The executive making this observation is not a skeptic. Bryan Catanzaro is NVIDIA's vice president of applied deep learning — he runs one of the most compute-intensive teams at the company that makes the hardware the entire AI industry runs on. When he tells Axios that compute costs for his team dwarf employee costs, he is describing the real operational economics of frontier AI work, not a hypothetical.
It is a number that matters because NVIDIA has spent two years being the most bullish voice in the room about AI's transformative potential. Catanzaro is not contradicting that view — he is adding a cost dimension that is often absent from the broader conversation. Capability is real. So is the bill.
The MIT Data: 77% of Roles Still Cheaper With Humans
A 2024 MIT study on AI automation reached a conclusion that cuts against the dominant narrative. Across vision-based tasks — the category of work most amenable to current AI capabilities — AI automation was only economically viable in 23% of roles. In the other 77%, humans were cheaper.
The study was deliberately narrow in scope: vision tasks, not all knowledge work. But the finding is a data point that the business case for AI automation at scale is not yet as clear as the capability story suggests.
Uber's AI Budget: Already Gone
Uber's CTO disclosed in early 2026 that the company's annual AI budget was already exhausted — and it wasn't even summer. The primary driver cited: coding tools. Claude Code was named specifically.
This is a different kind of cost story than Catanzaro's. NVIDIA is burning compute on deep learning research. Uber is burning it on developer productivity tooling. The cost profile is the same: token-based pricing that scales with usage, billed continuously, with no natural ceiling unless you impose one.
Token-based pricing turns AI inference into a recurring operational expense that compounds with adoption. The more your team uses AI tooling, the higher the bill — and unlike SaaS seat licenses, there is no flat rate to budget against.
$740 Billion in AI Capex This Year
Big Tech collectively committed approximately $740 billion in AI capital spending in 2026 — a 69% jump from 2025. That number covers data centers, chips, power infrastructure, and the physical buildout required to run inference at scale. It represents a bet that demand for compute will continue outpacing supply for years.
The scale of this investment creates a structural pressure: the companies spending $740 billion need the ROI story to hold. That creates incentive to lead with capability and trail with cost. Understanding both requires reading past the press releases.
The Gartner Projection: 90% Cheaper in Four Years
Gartner projects inference costs could fall by more than 90% over the next four years as hardware efficiency improves and model architectures become leaner. If that holds, the economics Catanzaro describes today — where compute vastly exceeds employee cost — could look very different by 2029.
The trajectory of inference costs matters for security teams evaluating AI tooling. Tools that look cost-prohibitive at current pricing may become viable. Tools that are affordable today may see usage-driven cost creep as adoption scales. The current cost structure is a snapshot, not a steady state.
Gartner projects >90% reduction in inference costs over four years. Hardware efficiency gains and model compression are the primary drivers. The economics of AI tooling procurement today will look different by 2028–2029.
What This Means for Security Teams
Security organizations evaluating AI-assisted tools — code review, threat detection, vulnerability research, incident response automation — are operating in the same cost environment Uber described. Every query has a price. Usage scales faster than budgets. The teams that deployed Claude Code or similar tools in 2025 are already discovering this.
A few practical considerations from the current economics:
- Budget AI tooling on consumption, not seats. Token-based costs do not behave like SaaS licenses. A 10x usage increase produces a 10x cost increase. Model your AI tool spend the same way you would model cloud compute: with usage estimates and guardrails.
- Vendor pricing is not stable. The current cost structure reflects a market still finding equilibrium. Both upward pressure (demand) and downward pressure (competition, efficiency gains) are active. Lock in terms where possible; assume variability where you can't.
- The ROI calculation changes as costs fall. Gartner's 90% projection means tasks that don't make economic sense to automate today may make sense in 18 months. Revisit AI tooling evaluations on a shorter cycle than traditional software procurement.
- Compute cost is now a security risk vector. Teams spending AI budgets on productivity tooling may have less capacity for security-specific AI investment. Track the allocation explicitly.
Catanzaro's statement is not a warning about AI. It is a calibration. The capability is real. So is the cost. The organizations that manage both well will have a structural advantage over those that only tracked one.