Reducing AI-Related Costs with an Active Archive

AI can be expensive – very expensive. Consider a few of the costs:

• The large language models (LLMs) that spit out data and answers from the likes of ChatGPT must train on hundreds of billions of data points for months on end before they are accurate enough for broad usage.
• Extremely expensive graphics processing units (GPUs) and processing accelerators are needed in large numbers to crunch the numbers.
• New data centers must be designed, built, and customized specifically for AI applications.
• One Generative AI (GenAI) query to ChatGPT is estimated to generate 100 times more energy than a traditional Google search. With 10 million queries per day, daily power consumption from this one application is about 1 GWh. Training of one LLM requires around 10 gigawatt-hours (GWhs). That amount of power would cost something north of a hundred thousand dollars in the U.S.
• AI needs highly dense racks. The more density you pack in, the higher the cost per rack. AFCOM’s State of the Data Center 2025 report found that rack density has soared from 7.5 kW a couple of years ago to 16 kW today. Some data centers already have racks of more than 50 kW.

There is no escaping many of these costs. GenAI may currently only account for 3% of total IT spend, but that will rise to 12% by 2032 – or $1.3 trillion per annum.

“AI is fundamentally changing the way IT operates, and something will have to give in terms of how IT operates,” said Rich Gadomski, Head of Tape Evangelism at Fujifilm Media USA, during the recent Archive Alliance Virtual Conference, Why AI Needs Active Archive. “One area where expenditures for AI can be drastically reduced is storage and archiving.”

Source: The Sustainable Advantage of Enterprise Data. Source: Monroe and Johns, Jan 2024

If storage costs for retaining AI data on disk are already astronomical, the price tag will only increase as the amount of data retained for AI purposes mushrooms. AI workflows, after all, will require multi-petabyte storage. A review of five-year costs per TB indicates that the current gap between the cost of SSD, HDD, and a tape archive will only widen. By 2035, an active archive will be 8X cheaper than tape and 33X cheaper than SSD.

“An active archive can leverage low-cost storage tiers including tape, economy disk, optical technologies, and cost-effective cloud options,” said Gadomski.

The Global Datasphere
Pete Paisley, Business Development at Magstor, emphasized the importance of storage economics by laying out the predicted expansion in the volume of data stored digitally across the world – known as the global datasphere.

AI is creating massive data sets. Organizations are figuring out how to scale for the zettabyte era while maintaining security and providing AI applications with timely access to data.

“Energy demands are the primary lifecycle cost, and a tape archive provides the lower cost per TB,” said Paisley. “The TCO for tape is a fraction of cloud and disk while offering superior air-gapped security.”

However, if most data exists on disk, the migration of data to tape becomes a serious operational challenge. It involves a lot of labor and solving problems such as obsolete formats and backward compatibility, as well as how to move email archives onto tape.

“Many companies are asking for our help on how they can best repatriate cloud data to tape,” said Paisley.

For some, the solution may be investing in the hardware and infrastructure needed to set up their own tape and archiving systems internally. For others, subscription-based services are becoming available for non-critical data access at very low frequency.

What is clear is that the resurgence of interest in tape storage due to the advent of active archives is about to get another boost courtesy of the AI revolution. The data sets required for effective AI solutions necessitate a complete rethink of the storage architecture. That will lead many to low-cost alternatives to archive little-used data while keeping it close at hand should AI need access.

“Many data center managers will soon be forced to use tape and various enterprise emerging technologies as ultra-low-cost, sustainable storage alternatives,” said John Monroe, Principal at Furthur Market Research.