The narrative that “the HDD is dead” is once again challenged by data. During a press conference in Tokyo, Irving Tan, CEO of Western Digital (WD), argued that hard drives not only do not disappear with artificial intelligence, but grow as a key component of mass storage in data centers. The executive shared figures, technologies, and timelines: 32 TB today, 36 TB by mid-2026 with ePMR, and a jump to 44 TB with HAMR in the second half of 2027.
Beyond the roadmap, Tan explained the reason: in hyperscale environments — those of cloud and AI giants — about 10% of data resides in NAND/SSD, 80% in HDD, and 10% in tape. This isn’t technological nostalgia, he asserted, but data economics: cost per terabyte, reliability, durability, and watts per capacity continue to favor the hard drive as a repository for cold and warm data, while tape maintains its niche for long-term legal archiving.
The data economy: hot, warm, and cold
The example Tan used is everyday: a video of a cat playing with a ball on a social platform. While the content becomes viral, the concentrated access and latency requirements justify using SSD. After the peak week, maintaining it on SSD becomes less optimal: the cost per GB and energy per capacity penalize. Migrating it to HDD becomes the total cost of ownership (TCO) decision. And when legislation mandates storing years of information with almost null access — e.g., financial records — tape remains the most suitable option.
AI doesn’t break this balance; it intensifies it. According to Tan, with AI generating more data from existing data, estimates suggest tripling the overall volume of information by 2030. This context pushes hyperscalers to optimize costs and energy per storage level, rather than indiscriminately replacing technologies.
WD: from consumer to hyperscale
Western Digital — founded in 1970 — was for decades synonymous with disks for PC and devices (such as the first iPod). But the center of gravity shifted: after strengthening its data center business, Tan asserts that in the last quarter, hyperscale customers already accounted for around 90% of the company’s revenue. In this tier, customers don’t buy hundreds of units but tens of thousands per wave, with reliability demands that influence all technological decisions.
36 TB with ePMR in 2026 and 44 TB with HAMR in 2027: why not sooner
The roadmap is broadly known, but WD provided specific dates:
- Today: drives of 32 TB.
- Mid-2026: drives of 36 TB based on ePMR (energy-assisted Perpendicular Magnetic Recording).
- Second half of 2027: drives of 44 TB with HAMR (Heat-Assisted Magnetic Recording).
The obvious question — if HAMR is ready in labs — is why **wait**. Tan’s answer isn’t technical, but industrial: it’s not enough to be able to send a few 44 TB units to pilot clients today; one must guarantee a ramp of hundreds of thousands and then millions of units per quarter, with reliability matching what the market already expects from ePMR. Achieving this demands that R&D and manufacturing synchronize from the start.
Changing development: from sequential pilot to parallel inline processing
To ensure this, WD is modifying its process. Traditionally, the company developed sequentially: pilot line, testing, and if all went well, transfer to high-volume line. With HAMR, validation occurs directly on production lines, in parallel: manufacturing at volume while confirming quality and reliability parameters, speeding up the time to scale capacity.
“Until now, development was sequential; now we go in parallel. Can we send a 44 TB with HAMR today? It’s not impossible. Can we manufacture a million? No. We’ll move into HAMR once we achieve a product with reliability comparable to ePMR and volume capacity,” Tan emphasized.
Beyond density: energy, performance, and UltraSMR
WD’s plan isn’t limited to capacity. The executive also outlined efforts to reduce consumption and improve throughput:
- Energy: two fronts. More efficient motors (with Japanese partners) and lighter, more effective processors/firmware inside the drive. “Soon we’ll be showing innovations in energy efficiency,” he hinted.
- Performance: ongoing research with an aim to announce during the first quarter of 2026.
As an immediate leverage, WD highlighted UltraSMR: a recording scheme that, compared to CMR, adds ~20% capacity; versus conventional SMR, around ~10%. It’s especially useful for read-heavy workloads typical of large repositories of content and analytics.
Japan, a strategic partner: investing, talent, and spintronics
WD’s Japan Country Officer, Kimifumi Takano, highlighted that Japan is “a key investment destination” for the company. In fiscal 2025, expenditure with suppliers and partners in Japan exceeded $1.5 billion (approximately 2.21 billion yen at the indicated rate). Moreover, the company plans to invest $1 billion (1.47 billion yen) in R&D over the next five years.
The Fujisawa center — with a legacy of HGST and IBM — hosts part of this talent, and WD plans collaborations with universities and the NIMS (National Institute for Materials Science) to advance spintronics applied to sensitive read-heads, key for increasing densities without sacrificing reliability. The company reaffirmed its innovation history — helium-sealed drives now standard in the industry — as proof of its intent to lead in this new phase as well.
Why HDD remains dominant in hyperscale
WD’s reasoning can be summarized in four points, which together explain the 80% market share of HDD in hyperscale:
- Massive capacity and cost: €/TB and $ /TB still favor drives as volume grows exponentially.
- Reliability and durability: read/write cycles and MTBF suitable for persistent services at large scale.
- Energy per capacity: in watts/TB, HDD — in idle and warm access modes — maintains advantages over SSD, which reserve their benefits for hot data.
- Tiered architectures: major operators already orchestrate automatic migrations from hot to warm to cold, based on access and data value, with no user friction.
The AI era adds a layer of nuance: not all data has the same value or relevance over time. Training, tuning, and serving models demand different latencies and broadband. In that chain, HDD retains — and expands — its role as a massive repository for cycled and less active artifacts, just as tape sustains long-term archives mandated by regulation.
Challenges and uncertainties: HAMR ramp-up and client timelines
WD’s roadmap pushes competitors and excites the market, but it also bears conditions:
- Volume ramp-up: HAMR must reach a capacity to produce millions of units per quarter.
- Reliability: hyperscalers won’t tolerate a degradation compared to ePMR; data is business.
- Customer schedule: for many operators, adopting new technologies depends not only on the offer, but also on their own deployment cadence and capex.
The company, aware of this, has rethought its development process — validating inline and parallel — precisely to synchronize product and volume.
Market context: neither SSD “yes” nor HDD “no” — layering and economics
The “SSD yes / HDD no” polarization doesn’t hold when examining hyperscale figures. SSDs and HDDs are not perfect substitutes; they are complementary in a decision tree where latency, bandwidth, cost, and energy dominate. Add tape for long-term legal custody, and the picture WD described emerges: 10% NAND, 80% HDD, and 10% tape. If AI triples data volume by 2030, as various projections suggest, the pie grows for all levels. The challenge is to manufacture and operate at the speed demanded by demand.
Implications for data center customers
- Capacity planning: assess scaling to 36 TB (ePMR) in 2026 and 44 TB (HAMR) in 2027, with reliability and power consumption tests in pilot fleets.
- Energy: develop watts/TB and TB/rack metrics to quantify improvements; WD anticipates efficiency innovations.
- Performance: track throughput announcements for 1T26; dimension caching and tiers of SSD according to new bandwidths in HDD.
- Layered architecture: automate hot-warm-cold cycles based on access and data value; review migration policies.
FAQs
Why does Western Digital state that HDDs will remain crucial in AI data centers?
Because in hyperscale environments, the cost-energy-capacity balance favors HDDs for warm and cold data. WD estimates roughly a 10% NAND/SSD, 80% HDD, and 10% tape split. AI triples data volumes and forces optimization by level, not disk elimination.
What’s the difference between ePMR and HAMR in WD’s roadmap (36 TB vs. 44 TB)?
ePMR (energy-assisted PMR) is the evolution of perpendicular recording with energy assistance, supporting the 36 TB planned for mid-2026. HAMR (heat-assisted MR) uses heat to increase density and enable 44 TB in the second half of 2027. WD postpones HAMR until it can produce millions of units per quarter with reliability comparable to ePMR.
What is UltraSMR, and when does it make sense over traditional CMR or SMR?
UltraSMR is a recording scheme that adds roughly 20% capacity over CMR and about 10% over standard SMR. It’s particularly advantageous for read-heavy workloads where increased TB per unit outweighs the cost of less frequent writes.
What investments does WD plan in Japan, and why are they important for HDD innovation?
WD invested over $1.5 billion with suppliers in Japan during fiscal 2025, and plans to invest $1 billion in R&D over the next five years. The Fujisawa center — legacy of HGST and IBM — collaborates with universities and the NIMS (National Institute for Materials Science) to develop spintronics for more sensitive read-heads, crucial for higher densities while maintaining reliability. WD emphasizes its history of innovation — including helium sealing drives now standard industry-wide — as proof of its leading role in future developments.
via: PC Watch