NVIDIA and Dell Push Enterprise AI Out of the Cloud

The story at Dell Technologies World this week was not really about new servers. It was about where enterprise AI spending is heading, and who controls the infrastructure that runs it. Jensen Huang and Michael Dell made the same argument in complementary ways: the cloud was the right place to experiment with AI, but it is the wrong place to run AI at scale. The cost math has turned. For companies running heavy, predictable AI workloads, owning the compute is now substantially cheaper than renting it. The numbers are not marginal. Dell's own research found that on-premises AI can be up to 62% more cost-effective than public cloud for running large language models, and independent analysis puts the advantage at up to 18 times cheaper per million queries when comparing owned infrastructure against cloud API pricing. On-premises now reaches a cost breakeven in under four months for high-utilization workloads, down from 12 to 18 months just two years ago. This is what makes the hardware announcement significant for non-technical operators. NVIDIA's new Vera Rubin platform, which ships in the second half of this year, cuts the cost per AI query to one-tenth of what the previous Blackwell generation cost. That is not a small improvement. It means that if you were running 100 AI queries at a certain cost last year, you can now run 1,000 for the same price, or run the same 100 for a fraction of the budget. Dell is packaging this hardware into what it calls the AI Factory: a fully integrated system of compute, networking, and storage that arrives pre-engineered rather than requiring a company to assemble components itself. The company has already deployed this to more than 5,000 enterprise customers. AI server sales went from $10 billion in early 2025 to $25 billion today, with $50 billion projected for the full year. The customer list on stage tells you which industries are moving fastest. Lilly is running AI for drug research at scale. Honeywell moved workloads from cloud back to on-premises for industrial automation and digital twins. Samsung is using it for chip design. Hudson River Trading is running AI-driven financial research on dedicated Dell hardware. The security argument is the one that will matter most to regulated businesses. NVIDIA is building confidential computing directly into the hardware at full-rack scale, meaning that AI models and the data they process stay encrypted even while the system is running. This matters for any company dealing with patient records, financial data, proprietary manufacturing processes, or government contracts. Running that kind of work through a shared cloud environment carries real governance and compliance risk that on-premises infrastructure eliminates. The practical limit of all this is that it requires real infrastructure. The new Vera Rubin hardware uses liquid cooling, requires proper data center facilities, and arrives in the second half of 2026. Most mid-sized businesses are not going to buy a full rack of this on day one. The realistic path for most operators is hybrid: use cloud for experiments and variable workloads, move steady high-volume inference on-premises once the workload is predictable enough to justify the capital commitment. The breakeven is now fast enough that the decision is worth making earlier than most finance teams assume. The direction is clear. The cost of running AI in-house is dropping faster than anyone expected. The cost of running it in the cloud is rising as usage scales. The companies that get ahead of that equation in the next 18 months will have a structural cost advantage over the ones that wait.