Public Cloud vs. Private Cloud for AI

Future-proofing your AI architecture with Equinix

Kaladhar Voruganti
Public Cloud vs. Private Cloud for AI

Now that generative AI has gone mainstream, enterprises are investing heavily in AI and developing their AI strategies. In fact, IDC predicts that annual spending on AI—including generative AI—will go from $175.9 billion in 2023 to $509.1 billion in 2027, a compound annual growth rate (CAGR) of 30.4%.[1] As the generative AI industry expands, organizations are increasingly using AI models created by others instead of building their own AI models from scratch. Many key players in the industry have developed large language models (LLMs)—both commercial and open source—that are available to enterprises. Some are general, while some are designed for specific industries or use cases.

Taking advantage of existing AI models helps companies jumpstart their AI initiatives. Many organizations are now starting with foundation models—that is, models trained on a broad range of generalized and unlabeled data—and customizing them to meet their needs. Here are three common ways companies are building on top of or customizing existing models:

  • Finetuning: Finetuning involves retraining an existing model using an organization’s own data.
  • Retrieval augmented generation (RAG): The RAG approach can be used to create better input prompts to improve the accuracy of results from generative AI queries. With this approach, instead of retraining the model, you provide more relevant context data as part of the input prompt.
  • AI agents: An AI agent is software that monitors the mood of the end user and dynamically helps to create generative AI prompts. For example, many GPS systems already know where a user is likely to travel at certain times of the day based on past travel history.

As organizations explore opportunities to launch new AI initiatives, they need to think about designing the right infrastructure to support them. Enterprises that are planning to use existing AI models can either move their private data to where the AI model is (in the public cloud) or move the AI model to where their data exists (in a private cloud). These two approaches both offer advantages, and there’s no one-size fits all method. A recent Enterprise Technology Research survey showed that 32% of enterprises use only a public cloud approach, 32% use only a private cloud approach, and 36% use both.[2] For the latter group, the choice between public and private cloud for AI is based on the use case.

Public cloud for AI

When organizations adopt the public cloud approach for AI, it involves moving their data to the model. In other words, companies take their data, upload it to the public cloud and then leverage AI models in the cloud.

There are several reasons why an enterprise might choose public cloud for AI:

  • Lack of AI talent in house: If you don’t have the AI skills in your company, public foundation models can help you fast-track your AI initiative.
  • Low barrier to entry: Using the public cloud for AI—by leveraging GPUs in the cloud—can help you get your AI initiative up and running quickly. The models are already out there being used by lots of companies.
  • Higher quality models: It’s often, but not always, true that publicly available AI models are currently better performing, more accurate and higher quality than those organizations develop themselves. Unless you have a lot of AI talent, commercial models like Bard, ChatGPT and so forth, can offer superior results.

Private cloud for AI

Companies using a private cloud approach to AI move the model to their data that’s stored in private infrastructure. In other words, organizations take a foundation model developed somewhere else and download it to run on premises where their private data exists, so there’s no need to upload data to the public cloud.

Again, there can be important reasons for choosing a private cloud for AI:

  • Data privacy and security: In certain use cases, and in highly regulated industries, organizations need a higher level of data privacy and security. By using private cloud, you can host the model inside the company firewall and ensure compliance with data regulations.
  • Model lineage control: Moving the AI model into a private cloud also strengthens your ability to manage, monitor and extend it.
  • Lower cost at scale: Cost-wise, it can be useful to test out a couple of generative AI models in public clouds as you’re evaluating AI use cases. However, if you’re using many generative AI applications across many departments, leveraging AI models in the clouds via APIs will quickly become expensive. Similarly, data egress costs for moving query results out of the clouds will also become expensive.
  • Smaller, faster models: AI models in a private cloud can be smaller and faster, depending on the use case. There’s a bifurcation in the foundation model space. You can find very large foundation models that can satisfy multiple use cases, or you can choose AI models for their specific use case that are smaller in size, and thus consume fewer infrastructure resources.

Choosing infrastructure to support your AI strategy

As the world’s digital infrastructure company®, Equinix is playing a strategic role in helping organizations with both public and private cloud approaches to AI. We offer a range of solutions to support your generative AI initiatives.

Public cloud for AI at Equinix

If you choose public cloud for AI, you can still store data that’s generated outside the cloud, as well as legacy data sources, near the cloud. A cloud-neutral facility like Equinix supports a multicloud posture and delivers a more predictable storage cost model. With cloud adjacent storage at Equinix, you can keep that data at Equinix, either in colocation or Equinix Metal®, our Bare Metal as a Service (BMaaS) solution, and ingress it to clouds as needed. This way, you’re not locked into a single cloud provider. Equinix partners with all the major storage vendors, allowing you to both deploy your storage hardware in colocation or consume it as a service.

Cloud adjacent storage gives you greater control of your data while ensuring flexibility and access to all major clouds via cloud on-ramps. You can keep both your raw data as well as vector databases (when using the RAG AI inferencing approach) at Equinix and then access AI models in public clouds via APIs. Cloud adjacent storage also gives you the flexibility to switch between different AI models based on performance, accuracy and cost. And you can use foundation AI models from different providers (e.g., a weather model, finance model, traffic model, etc.) quickly and securely. Furthermore, in the era of GPU shortages, you can switch between GPU as-a-service providers by ingressing your data from a neutral cloud-adjacent location like Equinix. Equinix data centers have high-speed and low-latency connectivity to multiple public cloud data centers.

If you’re using public cloud for AI, you’ll also need an efficient multicloud connectivity solution. In many cases, companies need to dynamically access data from external sources in real time as they’re constructing the AI query response. Fabric Cloud Router—the Equinix multicloud networking solution—helps you to access external data sources in a highly performant, cost effective and secure manner. And Equinix Fabric® virtual routing can help you connect to multiple cloud providers over a private network with built-in resiliency.

Private cloud for AI at Equinix

If you’re concerned about data privacy and the lineage of your AI model, Equinix provides a private AI approach that offers data privacy, predictable costs and lower latency performance. In the private cloud model, AI compute infrastructure (CPUs, GPUs, FPGAs) and storage would be installed at Equinix, and foundation models are then accessed from public clouds or AI model marketplaces (e.g., HuggingFace) and subsequently deployed on the Equinix AI infrastructure.

Increasingly, organizations are worried about the lineage of public cloud foundation models because of pending lawsuits between data providers and model creators. Using a private AI infrastructure provides you with greater control over what data is in the AI model.

Generative AI solutions also have substantial power and cooling requirements. Some companies are deploying private AI infrastructure at Equinix instead of in their private data centers because they can’t handle the power and cooling requirements of AI, which can exceed 30KW per rack. Furthermore, Equinix offers high-bandwidth, low-latency, secure access to data sources in the clouds.

Doing private AI at Equinix also gives companies access to network service providers that bring in traffic from users and devices at the edge, as well as to a robust ecosystem of 10,000+ enterprises where they can exchange data using Equinix Fabric.

Our new Private AI solution with NVIDIA DGX allows you to create a fully managed private cloud using NVIDIA DGX AI supercomputing infrastructure to build and run custom generative AI models.

In conclusion, you can future-proof your AI architecture by keeping your data at a cloud-adjacent location like Equinix. This will give you the flexibility to pursue either public or private cloud AI architectures. If you’re looking for a place to fast-track your generative AI initiatives, Equinix can help with a range of solutions to support your AI strategy. To learn more about how organizations are thinking about infrastructure for AI, download the Equinix 2023 Global Tech Trends Survey.


[1] Rick Villars, Karen Massey, Mike Glennon, Eileen Smith, Rasmus Andsbjerg, Peter Rutten, Ritu Jyoti, Jason Bremner, David Schubmehl, GenAI Implementation Market Outlook: Worldwide Core IT Spending for GenAI Forecast, 2023–2027, IDC Market Note, Doc # US51294223, October 2023.

[2] Enterprise Technology Research, “July 2023 Macro Views Summary,” July 2023. In Weighing the Open-Source, Hybrid Option for Adopting Generative AI, Harvard Business Review Analytic Services, November 2, 2023.

Subscribe to the Equinix Blog