Building and evaluating new artificial intelligence (AI) technologies is critical for long-term technological success, but it’s not the whole picture. At Red Hat, we also spend time creating reusable patterns and refining the way AI models can be consumed. We do this so we, and our customers, can adopt AI with confidence.
We're also making sure that we don't lose sight of the practical side of AI. For example, how in the world are we going to give everyone private access to AI models while minimizing costs? How can we harness the value and scale of the “next big thing”?
One key answer: Models-as-a-Service (MaaS).
Not everyone wants to be an AI expert, and let’s be honest, no organization needs everyone building their own models. For now, the goal should be enabling as many people as possible to use the AI we already have. That’s where MaaS comes in.
Here’s how we landed on MaaS. Our users tend to fall into two groups:
- AI enthusiasts—the AI builders, developers, dabblers and pros who live and breathe AI
- Everyone else—those who just want to use AI without diving deep into the technology
MaaS is an approach to providing AI models as consumable resources—in the form of API endpoints—to your organization. It’s your IT team, or your AI platform engineers, serving models behind an API gateway to empower developers and, by proxy, all users so they can use AI in their day-to-day tasks.
Think about it this way—the AI giants don’t give you access to their hardware. You get access to their app or API endpoints. They provide MaaS. The only difference is that in this situation, instead of renting their services, you run your own MaaS to maintain control over costs, access and speed of innovation.
Sounds pretty great, right?
It is! We’ve been running MaaS internally at Red Hat for about a year now. Here’s what we noticed:
More innovation with reduced costs
Every time a new model comes out, hundreds of Red Hatters want to deploy it immediately. MaaS helps us deploy the model once and preserve our budget! Gone are the days of sourcing 10, 20, 50+ GPUs because 10, 20, 50+ people want to try a new model.
It’s a win-win. Our developers can try new models and focus on building new tools without breaking the bank.
Speed to innovation
We are able to test any new model hitting the market on our own timeline. Remember when DeepSeek disrupted the AI market? We had DeepSeek R1 running and available for everyone shortly after its release. Same for Granite, Llama 4, Phi—you get the picture.
Privacy and security
Sensitive data requires full and careful control. With MaaS, you become your own private AI provider able to closely safeguard your digital assets. You're not obligated to use public-facing API endpoints. In fact, many of our customers run their own models in fully air-gapped data centers.
Enterprise use
The API gateway powering MaaS gives you the scalability you need to reach every associate, the flexibility you need to keep pace with innovation and the enhanced security and observability tools you need to efficiently deploy AI models on your terms.
Reduced costs, again
MaaS reduces costs by directly using a shared resources model. You’ll find that fewer GPUs are required to achieve the same result, and GPU utilization metrics will improve. As models get better and smaller over time, you’ll get even more from this footprint. You could even use open source large language model (LLM) compression tools to balance model performance and size to suit your own requirements. In short, MaaS helps you to optimize your footprint and your models for maximum gain.
MaaS is about creating the right foundation while preparing for the future. Take AI agents for example. Agents aren’t one-shot, question-and-answer applications. They’ll keep searching for the answer, if you let them. What does that mean? Tokens. So many tokens. If you want scalability and the ability to more accurately project your costs, you should consider running MaaS in house.
AI is here to stay, and it’s time to get practical about cost, speed of innovation and privacy. Models-as-a-Service is a promising solution and one Red Hat is committing to. If these are also priorities for you, MaaS is worth considering.
Check out this interactive demo where Parasol, a fictitious insurance company, powers three AI applications using MaaS! For those AI enthusiasts out there, here’s the MaaS GitHub repository. Lastly, if you are attending Red Hat Summit this year, make sure you register for the How to become the hero of your artificial intelligence story and LLM-as-a-Service for enterprise: Building a customer large language model platform sessions to learn more.
关于作者
Karl Eklund is a Principal Architect aligning customer goals to solutions provided by the open source community and commercial vendors within the Red Hat OpenShift Data Science platform. Prior to joining Red Hat, Karl advised technology leaders on enterprise data and technology strategies and built machine learning models across multiple academic disciplines.
