We’ll dive into three major trend publications from 2024, starting with NVIDIA’s State of AI in Financial Services 2024. We’ll see if any trends appear multiple times and think about what to bring with us into 2025. But before we get to that, there are two things you need to hear. One is about building AI models. The other is about what just happened at CES 2025.
It’s 2025. Start building your own models.
It’s time. And it’s getting easier every day. I’m more serious about this than I was about reposting a promoted post about why an efficient $2B institution is way cooler than a typical $20B institution. This is about the magic, the magic that makes an institution what it is. Do not buy. Do not partner. Build.
Building machine learning, deep learning, and generative AI solutions is difficult. And then all of the sudden, it wasn’t. At least not as difficult, more like a challenge. A fun challenge. The tech hasn’t changed, but the packaging and perception just did. Last week, our ships rose with the tide, even if you didn’t notice. Building models is about to become as common as answering a call or merging a pull request. We’re no longer afraid.
You do not want to be caught flat-footed here. And by that, I mean you should think three times before signing that multi-year contract with a vendor claiming they have the best application of the best models with the best team. They don’t, and they can’t. Because they don’t have your data, at least not the most important data.
And they never will, nor should they.
Close the marketplace portal. Ping your development manager, email your trusted IT service provider. It’s time to start building your own future.
Take Our Money, NVIDIA—Again.
We might as well set up automatic payments at this point. CES 2025 was a banger for Team Green, and no one really saw it coming, because no one thought it was necessary. Oops.
Desire Athow, Editor-in-Chief at TechRadar, summed up NVIDIA’s unexpected innovation perfectly: “It’s time for Jensen’s law: it takes 100 months to get equal AI performance for 1/25th of the cost. NVIDIA’s desktop supercomputer is probably the greatest revolution in tech hardware since the IBM PC.”
I couldn’t agree more.
The announcement of the DIGITS AI supercomputer did two things: it made me immediately fill out the “Notify Me” form, and it finally gave the AI development lifecycle a physical face that enterprise leaders can love (quantify). I’m certain the second part was intentional. Jensen Huang knows exactly how to stop the executive hand-wringing over in-house AI development: boxes on desks. I’ll take two, please.
I love boxes on desks. Let me tell you why.
Even if you end up with an alternative platform, Digits has demystified what a day in the life of an AI developer will look like. It’s an iPhone moment. When you think of an AI developer’s workstation, it’s no longer an abstract concept. You can actually see it: the same old laptop, but now paired with what looks like an external hard drive circa 2005. Not exactly intimidating, no hidden cost monster lurking in some far-off server. It’s sleek, it’s local, priced right and it’s sitting right there next to Gary’s laptop. Oh, look! Gary put a sticker on it.
Previously, explaining the variable cost of an enterprise AI developer’s hardware and software toolset was a bit fuzzy, at best. Mastering an ML/DL development stack was no efficient task for most businesses due to the limited compute of common workplace laptops (a $5K Mac Studio being one way around that) and slow office networks (still not getting around the 2GB gRPC max message size). Everything was done in the cloud or in the datacenter.
And now, almost overnight, that complexity has been stripped away. Thank you, NVIDIA.
Take my money.
Ok, let’s look back at those NVIDIA 2024 trends
Here’s the link to the full report. We’ll look at two of the four trends and two charts that helped inform them.
“As the complexity and size of AI models have grown, data concerns have eclipsed recruiting experts as the biggest challenge to achieving AI objectives.”
Generally speaking, there was a significant increase in articles and knowledge sharing during 2024. That’s why recruiting experts is on the wane right now. We probably won’t see hiring pick up again until 2026. C-suites are having normalized discussions about AI and that’s allowing more freedom to address difficult problems around data usefulness and data governance.
The new complexity lies in putting data and services together, in the same place, at the same time. That’s a challenge and not always possible, or cost effective. At the end of the day, all things being solved completely, we’re still dealing with an asymptotic curve of technical desire against the wall of regulatory compliance. We’re going to need new techniques and this is why it’s time to start building your own solutions.
“Financial services firms are exploring new ways to build and deploy trustworthy, secure AI, including federated learning and confidential computing.”
When we focus on data-driven AI concerns it means we need to focus on trust and security; which is to say we need verifiability. We’re all gravy as long as we can verify what goes in is secured and verify what comes out is trustworthy. Easier said than done. More than likely we’ll have to build our own verifiability, even if we’re using someone else’s models and data.
How can you verify the accuracy of your vendor’s nifty new fraud detection product and not give them all of your data? You can’t. So, let’s not. That means a fresh mix of responsibilities are assigned to the CISO and some new quality assurance processes should be created. All in all, we’re going to need a bigger boat, a party boat. A federated learning, confidential computing party boat!
Let’s assume we have the budget, and agree that the most trustworthy and secure solutions are the ones we build—or fine tune—for ourselves. Let’s take a look at what type of workloads we were utilizing in 2024 that might inform what we should be building in 2025.
We need to build in the NLP, LLM, and GenAI space, but not because they trend lower. It’s where we take action on our own analytics and where differentiation will happen. You’re going to be good at it, or you’re not.
When the results of quality prompts are no longer helpful then it’s time to seek a refined LLM. When you have to conditionally approve an auto loan until proper documentation is verified by a human then you’re probably in the market for a sweet NLP solution. Doing either of these means you’re working in the GenAI space.
Wrapping up
I expect to see a large jump in utilization of GenAI workflows in 2025 with a very sharp increase of development in 2026, many writing in-house GenAI solutions. We’ll start to see an increase in development of techniques like retrieval-augmented generation (RAG), federated learning, and confidential computing. In the meantime we’ll be eagerly awaiting the unboxing of our own NVIDIA Digits box.
HiFin Technology, Inc. is an NVIDIA Inception partner. This content is not a paid promotion. For more information about HiFin’s NVIDIA services and solutions you can email info@hifintech.com.
