I saw a study from MIT the other day. It said that ninety-five percent of enterprise generative AI projects were judged failures. Ninety-five percent. Now, you’ve probably already heard the excuse from the industry chorus: “They didn’t buy enough (of my) infrastructure.” The claim goes something like this: if you didn’t refresh your servers, your storage, your networking stack, if you weren’t running the latest GPU farm from their catalog, then of course your AI project was doomed to fail. And always with that tired refrain: “it didn’t scale.”
But what does that even mean? What does scale mean if you can’t define it, if you can’t connect it to the outcomes your enterprise actually cares about? Let me offer a different perspective. MIT’s finding—that enterprise AI fails ninety-five percent of the time—has very little to do with vendor choice, infrastructure stacks, or racks of blinking lights. It has everything to do with two things: people and time.
First, people. Human capital. Do you have the deep computer science expertise to build a foundational model? To train it, tune it, refine it until it becomes a competitive advantage? Or are you better off licensing a model built by someone who has already done it better, at global scale? Think about it. Enterprises no longer build their own payroll systems. They don’t roll their own storage or fabricate their own servers. They don’t write their own CRM. Why? Because those are contextual technologies, not core to the mission. The same is true for foundational AI models. Most enterprises will get more leverage using a mature and hardened one than trying to build them from scratch.
Second, time. Or more specifically, latency in preparation. AI at scale doesn’t just happen. If you wanted to run large-scale AI in your enterprise today, the truth is you needed to have started preparing seven years ago. Seven years ago you should have been constructing 100-megawatt data centers and signing massive power contracts. You should have ordered the Solar Turbines Titan 250/350s with their seven-year lead-times. Three years ago, you should have been planning for 250- to 500-kilowatt racks, immersive cooling, and cutting-edge heat dissipation architectures. You should have architected your networks to deliver 400, 800, 1600 gigabits per rack, or more. If you did that three to seven years ago, then today you’d be ready. You’d have the core infrastructure, the power, the cooling, and the capacity to make the required infrastructure investment in enterprise AI real.
But here’s the truth—most enterprises didn’t plan for AI seven years ago. And that’s okay. Because the other path forward doesn’t require waiting 84 months for a new data hall or rebuilding your entire IT stack from scratch. The other path is the cloud. With the cloud, you eliminate the latency of infrastructure lead times. Your AI project can start tomorrow, not years from now if you try to build everything yourself, and not months from now if you waste time replicating petabytes of data before you can even begin. The path forward is software: connecting the data you already have—all of it, not just the 10 percent in databases, but the other ninety percent locked away in file and object stores. It is unifying that data across SaaS applications, cloud services, and on-prem systems, and reasoning across it immediately. Whether you’re running AI analytics platforms like Palantir, Glean, or Databricks that thrive on structured and semi-structured data and need to be able to consume unstructured data where it lives, or combining licensed models from Anthropic or OpenAI with internal datastores, or linking on-premises repositories directly to accelerated computing provisioned instantly in AWS, Azure, GCP, or OCI, you can rent, test, and train now. You can experiment across clouds, learn what works best for your business, and scale when you’re ready—without delay.
That is the real opportunity. It doesn’t depend on a forklift upgrade of your infrastructure. It doesn’t depend on re-architecting a local power grid. It depends on seeing your enterprise for what it already is: a treasure chest of data. Connect it. Unify it. Reason across it. That is where competitive advantage lies.
So the question isn’t whether you planned seven years ago. The question is whether you’re willing to act today. Will you keep chasing the myth that AI requires you to rebuild your entire infrastructure stack? Or will you seize the tools now available—tools that let you run on-premises, in the cloud, across every system you already own—and turn your data into something far more powerful than a sunk cost?
The future doesn’t belong to those waiting years on power grids and turbine deliveries or multiple quarters for GPUs and network infrastructure. It belongs to those who can harness the data they already have, wherever it lives, and turn it into insight, intelligence, and action. That’s the future worth building, and it is achievable today – it’s just a Terraform script away.
