The wave of deployments for Inference AI has raised alarms across the storage supply chain. According to industry sources cited by Taiwanese press, nearline hard drives (HDDs) for data centers are experiencing a persistent shortage, and lead times have extended beyond two years. Facing this bottleneck, major Cloud Service Providers (CSPs) in North America and China are urgently increasing orders and signing long-term contracts to secure supplies of both high-capacity HDDs and enterprise SSDs.
This imbalance has triggered an immediate secondary effect: the 2026 QLC NAND production capacity at several manufacturers has already been pre-sold, and the market predicts that by 2027, QLC bits could surpass TLC globally. Simultaneously, there is a noticeable shift in specifications: from the usual corporate SSDs (8–16 TB) to 64–128 TB in China, and up to 128–256 TB QLC in North America.
Why are HDDs in short supply (and why won’t it be resolved quickly)?
The accumulated demand for “hot” capacity for inference — large volumes of “near-line” data for serving models — mainly lands in the nearline segment where HDDs still dominate cost per terabyte. However, the disk manufacturing industry has consolidated into very few players and primarily operates by order (build-to-order), which limits response flexibility. The result: backlogs between CSPs and suppliers, growing lead times, and contracts extending into 2026 to ensure supply.
While major HDD manufacturers have improved revenue guidance, not all are investing aggressively in CapEx. For example, Seagate considers a +5% annual investment focused on technological upgrades (to boost performance and density), while Western Digital maintains a prudent expansion stance. In other words: there will be some capacity release from efficiency gains, but no rush of new factories in the short term. Additionally, an extra vector of risk exists: although the current shortage does not stem from raw materials, rare earth elements remain critical for HDDs; any export tensions could add uncertainty to supplies.
The cloud’s “Plan B”: more SSDs… despite the price
With HDDs bottlenecked, CSPs have no choice but to increase purchases of enterprise SSDs with large capacities, even though the price per TB remains 3–4 times that of nearline HDDs. This pivot is accelerating the transition towards QLC NAND in data centers, with lines up to 128–256 TB for AI racks. The obstacle is that not all QLC is ready yet, and there aren’t enough 2 TB dies to support massive volumes: a 256 TB SSD requires very dense configurations (e.g., 32×32 dies of 2 TB), and steady availability of such dies remains limited.
The immediate consequence is a price rally: in the Q4, enterprise SSDs and NAND for servers rise over 20%; some models see increases of 30–40%. Suppliers are also acting swiftly: Micron has shifted from quarterly to weekly negotiations, Kioxia directly warns of inventory shortages, and Samsung is transitioning from V6 to V8 in its Xi’an plant, with production contained in the first half of the year and ramping up in the second.
All of this is happening against a backdrop of strategic capacity reductions (around 10–15%) implemented by several manufacturers since late 2024, which have not yet been fully reversed. The prioritization of capacity towards servers is displacing capacity from consumer devices (e.g., smartphones), and factory utilization will gradually increase, but not fast enough to close the nearline gap.
A hot market until at least 2026
The situation is unprecedented: HDD, NAND, and DRAM are all under stress simultaneously, with three consecutive quarters of price increases and a high floor that — according to industry estimates — could persist through 2026. In NAND, scarcity is accelerating the adoption of QLC (better cost per bit, increasing densities, yield improvements, and more layers per stack), pushing the horizon for PLC (5 bits per cell) towards 2027–2028, initially in specific products with clear reliability and lifespan challenges.
Practical implications for planners and buyers
1) Adjusting the technology mix (and storage SLOs)
- Where latency and sustained performance matter (active inference setups), QLC SSDs may be the only realistic path while HDDs are scarce.
- For retention and archival, prioritize TLC with high density or contract-backed HDDs; consider tapes as capacity cushions at predictable costs.
2) Negotiations and contracts
- If your project depends on nearline HDDs, seal deals now for 2026 with suppliers; include priority clauses and scaling provisions.
- For NAND, consider price indexing or flexible agreements due to volatility (e.g., Micron’s weekly negotiations).
3) Capacity and efficiency
- Reassess tiering strategies: larger SSD caches to offload HDDs where available, or compress/deduplicate based on load (mind CPU impact); enable cold moves to cheaper media more often.
- Implement strict deletion policies and lifecycle management: each GB of unused storage today can be paid multiple times by 2025–26.
4) Reliability engineering for QLC
- For mass QLC deployments, use generous over-provisioning and write shaping; enhance wear monitoring (extended SMART/telemetry) and firmware mitigations.
- Plan for RAID/erasure coding with quick rebuilds (and prepared networks) to prevent long degradation windows.
5) Budgeting and TCO
- Anticipate price scenarios with +20–40% increases in enterprise SSDs and long lead times for HDDs.
- For 2025–2026, the TCO may favor hybrid solutions that reduce service risk despite higher €/TB.
What comes next? The role of photonics and density jumps
The pressure from AI is pushing the industry toward new interconnects within racks (integrated optics, CPO) and towards higher NAND densities. If HDDs maintain a cost per TB that remains competitive, they will keep a central role in nearline retention. However, gaps in availability and supply chain inertia could leave QLC SSDs as a bridge solution longer than expected. In any case, the mid-term points toward a more aggressive layered architecture, with automatic movements and smart caches to balance cost, service, and risk.
Frequently Asked Questions
Why are HDD lead times extending beyond 2 years?
Because manufacturing is highly concentrated, production is order-based, and nearline CSP demand has surged beyond what supply can quickly absorb. Without new factories or massive CapEx, adjustments depend on performance improvements and backlogs.
If SSD costs 3–4× per TB, why is QLC so heavily purchased?
Because enough HDDs are not available, and inference AI projects require active capacity now. QLC offers the highest density available in SSDs, and with proper engineering, it meets the needs of “warm” volumes.
What short-term price increases are to be expected?
In Q4, enterprise SSDs and server NAND are expected to rise >20%, with some models climbing 30–40%. Looking ahead to 2026, analysts estimate SSD bit demand for servers will grow about 50% over 2025, maintaining market tension.
When will QLC overtake TLC?
Based on current trends, industry projections indicate QLC bits could surpass TLC between 2027 and 2028, earlier than last year’s expectations. The introduction of PLC (5 bits per cell) might begin in niche applications within the same timeframe, facing clear reliability and lifespan challenges.
via: digitimes