It goes without saying that enterprises need an AI strategy in order to thrive in this new business era. The question now is what that strategy should look like to capture the full benefits of AI and minimize the risks.
In a recent blog post, my colleague Ruth Faller introduced the concept of private AI, where an organization builds or fine-tunes its own AI models, hosts those models inside a protected environment, and feeds proprietary data into the models. Ruth argued that a private AI approach is essential for any organization that wants to move past basic use cases and scale up a holistic enterprise-class AI strategy.
Enterprise IT leaders increasingly recognize that private AI offers the control and protection they need, while public AI services such as ChatGPT could place them at risk.
Keep your data private
Using public AI is risky for enterprises because of the potential for data leakage. Any training or inference data you feed into a public AI service could be accessed and stored by the service provider. This means that your proprietary data will no longer be under your control: Copies of your data could be leaked or sold to anyone, and you’d be powerless to stop it from happening. In contrast, when you build private AI models, no one outside your business has access to your models or the data you feed into them.
As enterprises increasingly seek to harness AI for competitive advantage, they often need a non-public environment where they can retain full control over their sensitive and proprietary data.”- Dave McCarthy, Research Vice President, Cloud and Edge Infrastructure Services, IDC
As businesses look to embed data privacy into their AI strategies, they must also distinguish between classical AI models—which have existed in some form for decades now—and generative AI models, which are relatively new. GenAI may be getting all the hype these days, but classical AI use cases such as predictive analytics are still important. These different varieties of AI have different infrastructure requirements and challenges. Building private AI models can be helpful for both GenAI and classical AI use cases, but the benefits will show up in different ways.
Let’s consider an example: One common enterprise use case for GenAI is employees using chatbots as virtual personal assistants. Employees use these bots to help with everyday tasks such as writing, brainstorming and research. When they do this, they’re giving the underlying AI models access to all the same data they access, including sensitive proprietary data.
In the very early days of generative AI, there were several high-profile instances of heedless users causing data leakage that put their companies at risk. For instance, Samsung discovered that some of its software engineers had effectively leaked confidential code by sending it to ChatGPT when searching for a bug fix. In response to this and other incidents, some companies took the proactive step of banning their employees from accessing ChatGPT at work. This was an early sign that enterprise leaders were beginning to recognize the need for private AI models.
Reduce your regulatory risk
The question of how AI will be regulated globally is something that’s still being decided today. We can assume that some jurisdictions will be stricter than others when it comes to setting requirements for data sovereignty and privacy, but enterprises need to be ready for anything. They must ensure they can meet the strictest compliance requirements across their global operations, and they can’t do that when they’re using public AI.
A compliance strategy for AI needs to incorporate both training and inference workloads. Even if you do everything right from a training perspective—using private data centers to make sure your data stays within the right borders—public AI could still put you at risk when the time comes to do inference.
Public AI uses the internet to move data. From the moment your data hits the public internet, it’s no longer under your control. The service provider for the model could create a copy of your data and store that copy anywhere they want. This means that your business may no longer be complying with data sovereignty requirements, and you wouldn’t know it until well after the fact.
This is in contrast to private AI models, which use only private, dedicated network connections to move data. And since you’re in complete control over your own private AI models, you can perform the necessary due diligence to ensure that those models aren’t moving or storing data anywhere they shouldn’t be.
Also, GenAI is based on foundational models trained on data crawled from the public internet. Even if you’re fine-tuning a public model using your proprietary data, that model has already been pre-trained on publicly available data, some of which may be copyrighted material. This means that you have no control over what data the model uses to service your requests.
The public model could leverage data sets that your company isn’t legally allowed to access. Even though you didn’t choose to access those data sets, your company could still be held accountable for the fact they were used on your behalf. This could even make the company vulnerable to future legal action. By using only private AI models, you can fully remove this risk.
Optimize your costs and performance
AI infrastructure should be distributed in order to balance the requirements of compute-intensive training workloads and latency-sensitive inference workloads. However, deploying these distributed components while balancing both costs and performance can be challenging, especially when you rely on public AI infrastructure.
Businesses can accrue high inference costs by using public AI models for GenAI use cases. These costs may not seem significant on a per-use basis, but they could very easily grow out of control if all employees from across the organization are allowed to use public LLMs as much as they want. This is yet another reason it may be best for enterprises to avoid using public AI. The cost could be even higher when a corporation builds customer-facing applications (such as a customer service chatbot) with public AI.
To get the AI infrastructure they need while keeping costs low, enterprises often turn to the public cloud. However, this might hurt performance for their AI workloads. Even cloud workloads have to live somewhere; in the case of public clouds, they’re typically hosted in regions that offer the cheapest energy. (This benefit isn’t always passed on to cloud customers, since public cloud often costs more than expected.)
When companies host their AI workloads in the public cloud, they won’t be able to ensure proximity between data sources and compute locations. In turn, this means they won’t be able to effectively support latency-sensitive inference workloads.
Hosting models in private environments can help enterprises simultaneously avoid the high costs and network latency that could come from using public AI infrastructure. This is why it’s the best choice for enterprises looking to ensure predictable costs as they scale their AI strategies.
It is true that network latency isn’t that big of a concern for GenAI, since compute latency is typically the main source of delay for these use cases. However, some classical AI use cases—such as high-frequency trading—are extremely sensitive to network latency and can thus benefit from being hosted in private environments to ensure proximity.
When enterprises adopt private AI, it doesn’t mean that they can’t incorporate public cloud services at all. Rather, it means they should do so as part of a hybrid infrastructure that allows them to minimize the potential downsides. An ideal hybrid infrastructure would meet the needs of different AI workloads by offering:
- Private compute infrastructure at the digital edge to support latency-sensitive inference workloads.
- A cloud adjacent data architecture to support workloads that are better suited for public cloud. This architecture places data near the cloud without putting it in the cloud. This allows for multicloud access on demand, without the potential cost and performance drawbacks.
Start building your private AI infrastructure today
The value of a private AI strategy is clear. Now, the question becomes how to build the infrastructure needed to execute that strategy.
The future of enterprise IT will be built around the hybrid multicloud model, and AI is no exception. AI will happen in private environments, in the public cloud and everywhere in between. This is why it’s so important for enterprises to work with a digital infrastructure partner that can bring together these different environments on a single interconnected platform.
At Equinix, we work with leading ecosystem partners to help our customers access the infrastructure they need to do private AI right, whether that means privately hosted GPUs, cloud adjacent data storage or low-latency multicloud access. We also offer global reach and on-demand private interconnection capabilities, so that you don’t have to rely on the public internet to move your most sensitive data into AI models.
To learn more about how hybrid infrastructure can help you succeed with private AI and other digital transformation imperatives, read the leader’s guide to hybrid infrastructure.
 Emily Dreibelbis, Samsung Software Engineers Busted for Pasting Proprietary Code Into ChatGPT, PCMag, April 7, 2023.
 Aaron Mok, Amazon, Apple, and 12 other major companies that have restricted employees from using ChatGPT, Business Insider, July 11, 2023.