Every photo you’ve uploaded to the cloud, every show you stream, every file sitting in your inbox—where does it actually live?
Most people picture something out of the movies: racks of SSDs, optimized for performance.
The reality is different. And more revealing.
At scale, the cloud is still built largely on spinning disks. That becomes even more true in AI systems, where data volumes grow continuously over time. The same basic technology your parents’ old desktop had that was invented in 1956—back when a computer took up an entire room and IBM thought a 5MB hard drive was a marvel.
Today, precision-engineered and nano-fabricated hard drives are still the bedrock of storage of cloud and AI. It isn’t an accident. It is because the storage systems are designed for scale. And scale matters a whole lot.
Scale breaks intuition
Think about how you’d pick storage for your laptop. You’d likely go for an SSD. They’re faster, they can be more resistant to shock, and the price difference versus a hard drive is much more manageable when you are only buying half a terabyte. A good decision balancing performance and cost.
When an enterprise is picking storage for a critical database, they would still likely choose an SSD. Databases typically live and die on fast little read-and-write operations, which is SSD’s primary advantage over HDDs. Also, a good decision.
But at hyperscale, the problem shifts.
You are no longer thinking in terabytes, you are thinking in exabytes, 1,000,000 times bigger than a laptop and 1000 times bigger than a typical database. A typical cloud and AI storge system has in fact many hundreds of exabytes. At that scale, economics become architecture.
Amazon S3 alone stores more than 500 trillion objects across hundreds of exabytes of data—and serves more than 200 million requests every second, globally1. At that size, a one-cent price difference per gigabyte works out to $10M (yes, ten million dollars) across a single exabyte. Across a hundred exabytes, it’s a billion. Suddenly “faster” matters a lot less than “cheaper per terabyte” and because you have thousands of HDDs in a cluster, you spend money on solving performance through software.
What looks like a marginal decision at small scale becomes the difference between a sustainable business and one that simply doesn’t work economically and is bankrupting itself on hardware.
Without this balance of cost and performance, the cloud storage foundation goes away, and the cloud economy becomes more fragile.
You may ask, what does this have to do with AI—is that different? So, let’s take a look.
AI is a data system, not a single workload
AI isn’t one thing—it’s a sequence of workloads working together as part of a data system.
It starts with data ingestion, bringing data in from the real world. Then, different compute tasks do the preparation, cleaning, and organizing of that data so it can be used. Then training—where models actually learn from that data. And ultimately, inference—where those models are applied in real time. That’s the point where AI becomes economically meaningful, justifying the infrastructure investment.
But the important part is what sits underneath.
All these workloads depend on data stored in object stores and parallel file systems. That’s the foundation the system is built on.
And these systems aren’t static, they’re continuously learning. Every interaction, every model update, every new dataset adds to the continuously growing data footprint.
Multiply that across billions of users, and the scale becomes very real, very quickly.
AI is creating a new layer
If this is all true, why is flash growing so quickly? AI needs the bulk object stores but it also needs something new that the cloud did not need.
AI doesn’t just store data—it needs to understand it. And that requires a different kind of representation.
Take a simple example: a photo stored in an object store. The system doesn’t “look” at the image every time you search for it. Instead, it creates a compact mathematical representation—a vector—that describes what’s in the image: a dog, a beach, a sunset.
At scale, this creates an entirely new dataset—one that grows alongside the original data. Every image, every document, every video can generate multiple vectors.
These vector datasets behave very differently from the underlying data they describe. They are small, accessed frequently, and queried in highly random patterns during inference. They’re small—but they’re hit constantly.
That makes them a natural fit for SSDs.
This is where much of the growth in flash demand is coming from2—not replacing hard drives but serving a new layer of the stack that didn’t exist before.
The more data you store, the more vectors you generate—and the more effectively AI systems can retrieve and reason over it. The more capable the AI model, the more content is created, which drives more data, which drives more vectors.
Storage growth is no longer linear, growing with real-world data only. It compounds real-world and generated data together.
The system is designed around the at-scale workload
On your laptop, files live in folders, folders live inside other folders, and the whole thing is a neat hierarchy. That works great when you’ve got a few thousand files. It starts falling apart at a few billion. And at a few trillion, it’s completely hopeless.
So, the cloud doesn’t use folders at all. Instead, it uses the object stores mentioned earlier. They are a giant flat pool where every item has a unique name, and you just ask for things by name. And crucially, you can’t edit things in place. If you want to change a file, you replace the whole thing in big chunks.
Hard drives struggle when the drive head must jump around to different parts of the platter to read and write. But they shine at large, sequential reading and writing of data in big chunks, especially higher-capacity SMR drives. By design, object storage aligns naturally with that behavior—you could say they were designed for it. You’re not modifying files in place; you’re writing and retrieving complete objects.
So, the infrastructure and the workload evolve together. This is not by accident, but by co-design. At scale, infrastructure, economics, and workload are inseparable; the system is designed as a whole.
In the data center, every tier of data memory/storage serves a different need—SSDs, HDDs, and tape each do what the others can’t. The system works because they’re not competing for the same role and occupy different spaces. They’re teammates.
And this co-design is a direct consequence of designing at-scale and for affordable cost.
The economics can’t be ignored
At hyperscale, even small differences in cost per terabyte become significant.
When you’re dealing with exabytes of data, the gap between storage technologies translates into very large numbers. And for most workloads—especially those designed around object storage—the incremental performance benefits of all-flash don’t justify the cost.
Using VDURA’s Flash Volatility Index3, in Q1 2026 a 30TB TLC SSD cost $17,500 ($583.33 per TB) and a 30TB HDD cost $668 ($22.26 per TB). Now stretch that across a single exabyte. You’re looking at a cost difference of $5617,070,000 for 1EB, yet hyperscalers have 100s of EBs.
Despite years of industry predictions that SSD pricing would close the gap, the evidence points the other way. Using the same index, Q2 2025 SSDs were 4.9x more expensive than HDDs per TB. By Q1 2026 that figure had grown to 22.6x. The gap isn’t closing, it’s widening.
Put it another way: cloud storage has become economically viable because of hard drives, not despite them. If the industry had bet on SSDs for bulk storage, we’d all be paying vastly more for cloud services, including our AI subscriptions—or those services simply wouldn’t exist in the form we know them.
The growth of SSDs doesn’t contradict this reality, it reinforces it.
Flash hasn’t become dominant by matching hard drive economics. It has grown by finding workloads where its performance characteristics justify the cost—particularly in AI systems that require fast, repeated access to compact representations of data.
Takeaways
The question of whether SSDs are “replacing” HDDs assumes they’re competing for the same role. At cloud and AI scale, that’s not how the system works. They’re complementary technologies, each doing what it was designed to do well, especially in AI systems that require both high-performance access and cost-efficient bulk data retention.
SSDs are not replacing hard drives—they are expanding the system. AI has introduced a new layer, where flash excels, built on top of a foundation where hard drives remain unmatched. SSDs focus on fast-access, low-latency, highly transactional workloads. And by doing that, they allowed HDDs to focus on what they uniquely deliver: storing enormous volumes of data, reliably and efficiently, at a cost structure that makes the modern internet and AI viable.
Those 500 trillion objects in Amazon S3 aren’t on spinning HDDs despite the cloud era. They’re on spinning HDDs because of it. Every photo upload, every backup restore, is built on an economic and engineering reality: a magnetic recording technology that continues to evolve, to earn its place because it solves a problem no other technology can at that scale.
This isn’t nostalgia, it’s pragmatic at-scale system design.
As AI systems continue to generate and retain ever-growing volumes of data, the technologies that can store that data efficiently at scale become foundational.
The future of AI infrastructure won’t be defined by a single technology—but by how effectively these systems are designed to use each technology in its sweet spot, together.
- Twenty Years of Amazon S3 and Building What’s Next. https://aws.amazon.com/blogs/aws/twenty-years-of-amazon-s3-and-building-whats-next/
- Generative AI spurs new demand for enterprise SSDs. https://www.mckinsey.com/industries/semiconductors/our-insights/generative-ai-spurs-new-demand-for-enterprise-ssds
- VDURA Flash Volatility Index. https://www.vdura.com/flash-volatility-index-and-storage-economics-optimizer-tool/
