Sponsored by Dell AI Factory with NVIDIA
At the dawn of agentic AI, off-the-shelf models seemed to eliminate the complexity that was coming at enterprise leaders at breakneck speed. Now that enterprise AI dexterity has matured, it’s become clear that generic models make for generic agents, and that better outcomes will result from smaller, smarter models trained on your data and on AI infrastructure you own.
Dr. Jon Krohn and Frank Basso of Lightning AI break down this mindset shift in terms that will resonate with leaders: ROI that shows up in business outcomes and reduced token usage. They also detail how the Dell AI Factory with NVIDIA powers the principles through full-stack, integrated infrastructure that improves utilization.
Learn how the Dell AI Factory with NVIDIA can power your way to AI here: https://dell.com/YourWayToAI
Explore other Dell AI Factory with Nvidia stories:
Transcript:
The smartest path to better tokenomics isn’t haggling over the price of a token. It’s building smarter, smaller models.
Lightning AI is the AI-native cloud built to allow organizations to own their own AI. There’s been this evolution where initially we had Gen-AI models and you went to whichever was the best Gen-AI frontier provider. But what we’re realizing now is that that can be expensive, you’re locked into a particular system, and they can change things on you.
So the shift now, very recently, is a shift to enterprises owning their own AI. It’s historically been hard in a pre-Lightning world to be able to train your own AI models.
The companies in the lead on training AI models are doing three things differently. The first thing is evaluation. So even before they start training a model, they have a good sense of what the workflow is that they want to impact. The second thing is that they treat training as a loop. The third thing is owning the AI infrastructure themselves.
As an enterprise buyer, it’s always a concern to derive the best value for every dollar you spend. You have limited budgets and you have to get the most out of it. With traditional hyperscale cloud providers, you are paying for a large amount of equipment, or a large amount of instances can be very expensive if they’re sitting there and running when you’re not using them.
As one of the first providers to deploy Dell’s GB300 rack-scale systems, this gives us an advantage in the marketplace for our scale-out larger customers that need large amounts of high-performance super-pod capacity on demand or in a dedicated function for either training or other AI workloads.
Generic models make generic agents. When you have an agent fine-tuned to your own data, it doesn’t need to spend a whole bunch of time in the background thinking, which takes time, which uses tokens, which costs money. Instead, it can quickly get to the answer because it doesn’t need to reason through it.
The economic wins are usually orders of magnitude as opposed to a few percentage points, and that’s because we get to stack a bunch of positive effects together. We’re getting better accuracy on our models, which means that we need fewer human interactions downstream. We have a smaller model, which means that we’re saving on those per-token costs, and we’re getting our utilization up because smaller models make better use of the GPUs that we have available.
Enterprises don’t just want access to AI. They want predictability and performance. They want to own their own data, and they want the best value for every dollar spent. The Dell AI Factory with NVIDIA makes it so easy for Lightning AI to get hardware deployed for ourselves, which has downstream benefits for our customers. So by providing us with all of the pieces as one working kit – the liquid cooling, the GPUs, the storage, the networking – all of these things being available as one package, as opposed to something that we have to figure out piece by piece, allows us to serve our customers faster: Lightning faster.
