When the biggest vendor in the room says your sentence, you have won the argument and lost the differentiator.

The sentence

Buried in a product announcement from VMware Explore is the clearest statement of our own product thesis we have read from anyone else:

> "Bring the model to the data, not the data to the model."

Broadcom's framing is that when the AI wave hit in 2023 the conventional path pointed outward to public cloud, and they took the opposite position: enterprise AI had to be built where enterprise data already lives. They now say the market has validated that bet.

We agree, obviously. We have been selling it since before it was comfortable to say in a Hungarian boardroom. So the interesting question is not whether they are right. It is what it means that they are saying it.

The three forces they name

Their account of why workloads are moving is worth repeating, because none of the three reasons is about ideology:

• Hardware supply. AI demand has created a hardware supply crisis, driven specifically by the cost of DRAM. In their words, wait a week and a purchase order may be cancelled and reissued at a higher price.
• Security. Frontier models are accelerating the discovery of vulnerabilities, so that "attacks can happen in hours not weeks". This is the same force the Hungarian security profession has put on next week's ITBN agenda.
• Cost and IP. Customers are worried about token costs and about protecting intellectual property, and are bringing AI workloads back on-premises, closer to their data.

Note what is missing from that list: nobody's values. These are procurement reasons.

Three things we have to say about those numbers

We are quoting a survey that supports us, so we owe you the caveats before the conclusion.

The company selling the answer commissioned the question. Broadcom sells VMware Cloud Foundation. The article containing these figures ends by recommending VMware Cloud Foundation. The methodology is disclosed and the sample is large — a blind survey of 1,800 senior IT leaders across eight countries, now in its second year — but a vendor-commissioned survey finding in favour of that vendor's product category is evidence, not proof.

The 14 September article misquotes its own survey. It renders the headline finding as "Last year, 56 percent of enterprises now run or plan to run inference and production AI on private cloud", which mixes two different statistics into one ungrammatical sentence. The June report states them properly and separately, and that is the version we quote above.

There is a good survey that disagrees. IDC, in December 2025, asked 304 application-platform decision makers and found AI application deployment split almost exactly evenly between private and public cloud, with 63.6% expecting the ratio to shift toward public cloud. The IDC analyst who ran it said he does not see a major push to repatriate. That finding is older and smaller than Broadcom's, but it is real and it is not ours to ignore.

What this costs us

Here is the part that is uncomfortable to publish.

If Broadcom ships VMware Private AI Cloud to every enterprise with a VMware estate — and in Hungary that is a very large share of the mid-market and above — then "we run AI privately" stops being a differentiator. It becomes a checkbox that the incumbent virtualisation vendor ticks by default, sold by every reseller already in the building.

Their VMware AI Factory claims to take a customer from bare metal to a first served model in hours rather than weeks, and VCF AI Model Validation covers more than 150 open-source and commercial models. That is a serious product, built by a company with more engineers than we have employees.

So we should stop saying "private" as though it were the product. What is actually left, once private is table stakes, is narrower and more defensible:

• Somebody operates it. A platform you must staff is not the same offer as a platform that is run for you, with an SLA, by people who answer the phone in your timezone.
• Open weights, and they are yours. Not merely running in your building, but a model you can keep, move and inspect, with no licence that evaporates.
• A documented exit. The measure of lock-in is not how good the onboarding is. It is what a migration away costs, and whether anyone has written it down before you need it.

That is a smaller claim than we were making in 2025. It is also the one that survives Broadcom agreeing with us.