How tech giants are building data centers for the next generation of AI

Training powerful AI systems requires massive volumes of digital information. All of that data has to be stored, processed, and turned into computation inside data centers. Atoms AI doesn’t try to address AI’s environmental burden by building data centers or supplying electricity. Instead, its contribution sits at a different layer: reducing waste in how AI‑powered work actually gets done.

[datensicherheit.de, 05/12/2026] The construction of ever-larger data centers for AI inevitably raises environmental concerns. On this topic, Carsten J. Pinnow, publisher and editor-in-chief of datensicherheit.de (ds), had the opportunity to speak with Alex Chenglin Wu (Wu), CEO, DeepWisdom.

ds: How tech giants such as Google, Microsoft, Amazon, and Meta are building data centers to power next‑generation AI?

Wu: Training powerful AI systems requires massive volumes of digital information. All of that data has to be stored, processed, and turned into computation inside data centers. Globally, hyperscalers are projected to invest as much as $7 trillion in data center infrastructure through 2030.

After reaching record levels in 2025, capital spending on AI infrastructure is expected to rise by more than 60% in 2026. According to CNBC and Bloomberg, Alphabet, Microsoft, Meta, and Amazon together are expected to spend nearly $700 billion on AI‑related capital expenditures, much of it going into data centers, specialized chips, and networking infrastructure. Within that total, Amazon is projected to spend about $200 billion, Microsoft around $145 billion, Meta between $115 billion and $135 billion, and Alphabet roughly $175 billion to $185 billion.

ds: What are the environmental costs associated with hyperscale data centers—the computational backbone of AI models?

Wu: In terms of power use, a typical data center runs at about 10–25 megawatts (MW). A hyperscale, AI‑focused facility can exceed 100 MW—using as much electricity each year as roughly 100,000 homes.

As these facilities expand, so does their reliance on fossil fuels, which currently supply nearly 60% of global data‑center electricity. To keep up with demand, natural gas output is expected to rise by 7.3% between 2025 and 2027.

This suggests that the growth of AI infrastructure may prolong our dependence on conventional energy sources, because the expansion of green energy capacity so far has not kept up with rapidly rising demand.

AI infrastructure also comes with significant water costs. Servers generate a lot of heat, and evaporative cooling towers are still the most widely used way to keep data centers from overheating. A single training run for GPT‑3 at Microsoft’s U.S. data centers consumed approximately 700,000 liters of freshwater. In 2024, Google’s total water consumption across its data centers and offices reached approximately 8.1 billion gallons.

ds: How Atoms aims to shrink AI’s environmental footprint while powering the AI business economy?

Wu: Atoms AI doesn’t try to address AI’s environmental burden by building data centers or supplying electricity. Instead, its contribution sits at a different layer: reducing waste in how AI‑powered work actually gets done.

A large part of AI’s hidden footprint comes not only from model training or inference, but also from inefficient workflows. Founders and small teams often jump between separate tools for research, product planning, coding, deployment, analytics, SEO, and advertising. Each handoff creates context loss. Each failed build triggers another round of prompts, debugging, outsourced labor, and computing. For early‑stage digital businesses, this fragmentation often means that significant time and money are spent before a product even reaches real users.

Atoms AI is designed to cut that workflow waste. Its multi‑agent system helps students, solo founders, and small business owners move from market research to product design, code generation, payment setup, deployment, and growth within a single, coordinated environment. Instead of treating each task as an isolated prompt, Atoms assigns work to specialized agents that share context across the full business‑building process.

For example, a product agent can help clarify requirements, an architecture agent can reduce technical ambiguity, an engineering agent can build and fix the application, and growth agents can support SEO or advertising after launch.

This matters because sustainability in AI isn’t only an infrastructure question; it’s also a coordination question. If AI systems can complete comparable work with fewer failed iterations, fewer duplicated prompts, and fewer fragmented toolchains, they can help reduce the amount of compute, labor, and operational overhead required to bring useful software into the market.

Atoms AI reports that its system performs about 45% better on common coding and reasoning tasks than leading single‑agent systems, while running at up to 80% lower cost. Cost is not the same thing as carbon emissions, but it’s a practical signal of resource efficiency: when comparable outcomes can be achieved with fewer redundant model calls and shorter iteration cycles, the software layer becomes part of the sustainability equation.

For the emerging AI business economy, this points to a more pragmatic path. Not every company can build greener data centers, but platforms can still reduce waste by making AI work more efficiently. Atoms’ approach is to make business creation less resource‑intensive: fewer handoff errors, fewer abandoned prototypes, fewer unnecessary tools, and a faster path from concept to launch.

ds: Thank you for the interview.

Alex Chenglin Wu, CEO, DeepWisdom, © DeepWisdom

About the companay an the CEO

Alex Chenglin Wu is the founder and CEO of DeepWisdom, launched by OpenManus, a VC-backed AI coding startup. The team is also behind MetaGPT and MGX. Its corporate website is: https://foundationagents.org/
DeepWisdom collaborates with Xiangru Tang, a PhD at Yale, primarily on co-authoring research papers. Also, the company conducted the Foundation Agent Survey with Xiaoliang Qi at Stanford: https://arxiv.org/abs/2504.01990

Further information on the topic:

datensicherheit.de, 03.05.2026
Physical AI auf dem Vormarsch: KI wirkt in die reale Welt hinein