Bryan Catanzaro, a senior executive at Nvidia, recently highlighted a structural shift inside modern AI organizations: compute has become the dominant cost center, overtaking human labor expenses. In his framing, teams are now spending more on GPUs, data center capacity, and inference workloads than on salaries for researchers and engineers.
This is not a marginal accounting change—it signals a fundamental reordering of how AI companies allocate capital and where value is created. For decades, the canonical tech cost structure was labor-heavy. Software companies scaled through headcount: more engineers meant more features, more velocity, and more revenue. Compute was relatively cheap and predictable, often outsourced to cloud providers as a manageable operational expense.
That balance is now inverted. The rise of large-scale foundation models has turned compute into the primary production input, comparable to industrial energy costs in manufacturing. At the center of this shift is the GPU economy, heavily shaped by companies like NVIDIA. Training frontier models requires thousands to hundreds of thousands of accelerator hours, often running continuously for weeks or months.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab.
Inference at scale—serving billions of user queries—can exceed training costs over time. As a result, compute is no longer a background utility; it is the binding constraint on product velocity, model quality, and market expansion.
This inversion has profound financial implications. When compute exceeds salaries, traditional startup efficiency metrics break down.
Headcount-based burn analysis becomes misleading because marginal progress is no longer primarily determined by additional engineers but by additional FLOPs. A small team with massive compute budgets can outpace a large team with constrained infrastructure. This is why capital markets have increasingly begun evaluating AI firms less like software companies and more like infrastructure operators.
The economic structure also resembles a shift toward AI factories. In this model, GPUs are not tools but production machinery, continuously converting energy and capital into intelligence outputs. Salaries become fixed overhead, while compute becomes variable but dominant. The most important strategic question is no longer how many engineers do we have but how much compute can we sustainably deploy per unit of revenue?
This dynamic also introduces volatility. Compute pricing is sensitive to hardware supply cycles, energy costs, and cloud provider margins. A surge in demand for inference can instantly compress availability, forcing companies into bidding wars for GPU capacity. Unlike salaries, which scale linearly and predictably, compute costs can spike nonlinearly with usage, especially when products achieve viral adoption or sudden enterprise demand.
The implications extend to competitive dynamics. Firms with privileged access to compute—either through long-term contracts, proprietary data centers, or vertical integration—gain a structural advantage.
Meanwhile, smaller players face a steep marginal cost curve, where each additional model improvement requires disproportionately more capital. This creates a winner-takes-most environment not just in model quality, but in infrastructure control. At a broader level, Catanzaro’s observation reflects a redefinition of productivity in AI systems.
Intelligence is becoming something you rent from silicon rather than something you simply design with human effort. In that world, compute is not just a cost line—it is the core determinant of economic output. The labor force still matters, but it increasingly orchestrates systems whose true limiting factor is physical infrastructure rather than human ingenuity.
As AI systems continue scaling, this imbalance between compute and labor costs is likely to deepen. The central question for the next decade is whether compute becomes abundant enough to re-normalize costs—or whether AI development permanently shifts into a capital-intensive regime where intelligence is priced like industrial output rather than software.