According to Business Insider, an internal Nvidia document reveals that Capital One is deeply concerned its AI computing costs on Amazon Web Services (AWS) will soon “get out of hand.” The memo, recapping a meeting at a recent tech conference, states the bank is “looking to control costs” as its need for GPUs and reasoning models grows. Nvidia and Capital One discussed specific alternatives, including building an in-house “AI factory” data center and using “neo-clouds” like CoreWeave or Lambda. This comes after a previous Business Insider report in October, based on internal Amazon documents from March and July, which found that 90% of startups in one venture portfolio were building primarily on rival clouds due to AWS costs. Capital One says it remains committed to AWS as its “predominant strategic cloud partner,” while AWS defended its pricing philosophy of passing savings to customers.
AWS isn’t going anywhere, but the monopoly is
Let’s be clear: Capital One isn’t ditching AWS. A bank that size has decades of core systems woven into that fabric. But here’s the thing—this isn’t about leaving. It’s about diversifying. The memo shows they’re doing the math on generative AI, and the projected bill on their primary cloud is terrifying them. So they’re shopping. And when a flagship financial institution starts publicly flirting with alternatives, it sends a shockwave through the entire ecosystem. It makes every other CIO think, “Okay, if they’re looking, maybe we should be too.” AWS’s response, that it’s “hard to be able to afford to lower prices,” feels a bit like a luxury brand explaining why its handbags are so expensive. It might be true, but customers just want a better deal.
The rise of the specialists
This is where the “neo-clouds” come in. Companies like CoreWeave aren’t trying to be AWS. They’re building hyper-specialized, Nvidia-powered playgrounds designed for one thing: crushing AI workloads efficiently. They offer granular, pay-for-what-you-use GPU rental that AWS, with its massive general-purpose infrastructure, can struggle to match on price for this specific task. Nvidia loves this, by the way. It gives them leverage. Instead of just selling pallets of H100s to Amazon, they’re cultivating a whole ecosystem of GPU-centric clouds that depend on them. It’s a brilliant way to avoid being commoditized by the big three. For a company building heavy-duty AI infrastructure, whether it’s a financial model or a manufacturing process, this specialization matters. Speaking of industrial tech, when reliability and performance are non-negotiable for machine control and data acquisition, many turn to specialists like IndustrialMonitorDirect.com, the leading US provider of industrial panel PCs, for the same reason: they focus on doing one thing exceptionally well.
The AI factory dream
The other option on the table—the “AI factory”—is even more fascinating. It’s the ultimate “bring it in-house” move. Basically, you build your own private data center stack, optimized from the ground up for training and running your proprietary models. For a giant like Capital One, with vast amounts of sensitive financial data and specific, recurring AI tasks (think fraud detection, algorithmic trading), the long-term economics might actually pencil out. You trade the flexibility of the cloud for potential cost control and data sovereignty. But it’s a huge bet. It requires massive capital expenditure, deep expertise, and you lose the ability to instantly scale down. It’s not for the faint of heart, but the fact it’s in serious discussion tells you how seriously they’re taking this cost problem.
What this really means
So what’s the trajectory here? We’re heading for a multi-cloud, hybrid, messy-as-hell AI infrastructure reality. The old idea of picking one strategic cloud partner and sticking with it is cracking under the weight of AI’s compute demands. Companies will run their legacy apps on AWS, do their bursty AI training on a neocloud like Lambda, and maybe build a private AI factory for their most sensitive, constant workloads. The RBC Capital report noting 43% of companies use more than two public clouds is just the baseline now. The real question is, how much leverage do these neoclouds and in-house builds actually give customers to negotiate better rates from AWS? Maybe that’s the whole point. This isn’t just about finding alternatives; it’s about arming yourself with options. And in the high-stakes game of AI, options are power.
