Storage is the last AI infrastructure layer to be repriced. GPUs went first. Networking followed. Power came third. Storage is being repriced now because it was the most boring and the most taken for granted. The 2000s internet buildout suggests boring infrastructure reprices last and recovers most durably.
Seagate reported fiscal third-quarter revenue of $3.11 billion, up forty-four percent year over year. Non-GAAP earnings per share hit $4.10, more than double the prior year. Non-GAAP gross margins reached forty-seven percent. Data centers accounted for eighty-eight percent of exabyte shipments and eighty percent of revenue. The stock rose more than ten percent in a single session. Bank of America raised its price target to $700. Rosenblatt doubled its target to $1,000.
Two days later, Western Digital confirmed this is not a single-company story. Revenue of $3.34 billion, up forty-five percent year over year. Gross margin of 50.5 percent — higher than Seagate's. Cloud revenue accounted for eighty-nine percent of total sales. The stock is up a hundred and thirty-four percent this year.
The market now values hard drive companies like semiconductor companies.
The Last Layer
In every technology cycle, the infrastructure layer furthest from the application gets repriced last. The AI buildout has followed this sequence precisely. GPUs were repriced in 2023 and 2024 — NVIDIA's market capitalization went from $360 billion to over $3 trillion as data centers scrambled for training compute. Networking was repriced in early 2025 — Arista and Ciena surged as bandwidth between GPU clusters became the constraint. Power was repriced in mid-2025 — nuclear restarts, gas turbine backlogs, and behind-the-meter generation all reflected the discovery that AI data centers need more electricity than the grid can deliver.
Storage is being repriced now, in 2026, because it was the least visible and the most taken for granted. Hard drives have been a commodity for decades — the same spinning platters from the same two companies, declining in price per terabyte on a curve everyone assumed would continue indefinitely. The assumption broke when AI data lakes started growing faster than storage density.
The Margin Breakthrough
The earnings beat is not cyclical. It is architectural. Seagate's HAMR platform — Heat-Assisted Magnetic Recording — enables drives above forty terabytes in volume production. HAMR uses a laser to momentarily heat a nanoscale spot on the disk surface, allowing data to be written on smaller magnetic grains than conventional recording permits. The result is more data per platter, which drives the margin expansion. This is not incremental improvement. It is a platform shift comparable to the transition from longitudinal to perpendicular recording in the mid-2000s.
The trajectory tells the story. Seagate's non-GAAP gross margins expanded from the mid-thirties to forty-seven percent across recent quarters, with guidance implying continued acceleration toward fifty percent. Western Digital guided fiscal Q4 gross margins to fifty-one to fifty-two percent. Both companies confirmed that nearline drive capacity is sold out through calendar 2026, with firm orders extending into 2027 and 2028.
The measurement unit is changing. Data center operators increasingly evaluate storage in terms of tokens served per watt consumed, not input/output operations per second or cost per terabyte. When the buyer's metric shifts from the storage industry's output to the AI system's output, the infrastructure layer has been reclassified. Storage is no longer a cost center. It is an AI productivity input.
The Duopoly
Seagate, Western Digital, and Toshiba account for more than ninety-five percent of global hard drive shipments. Seagate holds roughly thirty-one percent, Western Digital twenty-eight percent. Western Digital separated its NAND flash business into SanDisk in 2025, leaving it as a pure-play HDD company. The barriers to entry are extreme — HAMR manufacturing requires precision at the nanometer scale that no new competitor can replicate without a decade of investment. Unlike GPUs, where AMD competes and custom silicon from hyperscalers is growing, the storage market has no credible new entrant. The AI data lake creates demand. The duopoly sets the price.
The Precedent
The 2000s internet buildout followed the same repricing sequence, layer by layer. Applications crashed first — Pets.com, Webvan, and eToys vaporized in 2000. Equipment crashed second — Nortel laid off sixty thousand workers, JDS Uniphase took the largest goodwill write-off in corporate history, and more than five hundred billion dollars in telecom equipment investment evaporated by 2002.
Physical infrastructure crashed last. Eighty-five percent of fiber-optic cable remained dark through late 2005. Global Crossing went bankrupt in 2002. But the fiber did not disappear. It was acquired at pennies on the dollar by survivors. Crown Castle bought distressed tower assets while competitors panicked. Equinix fell to two dollars a share in January 2003.
Then bandwidth costs fell ninety percent from the fiber glut, and demand materialized on a longer timeline than investors expected. The boring physical layer — towers, fiber, data centers — became the backbone of the 2005 to 2020 mobile and cloud economy. Equinix trades above $1,100 today. Crown Castle built a $50 billion enterprise from the wreckage.
The pattern: each layer further from the application crashed later, recovered later, and recovered more completely. Applications went to zero permanently. Equipment largely went to zero permanently. Physical infrastructure crashed in price but the assets persisted, consolidated, and repriced upward when demand materialized. Storage is the tower company of the AI cycle — boring, duopolistic, physically scarce, and taken for granted until the constraint became visible.
What to Watch
Two predictions, each falsifiable. First, Seagate's non-GAAP gross margin will exceed fifty percent within two quarters. The trajectory from the mid-thirties to forty-seven percent, combined with Western Digital already guiding to fifty-one to fifty-two percent, suggests the entire sector is converging toward margins historically reserved for chipmakers. Second, at least one hyperscaler will disclose storage cost per token as a named metric in an earnings call by end of 2026. The measurement shift from bytes to tokens is underway — the question is when it becomes a reporting category.
The repricing sequence is not a coincidence. It is the market discovering, one layer at a time, that the AI infrastructure stack has no optional components. GPUs were obviously essential. Networking was obviously essential once training scaled. Power was obviously essential once data centers drew more than the grid could supply. Storage was not obvious — until every AI model needed a data lake larger than anything the industry had built before, and there were only two companies in the world that could fill it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)