Definitions first, because this is a field where the unit itself is a marketing choice. One exabyte (EB) is 1018 bytes — 1,000,000 terabytes, or 1,000,000,000 gigabytes, using the decimal convention every cloud bill uses. (Binary-purists' exbibyte is 15% larger; nobody bills you in it.) Why should a compute buyer care about a unit this size? Because it is now the operating unit of the industry: Seagate alone shipped 220 EB in a single quarter, and AI is the demand engine behind it — the company's CEO, Dave Mosley, attributes the "multi-year growth phase" to data that is created, retained and reused, with the shift from training to inferencing driving large-capacity drive demand for everything from model checkpoints to inference datasets.

What does an exabyte cost per month at retail?

Three rate cards bracket the retail market. At the top, Amazon S3 Standard charges $0.023 per GB-month for the first 50 TB in US East (N. Virginia), stepping down to $0.022 for the next 450 TB and $0.021 above 500 TB (AWS's published on-demand rates). At exabyte scale you get the volume tier: 0.021 × 109 GB = $21M per EB per month; the headline $0.023 rate gives the $23M figure most buyers actually experience at petabyte scale. At the bottom, S3 Glacier Deep Archive is $0.00099 per GB-month — about $1M per EB per month — with a 180-day minimum storage duration and retrieval fees on the way out. In between sits the value tier: Backblaze B2 at $6.95 per TB-month (Backblaze raised it from $6 effective 1 May 2026 and made standard API calls free in the same change), which is $6.95M per EB per month.

LayerRatePer TB-monthPer EB-monthCatch
S3 Standard (rack rate)$0.023/GB$23$23MFirst 50 TB tier; $0.021 above 500 TB
Backblaze B2$6.95/TB$6.95$6.95M3× monthly storage egress free, then $0.01/GB
S3 Glacier Deep Archive$0.00099/GB$0.99≈$1M180-day minimum, retrieval fees, hours to restore
Media capex, hyperscale (derived)≈$19/TB shipped, 5-yr life≈$0.32≈$0.32MOur arithmetic on Seagate FQ1 2027; excludes power, replication, staff

Every row of that table is verifiable from a vendor's own page or filing, and the spread between its first and last row is roughly 70×. Nothing else in the infrastructure stack — not GPUs, not networking, not power — shows a 70× gap between the list price of the service and the commodity cost of the physical input. Memory came close this year (our DRAM shock analysis tracked server DDR5 contract prices up 93–98% in a quarter), but that was a shortage moving everyone's price up together. The storage spread is permanent architecture, not a shortage.

What does it cost to get the data back out?

The storage bill has a second meter that only shows at the exit. AWS charges $0.09 per GB for internet egress after the first 100 GB a month (published S3 data-transfer rates). Read one exabyte out over the public internet and the toll is 0.09 × 109 = $90M — about four months of hot storage rent for the same capacity, charged for the privilege of leaving. Backblaze's model is the counter-offer: egress up to 3× your average monthly storage is free, and the overage is $0.01/GB — about $10M per EB, a 9× difference on the identical byte. This is why "storage is cheap now" is a half-truth: the storage is cheap, the departure is not, and any workload with a second home (training on one cloud, inference on another, a migration, a regulator's data request) prices egress long before it prices rent. The archive tiers add a third meter — retrieval fees on top of egress — plus minimum-duration charges if you delete early. Glacier Deep Archive's 180-day minimum is not a detail; it is a commitment device that makes "cheap" storage a contract.

What does the industry actually pay at the loading dock?

The cleanest public window into the dock price is Seagate's own filing. Its fiscal Q1 2027 (quarter ended 2 October 2026, reported 27–28 October) shows revenue of $4.233B, up 61% year over year from $2.629B, with total shipments of 220 EB (up 21% year over year, per the earnings call) and a record 198 EB delivered to data center customers — about 90% of volume. Divide one by the other and the company booked about $19.2 of revenue per shipped terabyte. A year earlier the same arithmetic on its disclosed figures gives $2.629B ÷ 182 EB ≈ $14.4 per TB — a 33% year-over-year rise in revenue per shipped terabyte, which is mix shift and price increases compounding (data center share rose, capacity per drive rose, and management describes a deliberate "value-based pricing strategy"). The gross-margin line confirms it independently: 57.3% GAAP versus 39.4% a year ago — for a hardware company shipping commodity-spinning rust, that is a price story, not a cost story.

Now convert capex to a monthly rate, which is the unit a buyer actually thinks in. Assume a five-year service life — consistent with the refresh cycles drive vendors plan around — and $19.2/TB of media amortizes to about $0.32 per TB-month, or $0.32M per EB-month. Be honest about what that number excludes: drive enclosures, servers, networking, the data center shell, power (a hard drive draws 5–10 W around the clock), replication and erasure-coding overhead, technicians, and Seagate's own profit margin, which is at a record high precisely because supply is disciplined — the company says the vast majority of its nearline exabyte supply is allocated under long-term agreements through 2028, with some cloud customers extended to 2030, and guided fiscal Q2 2027 to $4.8B ± $0.1B. Real deployed storage costs multiples of the media number. But the ordering does not change: even multiplied several times over, owning bytes at the dock is an order of magnitude below renting them at retail — which is exactly why the companies with exabyte fleets build their own, and why Backblaze can sell at $6.95 where AWS charges $23.

The scarcity side deserves its own sentence: this is a supply-managed market, not a glut. Seagate's Q2 guidance of $4.8B implies another ~10% sequential revenue rise, capacity per drive is climbing (Mozaic 4 at 44 TB ramping, Mozaic 5 above 50 TB planned for 2027-end qualification), and the CFO's stated plan is to cover rising component costs with further price increases. Every layer of the AI stack in 2026 has learned the same lesson — HBM displaced DDR5 (the memory bill), transformers queue for years (128 weeks from PO to power) — and storage is now pricing like the rest of them: allocation, long-term agreements, and list prices that only go one way.

Why does the 20–70× spread survive competition?

Because each tier sells a different promise, and the expensive tiers are expensive for reasons a disk cannot erase. Durability: 11 nines of design durability means redundancy — erasure coding across many drives and sites — so one stored TB consumes 1.3–1.5 TB of raw media before anyone flips a switch. Latency: "always-hot" storage keeps data on spinning media that answers in milliseconds; Deep Archive keeps it on tape-class economics that answer in hours. Operations: power, real estate, drive-churn logistics (a 24 TB drive fails on schedule, somewhere, every minute of a fleet's life), and staff. And margin: the top tier prices what the market pays for not thinking about storage. The spread is the market's honest accounting of how much of that burden you want to own versus rent. What has changed in 2026 is where the boundary sits: Backblaze's own $6.95 list price is still about 20× the derived media cost, meaning even the aggressive value tier captures the operations-and-durability multiple at scale — and the May 2026 price rise shows the input-cost wave (component prices, per the same logic as DRAM) pushing the whole ladder up, not just the top rung.

For the buyer, the honest framing is: the 70× headline is a ceiling illustration, not a procurement plan. You will not run your training dataset on Glacier Deep Archive, and you will not build an exabyte fleet to save 70%. The usable insight sits in the middle of the table: at petabyte scale — where most serious AI teams actually live — the difference between the $23M/EB and $7M/EB layers is $16M per month, which buys a lot of GPU-hours, and the difference between those layers and the dock is why every hyperscaler internalizes storage before it internalizes anything else.

What does this mean if you are buying compute?

Price storage where the compute sits, then never move it. The egress arithmetic makes data heavier every year it accumulates: an exabyte that costs $23M a month to rent also costs $90M to read out. That asymmetry is the mechanism behind data gravity — once your checkpoints, embeddings and video corpus live next to a particular GPU fleet, the cost of testing a second fleet includes the toll of moving or re-generating the data. Ask any prospective GPU provider for their storage line item and their egress policy in the same breath as their hourly rate; a cheap GPU-hour attached to an expensive byte is the same trap as a cheap GPU attached to an idle-hour.

Size the checkpoint bill before it sizes you. Our checkpoint tax analysis put a 405B-parameter run at 6.5 TB per save and one save every ~13.6 minutes at the optimum — and checkpoint retention is a pure storage-tier decision. Training keeps the last few checkpoints hot (Standard), keeps the daily lineage warm (value tier), and should push anything older than a restart-horizon into archive with an explicit retention window — remembering that Deep Archive's 180-day minimum turns "we might delete it sooner" into a fee. The storage tier is not an infrastructure detail; it is a training-reliability parameter, because a job that loses its checkpoint budget to a retrieval queue is a job that stops.

Read a storage quote the way you read a GPU quote. The discipline is the same one that built the agent token bill: decompose the line into its meters. Storage has four — capacity-month, request/API count, retrieval, and egress — and vendors quietly re-weight them. Backblaze made API calls free in May 2026 while raising capacity 16% ($6 → $6.95); AWS keeps headline storage unchanged while egress at $0.09/GB funds the margins. Neither is a scam; both are price structures. The winning question at exabyte-adjacent scale is not "what does storage cost" but "what does it cost to keep, what to touch, and what to leave" — and in 2026 those three numbers are $23M, $90M and zero-for-so-long-as-you-stay, respectively, per exabyte per month. Storage is no longer where data goes after the interesting part; it is the fourth bill, it is compounding, and it has the widest spread of any line on the invoice.