Something unprecedented is unfolding at the intersection of capital markets and computing infrastructure — and its consequences will be felt far beyond the United States. The sums that AI companies are now committing to hardware and energy would have seemed implausible just two years ago, and they are accelerating.
Anthropic has placed an $11 billion order for Google TPUs. OpenAI has secured a 10-gigawatt energy agreement — sufficient electricity to power roughly 7.5 million American homes, or the entirety of Portugal — to feed its expanding data centre footprint. Meta has issued aggressive capital expenditure guidance for 2026 that analysts describe as a generational infrastructure bet. NVIDIA, meanwhile, has unveiled its Vera Rubin platform, the next step in a hardware roadmap that has made the company the most consequential chokepoint in modern technology — and a central concern for industrial policymakers from Brussels to Beijing.
This is not incremental investment. This is an arms race with civilisational stakes.
A Race Reshaping Global Power
The competitive dynamics driving this surge are not confined to corporate boardrooms. Governments around the world are watching closely — and responding. The European Union has launched its AI Continent Action Plan, pledging to attract investment in AI gigafactories and data infrastructure as a counterweight to American dominance. China has accelerated state-backed investment in domestic chip development, seeking to reduce its dependence on NVIDIA hardware following successive rounds of US export controls. The United Kingdom, Canada, Japan, and South Korea have each announced or expanded national AI compute strategies in the past twelve months.
The proximate cause of the private sector surge is competitive pressure. Every major AI laboratory understands that training frontier models requires not just better algorithms, but raw computational throughput that dwarfs what was needed even eighteen months ago. The relationship between compute and capability — long theorised through scaling laws — has proven durable enough that no serious player can afford to fall behind on infrastructure.
But there is a deeper driver: current AI systems still have significant capability gaps that demand continued research investment. Work from Berkeley Artificial Intelligence Research (BAIR) illustrates the challenge concretely. The Visual Haystacks benchmark, which tests large multimodal models on sets of images rather than single inputs, found that state-of-the-art proprietary models — including GPT-4o, Claude-3 Opus, and Gemini-1.5-pro — drop to roughly 50% accuracy (effectively random guessing) when processing just 50 images in multi-needle tasks. LLaVA, a widely used open-source model, shows performance drops of up to 26.5% depending on where the relevant image appears in the input sequence.
These are not edge cases. Long-context and cross-image reasoning are central to enterprise deployment at scale — whether in a hospital in Singapore, a logistics firm in the Netherlands, or a financial institution in São Paulo. Closing these gaps requires better architectures, and training better architectures requires more compute.
The Hardware Layer as Strategic Moat
Anthropic's decision to anchor its compute strategy around Google TPUs rather than NVIDIA GPUs carries strategic significance well beyond one company's procurement decision. It reflects both the maturation of alternative accelerator ecosystems and the imperative to diversify supply chains — a lesson the global technology industry absorbed painfully during the pandemic-era chip shortages that disrupted manufacturing from Detroit to Wolfsburg.
When a single laboratory places an $11 billion order for a specific chip architecture, it is not merely buying compute — it is shaping the supplier's hardware roadmap and locking in preferential capacity for years. For nations attempting to build sovereign AI capability, the message is pointed: the window for securing meaningful infrastructure access may be narrowing.
OpenAI's energy play operates on a similar logic. By securing 10 gigawatts of power capacity, the company is effectively becoming an infrastructure utility — one with direct implications for national grid planning in whichever jurisdictions host its data centres. The energy demands of frontier AI are already straining power networks in Ireland, where American hyperscalers have concentrated European data centre operations, and sparking regulatory debate in Denmark, Sweden, and the United States over how AI's electricity appetite should be managed and priced.
What This Means for the Rest of the World
For countries outside the current AI superpower duopoly — the United States and China — the infrastructure arms race presents both risk and opportunity. Nations that can offer reliable energy, skilled workforces, and stable regulatory environments are positioning themselves as attractive hosts for the data centre capacity that AI demands. The Gulf states, particularly the UAE and Saudi Arabia, have moved aggressively in this direction, leveraging sovereign wealth funds and energy abundance to attract hyperscaler investment. Singapore and Malaysia are competing fiercely for Southeast Asian data centre mandates. Poland, Spain, and the Nordic countries are emerging as significant European hubs.
Yet access to infrastructure does not automatically translate into access to the models trained on it. The concentration of frontier AI development in a small number of American laboratories — with Chinese counterparts increasingly walled off by export controls — raises legitimate questions about whether the benefits of the current investment wave will be broadly distributed, or whether they will consolidate advantage among an already narrow set of actors.
The billions being spent today are not just buying chips and kilowatts. They are buying influence over the trajectory of a technology that will touch virtually every sector of the global economy. How that influence is distributed — and governed — is perhaps the defining industrial policy question of this decade.
Sources:
1 News Report, "Are We Ready for Multi-Image Reasoning? Launching VHs: The Visual Haystacks Benchmark!"
2 Globe Newswire, "Global AI-Powered Humanoid Robots Market Size Expected to Reach $7.73 Billion as Engineering Drastic" (December 08, 2025)
3 Yahoo Finance, "Asia-Pacific B2B Payments Report 2025-2030: A $1.15 Trillion Market Featuring Leading Competitors - " (February 05, 2026)
4 Yahoo Finance, "B2B Payments Global Report 2025: A 15.88 Trillion Market by 2030 from $11.69 Trillion in 2024 with C" (February 05, 2026)
5 Yahoo Finance, "Broadcom (AVGO) Q4 2025 Earnings Call Transcript" (December 12, 2025)