It is reshaping the world through scale: the unprecedented scale of data and compute, and the scale of its impact on our industry and the global economy.
Some of that change comes from the sheer scale of capital. Microsoft, Oracle, Google, and Amazon now collectively hold $2.3 trillion in contracted backlog, up 16% in just the past quarter. Bank of America projects hyperscaler capex will rise 80% this year to $860 billion, with a path to $1.2 trillion in 2027. When capital is deployed at that scale, every architectural decision gets re-examined—including storage.
For most of the last decade, hyperscale storage followed a simple recipe: hard drives store the bytes, while a thin flash tier absorbs the hot data. That recipe is now changing rapidly.
Applications are getting hungrier for performance per terabyte. A recent CMU study with Google and Microsoft, drawing on production telemetry from two hyperscalers, illustrates a disruptive trend: IOPS-per-TB requirements are up roughly 25% over the last half-decade, with the sharpest rise occurring in the past three years as AI/ML analytics spread. Sixty-eight percent of workloads got warmer, and the trend is accelerating.
Although hard drive capacities keep growing, their performance density is moving in the wrong direction. A spinning disk delivers roughly the same IOPS it did years ago. The result is a linear decline in IOPS/TB going back more than a decade: each generation delivers less performance per byte than the one before.
With demand for performance density going up and HDD performance density going down, the gap must be filled by flash. At hyperscale economics, that means Quad-Level Cell (QLC) flash, and it is why QLC is suddenly in explosive demand.
What hyperscalers must do now
Once a hyperscaler commits to QLC at scale, three jobs become urgent: secure supply from as many QLC vendors as possible (since no single supplier can feed an exabyte-scale buildout); rapidly qualify QLC across a diverse and constantly changing set of workloads; and maximize efficiency, because at this scale every wasted bit of raw NAND represents real money.
That leads to the question hyperscaler storage teams are wrestling with: buy or build? The default choice is buying commercial SSDs, but these are fixed-function devices that present NAND as if it were a hard disk, locking in one set of trade-offs at manufacturing time. They sacrifice efficiency, struggle with shifting workloads, and force a long qualification cycle to restart with every new vendor and flash generation. Building custom drives in-house works eventually, but it demands years of investment across multiple NAND suppliers, flash generations, and hundreds of applications. The result is massive R&D spend and a significant opportunity cost.
Everpure enables a better way
There is a third option, and it’s at the core of our Hyperscale Solutions business: partner. Two of the world’s top hyperscalers have already made that choice, selecting us for their storage infrastructure. Our DirectFlash technology, with its software-driven architecture, serves as the engine powering our Hyperscale Solutions. Built flash-native from day one, DirectFlash exposes flash media’s full capabilities to maximize efficiency—never treating NAND like a spinning disk. A single architecture can qualify and run NAND from multiple suppliers, an advantage no SSD vendor tied to its own fabs can match. Because flash policies live in software rather than being frozen into firmware, the architecture adapts rapidly as workloads evolve, which matters more right now than ever.
Hyperscalers partner with us to co-design NAND management policies for their environment, drawing on Everpure’s fifteen-plus years of flash engineering and more than 1 million DirectFlash Modules shipped. They gain the control of building in-house along with the speed-to-market of buying. The benefits are concrete:
- Supply at scale. Our three strategic NAND partners together represent roughly 40% of global flash capacity—providing the most diversified supply available through a single architecture.
- Industry’s densest drives, 10-20% greater storage efficiency. 150 and 300 TB DirectFlash Modules deliver more capacity per rack than any SSD or HDD, and 10-20% more usable capacity from every raw bit of NAND.
- 2x longer media lifetime and 2–5x lower failure rates. Managing flash in software eliminates the error-prone complexity buried in black-box SSDs.
- Flexible, software-defined policies. Adapt rapidly to changing workloads instead of stranding usable capacity behind fixed functions.
- Faster qualification of new platforms. Maintains a consistent underlying architecture even as the media changes, shortening qualification cycles.
Consolidating hot and warm workloads on a common flash tier lowers TCO outright, and cold tiers will follow as flash economics allow. Replacing an equivalent disk footprint frees roughly 80% of storage power, unlocking enough energy capacity to power approximately 1,000 additional GPUs per exabyte of storage deployed. In an era where every megawatt is contested, storage efficiency is GPU capacity.
One architecture, everywhere
DirectFlash is a key building block of our storage systems portfolio, refined across every flash generation since 2015. The engineering innovation our hyperscale customers demand flows directly back into our enterprise portfolio, and vice versa.
Computing innovations spread. Cloud architecture started at AWS and Google and reshaped on-premises infrastructure everywhere, and we expect flash at hyperscale to follow the same path. The market has spoken, and we’re building for where it goes next.
