Susana Schwartz
AI data center math — does it add up?
A couple weeks ago, IBM CEO Arvind Krishna raised some eyebrows during a Decoder podcast in which he said there’s “no way” hyperscalers will turn a profit at the current rate of data center spending. He explained that a 1 GW data center costs about $80 billion, which means a company committing to 20-30 GW could face $1.5 trillion in spending, based on current costs.
Yesterday, Karthikeyan Natarajan, former Tech Mahindra exec and current CEO of Cyient, attempted to push a bit further into the math, with a post that shows what goes into a 1GW AI training facility in 2025:
→ GPUs & accelerators: $35-45B (50-60%)
→ Power delivery (substations, transformers, backup): $10-15B
→ Liquid cooling systems: $6-10B
→ Building retrofit/shell + networking: $5-8B
→ Land, permitting, misc.: $2-4B
The total came to $65-80 billion per GW, with an additional $3-4 billion/year in operating costs. Scaling that industrywide, the figure comes to about $8 trillion in capital for AI infrastructure. With 50+ GW of new capacity planned by 2030, what would it take for something like xAI Colossus or Oracle-Stargate to break even or turn a profit?
Susana Schwartz
Technology Editor
RCRTech
AI Infrastructure Top Stories
Dell, AMD partner for telco-specific AI innovation: Dell’s hardware, software, services and ecosystem are combined with AMD’s advanced CPUs and GPUs for a future-ready foundation that will help CSPs capitalize on AI.
Nokia surveys tech and business leaders: Nokia research says an AI boom will overwhelm today’s networks, with respondents broadly agreeing that current networks are not designed for the next phase of AI growth.
Cognizant talks of privacy vs. innovation: Telcos sit on one of the richest data environments in any industry, so how can they extract value from data without eroding consumer trust or compliance to privacy regulations?
AI Today: What You Need to Know
Gonka is integrating global GPUs in a decentralized manner: AI inference network Gonka has seen a substantial increase in computing power, with its network capacity soaring nearly 20x to surpass 10,000 NVIDIA H100 equivalents.
Musk’s sustainable abundance vision: Elon Musk says it’s critical to leverage solar-powered AI satellites to move AI to orbit, stating: “You can actually access over 1-billion-times more energy from the sun in deep space than you can on Earth.”
Oracle’s $10 billion Michigan data center on hold: Oracle’s largest data center partner Blue Owl Capital will not back a $10 billion deal for its next facility, as the software group faces increased concerns about its rising debt and AI spending.
Hut 8 and Anthropic deal for $7 billion data center: Anthropic has signed a deal to tap Hut 8’s AI DC in Louisiana’s West Feliciana Parish. Fluidstack will be part of the deal, delivering power and equipment for the 600-acre River Bend campus.
Texas Instrument’s $40 billion semiconductor factory: Production will begin production this week in Sherman, TX — the first of four planned semiconductor plants. The factory is expected to produce tens of millions of chips daily.
$275 million deal between ABM and WGNSTAR: ABM will acquire WGNSTAR, which has a highly skilled workforce of more than 1,300 employees, plus managed workforce solutions and equipment support services for the semiconductor industry.
