDEDA's numbers don't share: one is a spreadsheet problem, the other is a physics problem.

Edge computing has been "about to arrive" before

Coverage of this cycle carries an honest complication: this isn't the first time edge computing has been declared imminent. Trade press covering the 2026 push noted plainly that "edge computing has been 'about to arrive' for years" — a reference to earlier IoT-driven predictions that didn't reshape enterprise infrastructure on the timeline analysts expected. What's different this time is the specific driver. Earlier edge cycles were sold on general IoT device growth, a broad and slow-moving trend. This one is tied to a narrower, more measurable behavior: generative and agentic AI models being deployed for tasks with hard latency requirements, which is a smaller and more falsifiable claim than "the internet of things will need local compute eventually." Gartner and IDC's 2026 predictions could still overshoot, the way earlier edge forecasts did — but they're pinned to something concrete enough to check against in a year or two.

Which move actually applies to you

If the question is "should we stop renting so much cloud," the Broadcom data says look at utilization and governance requirements first, because that's what's actually driving the enterprises already doing it. If the question is "should our AI run physically closer to where the data is created," the answer depends entirely on whether a delay of even a few hundred milliseconds breaks the use case — and for most software products, it doesn't. The two moves will keep getting described as one trend because "AI is leaving the cloud" is a cleaner headline than "some inference is getting cheaper somewhere else while a smaller amount of inference is getting physically closer to sensors for unrelated reasons." The second sentence is the accurate one.

FAQ

Is edge computing the same thing as moving AI workloads to private cloud?

No. Private cloud repatriation moves inference from a public cloud provider to a company's own or leased data center, which is still centralized — just not rented from a hyperscaler. Edge computing moves compute physically close to where data is generated, such as a factory floor or a cell tower. Broadcom's 2026 survey measures the first trend; ZEDEDA's 2026 survey measures the second.

Why are enterprises moving AI inference off public cloud?

Mainly cost and governance. Broadcom's Private Cloud Outlook 2026, based on a Radius Tech survey of 1,800 IT leaders, found 62% of respondents very or extremely concerned about AI infrastructure costs, with 97% reporting some cloud spending waste and over half estimating that waste at more than a quarter of their budget.

If enterprises are leaving public cloud, why is hyperscaler spending still growing so fast?

Because the two data points describe different customers. Enterprise survey data reflects companies repatriating their own steady-state workloads, while hyperscaler capex — projected around $700 billion combined for 2026 across Amazon, Microsoft, Alphabet, and Meta, per CNBC and Futurum Group reporting on 2026 earnings calls — is largely aimed at model training and at selling capacity to AI labs like OpenAI and Anthropic, whose usage doesn't resemble a single company's routine inference load.

What kinds of AI workloads actually need to run at the edge?

Ones where a delay of even a few hundred milliseconds breaks the application — industrial defect detection, autonomous vehicle perception, real-time equipment monitoring. ZEDEDA's 2026 survey found computer vision and customer-experience use cases leading current production edge deployments, each cited by 45% of enterprises with active deployments.

How fast is edge AI adoption actually growing?

Gartner predicts more than two-thirds of enterprises globally will deploy edge AI by 2029, up from about 10% in 2025. ZEDEDA's February 2026 survey of 600 IT and operations leaders found only 24% of enterprises still rely primarily on centralized cloud infrastructure for AI, with 47% already running hybrid cloud-edge architectures.

Is it cheaper to run AI inference on your own hardware than to rent cloud GPUs?

It depends on utilization. GPU cloud providers price H100 GPU-hours at roughly $2 to $8 depending on configuration, per pricing data GMI Cloud published in April 2026. A workload running continuously at high utilization can reach a break-even point for owned hardware; a bursty or unpredictable one usually doesn't, which is why cloud remains standard for experimentation even at companies repatriating other workloads.