Artificial intelligence is not just chewing through GPUs. It is devouring storage. And so, Seagate has pulled the wraps off its Mozaic 4+ platform, and the headline number is hard to ignore… 44TB per hard drive! Not theoretical. Not coming someday. Shipping in volume to two hyperscale cloud providers right now.
Let that sink in. Forty four terabytes on a single 3.5-inch drive.
The technology behind it is HAMR, short for heat-assisted magnetic recording. If you have followed storage over the years, you have probably heard about HAMR as the future that never quite arrived. Seagate is making it clear that future is here.
Mozaic 4+ is the company’s next-generation HAMR platform, and it is already qualified at scale inside hyperscale environments. That matters. Hyperscalers do not experiment with fragile tech. If it is in their racks, it has passed some very serious testing.
And Seagate is not stopping at 44TB.
The roadmap calls for scaling from today’s 4+TB per disk toward 10TB per disk. That opens the door to 100TB hard drives down the line. One hundred terabytes. On spinning rust.
While the AI conversation is usually dominated by NVIDIA GPUs and trillion-parameter models, there is a quieter reality underneath it all. AI runs on data – massive archives, training corpora, synthetic outputs, video, logs, snapshots, you name it.
Most of that does not live on expensive flash. It lives on high-capacity hard drives because the economics demand it.
Seagate says that in a one-exabyte deployment, Mozaic can improve infrastructure efficiency by about 47 percent compared to standard 30TB drives. The company estimates that could reduce required data center footprint by about 100 square feet and lower annual energy consumption by roughly 0.8 million kilowatt-hours.
In hyperscale math, those are not rounding errors.
Dave Mosley, Seagate’s chair and CEO, put it this way: “Data has become one of the most valuable assets for enterprises, fueling business insights, enhancing productivity, and enabling competitive advantage. As the foundation of modern data center infrastructure, data storage solutions are essential to manage ever-increasing data volumes and maximize returns on investments in today’s AI driven-world.”
That may sound like executive language, but the core point stands. AI models do not exist in isolation. They are trained on oceans of information, and those oceans keep rising.
Technically, HAMR works by using a tiny laser to briefly heat the disk surface during writes. That allows for much higher recording densities without sacrificing long-term stability. Seagate says its custom-designed and manufactured laser technology reflects years of investment in nanophotonic engineering and gives it tighter control over yield, reliability, and supply chain resilience.
Vertical integration is not just a buzzword here. In a world where component shortages can derail entire product lines, owning key pieces of the stack matters.
Bob O’Donnell, president of TECHnalysis Research, highlighted the AI angle directly: “As AI models have evolved and GenAI-powered applications have expanded their capabilities and reach, it’s become abundantly clear that the need for massive amounts of data – both real and synthetically generated – are essential to keep AI advancements moving ahead. Whether for large-scale model training or sophisticated fine-tuning, companies who build and use these AI models have found that high-capacity hard drive innovations like HAMR have become critical to quality and speed of their outputs.”
For years, some critics questioned whether HAMR would ever be viable at scale. Now Seagate says the majority of the world’s largest cloud storage providers are already qualified on the Mozaic platform, with broader availability planned as production continues to scale.
We are watching the early chapters of the 100TB hard drive era. It sounds wild until you consider that AI-scale data growth is even wilder.
Storage may not grab headlines the way GPUs do. But without it, none of this AI hype works. Seagate seems determined to make sure it does.
You cannot translate 0.8 million kWh into 800 MWh, which is actually readable without calculating the prefixes yourself?
x 1000/1024: At least one prefix up. Not that hard.