AI’s Power Problem: Chips, Open Models, and the Bubble Question in 2026

The story of AI in 2026 is not just about what models can do; it is about what the world can power, manufacture, and finance. Data center analysts estimate that global capacity for AI workloads crossed roughly 10 gigawatts by early 2025, driven by shipments of millions of high‑end accelerators. AI‑focused servers can draw 10 kW or more each, making AI training and inference far more power‑hungry than typical cloud workloads. At the same time, investors have pushed the valuations of key chip and infrastructure firms into multi‑trillion territory on the assumption that this growth continues.

Layered on top of this are politics and open source. Chinese labs have embraced open models as a way to accelerate domestic ecosystems and project influence abroad, while US and EU regulators worry about safety, sovereignty, and export controls. India, meanwhile, is trying to turn AI into leverage for services and multilingual access rather than a race to build the largest model. The result is an AI economy that is both global and bloc‑like.

This post is Part 2 of a two-part series on AI in 2026.
Part 1, From AGI Promise to Agent Reality focuses on agents, supervision, and how work is changing.

Table of contents

• The energy footprint of intelligence
• Efficiency vs brute force
• GPUs as oil, fabs as refineries
• Open source as both deflation and geopolitics
• Europe and India: sovereignty and services
• Two narratives in the market
• Global valuations: the US vs China
• A quick snapshot of AI blocs
• Three questions that matter more than hype
• Conclusion

The energy footprint of intelligence

Building and running state‑of‑the‑art AI now has a visible physical footprint. Industry analyses estimate that from 2021 through 2024, Nvidia‑class accelerators shipped in volumes equivalent to several million H100‑grade GPUs, with AI data center demand crossing around 10 GW of capacity by early 2025. At typical utilization and efficiency levels, a single 20,000‑GPU cluster can draw on the order of 30 megawatts of power, consuming hundreds of thousands of megawatt‑hours per year - electricity on the scale of a small city.

Broader data center energy studies tell a similar story. Global data center consumption was in the few hundred terawatt‑hour range in 2022, and US scenarios now project a substantial rise by 2028, with AI servers as the main driver. While networks and storage have become more efficient, the servers and GPUs at the heart of AI workloads are responsible for most of the expected increase, concentrating both risk and opportunity in the energy profile of AI hardware.

Efficiency vs brute force

Hardware roadmaps suggest that accelerators will become even more power‑hungry on a per‑chip basis, with 1,000-1,500 W devices on the horizon, even as performance per watt improves. Companies respond by building denser clusters and more efficient cooling and by chasing cheap or stranded energy, but the simple fact remains: at scale, every point of extra model quality and uptime now has a measurable carbon and grid cost.

This has spurred a countertrend focused on efficiency. Alternative accelerators and model designs are pitched as delivering similar accuracy at a fraction of the energy and cost, and licensing deals around efficient chips underscore how seriously large buyers take power constraints. For investors and policymakers, the key question is no longer whether AI can be scaled, but whether it can be scaled within acceptable environmental, regulatory, and capital limits.

This article highlights how hyper-efficient models like DeepSeek (MoE, Chain-of-thought reasoning, etc ...) are enabling the possibility of general purpose robots

GPUs as oil, fabs as refineries

The analogy between GPUs and oil is tempting: a few suppliers dominate a critical input, and infrastructure players scramble to secure long‑term access. Nvidia’s data center business has become the fulcrum of the current AI wave, with analyses describing multibillion‑dollar orders from hyperscalers and sovereigns and highlighting the company’s role in driving AI data center power demand beyond 10 GW in 2025. Roadmaps point to more powerful, more power‑hungry accelerators and a growing ecosystem of networking and software around them.

Underneath Nvidia, the supply chain runs through a small number of advanced fabs and packaging facilities, making geopolitical risk and export controls central topics for AI planning. Governments that worry about dependence on foreign fabs are increasingly willing to subsidize local manufacturing and to selectively restrict high‑end chip exports, especially toward rivals viewed as potential military competitors.

Open source as both deflation and geopolitics

When marginal costs fall to zero, prices follow. - Chris Anderson

Open‑source models have shifted from hobby projects to strategic tools. Over 2024-2025, labs and companies released increasingly capable open models, enabling startups and enterprises to deploy “good enough” systems without paying for premium proprietary APIs. This has a deflationary effect on inference costs and pushes major providers to compete on performance, reliability, integration, and hardware rather than raw access to basic capabilities.

China has leaned particularly hard into this direction. Analysts of the Chinese ecosystem describe almost every important Chinese model since 2025 being released as open source, framed as part of an “AI+” strategy to accelerate domestic adoption and export influence. At the same time, Beijing has introduced detailed rules on generative AI and content labelling, though enforcement is uneven and unlabelled AI content remains common in practice. Export controls on advanced chips from the US and allies push Chinese firms toward more efficient models and alternative hardware, reinforcing the incentives for open, lighter‑weight systems.

Europe and India: sovereignty and services

Europe’s AI conversation revolves around sovereignty and safety. The EU AI Act, moving through finalization and implementation, categorizes AI systems by risk level and imposes documentation, transparency, and oversight requirements on high‑risk use cases. This approach encourages European firms to adopt AI in ways that maintain clear human accountability and often motivates interest in EU‑hosted or EU‑developed models, including language‑specific and domain‑specific systems backed by public funding.

Europe must be a place where AI thrives, grounded in trust and safety. - Ursula von der Leyen

India approaches AI from a different angle. Rather than racing to build the single most powerful model, Indian policymakers and firms focus on using AI to extend the reach of digital public infrastructure and services - from multilingual interfaces for government platforms to AI‑augmented code and support delivery in IT and BPO. The emphasis is on mobile‑first, multilingual, low‑cost access, with Indian services firms increasingly marketing AI‑enhanced offerings to global clients while experimenting with open and proprietary models behind the scenes.

India wins by building population-scale digital infrastructure, not proprietary monopolies. - Nandan Nilekani

Two narratives in the market

Financial markets have embraced AI as the dominant theme of this cycle. Chipmakers, cloud platforms, and a handful of foundation‑model and agent vendors have seen valuations soar, underpinned by rapid revenue growth and massive capital expenditure plans. Venture capital surveys at the end of 2025 describe “strong enterprise AI adoption next year - again,” with particular enthusiasm for agents, vertical applications, and infrastructure.

Every industry will be transformed by AI. - Jensen Huang

At the same time, investors and commentators are debating whether the sector already has bubble characteristics. Arguments for a bubble point to concentration of profits in a few infrastructure providers, high failure rates for AI pilots and projects, and waves of “agent‑washing” where ordinary automation tools are relabelled as autonomous intelligence. Arguments against point to real productivity gains in software development, marketing, customer support, and design, as well as structural demand from governments and enterprises that view AI as a long‑term capability rather than a passing fashion.

Global valuations: the US vs China

One striking detail is the valuation gap between US and Chinese AI companies. Coverage of cross‑border investment notes that foreign investors increasingly see Chinese AI valuations as cheaper and less bubbly: startups and platforms with similar or higher revenues to US peers can trade at roughly one‑quarter of the valuation multiples, aided by lower R&D and labor costs. This makes China attractive to investors willing to navigate regulatory and geopolitical risk.

A simple way to think about it is in terms of where the premium sits. In the US and parts of Europe, the premium sits on frontier labs, GPUs, and integrated platforms; in China, it often sits on ecosystem reach and state alignment, while multiples stay lower. For long‑term investors, the key question is whether profits will remain concentrated in a narrow set of global infrastructure firms or gradually spread to application and integration layers.

Productivity isn’t everything, but in the long run it’s almost everything. - Paul Krugman

A quick snapshot of AI blocs

Three questions that matter more than hype

Across energy, chips, models, and markets, three questions stand out for the next two years.

1 - Can AI be scaled within the limits of grids, climate goals, and political patience for large new power demands? Data center energy projections and hardware roadmaps suggest that tradeoffs between model size, deployment breadth, and environmental impact will become harder to ignore. Efficiency‑focused chips and smaller, specialized models may end up as important as frontier systems in practice.

2 - Who captures the value? Today’s profits cluster around a narrow set of chipmakers and cloud providers, but open‑source models, regional strategies, and services‑led approaches in places like India hint at a more distributed landscape. Whether the AI boom ends up looking more like the early internet - where value eventually spread outward - or like a capital‑intensive utility sector remains an open question.

3 - How much autonomy do regulators and societies actually want to grant AI systems before insisting on hard constraints? The answer will shape not just which models get trained, but which agents and applications are allowed to operate, in which sectors, and under what supervision. What exists today is not AGI, but a complex stack of agents, hardware, and institutions trying to figure out how far to push.

Conclusion

AI is entering its infrastructure phase. Open models push prices down, while geopolitics pushes power up. Europe builds rules, India builds services, the U.S. prices optionality, and China scales execution. The question is no longer whether AI works - but who can afford to run it, who controls its deployment, and who is accountable when it fails.

Part 1, on agents and work, is here