The previous article: hyperscalers bet on capture, funded by cloud AI revenue. But the cloud providers’ expansion thesis has an assumption: enterprises will use cloud AI. This article asks: is that assumption right?
This question is worth answering because if the most valuable enterprises build their own local AI (to prevent data outflow), the cloud providers’ data centers are underutilized, and the previous article’s bet loses.
The main answer: partly right, partly wrong. Most enterprises use the cloud (convenient, no self-maintenance, no utilization risk), but data-sensitive enterprises (IP, defense, regulated) tend to build their own. Public cloud primary production inference 56%→41%, private cloud 56%. The expansion thesis is right for “most enterprises” and wrong for “data-sensitive enterprises.”
1. Question: will enterprises use cloud AI?
The conclusion of the previous article: hyperscalers bet on capture, funded by cloud AI revenue.
But the cloud providers’ expansion thesis assumes “enterprises will use cloud AI.” If that assumption is wrong, the cloud providers’ data centers are underutilized, and the bet loses.
This article asks: will enterprises use cloud AI? Which will, which won’t? Why? This determines whether the cloud providers’ expansion can recover its cost. This article only handles “enterprise cloud vs on-prem choice”; it does not handle “whether AI is actually useful” (handled in the first article).
2. Evaluation framework: cloud vs on-prem depends on data sensitivity and utilization
Whether an enterprise chooses cloud AI or builds its own local AI depends on two things: data sensitivity and the economics of utilization.
Low data sensitivity + low utilization: use the cloud. The cloud is convenient, needs no self-maintenance, and carries no utilization risk. Experimental, bursty, non-sensitive workloads are economical on the cloud.
High data sensitivity (IP, defense, customer data, regulated): build your own. Sending data to a third-party cloud (especially US hyperscalers) carries data outflow risk. These enterprises build local AI; the data stays on-site.
Why do data-sensitive enterprises build their own? Because their data is a core asset. Semiconductor process data, defense-sensitive data, financial customer data — sending it to a third-party cloud (especially US hyperscalers) carries outflow risk. These enterprises prefer to build their own and keep the data in-house.
High utilization: building your own is more economical. Owned GPUs break even against cloud rental in about 4 months once utilization holds above 20%[4]. So steady, high-utilization inference workloads are cheaper on-prem than on the cloud. Conversely, low-utilization workloads (experimental, bursty) are not economical to build your own; the cloud is more reasonable.
So the enterprise’s choice is not a binary “cloud vs on-prem” but “stratifying by workload” — non-sensitive on the cloud, sensitive on-prem.
So: the cloud providers’ expansion thesis is right for “low data sensitivity, low utilization” enterprises and wrong for “data-sensitive” enterprises.
3. Comparison: the actual cloud vs on-prem split
2026 data:
| Metric | Value | Source |
|---|---|---|
| Public cloud as primary production inference (share of enterprises) | 56%→41% (falling) | LLM Configurator[1] |
| Private cloud (running or planning) | 56% | VMware[2] |
| Sovereign cloud market 2026 | $80B (YoY +35.6%) | NeuralTrust[3] |
| Owned GPU break-even (utilization >20%) | about 4 months | Gain America[4] |
Three observations. First, public cloud primary inference is falling: 56%→41%, down 15 percentage points in a year. Second, private cloud (on-prem) is rising: 56% of enterprises run or plan to run production inference on private cloud. Third, sovereign cloud is growing: $80B in 2026, YoY +35.6%, as enterprises move regulated AI workloads off US hyperscalers.
These three observations point to the same thing: enterprises are “stratifying” — the low-data-sensitivity ones stay on the cloud, the data-sensitive ones repatriate.
Note: the 56% private cloud is “running or planning,” not all on-prem. So “enterprises all go on-prem” is an exaggeration, but “data-sensitive enterprises tend to build their own” is real.
So the “enterprises go all-in on cloud” narrative has wobbled in 2026. Data-sensitive enterprises are repatriating.
4. Judgment: the cloud providers’ expansion thesis is half right
For “most enterprises” (low data sensitivity, low utilization): right, they use the cloud. Experimental, bursty, non-sensitive workloads are convenient and need no self-maintenance on the cloud.
For “data-sensitive enterprises” (IP, defense, regulated): wrong, they build their own. These enterprises are “the most valuable customers” — their AI workloads are large, their data is sensitive, and they are willing to pay. When they build their own, the cloud providers’ data centers lose a chunk of demand.
So the cloud providers’ data centers have demand from “most enterprises” but may be underutilized for “data-sensitive enterprises.”
For the cloud providers, this means: their expansion has demand from “most enterprises” but may be underutilized for “the most valuable data-sensitive enterprises.” If data-sensitive enterprises keep building their own, the cloud providers’ data center utilization drops, and the previous article’s bet becomes harder to win.
This also explains why sovereign cloud is growing — governments (the most data-sensitive customers) move AI workloads to sovereign cloud, not to US hyperscalers. Sovereign cloud reaching $80B in 2026, YoY +35.6%, is the embodiment of this trend. Practical recommendation for cloud providers: expansion should target “most enterprises’” needs (low data sensitivity, low utilization), while offering “data stays on-site” options (private cloud, sovereign cloud) to data-sensitive enterprises. Betting only on “all enterprises go to the cloud” loses to on-prem.
Here, distinguish two things. Hardware demand (NVIDIA/memory) is robust — cloud and on-prem both need GPUs and memory, so the memory price increase in the second article is unaffected. But the cloud providers’ cloud revenue is at risk — when enterprises build their own, the cloud providers’ data centers are underutilized, and the previous article’s bet loses part.
Applicable conditions: enterprise data sensitivity stratification, utilization differences.
Limitations:
- Most enterprises still use the cloud (the 56% private cloud is “running or planning,” not all on-prem). So the cloud providers’ base demand remains.
- Building your own has a barrier (buy GPUs, maintain, hire talent). Small enterprises can’t build their own and still use the cloud.
Ending
Back to the question: will enterprises use cloud AI? The answer: most will (low data sensitivity, low utilization); data-sensitive enterprises won’t (they build their own). The cloud providers’ expansion thesis is half right — right for most enterprises, wrong for the most valuable data-sensitive enterprises.
This also echoes the previous article: for the cloud providers’ bet to win, it must rely on “most enterprises’” needs, not “all enterprises.”
Applicable conditions: enterprise data sensitivity stratification, utilization differences. Limitations: most enterprises still use the cloud; building your own has a barrier.
Next article: the real bottleneck is not GPUs, it is electricity — grid interconnection 5 years, transformers 2.5 years; the “crazy expansion” is actually a paper boom.