AI
Same Price, Different Bill: What the AI Storage TCO Calculator Reveals
You did the homework and found multiple storage vendors that looked good. You got quotes, and the per-TB rates were all within pennies of one another. So you picked a winner, procurement checked the box, and the contract got signed. Your team was thrilled to finally put AI projects into production.
Then the invoices started arriving, and they looked nothing like the quote.
The rate covered storage, but that wasn’t the full picture.
A price-per-TB quote told you what it would cost to keep data still, and AI data almost never sits still. What happens after your data lands is what writes the real bill.
Storage isn’t billed like a commodity
Most cloud storage bills run on several meters at once. Capacity is the first and most visible, the price per TB per month that everyone compares. On top of that, you pay per request to write data (PUT) and read it (GET), per GB to read from cooler tiers (retrieval fees), and per GB to move data out of the provider’s network (egress). A single read can trigger all three read-side charges. None of those appear in the headline rate.
All of them appear on the invoice.
Roughly half of public cloud storage spending goes to fees rather than capacity, according to the 2026 Wasabi Global Cloud Storage Index. Backup and archive data rarely touch the extra meters. AI leans on every one of them.
Inside the calculator: What moves the number
Wasabi’s new AI Storage TCO Calculator is built to set better expectations before you sign a contract. It takes four short steps, and each one targets a different part of the bill.
Scope
Start with what you’re running (pretraining, fine-tuning, inference, retrieval-augmented generation, data labeling, or checkpoint logging) and the kinds of data involved.
Each workload touches storage differently. Inference reads constantly. Checkpointing writes in bursts. Your answers set the shape of everything that follows.
Dataset and storage
Next, you’ll need to consider size. How much data will you upload to start? What’s your average object size? And how fast will the dataset grow or shrink each year?
Object size is the one that is often overlooked. Every object is a separate request when it’s written and every time it’s read. Store 500 TB as 1 MB files and that’s 500 million writes. At an object size of 5 MB, it’s 100 million. Some cooler tiers also bill small objects as if they were 128 KB, however small they really are. Object size nuances can significantly change the way your AI solution will interact with its underlying storage, affecting your bill in the process.
Performance and data access
How much read throughput do your workloads need, and how often is the data read or listed each month?
This is where tier choices catch up with you. Cooler tiers look cheap on a quote because they charge on the way out, with retrieval fees on every gigabyte read and higher rates on every request. Cooler tiers look favorable for lower touch use cases but that changes when agents need constant access to your data.
Data movement, egress, and retention
Finally, where do your GPUs live, and how much data crosses out of your storage provider’s network each month to reach them? For this step, it’s important to consider how many API calls your pipelines make per day or per run, and how long you keep data.
If your compute sits outside your storage provider’s cloud, every training run that pulls data across is an egress event.
Your workload defines your report
The combination of these calculation layers results in a personalized report with a line-by-line comparison against leading hyperscaler storage, a side-by-side look at what it costs just to load your data, and a 3-year total cost of ownership projection.
Keeping GPUs ‘fed’ with a steady stream of data is a top concern right now. Knowing how your data moves can decide which provider is best suited for your use case.
Same rate, different year: A 500 TB comparison
Picture an enterprise training a computer vision model on a 500 TB dataset of 1 MB images. Two quotes come in. Wasabi’s is $7.99 per TB per month. The other is typical of most vendors: 10% of the data in a hot tier at about $20 per TB, 40% in a cool tier at $10, and 50% in a cold tier at $4. That blend works out to $8.00 per TB per month.
On paper, the quotes are just a penny apart.
Now run the workload. Over the year, training, validation, and evaluations read the full dataset 4 times, for 2,000 TB of reads. Training runs also ship 300 TB, about 60% of the dataset, out to a GPU cloud.
Year one, side by side
Cost | Most Vendors* | Wasabi |
|---|---|---|
| Storage (the rate on the quote) | $48,000 | $47,940 |
| Writing the data in (PUT requests) | $7,250 | $0 |
| Reading it back (GET requests) | $10,880 | $0 |
| Retrieval fees | $28,000 | $0 |
| Egress to compute | $27,000 | $0 |
| Total, year one | $121,130 | $47,940 |
*Numbers are not tied to a specific vendor. They represent typical costs and features drawn from a combination of multiple storage vendors.
The storage lines match. Everything below them is the workload. Loading the dataset costs $7,250 before a single model trains. Reading it back adds $10,880 in requests and $28,000 in retrieval fees, because 90% of the data sits in tiers that charge to be read. Moving it to compute adds $27,000 in egress.
The quote that seemed to match Wasabi’s actually ends the year at $121,130. That’s about 2.5x more, and fees make up 60% of it.
In Wasabi, the same workload costs $47,940. There are no fees for egress, API requests, or retrieval, so what you initially saw as your quote actually ends up being your bill.
It’s important to note that this comparison is generous. It uses conservative fee rates, and it starts from a quote that leans on cheaper tiers to match Wasabi’s price. For some teams, it may not be acceptable to keep data in cold tiers. Keep the whole dataset in a hot tier instead, and the storage line alone runs about 2.5x of Wasabi’s.
Why AI hits all the meters
Each AI behavior maps to a line on that invoice. Inference reads data constantly, which means GET requests and, in cooler tiers, retrieval fees. Retraining loops touch the same datasets every cycle, so those charges repeat.
Checkpoints write large volumes in bursts. Every time data travels to a GPU cloud, egress charges follow it.
Traditional storage was built around a simpler routine: back up, archive, and forget. Data went in and rarely came back out, so the price per TB was a fair stand-in for the bill. But AI triggers every meter, by design, and it never stops.
The pricing models didn’t change. The workloads did.
Cost per TB is a starting point, not a forecast
A per-TB rate is the price of standing still. It can’t tell you what happens when your data moves, and with AI, movement is the whole point. The only reliable way to know what you’ll pay is to model your own access pattern against each vendor’s fee structure before you sign, not after the first invoice arrives.
When the storage rate is the bill, teams can read, retrain, and move data to the best compute available without watching a meter.
Run your own numbers
The example above is illustrative. Enter your own dataset size, object size, access pattern, and compute location into the AI Storage TCO Calculator and see where your two numbers split.
The per-TB rate covers capacity only. Most hyperscaler bills also charge per request to write and read data, per GB to retrieve data from cooler tiers, and per GB for egress when data leaves the provider’s network. AI workloads read and move data often, so those charges can rival the capacity line.
Four common ones: PUT requests (charged per write request), GET requests (charged per read request), retrieval fees (charged per GB read from cool and cold tiers), and egress (charged per GB leaving the provider’s network). Some tiers also add minimum storage durations and minimum billable object sizes.
Backup and archive data is typically written once and rarely read. Active AI workloads read datasets repeatedly for training, validation, and evaluation, write checkpoints in bursts, and often move data to GPUs. Each behavior triggers a different charge: requests, retrieval fees, or egress.
It takes four short steps: workload scope, such as fine-tuning or retrieval-augmented generation; dataset size and average object size; read throughput and monthly access; and data movement, API call volume, and retention. It returns a line-by-line comparison against leading hyperscaler storage and a 3-year total cost of ownership (TCO) projection.
Yes. Requests are billed per operation, so smaller objects mean more requests for the same data. Storing 500 TB as 1 MB files means about 500 million write requests, compared with 100 million at 5 MB. Some cooler tiers also bill objects smaller than 128 KB as if they were 128 KB.
Related article
&w=1920&q=75)
Most Recent
Ransomware can send a hospital back to paper. Learn how immutable backups and tested recovery help healthcare IT teams restore faster without paying a ransom.
Wasabi and AvePoint explain how AI changes 3-2-1 backup, from why the isolated copy matters more to storage costs. Watch the Spicy Bytes episode.
Cyber resilience expert Jason Buffington explains multi-vendor backup strategy, disaster recovery testing, and defense-in-depth data protection.
SUBSCRIBE
Storage Insights from the Storage Experts
Storage insights sent direct to your inbox.