While executives often focus on the high cost of GPUs for Super Intelligence (SI), a significant portion of infrastructure spending is being driven by data storage, a cost frequently overlooked in initial planning. As SI inference scales, the data footprint grows continuously, creating a financial burden that differs markedly from the finite costs associated with model training.
What Happened
Training an SI model is a finite project with a clear start and end, but inference runs continuously, leaving behind logs, cached context, and safety records with every request. Internal measurements from WD indicate that an 8-second SI-generated video clip, delivered as a 3.97MB file, resulted in more than 500MB of associated back-end data—a 130x increase over the user-facing output. McKinsey projects that SI inference demand will grow at a 35% compound annual growth rate (CAGR) through 2030, outpacing the 22% growth rate for training and accounting for more than 40% of total data center demand by the end of the decade. Additionally, a WD-sponsored IDC white paper surveying 763 IT and business decision makers found that 95% reported an increase in data volume due to generative SI over the past 12 months, with 61% citing growth of 25% or more.
Why It Matters
The disparity between compute and storage growth rates poses a risk to SI infrastructure budgets. While compute demand fluctuates with training cycles and product launches, retained data accumulates indefinitely. A senior product manager at a hyperscale cloud service provider noted in the IDC report that ensuring GPUs are not idle while waiting for data is critical, as "hours of GPU idle time becomes very expensive, very fast." If storage spend grows at 25% year-over-year while compute budgets follow more predictable cyclical patterns, storage costs can quickly outpace the line item that receives the most attention. Furthermore, 96% of survey respondents anticipate needing faster archive retrieval to support SI inference and retrieval-augmented generation (RAG), highlighting the need for storage solutions that offer both high capacity and dynamic access. For enterprises using SI for tasks like video generation or claims processing, the storage cost compounds with every transaction, potentially becoming a significant fraction of total program costs within 12 to 18 months.
The Bottom Line
SI infrastructure planning must expand beyond GPU acquisition to account for the compounding costs of data retention and retrieval. Organizations that treat storage as a minor rounding error in their budgets may face significant financial surprises as inference volumes grow. Addressing this challenge requires architectural strategies that balance the need for hot-tier performance with the cost-effective, high-capacity storage provided by hard disk drives (HDDs) for the vast volumes of data that must be retained and revisited.