Why the Memory & Storage Crisis Is Reopening the Repatriation Debate

For most of the past three years, the enterprise IT narrative on infrastructure ran in one direction: pull workloads back from the public cloud. Egress fees, unpredictable bills and a desire for control pushed a large share of CIOs toward on-premises and private infrastructure. A Barclay's CIO study found that 83 percent of enterprise IT leaders planned to shift at least some workloads off public cloud, and other surveys put the share planning some form of repatriation even higher.

That momentum has not stopped, but a new variable has entered the calculation and it is forcing IT leaders to rerun the math on where data and workloads should live. The hardware side of the ledger became much more expensive in 2026.

AI’s Impact on Flash Storage Pricing

A global memory shortage, driven largely by demand from hyperscalers and large AI companies monopolizing high bandwidth memory, has pushed enterprise flash storage prices to new heights. The current price of a 30TB TLC enterprise SSD is $22,600, compared with $3,460 in Q3 2025, while a 30TB QLC SSD now costs $18,080, up from $2,768 during the same period, according to Storage Review. An equivalent hard drive costs $1,216, putting flash at 18.6 times the price of disk per terabyte.

Gartner has estimated that server costs overall climbed more than 125 percent in the first half of 2026.  None of this looks like a short cycle that corrects itself. Major suppliers like Micron have alluded to supply and cost pressures continuing through next year.  

Meanwhile, the major cloud service providers have not raised prices for enterprise customers yet. Even while their own spending has gone through the roof, the Big 3 are absorbing the costs for now, according to CNBC. Cloud executives likely see an opportunity to expand further, amid a year of strong growth.

Assessing Cloud in 2026

“We will certainly see an increase in public cloud adoption this year and next as cloud prices have either not increased, or have marginally gone up,” said Shrish Pant, director analyst at Gartner (News - Alert). “Accessibility for cloud has improved because of availability, immediate availability versus long lead times for server procurement, and a different cost trade-off.”

Of course, cloud regions are built from the same overpriced DRAM and NAND and providers will allocate components to the workloads that are more profitable. Right now, this is AI inferencing rather than general enterprise compute.

There are early signs that enterprises are already shifting more spend to the cloud. In the first quarter of 2026, Google Cloud grew 63 percent year over year, Microsoft (News - Alert) Azure 40 percent and AWS 28 percent. While there is no single cause for growth, this indicates a shift in sentiment from repatriation in recent years.

The logic is clear. When an on-premises server costs roughly four times what it did a year ago, cloud consumption starts to look like a hedge against volatile capital costs rather than simply an operating expense.

How to Approach a Hybrid Strategy

It is always best practice to regularly evaluate data and workload placements across all infrastructure; departmental needs are often in flux as is pricing, new products, vendor changes and regulations. With hybrid infrastructure more prevalent, IT needs a granular understanding of data usage and growth, performance and security requirements by department and data types, costs across storage and long-term needs.

Consider these key decision points:

Separate workloads by pattern.

Steady state, high utilization, predictable systems, including many regulated data sets and continuous AI inference jobs, often still make sense on owned or leased infrastructure where the economics are known in advance. Unpredictable, variable and globally distributed workloads are generally better suited to public cloud, where elasticity is a benefit. Build a full total cost of ownership model that includes storage, retrieval and egress fees, IOPS, redundancy, staffing and support. These are the factors that most impact your total costs.

Don’t lock yourself out of options.

If you decommission cloud capacity after repatriation, you lose the option to route it back if on-premises hardware costs keep climbing or if utilization patterns change. Keeping a standby cloud footprint for burst capacity or disaster recovery has a defined, calculable cost that preserves flexibility. Ensure that your data is never locked into storage, as well. This happens when using storage vendor block-based methods for moving and managing data, which makes it difficult and costly to move data to new storage due to rehydration requirements. Storage-agnostic unstructured data management platforms avoid this problem with file-level data movement.

Match data placement to real access patterns.

Hot, frequently accessed data belongs on the fastest available tier, whether on-premises flash or cloud object cost disk. Meanwhile, data kept mainly for compliance or audit belongs in archive tiers where storage cost is lowest.

Balance cost against AI needs.

AI workloads have distinct requirements, including data locality for training, throughput for inference and long-term retention of unstructured data for future model use. These needs sometimes argue for keeping data close to compute and sometimes argue for the elastic capacity cloud offers, so making decisions by workload is best rather than as blanket policy.

Security and governance.

A hybrid strategy needs consistent identity management, encryption and monitoring across every environment, plus a current inventory of where regulated data actually lives, since data residency rules increasingly dictate which region or provider is even an option. AI requires additional protections, including sensitive data detection to exclude protected data from AI pipelines and auditing mechanisms to ensure compliance with internal and external regulations.

These are a few of primary things to keep in mind, but there are other considerations and trends that deserve a place in any storage recalibration plan.

• Energy constraints are a growing cost and risk factor. Rising energy demands and growing energy consumption are now core factors in infrastructure planning. Whether procuring hardware, leasing data center space or calculating cloud footprints, designing for energy efficiency is key to control costs and to be sustainable.
• Data sovereignty and residency requirements including the EU’s AI Act and GDPR and China’s PIPL are making sovereign cloud regions an important distinction when deciding where to host governed workloads.
• Tiering is a game changer as prices continue to rise for high-performance storage whether on-premises or in the cloud. Adding HDD, cloud object storage and even tape are ways to balance cost appropriately across cold, warm and hot data. Tape storage is having a comeback moment with enterprise purchases growing at 11% right now.

The global memory shortage has created a new wrinkle in the cloud versus on-premises debate. For IT leaders already running a hybrid environment, continue assessing and testing where each workload and each tier of data actually belongs. If running a largely on-premises environment, now is the time to consider moving some of your data out of the data center to cloud.

Spreading the risk across hybrid infrastructure while retaining flexibility for your enterprise data are core storage procurement tenets for the foreseeable future. Understanding unstructured data usage, requirements and costs across storage and being able to move it without penalty and without disruption go hand in hand as you design a stable, resilient and cost-efficient storage infrastructure for the near and long term.

About the author: Randy Hopkins is the vice president of global systems engineering and enablement at Komprise. Over a tech career spanning more than 30 years, Hopkins has developed expertise in building and running systems engineering organizations, include pre-sales, technical operations, channel and product introductions. He has a Bachelor of Science degree in Information Technology from California State University, Fresno.