A petabyte used to be a hyperscaler problem. In 2026 it is a mid-size company problem.
One year of 4K video from a media team, a few rounds of model checkpoints, or a retained training corpus for a single foundation model fine-tune will get you there faster than most budget cycles anticipate. And the moment it arrives, a simple-sounding question lands on the CTO's desk: what does it actually cost to keep it?
The honest answer is that list prices for storing one petabyte in 2026 range from $11,880 a year to $258,600 a year before you move a single byte. Add realistic egress and the top of that range climbs past $352,092. The spread is 22x from cheapest to most expensive tier, which means the storage class and provider you pick matter more than any discount you will negotiate.
This guide prices one petabyte across the hyperscalers, the zero-egress alternatives, archive tiers, and buying the drives yourself, using list prices published or reported in 2026. Every model assumption is stated so you can swap in your own numbers.
For this analysis, one petabyte means 1,000 TB, or 1,000,000 GB, stored for a full year at US list prices with no committed-use discount. Requests and API calls are excluded because they depend entirely on object size and access pattern. Here is the storage-only bill.
AWS S3 Standard in US East costs $21,550 a month for a petabyte once its volume tiers are applied: $0.023 per GB for the first 50 TB, $0.022 for the next 450 TB, and $0.021 above 500 TB, according to AWS's published pricing as summarised by CloudZero in 2026. That is $258,600 a year. Google Cloud Storage Standard in a single US region lists at $0.020 per GB, or $240,000 a year. Azure Blob Hot on locally redundant storage is widely reported at about $0.018 per GB, or $216,000 a year.
The alternatives land far lower. Cloudflare R2 Standard is $0.015 per GB, which is $180,000 a year. Wasabi lists at $7.99 per TB per month, or $95,880 a year. Backblaze B2 lists at $6.95 per TB per month, or $83,400 a year. At the bottom sits S3 Glacier Deep Archive at $0.00099 per GB, which works out to $11,880 a year for a full petabyte.
Disclosure: Backblaze is a DataStorage.com partner. Its figures here are its own published list prices, and it is ranked by the same method as every other provider.
S3 Standard is the price most teams anchor on, and at petabyte scale the volume discount barely registers: you pay about 9 percent less per GB above 500 TB than for your first 50 TB. The real lever on AWS is storage class. S3 Glacier Deep Archive is 23 times cheaper than Standard per GB, but it comes with a 180-day minimum storage duration, retrieval measured in hours, and a standard retrieval fee of $0.02 per GB. Restoring the full petabyte costs $20,000 in retrieval fees alone, before any data leaves AWS.
Azure's Hot tier looks about 22 percent cheaper than S3 Standard, but the comparison is not like for like. Locally redundant storage keeps copies inside a single data center, while S3 Standard spreads data across at least three availability zones. Matching S3's resilience on Azure means zone-redundant storage, which costs more. Azure list prices also vary noticeably by region, so confirm yours in the calculator before you commit.
Google Cloud Storage Standard at $0.020 per GB undercuts S3 on paper. Its Archive class at $0.0012 per GB ($14,400 a year per petabyte) is slightly more expensive than Deep Archive and carries a 365-day minimum duration, twice AWS's. Google also moved prices in both directions in 2026, raising multi-region Nearline from $0.010 to $0.015 per GB and cutting multi-region Archive in the US and EU from $0.004 to $0.0024, according to Finout's 2026 pricing comparison.
The zero-egress and flat-rate providers compete on one simple idea: one hot tier, one price, and little or no charge to read your own data. At a petabyte, that model saves real money even before egress enters the picture.
Backblaze B2 charges $6.95 per TB per month on pay-as-you-go, with egress free up to three times your average stored data and $0.01 per GB beyond that, plus free egress through partners including Cloudflare, Fastly, CoreWeave, Vultr and phoenixNAP, per Backblaze's pricing page. Standard API calls are free. For throughput-heavy AI work, B2 Overdrive starts at $15 per TB per month with unlimited free egress, but requires a multi-petabyte commitment.
Wasabi charges $7.99 per TB per month with no fees for egress or API requests, per its pricing page, subject to a fair-use policy that expects monthly downloads to stay within your stored volume. Cloudflare R2 costs more per GB than either at $0.015, but charges nothing for egress to the internet and bills operations separately. R2 is the natural pick when a petabyte is actively served to users; B2 and Wasabi win on data that mostly sits.
IONOS has also pushed European object storage pricing toward the $5 per TB mark, a strategy its team walked through on Episode 3 of The DataStorage.com Podcast. For teams with EU data residency requirements it belongs on the shortlist.
Storage-only comparisons flatter the hyperscalers because they leave out the fee that actually decides the bill. Model a modest, realistic workload: the petabyte sits in hot storage and 10 percent of it, 100 TB, is read out to the internet each month by customers, partners, or a GPU cluster in another cloud.
On AWS, 100 TB of monthly internet egress costs about $7,791 at list rates: the first 100 GB free, then $0.09, $0.085 and $0.07 per GB as volume tiers step down. On Google Cloud's default premium network tier, the same 100 TB costs about $8,310. On R2, Wasabi and B2, it costs nothing, because 100 TB sits inside each provider's free allowance.
Annualised, the S3 petabyte now costs $352,092 a year and the Google Cloud petabyte $339,720. Backblaze B2 stays at $83,400. That is a 4.2x difference for the same bytes doing the same job. Egress is the toll booth of the cloud: ingress is free, and every attempt to use your data somewhere else is taxed.
| Provider | Storage / month | Egress / month (100 TB) | Total / year |
|---|---|---|---|
| AWS S3 Standard | $21,550 | $7,791 | $352,092 |
| Google Cloud Storage | $20,000 | $8,310 | $339,720 |
| Cloudflare R2 | $15,000 | $0 | $180,000 |
| Wasabi | $7,990 | $0 | $95,880 |
| Backblaze B2 | $6,950 | $0 | $83,400 |
One exception is worth knowing. Since 2024, AWS has waived internet egress fees for customers moving all of their data off AWS, after a request to support. The waiver covers a full exit, not the partial, recurring moves that multi-cloud and multi-GPU strategies generate, so it does not rescue the model above.
The hyperscalers held their headline storage prices steady this year. The low-cost providers did not. Backblaze raised B2 from $6 to $6.95 per TB per month in 2026 while making standard API calls free, and Wasabi's pay-as-you-go rate now starts at $7.99 per TB. Both run on hard drives, and hard drives are in the middle of a genuine shortage.
Western Digital told investors in February 2026 that it was "pretty much sold out for calendar 2026," with long-term agreements already signed into 2027 and 2028, as Tom's Hardware reported. VDURA's Flash Volatility Index, covered by Blocks and Files in August 2026, puts a 30 TB hard drive at $1,216, up from $495 a year earlier. A 30 TB TLC enterprise SSD now costs $22,600, making flash 18.6 times more expensive per terabyte than disk.
Treat the index with appropriate scepticism: VDURA sells hybrid flash and disk systems, so a series showing flash as unaffordable supports its pitch, and it has already revised its own historical figures once. Tom's Hardware also notes that buyers without long-term supply contracts pay spot prices above the index. But the direction is corroborated by the manufacturers themselves. The practical read for buyers is that the cheapest capacity on the market is priced off hard drives, and hard drive costs are still rising.
The shortage is also reshaping where AI clouds put their data. In June 2026 Backblaze signed a five-year, multi-exabyte agreement with CoreWeave worth about $335 million to supply HDD-based tiers inside CoreWeave AI Object Storage, according to Blocks and Files. Even the neoclouds are tiering: flash for the hot working set, disk for the petabytes behind it.
At petabyte scale, someone always asks whether owning the hardware is cheaper. Start with the drives. At VDURA's August 2026 index price of about $41 per TB, one usable petabyte with a typical 1.5x erasure-coding overhead needs 1,500 TB of raw disk: 50 drives of 30 TB, costing about $60,800. Spread over a five-year drive life, that is roughly $12,160 a year.
| Media | Price per 30 TB drive | Drives needed | Drive cost |
|---|---|---|---|
| 30 TB HDD | $1,216 | 50 | $60,800 |
| 30 TB TLC SSD | $22,600 | 50 | $1,130,000 |
Now compare: S3 Glacier Deep Archive stores the same petabyte, fully managed and replicated across availability zones, for $11,880 a year. The raw drives alone cost about the same as AWS's entire archive service, before a single server chassis, rack, power bill, network switch or engineer. For cold data, doing it yourself rarely wins in 2026.
The same build on TLC flash would cost about $1,130,000 in drives. That is why nobody should put a cold petabyte on NVMe this year, and why storage for AI workloads is now a tiering decision first and a vendor decision second.
Owning hardware can still win for hot, heavily read data where egress would otherwise dominate, especially in colocation next to your own compute. That is a repatriation model with power, space and staffing in it, and it deserves its own worked case rather than a drive-price shortcut.
GPU capacity in 2026 is bought where it is available: AWS this quarter, a neocloud the next. Every switch drags the dataset along. Moving one petabyte out of S3 to a GPU provider in another cloud costs about $53,791 in internet egress at list prices, and pulling it from Deep Archive first adds $20,000 in retrieval. A price advantage on GPU hours can disappear in a single migration.
This is data gravity in dollar terms. Sunny Smith, founder and CTO of Massed Compute, made the point on Episode 7 of our podcast: storage has to sit next to the GPUs, and moving a petabyte between data centers is prohibitive. The architecture answer is to land the canonical copy of large training data in storage with free or partner-free egress, then feed whichever GPU cloud is cheapest that month. That is not a sales angle; it is the only way multi-provider GPU strategies keep their savings.
Use this sequence before you sign anything or renew a commit.
A petabyte is cheap to store and expensive to move. Price the egress first, and the right provider usually picks itself.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
