Bryan Catanzaro, Nvidia's vice president of applied deep learning, told Axios last week that for his team "the cost of compute is far beyond the costs of the employees," a statement that drew over 2,500 upvotes on Reddit because it named publicly what a growing number of enterprise finance teams are discovering privately: the fully loaded cost of AI automation frequently exceeds the fully loaded cost of the human work it displaces, at least right now.
The comment requires context to interpret correctly, and Catanzaro was careful about the framing. He was describing his own team's experience, a deep learning research and applied engineering group at one of the most GPU-intensive organisations in the world, not a universal claim about AI economics across all sectors. The workloads that his team runs on AI infrastructure, training experiments, inference pipelines, evaluation suites, iterative model testing, are inherently compute-heavy in ways that an administrative workflow or a customer support queue is not. His statement is most accurately read as a description of what happens when a sophisticated AI engineering team internalises the cost of compute that most product teams currently externalise to an API provider's pricing model. What you pay OpenAI per month per seat does not reflect the infrastructure cost behind that seat. The margin compression at frontier AI providers, which several have acknowledged losing money on heavy users under flat subscription pricing, is the gap between what the API costs to serve and what you are paying for it. When that subsidy ends, or when you are running your own inference rather than renting capacity, the compute cost becomes visible in a way it currently is not for most buyers.
The MIT study that Catanzaro's comments have been framed alongside is worth reading precisely. Published in 2024, it examined 1,000 visual task categories that could theoretically be automated and found that only 23% were economically viable for AI replacement at current cost structures. For 77% of tasks, human workers were still cheaper. The study's methodology was narrow, restricted to visual inspection and image-based tasks, and its findings should not be extrapolated broadly across knowledge work or multimodal workflows. But the number is useful as a corrective to the assumption that AI cost efficiency is universal. It is not. The economics depend on task type, volume, error tolerance, supervision requirements, and the availability of models fine-tuned well enough to perform reliably without significant human review of outputs. Each of those variables shifts the cost comparison in ways that most automation ROI analyses do not adequately model.
The startup sales narrative problem is where this lands hardest for founders. The dominant commercial pitch in enterprise AI right now is labor replacement: deploy our agent, reduce your headcount or avoid adding it, capture the cost savings as margin. That pitch is landing because enterprise buyers are motivated by cost reduction and because the demos are genuinely impressive. What the pitch often elides is the total cost structure required to make AI outputs production-reliable. Inference costs are visible. Orchestration infrastructure, evaluation pipelines, error handling, human review of low-confidence outputs, incident response, and the engineering time required to maintain prompt quality and model performance over time, are frequently not modelled upfront. The enterprise buyer who signs a contract based on a cost-reduction thesis and then discovers eighteen months later that their IT budget has been reallocated from headcount to compute, without a corresponding reduction in total operational cost, is the buyer who writes the critical case study that damages the next deal in that sector.
The valuation of speed, scale, and consistency is the legitimate counter-argument, and it is a strong one in specific contexts. An AI system processing 50,000 insurance claims per day is not competing with a human team on a per-claim cost basis in any meaningful sense, because the human team does not exist at that volume at any price. A coding assistant that eliminates the three-day turnaround on a routine data pipeline request is not competing with an offshore contractor; it is competing with the opportunity cost of a senior engineer's time. When AI is doing something that human labor could not do at that scale or speed regardless of cost, the comparison is not cost-per-unit, it is capability unlock. The mistake is applying the capability-unlock framing to every automation use case when many of them are genuine cost comparisons where the human alternative is entirely viable and frequently cheaper.
The disclosure question is the one that will shape enterprise sales practice over the next two years. Right now, most AI startup sales decks present ROI projections based on headcount reduction without disclosing the compute, maintenance, and supervision costs required to achieve that ROI. As enterprise buyers become more sophisticated and as IT departments build internal experience with AI cost structures, that omission will become more visible and more damaging to trust. The startups that will win durable enterprise relationships are the ones that present total cost of ownership comparisons that include inference, orchestration, evaluation, and supervision, rather than just the headline labor cost offset. Some of those comparisons will still be favourable. Many of the use cases that are genuinely well-suited to AI automation at scale will show positive ROI even under honest full-cost accounting. But presenting the honest number rather than the flattering partial number is the difference between a customer who renews and expands, and a customer who goes quiet after the first year and serves as a cautionary example to the next prospect.
Big Tech has collectively committed approximately $740 billion in capex for AI-related infrastructure in 2026, a 69% increase over 2025 according to McKinsey data. That spending is not being driven by demonstrated unit-economics improvement across most enterprise deployments. It is being driven by strategic positioning, competitive anxiety, and the expectation that costs will decline fast enough to validate current investment by the time the contracts for future capacity come due. At the individual startup level, the equivalent behaviour is selling labor replacement before the labor replacement math actually works. Catanzaro's comment resonated on Reddit because practitioners, not analysts, are the ones who see the AWS bills.
Also read: Nvidia Backs DeepInfra's $107 Million Series B and the Investment Is About More Than One Inference Startup • If You Downloaded Gemma 4 GGUFs at Launch, You Need to Redownload Them and the Reason Why Matters More Than the Fix Itself • Six Intelligence Agencies Just Told Enterprise Builders That Agentic AI Is a Live Security Risk and the Guidance Is More Specific Than Anyone Expected
