For a brief window, going all-flash was the easy call. NVMe SSDs were affordable, plentiful, and fast. Why bother with tiering when you could throw flash at every storage problem and move on to the next GPU order?

That window has closed. And the bill is coming due.

The Numbers That Changed Overnight

In January 2026, TrendForce projected NAND flash contract prices would rise 33–38% quarter over quarter. By February, The Register estimated upward to 55–60%. DRAM forecasts moved even more dramatically, from an initial 55–60% increase to 90–95% in a single quarter.

These are not gradual shifts. Phison’s CEO has described NAND production capacity as effectively allocated through 2026 and has warned that tight supply could persist for a decade. Silicon Motion’s CEO called the current environment unprecedented, with HDD, DRAM, HBM, and NAND all constrained at the same time. Goldman Sachs projects double-digit price increases continuing quarter over quarter throughout the year.

For neocloud operators who built their entire storage layer on all-flash architectures, this is not a cost increase. It is a structural repricing of a core assumption their business model was built on.

Storage Was Supposed to Be the Quiet Line Item

When the neocloud wave crested in 2021 and 2022, storage typically represented about 10% of an AI infrastructure budget. Operators were laser-focused on GPU procurement. Storage was an afterthought, something you solved once and forgot about.

Today, storage is trending toward 20–30% or more in all-flash deployments, and it is the fastest-growing cost in the stack. A 122 TB QLC NVMe SSD runs roughly $47,000. A 3.84 TB drive from the same generation delivers comparable sequential throughput for around $2,500. The performance difference between the two is negligible. The only variable is how much cold data you are willing to park on premium media that delivers zero additional throughput for the privilege.

That is the question neocloud operators need to answer honestly: are you buying flash for performance, or are you buying it out of architectural inertia?

The Hyperscalers Already Figured This Out

Google, Meta, and Microsoft do not run all-flash storage. They never have at scale. They deploy mixed-tier architectures with intelligent tiering, using just enough NVMe flash to saturate GPU throughput requirements, then draining data to high-density HDDs as the workload allows. Flash handles the hot path. Disk handles everything else.

This is not a philosophical preference. It is an economic imperative driven by the physics of AI workloads. NVIDIA’s own DGX storage guidance specifies that text-based LLM training requires roughly 0.5 GB/s of read throughput per GPU. Even physical AI and visualization workloads top out at approximately 4 GB/s reads and 2 GB/s writes per GPU. You do not need an ocean of flash to meet those numbers. You need the right amount of flash in the right place, managed by software smart enough to keep data flowing between tiers without human intervention.

The operators who survive the coming consolidation will be the ones who adopted this playbook before the market forced their hand.

Consolidation Is Already Here

H100 rental rates have dropped over 60% from their peak. First-generation neocloud infrastructure from the 2021–2022 deployment wave is hitting depreciation limits, forcing fleet-wide replacements at today’s inflated component prices. The era of rewarding companies simply for stacking GPUs is over. The market now demands proof of return on invested capital.

In this environment, total cost of ownership decides everything. And storage is where the biggest TCO gaps hide. An architecture that needs three times the SSDs to hit the same throughput burns three times the power, takes three times the rack space, and carries three times the capital exposure to a NAND market that shows no sign of softening.

Meanwhile, Seagate’s CEO told investors that nearline HDD capacity is fully allocated through 2026, with customers already discussing 2028 supply assurance. Even disk is getting tight. The operators with flexible, mixed-media architectures have options. The operators locked into a single storage medium do not.

The Takeaway

The all-flash era was a product of temporarily favorable economics, not architectural wisdom. The organizations that recognized this early, built storage layers capable of riding independent cost curves for flash and disk, and treated intelligent tiering as a first-class engineering priority rather than a bolt-on afterthought, are the ones positioned to weather what comes next.

Everyone else is doing the math right now, and the math is not kind.

Sources cited in this article:
TrendForce  |  The Register  |  Tom’s Hardware  |  IDC  |  SiliconANGLE  |  MLQ Research  |  VDURA