Let’s face it, the AI infrastructure market and data center build out is moving at a speed and complexity that is bewildering. The tangle of issues facing us includes supply chain, national and international regulatory frameworks and tailoring workloads to specific hardware GPU vendor roadmaps to name just a few.

The way forward isn’t obvious and it’s tough to get a clear view of what will happen and what decisions those in the field should make. Still, there is a storehouse of wisdom that can help show us the way forward.

As any good Deadhead will tell you, “If you get confused, just listen to the music play.” (Franklins’ Tower, 1975) In this case, that means listening to what the market is telling us, letting what you see in the marketplace be your guide. Here are five trends that I’m seeing in the market and in discussions with customers that are likely to continue and intensify through 2026.

Trend 1: Cooling and hybrid data centers

Most greenfield data centers are already plumbed for liquid cooling due to the power and thermal limitations of existing data centers which simply cannot accommodate the latest GPUs. In fact, as early as April 2024, surveys already pointed to 22 percent of data centers using liquid cooling. Based on anecdotal evidence, virtually no new data centers that are AI focused or accommodate AI tenants are going to be purely air-cooled.

But many data centers currently cannot support that level of complexity, and that’s where alternate cooling strategies play a role.

From rear door chillers to sidecar cooling and new cooling fluids which are engineered to support two phase cooling solutions with lower boiling points, data center managers and infrastructure providers are looking for ways to adapt existing data centers.

Expertise as a trusted advisor in the liquid cooling space will evolve from a nice-to-have skill set for those selling hardware into AI data centers to a necessary knowledge base. Without a clear understanding of liquid cooling trends and technologies, data center operators and planners may miss out on the latest technologies.

Companies selling systems into the AI ecosystem would do well to take advantage of opportunities to become more expert on liquid cooling or partner with companies that do to successfully compete in the marketplace.

Trend 2: Regulation and traps for the unwary regarding cloud computing

Especially in the Enterprise or Fortune 1000 space, many industries and customers will face the decision of whether to locate their AI server infrastructure on premises, often in a colocation data center, or in the cloud. For many, particularly those in highly regulated industries, having everything in the cloud will not be an option due to privacy regulations in both the EU, UK, and many US states like California.

This entails everything from European laws on AI that restrict its use (prohibited use cases are quite broad and subject to interpretation but that’s for another discussion) to the increased intersection of privacy laws with AI in the US.

In addition, certain fields like the law will require on-premises deployments. In many cases, this will be necessary to satisfy confidentiality requirements and other rules that are specific to that industry.

In addition to regulatory complexity, the issue of cost plays an outsized role. As anyone who's used any of the cloud providers at hyperscale can attest, the cost of locating computing resources and data in the cloud can become very high, especially at large scale. These questions of regulatory compliance, cost and industry specific rules of content control will continue to play a role in the future.

Generating expertise internally in this area or partnering with those who already have regulatory expertise is an important strategy for success in the AI data center market. Moreover, it is a way to avoid costly traps for the unwary that will only be obvious to someone with a great deal of exposure and experience in the regulatory space.

Trend 3: Supply chain headaches

Many of the core components necessary for AI servers have become more costly, are impacted by shortages, or have very long lead times. DRAM prices have increased 171 percent year-over-year according to CTEE and are now outpacing the rise in the price of gold.

NAND Flash used in SSDs is also increasing in price, and GPUs and processors are now strained by the spikes in AI demand as well, as evidenced by shortages for even strategic customers.

If the demand for AI servers and their respective components remains at a fever pitch, these shortages, long lead times and lack of availability will continue to hinder the creation of new data centers.

Large OEM suppliers have long-term contracts with strategic suppliers of DRAM, NAND and CPUs and GPUs and can guarantee supply, delivery and pricing with appropriate lead-times. Given the crunch we are undergoing it makes sense to make strategic bets on components that are necessary by working with those who have already secured allocations and work together to close deals.

Trend 4: Horses for courses

Another key issue will be the optimization of locations for inference nodes throughout the user pipeline. This will include inference nodes at the edge for simpler inference tasks and larger more compute intensive nodes in the data center for the most complex inference models and higher throughput inferencing tasks.

The logic behind this is simple: reduce latency and cost while using appropriate nodes for appropriate workloads. Essentially, a ‘horses for courses’ approach where inference workloads should occur. That is, finding the optimal place to locate hardware based on its workload and latency for clients.

For example, in the quick-serve restaurant (QSR) business, AI agents translating drive-through interactions to kitchen orders can be automated with AI and processed locally at the restaurant to minimize latency with relatively light compute power. This trend will continue as the pressure to make inference workloads cheaper to deploy increases.

Trend 5: The growing importance of sovereign AI

Many countries see it as a national policy issue to have control over the deployment of AI in their countries as well as a source of technology and economic development. Moreover, many of these nations want to control the content output and speech associated with their AI based computing resources, this trend will likely continue.

In addition, generative AI LLMs optimized in local languages will perform better than generalized LLM models for local language text generation. This fact also drives demand for sovereign AI.

Planning to meet the needs of sovereign data center customers requires several special skill sets such as regulatory framework mastery and the ability to navigate export controls both in the US and abroad as well as localization expertise. Companies working in this space would do well to partner with those that have these skills to succeed in the sovereign AI space.

Conclusion

As we've seen, the AI and data center industry is rapidly developing and that development is characterized by flux, speed, and very rapid change. With the trends discussed previously in mind, my hope is the practitioners will be able to avoid the shoals, hidden dangers and sandbars that can complicate the AI journey.

The solution to these problems usually requires creating expertise in-house, partnering with others or making educated bets on the future. For many, partnerships with those that can fill these gaps will be the preferred starting point. This requires playing well with others in addition to making investments and taking risks.

As Jerry counseled, “May the four winds blow you safely home” as you make your AI data center journey. And be nice to each other, “If you plant ice, you’ll harvest wind.”