Artificial intelligence has been marketed for months as a leverage to reduce costs, automate tasks, and increase team productivity. The emerging reality within some tech companies is more uncomfortable: using advanced AI can also be very expensive. Not only due to model subscriptions but because of computing, tokens, memory, data centers, and all the infrastructure needed to support agents, copilots, and large-scale inference loads.
The best phrase summarizing this shift was said by Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA, in statements reported by Axios: “For my team, the compute cost is far above the employee cost.” The nuance is important. It doesn’t mean that every AI system is always more expensive than a worker, nor that automation is financially unviable. It means that, on the frontier teams that heavily use models, GPUs, and infrastructure, the technical bill is beginning to surpass items that were previously considered the heaviest: salaries.
AI isn’t free: someone pays for tokens, GPUs, and energy
During the initial adoption phase of generative AI, many companies mainly looked at potential savings: less coding time, fewer hours on documentation, faster customer service, quicker data analysis, and automation of repetitive tasks. But as AI moves from experimental to daily operational processes, unforeseen costs emerge.
Tokens cost money. Inference incurs expenses. Agents that make multiple calls, read repositories, invoke tools, generate plans, and retry on failures consume far more resources than an occasional chatbot. Adding advanced reasoning, long context, multimodal models, or private deployments increases the costs even more.
Axios highlights other sector examples: Uber’s CTO reportedly exhausted their AI budget for 2026 due to token costs, according to The Information, and other companies are beginning to boast or worry about rising bills linked to intensive model usage. The underlying message is clear: AI doesn’t eliminate work costs; it shifts them partially to infrastructure, model providers, and computational consumption.
This doesn’t mean AI isn’t valuable. It can deliver great value. But it compels us to measure that value differently. Companies can no longer simply claim that their employees use AI. They need to demonstrate whether this use genuinely reduces times, improves quality, generates revenue, or allows more work to be handled without increasing headcount. If computational spending grows faster than productivity, the promise turns into financial pressure.
Gartner raises the forecast: data centers lead spending
Gartner’s data shows the scale of this phenomenon. The consultancy predicts global IT spending will reach $6.31 trillion in 2026, a 13.5% increase over 2025. The fastest-growing category isn’t devices or communications services but data center systems, driven by AI infrastructure, advanced memory, and high-performance computing demand.
| Category | 2025 Spend | 2025 Growth | 2026 Spend | 2026 Growth |
|---|---|---|---|---|
| Data Center Systems | $505.634B | 51.6% | $787.990B | 55.8% |
| Devices | $791.663B | 9.7% | $856.189B | 8.2% |
| Software | $1,254.449B | 12.8% | $1,443.621B | 15.1% |
| IT Services | $1,715.650B | 6.2% | $1,870.197B | 9.0% |
| Communications Services | $1,296.409B | 3.3% | $1,358.553B | 4.8% |
| Total IT | $5,563.805B | 10.5% | $6,316.550B | 13.5% |
The message is clear: spending on data center systems is projected to rise from $505.634 billion in 2025 to $787.990 billion in 2026—an increase of 55.8% in just one year. Gartner attributes this surge to the momentum of AI infrastructure, growing workloads, and the rising cost of components like high-bandwidth memory.
The contrast with other segments is also telling. Device spending grows by 8.2%, communications services by 4.8%, and IT services by 9.0%. Software expands more significantly, by 15.1%, also driven by generative AI and new enterprise platforms. But nothing matches the pace of data center investments.
NVIDIA profits but also bears the costs
NVIDIA is the main beneficiary of this cycle. Its GPUs, networks, systems, and software have become fundamental to much of modern AI infrastructure. The company has transitioned from being mainly associated with gaming graphics cards and professional stations to playing a central role in data centers, model training, inference, robotics, simulation, and agents.
However, Catanzaro’s statement shows even the AI hardware leader faces the other side of the coin. NVIDIA sells the infrastructure everyone wants but also uses AI intensively in its own operations. When working at the frontier, internal compute costs can easily surpass traditional costs.
The paradox is intriguing. AI promises to save human labor but demands enormous machinery: GPUs, clusters, high-speed networks, storage, energy, cooling, software, maintenance, and specialized personnel. In some cases, the costs don’t disappear—they shift in the income statement.
This explains why business conversations are shifting from mere adoption to return on investment. Asking how many employees use AI is no longer enough. The right questions are: what’s the cost of each automated workflow, how many calls does an agent make, which models are used for each task, what can be run with smaller models, what context is truly necessary, and which processes generate measurable value?
The next battle: efficiency
Enthusiasm for AI isn’t slowing down anytime soon. Big tech companies continue investing in data centers, accelerators, networks, and energy. Hyperscalers compete over power capacity, land, chips, and supply agreements. Companies want to deploy internal assistants, development agents, document automation, semantic search, and smarter support systems.
But the free-for-all phase is starting to wind down. Just like cloud computing, many firms will find that rapid scaling without controls can lead to unjustifiable bills. Cloud practices like FinOps have emerged to monitor spending, allocate costs, and avoid underutilized resources. In AI, similar but stricter measures will be necessary: token management, dynamic model selection, caching, quantization, local or private inference, context limits, use-case monitoring, and continuous ROI evaluation.
Efficiency will shift from being just a technical detail to a key competitive advantage. Companies that use AI with appropriate models, concise prompts, well-designed agents, and reasonable architectures can achieve productivity gains without astronomical costs. Conversely, companies adopting AI as a trend without measuring consumption and results risk automating at a cost that exceeds the promised savings.
NVIDIA’s statement doesn’t dismiss the AI narrative. It matures it. Artificial intelligence could be one of the most important technologies of this decade, but it’s not cheap magic. Behind every answer is infrastructure. Behind every agent are calls, memory, and computing. And behind every productivity promise is an increasingly tough question: how much does it really cost to make AI work?
Frequently Asked Questions
Did NVIDIA say that AI costs more than human workers?
Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA, told Axios that in his team, compute costs are far above employee costs. This shouldn’t be interpreted as a universal rule for all companies.
Why are data center expenses rising so much?
Due to the demand for AI infrastructure, GPUs, advanced memory, high-performance networks, storage, energy, and cloud capacity for training and deploying models.
How much will global IT spending grow in 2026?
Gartner forecasts it will reach $6.31 trillion, a 13.5% increase over 2025.
Will AI remain profitable for companies?
It depends on the use case. While it can generate significant value, companies will need to carefully measure token, compute, license, infrastructure costs, and actual productivity gains to justify the investment.
via: wccftech