Compute costs staying high
Demand for AI is keeping data‑centre investment and GPU prices elevated, so compute costs are unlikely to ease soon. Hyperscaler backlogs and a 57% jump in data‑centre capex last year point to persistent capacity pressure that makes enterprise AI projects require clearer cost‑to‑value thresholds. ( )
AI was supposed to get cheaper as the boom matured. Instead, the basic ingredients of AI are getting more expensive. Data center capital spending rose 57% in 2025 to $726 billion, according to Dell’Oro Group, and the firm now expects another year of growth above 50%, pushing total spending past $1 trillion in 2026 years earlier than it had forecast before (networkworld.com). That kind of jump does not happen in a market that is settling down. It happens when the biggest buyers in tech are still scrambling for capacity. Those buyers are the hyperscalers, and their numbers show the pressure plainly. Amazon said its backlog reached $244 billion, up 40% from a year earlier, while Google reported a $240 billion backlog and said the number of billion-dollar deals signed in 2025 exceeded the previous three years combined (networkworld.com). Amazon plans to spend about $200 billion in capital expenditures in 2026, mostly on AWS, and Google has said it expects roughly $180 billion in capex this year (networkworld.com). When cloud companies spend at that scale, they are not preparing for a lull. They are trying to stay ahead of customers who already signed up. That demand is flowing straight into GPU prices. Silicon Data, a firm that tracks cloud GPU rentals, found its Neo Cloud H100 index rose from 2.20 to 2.64 over three months, a 20% gain, while its Neo Cloud B200 index climbed from 4.40 to 5.35, up 22% (africa.businessinsider.com). Even hyperscaler H100 pricing moved higher. Carmen Li, Silicon Data’s chief executive, said the market has broken the usual pattern in which new chips launch at high prices and then ease as supply catches up (africa.businessinsider.com). The old hardware is staying expensive because the new hardware is not arriving in enough volume to loosen the market. The cloud markup makes that worse for everyone outside the largest platforms. Silicon Data says H100s can cost almost three times as much to rent from hyperscalers as from specialized neocloud providers such as CoreWeave (africa.businessinsider.com). Nvidia is also not selling bargain silicon into this frenzy. Analysts cited by CNBC say Blackwell GPUs can cost about $40,000 each, and Nvidia said the four largest cloud providers have already bought 3.6 million Blackwell GPUs under its current counting method (cnbc.com). Faster chips may lower the cost per token. They do not lower the cost of getting in line. And the line is no longer just about chips. Power has become the next hard limit. The Financial Times reported that Microsoft CEO Satya Nadella recently said the company’s biggest issue is “not a compute glut” but power, as hyperscalers pour more than $400 billion into data centers and still struggle to energize new capacity fast enough (ft.com). Dell’Oro has already shown what that means inside enterprise budgets: accelerated servers now take about 35% of enterprise data center capex, up from 15% in 2023, and those AI servers can cost $100,000 to $200,000 each, versus roughly $7,000 to $8,000 for a traditional server (networkworld.com). That is why enterprise AI projects now need a much clearer answer to a simple question: what, exactly, is this workload worth?
