This allows you to create virtual connections to other clouds or new edge locations as the need arises. The ideal data architecture for AI or any other advanced data-driven use case should pair scalable storage solutions with agile, programmatic interconnection capabilities, as shown in the diagram below. Finally, once models have been trained, you’ll want to move them back to the edge to support latency-sensitive inference workloads. You may choose to leave these workloads in the core and use private compute resources or a Bare Metal as a Service offering to process them. However, your data engineering and tuning workloads may include sensitive data that you don’t want to expose to the cloud. This is why you need distributed digital infrastructure that can quickly move your data from the edge to the cloud and back.įor instance, you may choose to do model training in the cloud to take advantage of LLM services from a particular provider. In the case of AI, different workloads have different requirements, and should therefore be hosted in different locations. In that previous blog, I also described four data motion patterns that allow enterprises to take advantage of cloud services on demand while also minimizing the impact of data egress fees.īuild a data architecture that accounts for the unique requirements of distributed AI workloads Then, you should be able to move that data wherever it needs to go to support different use cases-either upstream to multiple cloud providers or downstream back to edge locations. You should be able to aggregate data into your core from a variety of sources at the digital edge. An authoritative data core is not a specific location, but rather a logical layer at the center of your data architecture. In a previous blog post, I described what an authoritative data core is, and why it’s so important for enterprises to maintain their data cores outside the cloud. These fees make something that should be a quick and simple business decision-like leaving one cloud provider for another that offers better services-into a complex, time-consuming and expensive endeavor. If you build your data architecture directly in the cloud, you’ll face onerous data egress fees that place artificial constraints on your business. There are many valid reasons you might want to use cloud services as part of your data architecture, but the risk is that you may lose custody of your data if you aren’t careful about how you access those services. Utilize the financial model that best meets your needs-OPEX or CAPEX.

Analyze it by feeding it into the appropriate tools.

Recover your data in the aftermath of an outage or disaster.

Secure your data against a variety of threats.

Audit your data and the hardware on which it resides.

When you’re the custodian of your data, you’re free to: This means that you need a place within your data architecture where you can store your data while maintaining complete control and ownership over it. The most important concept for enterprises to understand when it comes to building their data architecture for AI is that of data custody. In this blog, I’ll describe what it means to get your data architecture right: building an authoritative data core that allows you to move data from the edge to the cloud and back, without ever having to give up control over that data. As LLMs go from a novelty to an ingrained part of day-to-day enterprise operations, the drawbacks of having the wrong data architecture will only continue to grow. Looking toward the future, we have every reason to believe this trend will continue. Thus far, it’s been the enterprises with the most flexible, distributed data architectures that have been able to make the most of generative AI services, while those that depend on traditional, static data architectures have started to fall behind. In this way, LLMs are not so different from many other enterprise IT services that came before them. This means that how you build your data architecture is the single factor that will ultimately determine the success or failure of your AI initiatives. LLMs and other AI applications may seem magical to the uninitiated, but in reality, the results you get are dependent on the quality and availability of the data you feed into those models. Things seem likely to ramp up even faster now that OpenAI has announced an enterprise version of ChatGPT. From the enterprise side, businesses have been scrambling to be the first to capture the competitive advantages that LLMs offer. In the first two months after its launch, more than 100 million users had tried ChatGPT, making it the fastest-growing consumer app ever at the time.

Large language models (LLMs) have been available to support mainstream AI use cases for less than a year now, and we’ve already seen how much they’ve impacted our lives.

Data architect