Compute is moving into rented infrastructure. Data is moving out of it. The bill lands on the link between them.
Two numbers landed within two weeks of each other this summer, and they point in opposite directions. On 11 August 2026, CoreWeave reported second quarter revenue of $2.575 billion, up 112 percent year over year, with a contracted revenue backlog of roughly $104 billion. Enterprises, frontier labs and hyperscalers are renting GPUs faster than anyone can rack them.
Meanwhile, a February 2026 survey of 212 senior IT decision makers, commissioned by Cloudian and run by Centiment, found that 75 percent had moved at least some workloads back out of public cloud in the previous 24 months, and 89 percent planned to grow their on-premises footprint over the next two years.
Read separately, these look like contradictory trends: one says the cloud is winning, the other says it is losing. Read together, they describe something more specific and more expensive. Compute is moving into rented infrastructure. Data is moving out of it. We call this the Great Divergence, and the cost of it shows up in one place: the network link between where your GPUs live and where your data lives.
For fifteen years, the default cloud model bundled compute and data in the same place. You put your data in Amazon S3 and ran your EC2 instances next to it. The hyperscaler owned both, and in-region transfer between them was free. That bundle is now coming apart, and it is coming apart in opposite directions.
GPU compute is flowing toward rented capacity because it is capital intensive, fast depreciating and wildly uneven in demand. Data is flowing toward owned or neutral storage because it is steady, compounding and increasingly regulated. The two flows are rational on their own. The problem is that AI workloads need both in the same place at the same time.
The scale of the compute side is not in dispute. Synergy Research Group put enterprise cloud infrastructure spending at $143 billion in Q2 2026, up 43 percent year over year and the fastest growth rate in eight years. Generative AI specific cloud services grew 165 percent, and nine GPU-focused neocloud operators now rank among the top 40 cloud providers by revenue.
| Factor | GPU compute | Enterprise data |
|---|---|---|
| Direction of travel | Into rented cloud | Out to owned or neutral storage |
| Demand pattern | Spiky, launch driven | Steady, compounding |
| Asset lifespan | About 18 months | Years |
| Main driver | Capital cost | Cost and sovereignty |
The economics of owning GPUs have become financial engineering rather than procurement. Sunny Smith, founder and CTO of Massed Compute, told us on Episode 7 of the DataStorage.com Podcast that a single B300 server costs around $750,000. Few enterprises want that on the balance sheet when the next generation ships within 18 months and utilization swings with every model release.
Three forces keep pushing GPU demand toward rental:
The result is that GPU placement has become fluid. A team might run training on CoreWeave, fine-tuning on Lambda Labs and production inference on a hyperscaler, and move any of those next quarter if price or availability shifts.
Data has the opposite profile. It accumulates every month, it rarely shrinks, and a growing share of it carries sovereignty and compliance obligations. That makes it the workload where cloud pricing hurts most and where ownership pays back fastest.
The cleanest public example remains 37signals, maker of Basecamp and HEY. Its CTO, David Heinemeier Hansson, disclosed spending just under $1.5 million a year on S3 to host its files. The company bought 18 petabytes of Pure Storage capacity for about $1.5 million, which it expects to cost under $200,000 a year to run, a projected saving of $1.3 million annually. AWS waived roughly $250,000 in egress fees for the exit under its 2024 policy for customers leaving the platform.
Two caveats keep this honest. Data Center Dynamics noted that the savings math does not fully account for hardware refresh, added operations staff, power and cooling. And 37signals bought its flash before the 2026 NAND price shock, which matters a great deal, as we cover below.
Cost is not the only driver. In the Parallels Cloud Survey 2026, 84.2 percent of respondents reported concerns about compliance, residency requirements and risk tied to foreign data centers, and about 87 percent showed some intent to move part of their workloads away from public cloud this year. Flexera's 2026 State of the Cloud Report found 73 percent of organizations now run hybrid cloud, up three points from 2025.
A note on sources: Cloudian sells on-premises object storage and Parallels sells desktop virtualization that runs on private infrastructure, so both have a commercial interest in repatriation headlines. The direction is consistent with Flexera's independent numbers, but treat the exact percentages as vendor survey data, not census figures. Even the most aggressive surveys show very few organizations leaving public cloud entirely. This is selective relocation of data, not an exodus.
Here is the part neither trend story covers. A rented GPU is only useful when it can read your data at full speed. If the data now sits in your colocation cage or a sovereign region, and the GPUs sit in a neocloud three states away, every training run, every checkpoint restore and every retrieval index rebuild has to cross that gap.
Smith put it bluntly on our podcast: storage has to sit next to the GPU with high east-west bandwidth, and moving a petabyte between data centers is prohibitive. That is data gravity, and the Great Divergence makes it worse, because the two halves of the stack are now chosen by different teams on different contract cycles.
Take a 500 TB training and retrieval corpus that gets pulled in full to a GPU provider four times a year, which is a realistic cadence for a team that moves capacity between providers as prices and availability change. At AWS list prices in US East, internet egress runs $0.09 per GB for the first 10 TB in a month, stepping down to $0.05 per GB above 150 TB. A single full pull of 500 TB costs about $29,491 in egress. Four pulls cost about $117,965 a year, on top of roughly $135,782 a year to store the data in S3 Standard.
The same corpus in a zero or low egress object store changes the picture entirely. Backblaze B2 lists $6.95 per TB per month with free egress up to three times average monthly storage, and unlimited free egress to partners including CoreWeave, Vultr and phoenixNAP. Wasabi lists $7.99 per TB per month from July 2026, with no egress or API fees under its fair use policy. CoreWeave's own AI Object Storage charges no egress or request fees inside its platform.
| Storage option | Storage / yr | Egress / yr | Total / yr |
|---|---|---|---|
| AWS S3 Standard | $135,782 | $117,965 | $253,747 |
| Backblaze B2 | $41,700 | $0 | $41,700 |
| Wasabi | $47,940 | $0 | $47,940 |
Our model puts the annual total at about $253,747 on S3 Standard with four pulls, versus about $41,700 on Backblaze B2 and $47,940 on Wasabi. The storage line is the smaller part of the gap. The egress line is what turns a cheap GPU decision into an expensive one, which is exactly the trap behind the claim that a cheaper GPU hour is not a cheaper bill.
Repatriation business cases written in 2024 and 2025 assumed flash prices that no longer exist. TrendForce reported that enterprise SSD contract prices rose roughly 80 percent in Q1 2026, and mainstream NAND contract prices climbed another 70 to 75 percent in Q2. Revenue for the top five enterprise SSD brands roughly doubled quarter over quarter to $37.59 billion in Q2 2026.
VDURA's Flash Volatility Index, updated in August 2026, priced a 30 TB TLC enterprise SSD at $22,600 against $1,216 for a 30 TB nearline hard drive, a ratio of 18.6 to 1. VDURA sells hybrid flash and disk systems, so it benefits from that framing, but the direction matches TrendForce. Smith told us NVMe prices have roughly tripled and he expects scarcity to run to the end of 2027.
The practical consequence: an all-flash, on-premises copy of your AI corpus now costs several times what the same build cost a year ago. Repatriating the system of record onto dense disk with a modest flash tier still pencils out. Repatriating everything onto NVMe to feed GPUs that are not even in your building usually does not.
From 12 January 2027, the EU Data Act prohibits cloud providers from charging switching fees, including egress charges tied to moving to another provider. AWS already tells EU customers on its S3 pricing page that they can request reduced data transfer rates for eligible use cases under the Act. That will make one-time exits cheaper, and Q4 2026 is the right moment to review any contract that auto-renews past that date.
But the Act targets switching as a one-off event. Egress for ongoing, parallel multi-cloud use remains chargeable. The Great Divergence architecture, with data in one place and rotating GPU providers in another, is ongoing multi-cloud use by definition. Regulation lowers the cost of leaving. It does not lower the cost of commuting.
The goal is not to stop either flow. Renting GPUs and owning data are both sound decisions. The goal is to design the seam between them deliberately instead of discovering it on the invoice.
Separate the system of record (raw data, compliance archives), the working set (curated training corpus, embeddings, retrieval indexes) and the ephemeral layer (checkpoints, scratch). Only the working set needs to sit next to GPUs. The system of record can live wherever ownership and sovereignty are cheapest.
Hold the working set in S3-compatible storage that does not tax reads and has direct network paths to multiple GPU providers. That keeps your GPU choice reversible. If moving the working set to a new provider costs nothing but time, you can actually exercise the leverage that a shifting GPU supply market gives you.
When you evaluate a GPU provider, ask three things before price: do you own the GPUs, which object stores have free or private network paths into your data centers, and what does east-west bandwidth between storage and compute look like at our cluster size. A cheap GPU hour behind a slow or metered data path is not cheap.
For each workload, model storage cost per month, full-pull egress cost per move, expected moves per year and the GPU hourly saving that motivates the move. If annual egress exceeds the GPU saving, the move does not pay. Rerun the model each quarter, because both GPU rental prices and flash prices are moving fast in 2026.
Keep GPU commitments short to capture falling prices and new capacity. Commit longer on storage, where price stability matters more than flexibility. And if you have EU exposure, time any planned storage migration to land after 12 January 2027.
Rent the GPUs, own the data, and design the seam between them on purpose. Otherwise the seam becomes your biggest line item.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
