What topics does Aurora publish insights on?

Aurora publishes technical insights on private AI infrastructure — including GPU cluster deployment and management, S3-compatible object storage architecture, data residency and regulatory compliance for AI workloads, sovereign AI deployment, managed Kubernetes for AI, and the economics of operating AI outside hyperscaler platforms.

Who writes Aurora's technical content?

Aurora's content is produced by its engineering team — the same team that designs, deploys, and operates private AI infrastructure for enterprises, telcos, and governments. Aurora's engineering leadership includes one of the original architects of IPFS and Filecoin, joined by an EU-based team with decades of combined experience across data center builds, global infrastructure deployments, and sovereign AI systems.

What is the Aurora perspective on hyperscaler AI infrastructure?

Aurora's position is that hyperscaler infrastructure was designed to serve the hyperscaler's business model — not the organization running workloads on it. Egress fees, proprietary APIs, shared compute, and centralized data routing create structural constraints on AI economics, data sovereignty, and regulatory compliance. Private AI infrastructure is an architectural response to those constraints — not a political position.

What is sovereign AI and why does it matter?

Sovereign AI refers to AI infrastructure and workloads that operate within a defined jurisdiction — under local data residency requirements, local regulatory frameworks, and organizational control over encryption keys and access. It matters because regulations such as GDPR, the EU AI Act, and NIS2 create specific obligations around where data is processed and who can access it. Hyperscaler default architectures route data in ways that are difficult to constrain to a single jurisdiction. Private, in-region AI infrastructure is the practical solution for organizations with sovereignty requirements.

What are the economics of private AI infrastructure versus hyperscalers?

Private AI infrastructure delivers 50 to 80 percent lower total cost than hyperscalers at production scale. GPU pricing on Aurora runs 2 to 7 times below AWS, GCP, and Azure rates. Aurora Storage is priced at $5.99 per TiB per month with no egress fees — hyperscaler object storage generates additional billing events for every data movement to compute. At PB-scale data volumes and regular training or inference workloads, the cost difference between private and hyperscaler infrastructure is decisive, not marginal.

What does AWS S3 actually cost when you run AI against your data?

AWS S3 standard storage costs $0.023 per GB per month — approximately $23,500 per month at one petabyte. That is the storage line item. The costs that compound on top of it are: egress to the internet at $0.09 per GB, transfer between AWS services within the same region at $0.01 to $0.02 per GB, cross-region transfer at $0.02 per GB, S3 Glacier archive retrieval at $0.01 to $0.03 per GB, and API call charges above the free tier for PUT, COPY, POST, and LIST requests. Organizations running AI training or inference pipelines against large datasets find that data movement costs become a first-order budget item on top of storage.

How does hyperscaler GPU pricing compare to private infrastructure?

AWS p4d.24xlarge instances with 8x A100 GPUs run at approximately $32 per hour on-demand. H100 GPU instances on private infrastructure run at $2.50 to $3.00 per hour — 2 to 7 times cheaper than equivalent hyperscaler GPU capacity, before accounting for the egress costs eliminated by co-locating storage and compute on the same private network. At 1,000 GPU-hours per month — a modest inference or fine-tuning workload — the cost difference between $32 per hour and $2.75 per hour is $29,250 per month.

What does it cost to move a petabyte of data out of AWS S3?

Moving one petabyte out of AWS S3 to the internet costs $90,000 in egress fees alone at AWS's published rate of $0.09 per GB. This is not a hidden fee — it is disclosed in AWS pricing tables — but most organizations do not model it until the bill arrives. This structural cost makes migrating away from AWS S3 expensive once data volumes reach PB scale, which is why the choice of storage infrastructure is effectively an AI architecture decision.

How much does Aurora Storage cost compared to AWS S3 at petabyte scale?

At one petabyte, Aurora Storage costs $5,990 per month at $5.99 per TiB per month, with no egress fees, no API surcharges, and no retrieval penalties. AWS S3 standard storage at the same volume costs approximately $23,500 per month before a single byte moves anywhere. The savings on data movement between storage and GPU compute — which generates no billing event on Aurora's private network — typically exceed the storage savings by a significant margin on active AI pipelines.

What is hyperscaler storage lock-in and why does it matter for AI?

Hyperscaler storage lock-in occurs when an organization's data lives in a provider's storage layer and its AI workloads run on the same provider's compute — making every subsequent infrastructure decision constrained by that starting point. Switching providers at PB scale means paying egress fees on every byte moved. Hyperscaler APIs create additional switching costs if custom integrations are built on top of proprietary features. Private S3-compatible infrastructure removes the lock-in: full API compatibility means existing pipelines do not need to change, and no egress fees means data moves freely to any compute layer.

Why is the storage decision also an AI architecture decision?

The storage vendor determines where data lives, and where data lives determines what it costs to run AI against it. Organizations that store data in hyperscaler object storage and then try to run GPU training or inference against it encounter egress fees, cross-region transfer costs, and architectural constraints that grow with data volume. Organizations that build their data foundation on private, S3-compatible infrastructure with no egress fees arrive at the AI compute conversation in a fundamentally different position — data moves freely to compute, economics work at scale, and the architecture does not need to be rebuilt.

Aurora MarketingApr 23, 2026 9:30:37 AM3 min read

The Hidden Cost of Hyperscaler Storage (When You Start Running AI)

4:33

The bill arrives when you start using your data.

AWS S3 costs $0.023 per GB per month. That number looks fine in a spreadsheet. At one petabyte — which is not a large dataset for an organization running production AI — that is $23,500 per month. Still manageable. The problem is not the storage line item. The problem is everything that happens when you try to use that data.

Hyperscaler storage is cheap to write to. It is expensive to operate AI against. The pricing model is designed that way.

What you are actually paying for

Here is the cost structure nobody shows you on the getting-started page.

Egress out of S3 to the internet: $0.09 per GB. Run a 10 TB training dataset through a pipeline once, that is $900 in egress alone.
Transfer between AWS services within the same region: $0.01 to $0.02 per GB. Moving data from S3 to EC2 GPU instances costs money even when everything is in the same availability zone.
Cross-region transfer: $0.02 per GB. If your storage and compute are not in the same region — and they often are not — every training run crosses a billing boundary.
Archive retrieval: S3 Glacier charges $0.01 to $0.03 per GB to retrieve. Cold data that needs to warm up for a training run generates a retrieval bill before a single GPU instruction runs.
API calls above the free tier: S3 charges per PUT, COPY, POST, and LIST request above a threshold. High-frequency AI pipelines that read and write metadata constantly can generate meaningful API costs.

None of these are hidden fees in the sense of being undisclosed. They are disclosed in detail, in pricing tables that require careful reading to understand. The problem is that most organizations do not model them until the bill arrives.

Hyperscaler storage is cheap to write to. It is expensive to operate AI against. The pricing model is designed that way.

The GPU cost compounds everything

The storage costs are painful. The GPU costs are where the economics break down completely.

AWS p4d.24xlarge instances — 8x A100 GPUs — run at approximately $32 per hour on-demand. Reserved instances reduce that, but reserved GPU capacity requires 1 or 3 year commitments that most organizations are not ready to make when they are still figuring out their AI workload patterns.

H100 GPU instances on private infrastructure run at $2.50 to $3.00 per hour. That is 2 to 7 times cheaper than equivalent hyperscaler GPU capacity — before accounting for the egress costs you are no longer paying to move data between storage and compute.

At 1,000 GPU-hours per month — a modest inference or fine-tuning workload — the difference between $32/hr and $2.75/hr is $29,250 per month. Every month. That is before egress. That is before API fees. That is before the archive retrieval bills.

The lock-in makes it worse

The costs alone are enough to warrant a serious look at alternatives. The lock-in makes the decision more urgent.

Once your data lives in S3 and your AI workloads run on EC2, every infrastructure decision that follows is constrained. Moving data out of AWS costs money — at $0.09 per GB, moving a petabyte costs $90,000 in egress alone. That is not a switching cost in the traditional sense. It is a structural trap.

Private S3-compatible infrastructure removes that trap. Full API compatibility means nothing in your pipeline changes. No egress fees means your data moves freely to compute. No lock-in means your architecture decisions are yours to make.

What the alternative looks like

Aurora's storage pricing is $5.99 per TiB per month. No egress fees. No API surcharges. No retrieval penalties. Data moves freely between storage and compute on the same private network — no billing event, no cross-region transfer, no architectural constraint.

At one petabyte, that is $5,990 per month — versus *$23,500 for AWS S3 standard, before a single byte moves anywhere. The egress savings on a regular AI pipeline typically exceed the storage savings by a significant margin.

Start with a 1 TB free trial. Full S3 compatibility. No credit card required. The bill that does not arrive is often more persuasive than any pricing table.

*Pricing figures: AWS S3 pricing sourced from AWS public rate cards, April 2026. Aurora pricing current as published. GPU pricing indicative — confirm current rates. Figures are illustrative; actual costs depend on workload patterns and configuration.

Aurora Marketing