Milind Wagle, Chief Information Officer, Equinix

While the potential benefits of artificial intelligence (AI) speak for themselves, enterprises need a measured, strategic approach to pursue those benefits without putting their valuable intellectual property at risk. This is why many businesses are beginning to build their own, AI models, host those models on private infrastructure and use only proprietary data sets to train them. This concept is known as private AI.

Many enterprises now recognise that when they feed sensitive data into public AI services such as ChatGPT, that data becomes part of the model’s data. This means that the data could be accessed by anyone who uses the model in the future. Open AI’s own FAQ advises users not to share any sensitive information with ChatGPT as there is no way to delete specific prompts from a user’s history.

With private AI, businesses can extract insights from their data without compromising their privacy or control over that data. This article highlights four actions that should be incorporated into strategy to help businesses succeed with private AI.

Ensure private AI is right for you

It is important to realise that not all businesses will be successful with private AI, particularly if they do not start with a clearly defined vision of what success looks like for their specific situation. For businesses in highly regulated industries such as healthcare and financial services, the benefits of private AI are obvious. They must avoid any practices that could compromise their sensitive data, making private AI a natural fit.

Businesses in unregulated industries may still be able to benefit from private AI, but the value proposition may not always be as clear. Such businesses need to consider the trade-offs – both the risk of data leakage, and the cost and flexibility impact of implementing AI on public infrastructure. Some companies gravitate toward public cloud because they view it as an easy and cost-effective way to access scalable compute infrastructure for their AI model’s demand. However, accessing public cloud compute is often more expensive and difficult than expected, largely due to high data egress fees.

If the perceived benefits of public cloud infrastructure are not enough to outweigh the potential risks, then the business is well suited for private AI.

Incorporate data management into strategy

Given the rapid advancements in AI technology over the past several years, it is worth taking a step back to consider one fundamental fact: your AI models can only be as good as the data you feed into them. This is why effective data management is essential for the success of private AI.

There is also a need to deliver the right data to the appropriate destinations without delay. This can be challenging due to the highly distributed nature of AI infrastructure. This entails:

  • The collection of data from all applications – typically in a hybrid multi cloud architecture – to feed training models.
  • The deployment of inference workloads at the edge (locations where end users interact with AI models) to ensure proximity between data sources and processing locations. This is essential because inference workloads are very sensitive to latency, and distance is the primary driver of network latency.
  • The deployment of training workloads on core infrastructure to provide the substantial compute capacity that workloads demand.
  • Flexible, high-performance networking between diverse workloads to enable rapid and reliable movement of data from the source to various processing locations.

One optimal way to build an AI-ready data architecture is by using cloud adjacent storage. This enables the integration of public cloud services into the private AI strategy of a business while mitigating the potential risks, costs and complexity. It is like having the best of both worlds for AI infrastructure: you are close enough to the cloud to access services when needed, but you are also able to keep your authoritative storage environment separate from the cloud.

This approach means that a business can maintain complete control over its data, using it when and how it wants, without worrying about it being leaked through a public AI model or getting locked into a particular cloud. Ensuring this level of control over data is a hallmark of an effective private AI strategy.

Consider compute needs

The explosive growth of AI has led to increased demand for powerful graphic processor unit (GPU) hardware. While manufacturers are ramping up production to meet this demand, supply shortfalls are likely to persist in the foreseeable future. Limited hardware availability could prevent businesses from fully realising their private AI goals. However, there are ways to avoid this bottleneck and still get the required compute capacity.

While many people consider “GPUs” as synonymous with “AI hardware”, this is not necessarily true. Although GPUs are needed to support the most demanding training workloads, readily available central processing units (CPUs) can meet smaller inference workloads. In fact, businesses could use a bare metal as a service solution such as Equinix Metal® to help deploy CPUs on demand without high upfront costs.

Even for workloads that do require GPUs, there are options beyond self-deployment and management of hardware (after waiting months for delivery). For instance, Equinix recently announced a fully managed private cloud service in partnership with NVIDIA. This service expedites the acquisition of advanced AI infrastructure, packaged with the required colocation, networking and managed services to host and operate the infrastructure. The solution offers flexibility similar to a public cloud solution, while allowing control over data in a private environment.

Plan for sustainability and efficiency

Many people are rightly concerned that the current rush towards AI will derail the sustainability progress that some enterprises have made in recent years. It is true that AI workloads, and training workloads in particular, can be very energy intensive. To limit the carbon impact of these workloads, it is essential to execute them as efficiently as possible.

For instance, new liquid cooling technology for data centres is inherently more efficient than traditional air cooling. It will play an essential role in cooling high-density workloads such as AI in an energy-efficient manner. At Equinix, we have been testing liquid cooling extensively for years now and have begun using it to support production workloads.

In addition, it is important to consider how placement of workloads can impact sustainability. It is crucial to position workloads in locations where they can draw the least carbon-intensive energy from the local grid. One way to achieve this is by working with a digital infrastructure partner committed to investing in renewable energy.

At Equinix, we are well on our way to achieving our goal of 100 per cent renewable energy coverage globally by 2030. To help create greener grids, we have invested in power purchase agreements that support renewable energy projects across the world. Our customers can benefit from this increased renewable coverage as they work to meet their goal of implementing AI in a sustainable manner.