AI Data Centers Revive HDD Demand, but NAND Flash Is Not in Decline

Google News surfaced a striking storage claim: AI data centers are reviving hard disk demand as NAND flash approaches a decline. The first half is supported by supplier results. The second needs a major qualification.

Seagate and Western Digital are shipping more high-capacity storage into cloud environments while reporting stronger margins and multiyear customer commitments. AI training, inference, and content generation are expanding the volume of data that operators retain after expensive computing work ends.

However, HDD growth does not mean enterprise flash is collapsing. Current market evidence shows strong demand for both technologies, accompanied by shortages and higher contract prices. The real divide runs between storage tiers and customer segments, not between a victorious HDD market and a dying NAND industry.

That distinction matters for anyone budgeting AI infrastructure. Fast enterprise SSDs remain essential when GPUs need immediate access to data. High-capacity HDDs handle larger pools where cost, retention, and density matter more than response time. Consumer flash demand is weaker, but that weakness does not describe the full market.

What the Google News Headline Gets Right

AI has turned persistent storage from an afterthought into a capacity constraint.

The storage story began with an article distributed through Google News, but supplier disclosures provide stronger evidence than the headline alone. Western Digital reported 3.34 billion dollars in fiscal third-quarter 2026 revenue, an increase of 45 percent from the previous year.

Cloud customers generated 89 percent of Western Digital’s revenue during that quarter. Its quarterly capacity shipments reached 222 exabytes, up from 166 exabytes one year earlier. One exabyte equals one billion gigabytes, making that change more useful than a simple drive-unit comparison.

Western Digital’s earnings data also show rising profitability. Its non-GAAP gross margin reached 50.5 percent, compared with 40.1 percent in the year-earlier quarter.

Seagate reported a similar pattern. Its fiscal third-quarter revenue reached 3.11 billion dollars, up from 2.16 billion dollars one year earlier. GAAP gross margin increased from 35.2 percent to 46.5 percent.

Chief executive Dave Mosley described the change as structural growth supported by AI data creation. That remains a supplier’s interpretation, but the company’s results offer measurable support. Seagate generated 1.1 billion dollars in operating cash flow during the quarter.

The companies are benefiting from more than a temporary restocking cycle. Cloud operators are negotiating longer purchasing arrangements because high-capacity HDD supply cannot expand quickly. Building additional manufacturing capacity requires long planning periods, specialized equipment, and confidence that demand will persist.

High-capacity drives also take time to qualify. A hyperscaler cannot replace one model with another across thousands of racks without reliability testing, firmware validation, and workload analysis. Those qualification requirements slow supply responses and strengthen established vendor relationships.

The shift is especially visible in nearline HDDs. A nearline drive is a high-capacity disk designed for data centers that need frequent availability without SSD-level latency. These drives sit between actively accessed flash storage and offline archival media.

AI expands that middle layer. Training data, generated media, checkpoints, logs, retrieval indexes, and compliance records do not all need immediate GPU access. Operators still want to retain them because recreating the information can be costly or impossible.

This is why the HDD comeback reported through Google News is credible. AI increases compute spending first, but the resulting data eventually moves through a broader storage hierarchy. Hard disks remain difficult to replace when the requirement is storing many petabytes economically.

The headline becomes misleading only when it treats NAND flash as the losing side of a simple substitution contest. Enterprise infrastructure rarely chooses one medium for every job. It combines different media according to latency, endurance, capacity, power, and cost requirements.

AI Data Centers Need More Than Fast Storage

The AI storage problem is a data-lifecycle problem, not a race to identify one universal medium.

A training cluster needs fast access to active datasets because idle accelerators waste scarce computing capacity. Enterprise SSDs serve that layer with high throughput and low latency. HDDs take over when information becomes less active but remains valuable.

Consider a company training a video-generation model. Its active training samples and current checkpoints may sit on flash. Older model versions, raw footage, audit records, and less frequently used datasets can move to HDD-based object storage.

Inference adds another source of persistent data. Each interaction can produce prompts, retrieved documents, outputs, safety logs, user feedback, and operational telemetry. An individual record may be small, but a widely used service produces them continuously.

Agentic systems add longer chains of intermediate work. An agent might search documents, call software tools, generate drafts, and preserve traces for evaluation. Those records support debugging and governance even when they no longer require the fastest storage tier.

Physical AI produces another stream. Robots, autonomous systems, and industrial equipment generate sensor readings, images, and operational histories. Some data feeds immediate decisions, while larger volumes remain available for later model improvement.

This pattern explains why storage suppliers talk about AI data centers as data systems. The expensive GPU cluster captures attention, but the supporting architecture must ingest, classify, move, retain, and eventually delete information.

The process resembles a funnel. A smaller active set remains on high-performance flash, while a larger historical set migrates toward cheaper capacity storage. Tape may handle the coldest archival layer, particularly when recovery speed is less important.

HDD manufacturers are improving density to strengthen their position in that funnel. Seagate’s Mozaic platform uses heat-assisted magnetic recording, or HAMR, which briefly heats a tiny disk area before writing denser magnetic bits.

In March 2026, Seagate said its Mozaic 4+ platform had entered production with two hyperscale cloud providers. The platform supports drives with capacities up to 44 terabytes, with additional customer qualifications under way.

Seagate’s HAMR announcement says the company plans to move from more than four terabytes per disk platter toward ten terabytes. That roadmap would eventually support drives reaching 100 terabytes.

Those future capacities are company targets, not completed products. Still, current deployments show that HAMR has moved beyond laboratory demonstrations. Seagate says 44-terabyte products are shipping in volume to two cloud providers.

Western Digital is taking a different route while also working toward higher capacity. Its portfolio includes energy-assisted perpendicular magnetic recording and shingled magnetic recording, which overlaps data tracks to increase density.

The company has discussed 40-terabyte UltraSMR drives and a longer-term path toward 100 terabytes. It is also developing designs intended to improve HDD bandwidth, addressing an important weakness as AI systems access larger data pools.

Neither roadmap turns HDDs into SSD replacements. A denser disk still contains mechanical components, so random access remains slower than flash. Higher density primarily improves capacity per rack, power per stored terabyte, and total infrastructure economics.

The result is complementary growth. Flash keeps the GPUs fed, while HDDs retain the expanding body of information produced around those GPUs. Google News captured the HDD resurgence, but the underlying mechanism supports a mixed architecture.

HDD Demand Is Pressuring a Concentrated Supply Chain

Cloud buyers face a capacity problem because HDD demand is rising faster than manufacturers can safely expand output.

The high-capacity HDD market has few large suppliers. Seagate and Western Digital account for most nearline shipments, while Toshiba remains the other significant producer. That concentration limits how quickly customers can find alternative capacity.

Manufacturers also changed their behavior after the storage downturn of 2022 and 2023. Suppliers reduced production, managed inventory, and became more cautious about building capacity before receiving firm demand signals.

That discipline now meets an AI infrastructure cycle. Hyperscalers want more exabytes, but manufacturers prefer density improvements over rapid unit expansion. Adding capacity per drive protects margins while helping customers fit more storage into existing facilities.

The strategy is visible in Western Digital’s quarterly figures. Revenue rose 45 percent year over year while exabyte shipments increased about 34 percent. The difference suggests a combination of stronger pricing and a richer product mix.

Seagate’s margin expansion points in the same direction. Strong demand is allowing suppliers to emphasize higher-capacity products and maintain tighter commercial terms. Buyers cannot assume that storage costs will fall simply because drives become denser.

Long-term agreements reinforce the change. These contracts give suppliers clearer demand visibility and offer customers reserved capacity. They also reduce the freedom buyers once had to wait for cyclical price declines.

For cloud providers, the forced response has several parts. They must forecast storage requirements earlier, qualify higher-capacity models, and design software that moves data among tiers efficiently.

That final step is often underestimated. A storage hierarchy only saves money when applications know which information requires low latency. Keeping everything on flash wastes capacity, while moving active data too early can slow training or inference.

Procurement teams therefore need workload information from engineers. They must understand access frequency, retention rules, recovery targets, and the cost of rebuilding data. A nominally cheaper storage tier can become expensive if it creates repeated transfers or delays costly compute jobs.

The pressure extends beyond hyperscalers. Smaller cloud providers and enterprise buyers may struggle to obtain the same contractual priority. Large customers can offer predictable multiyear demand, while smaller buyers often purchase through less favorable channels.

This imbalance can shape AI competition. A company with secured storage allocations can preserve larger datasets and operate more experiments. A constrained buyer may delete useful information, delay deployments, or pay more for available flash capacity.

It can also affect software architecture. Storage platforms that reduce duplicate data, compress files, or automate tiering become more valuable when raw capacity is scarce. Knowledge workers see the same issue at a smaller scale when scattered AI outputs become difficult to retain and retrieve.

A structured AI knowledge base addresses the organizational version of this problem. Data-center operators face a larger physical version, but both cases depend on keeping valuable information findable without treating every item equally.

HDD supply pressure does not guarantee uninterrupted growth. Cloud companies can optimize retention, improve compression, or shift workloads. A slowdown in AI capital spending would also weaken long-range demand forecasts.

Yet the current evidence shows real pressure rather than a speculative future. High cloud revenue shares, rising exabyte shipments, and production qualifications all point in the same direction. HDDs have returned to infrastructure planning because AI makes long-term data retention more valuable.

NAND Flash Is Splitting, Not Approaching a Broad Decline

NAND flash is experiencing market divergence, with weak consumer demand offset by scarce enterprise supply and AI-driven orders.

The phrase “NAND flash decline” can refer to several different measurements. Unit demand, bit shipments, contract prices, supplier revenue, and consumer purchasing can move in opposite directions during the same quarter.

Smartphones and personal computers remain sensitive to component costs. When flash prices rise, manufacturers may reduce storage specifications or postpone orders. Retail buyers can also resist higher prices, weakening spot-market activity.

Enterprise customers behave differently. A cloud provider building an AI cluster cannot easily remove the storage needed to feed accelerators. It may accept higher SSD prices, negotiate a long-term agreement, or redesign the balance between SSDs and HDDs.

TrendForce reported that the top five NAND suppliers generated more than 38.9 billion dollars in revenue during the first quarter of 2026. That represented an 83.7 percent increase from the previous quarter.

Its NAND supplier data attributed the increase to stronger average selling prices and enterprise SSD demand. The research firm also said a shortage of HDDs had pushed some orders toward high-capacity QLC enterprise SSDs.

QLC NAND stores four bits in each memory cell. It offers greater density than alternatives that store fewer bits, although endurance and performance characteristics require careful workload design.

The HDD shortage therefore strengthens some flash demand rather than replacing it. When buyers cannot obtain enough disk capacity, high-density enterprise SSDs can absorb workloads that would otherwise use HDDs.

TrendForce estimated that Samsung’s first-quarter NAND revenue reached 13.51 billion dollars, rising 104.7 percent sequentially. SK hynix Group followed with approximately 7.53 billion dollars, including revenue from enterprise SSD supplier Solidigm.

Kioxia, Micron, and SanDisk also recorded sharp sequential revenue increases. TrendForce said SanDisk’s data-center business grew by more than 200 percent from the previous quarter.

These figures do not resemble a broad NAND collapse. They show a supply-constrained market in which AI infrastructure has become more important than traditional consumer demand.

The distinction becomes clearer when looking at prices. TrendForce expected overall NAND contract prices to rise by 70 to 75 percent during the second quarter of 2026. Enterprise SSD demand was one of the central drivers.

The firm’s contract forecast said meaningful capacity expansion was unlikely before late 2027 or 2028. Cloud providers were signing longer agreements to secure supply.

By July, the market had become more divided. Consumer buyers resisted further increases, and spot-market activity remained weak. Contract-price growth began flattening after the earlier surge.

That plateau is probably the source of the “approaches decline” interpretation. However, slower price growth after unusually large increases is not the same as falling enterprise demand.

TrendForce’s July market bulletin described strong infrastructure demand alongside weak consumer markets. It also highlighted longer contractual commitments intended to reduce suppliers’ exposure to historical memory cycles.

NAND manufacturers are concentrating production on higher-margin server products. That decision can restrict supply for mature flash types used in industrial equipment, vehicles, networking hardware, and embedded devices.

Some mature NAND products may face shortages even when the devices using them sell slowly. Suppliers can retire older production lines or convert resources toward advanced layers, leaving legacy customers with fewer options.

The result is not one unified flash cycle. Enterprise SSDs, smartphone storage, retail drives, industrial NAND, and removable media can follow different paths.

This also explains why HDD and NAND suppliers can report strong results together. AI data centers need fast flash near compute and inexpensive capacity behind it. Shortages in either layer can redirect limited workloads toward the other.

Google News readers should therefore treat the original headline as a market-divergence story. The evidence supports an HDD revival and consumer NAND weakness. It does not support an approaching decline across the entire flash industry.

The HDD Revival Still Has Technical and Commercial Limits

Higher demand does not remove the latency, power, reliability, and execution risks attached to large HDD deployments.

Hard disks remain mechanical devices. A read head must move to the correct track while platters rotate underneath it. That process creates latency that software cannot eliminate completely.

Sequential throughput can improve, and suppliers are exploring designs that access more than one part of a drive simultaneously. Still, SSDs retain a decisive advantage for random access and workloads that require rapid response.

This difference matters when accelerators wait for data. GPUs represent a large infrastructure commitment, so operators will not trade away utilization merely to save on storage media. Active datasets and frequently accessed model artifacts remain strong SSD candidates.

HDDs also consume rack space and power at scale. Higher-capacity models improve efficiency per terabyte, but an expanding data pool can offset those gains. Total power consumption may keep rising even as each stored unit becomes more efficient.

Seagate says a one-exabyte Mozaic 4+ deployment can reduce annual energy consumption by about 0.8 million kilowatt-hours compared with standard 30-terabyte drives. It estimates an infrastructure-efficiency improvement of approximately 47 percent.

Those figures come from Seagate’s internal calculations. They depend on deployment assumptions, workload behavior, and the comparison system. Independent operators may see different results.

Reliability requires similar caution. A higher-capacity drive concentrates more data in one device, increasing the amount that must be reconstructed after a failure. Storage systems need adequate redundancy and recovery bandwidth.

Large buyers test new drives before broad deployment for this reason. A product can meet laboratory specifications while presenting unexpected behavior in dense racks, mixed workloads, or extended operation.

Backblaze’s public drive studies offer one independent reference, although its fleet does not perfectly represent hyperscale AI environments. A recent analysis covered 341,263 drives and reported a fleetwide annualized failure rate of 1.39 percent.

Failure rates varied considerably by model and age. The comparison illustrates why buyers qualify individual products rather than assuming that every drive from one manufacturer behaves identically.

Supply concentration creates another risk. With only three major HDD manufacturers, production problems or qualification delays can affect a large share of available capacity. Customers cannot quickly shift purchases to a long list of interchangeable suppliers.

HAMR adds manufacturing complexity. Seagate must produce integrated recording heads, media, lasers, and control components at high yields. Its current hyperscale deployments are meaningful, but the company still needs to scale output across more customers.

Western Digital faces its own execution questions. It must deliver higher-capacity products while advancing performance features and managing its eventual transition toward HAMR. Roadmaps extending to 100 terabytes should be treated as direction, not guaranteed timing.

Demand forecasts are uncertain too. Storage suppliers assume AI applications will keep producing data worth retaining. Operators may counter that growth with better data selection, synthetic-data policies, compression, deduplication, or shorter retention periods.

Legal constraints can reduce stored volumes. Privacy rules, copyright disputes, and enterprise governance policies may force organizations to delete information rather than keep everything for future training.

Model development could also become more efficient. If smaller datasets, improved training methods, or more selective retrieval deliver comparable results, storage growth may lag the most optimistic projections.

None of these risks erase present demand. They define the boundaries around supplier claims. The safe conclusion is that AI has strengthened HDD economics, while the duration and scale of that benefit remain tied to customer behavior.

The same skepticism should apply to NAND forecasts. Current shortages and higher supplier revenue do not guarantee permanent pricing power. New capacity, weaker capital spending, or inventory corrections can eventually restore the industry’s familiar cyclicality.

Three Signals Will Test the Storage Reversal

The next phase will depend on actual shipments, enterprise SSD pricing, and binding cloud commitments rather than headline momentum.

The first signal is the pace of high-capacity HDD qualifications. Seagate says two hyperscalers are already receiving Mozaic 4+ products, while broader qualifications continue.

Investors and infrastructure buyers should watch whether 44-terabyte shipments expand beyond those initial customers during the second half of 2026. Wider qualification would strengthen the case for durable HAMR adoption.

Delays would weaken it. They could indicate manufacturing challenges, conservative customer testing, or limited demand for the newest capacity points.

Western Digital’s 40-terabyte qualification progress belongs in the same signal. Volume production across multiple cloud customers would show that HDD density gains are becoming a competitive market trend, not one supplier’s isolated achievement.

The second signal is enterprise SSD pricing and order volume. NAND flash cannot be described as broadly declining while cloud customers keep signing supply agreements and absorbing high-capacity SSD output.

A meaningful drop in enterprise contract prices, accompanied by weaker server orders, would challenge the current mixed-storage thesis. Flat consumer demand alone would not be enough.

The important split is between enterprise and consumer indicators. Retail SSD prices and smartphone specifications can weaken while data-center SSD demand remains tight. Reporting should identify which segment is moving before declaring a NAND downturn.

The third signal is the length and firmness of cloud purchasing commitments. HDD vendors have discussed agreements extending several years, while NAND suppliers are also seeking longer visibility.

Confirmed commitments through 2028 or 2029 would support the structural-growth argument. They would show that customers expect AI storage requirements to persist beyond one capital-spending cycle.

Canceled orders, reduced commitments, or rising supplier inventory would point in the opposite direction. Those changes would suggest that buyers overestimated demand or improved storage efficiency faster than expected.

Readers should also separate contractual demand from projected demand. A supplier’s discussion with a customer carries less weight than an enforceable purchase agreement. Revenue, shipments, and inventory provide the clearest confirmation.

The storage market entering late 2026 is not following a simple HDD-versus-NAND script. Hard disks are gaining importance because AI creates large retention pools. Enterprise flash is also benefiting because active AI workloads require speed.

Consumer NAND faces the sharper pressure. Buyers resist elevated component costs, and device makers can reduce capacity when demand is soft. That weakness coexists with tight enterprise supply.

The next Google News headline may compress these movements into another winner-and-loser claim. Readers should look beneath it for three facts: which storage tier is affected, which customer segment is buying, and whether the evidence comes from shipments or forecasts.

For enterprise buyers, the immediate action is practical. Map active and inactive data, test higher-capacity drives, and model procurement under constrained supply. Do not assume either HDD or NAND will become universally cheaper.

For developers and knowledge workers, the lesson is similar. AI creates information faster than organizations can classify it. Retaining everything in the fastest tier wastes resources, while deleting too aggressively destroys future value.

The storage reversal is real, but it is not a funeral for flash. It is a return to tiered architecture, driven by the uncomfortable fact that AI produces far more valuable data than one medium can handle alone.