Global AI Infrastructure Buildout Accelerates: 36 Data Center Projects Close Financing in One Week

The Megawatt Race Is On

DeepSeek's reported plan to deploy a 160,000-chip Huawei cluster in Inner Mongolia marks a watershed moment. Chinese AI startups, locked out of Western semiconductors and cloud infrastructure, are building their own data centers at scale. Meanwhile, GLP's 200MW facility in Foshan and Greenfield Partners' €360 million Milan investment show that this infrastructure arms race is genuinely global.

The geographic spread matters. Six of this week's 36 projects are outside North America—China (major clusters), Europe (Milan, Turkey), Malaysia, and the UAE. This is different from prior years, when US cloud providers held a near-monopoly on high-performance computing capacity. Now, regional players are investing to serve local demand and reduce dependence on foreign infrastructure.

DAMAC Digital and Vodafone Türkiye's announcement of the Aegean region's largest data center is a case study in this pattern. Turkey is building hyperscale capacity not to serve global clients (like AWS or Azure), but to anchor AI startups and software companies across Southern Europe and the Middle East. Similar reasoning explains the Malaysia facility being built by Key ASIC and CT Vision—Southeast Asian AI companies need somewhere to train their models.

North America still dominates (13 of 36 projects), but the concentration in Dallas, Texas hints at a shift too. Hyperscalers are moving beyond Silicon Valley and coastal metros, citing real estate scarcity, power grid limits, and cost. A 54MW joint venture in Dallas costs roughly a quarter of equivalent capacity in the Bay Area.

The GPU Shortage Is Forcing Hardware Decisions

Of the 36 infrastructure projects tracked, nine explicitly focus on semiconductor or GPU deployment. That's 25 percent—a far higher ratio than in previous years. Companies can no longer assume they'll source Nvidia H100s or similar chips from standard cloud providers. Instead, they're funding dedicated semiconductor facilities and custom chip design.

DeepSeek's Huawei Ascend 950DT cluster is the headline example, but it's part of a broader shift. Key ASIC's Malaysia project is building GPU-optimized infrastructure specifically. Humain's Oxagon facility in Saudi Arabia (part of the Neom project) is designed to handle proprietary silicon. These aren't generic data centers—they're built around specific chip architectures, since each AI vendor now has different hardware needs.

Quantum computing projects (3 announced this week) are still experimental, but they're competing for the same pool of venture and infrastructure capital. Each quantum project requires specialized cooling, power isolation, and networking—different from traditional data centers but equally capital-intensive.

Power Is the New Limiting Factor

Every project announcement this week included a megawatt specification: 200MW (Foshan), 160MW (estimated DeepSeek), 150MW (Humain), 100MW (DAMAC), 54MW (Dallas). Power availability—both generation and transmission—is now the primary constraint for AI infrastructure deployment.

This has regional implications. Europe's projects (Milan, Turkey) are tied to hydroelectric and renewable sources. Middle Eastern investments emphasize proximity to low-cost solar and natural gas. Malaysia's facility leverages tropical hydropower. Meanwhile, China's Inner Mongolia cluster is co-located with coal and wind capacity.

The power limitation also explains why data centers are moving further from major cities. Urban real estate boards and local governments are resisting 200MW facilities in metropolitan areas. Dallas, Foshan, the Aegean coast, and Inner Mongolia were chosen partly because they're far enough from population centers to source or build power infrastructure without political opposition.

Who's Funding This?

The investors are a mix of familiar names and newcomers. Greenfield Partners and Finsbury Infrastructure (existing European infrastructure managers) are diversifying into AI compute. GLP (Global Logistic Properties, a real estate company) is converting industrial space. DAMAC (Dubai luxury developer) is pivoting to tech infrastructure. Meanwhile, private equity and sovereign wealth funds are making discrete bets on individual projects.

What's striking is the absence of a single dominant player. Unlike cloud infrastructure (dominated by AWS, Azure, Google Cloud), AI compute infrastructure is fragmenting. Some companies are funding captive facilities for internal model training. Others are building to rent capacity. Still others are hedging: investors like Greenfield are backing multiple geographies simultaneously, betting that no single region will monopolize AI training.

DeepSeek and similar Chinese startups are forced into captive infrastructure because they can't reliably access foreign cloud capacity. But even Western startups (reflected in the VC-backed projects in this week's data) are considering dedicated facilities to control costs and performance—a major shift from the "just rent from AWS" era.

What This Means for AI Economics

Large language models and AI services are about to get cheaper and more distributed. Right now, a startup in Manila or Lagos training a model must either build their own data center (capital-prohibitive) or pay premium rates to rent Nvidia-powered capacity in Singapore or the US. This week's 36 projects are the first wave of regional infrastructure that will eliminate that penalty.

In 12-18 months, when facilities like Milan, Malaysia, and Oxagon come online, model training costs in Europe and Southeast Asia will fall to parity with US rates. This democratizes model development but also fragments the competitive landscape. A Chinese startup can train on Huawei chips at Inner Mongolia rates. A European SaaS company can use Milan data center capacity. Neither needs to negotiate with American cloud giants.

Paradoxically, this creates new dependencies. A startup's choice of data center locks it into specific silicon (Huawei Ascend, Nvidia H100, custom ASICs). Migration costs rise. Switching from one regional facility to another is no longer trivial. The "data center fragmentation" era has its own lock-in dynamics—just different from AWS.

Governments are also watching closely. Investments like Neom's Oxagon and China's Inner Mongolia cluster come with implicit geopolitical dimensions. The US will continue restricting chip exports to some nations. Europe will push for EU-hosted infrastructure, partly for data sovereignty, partly to reduce cloud provider dependence. India and Southeast Asia will see multiple regional players competing for market share.

The Buildout Accelerates

Infrastructure investment in AI compute is now the rate-limiting step for model training and deployment. This week's 36 projects—billions in committed capital across six continents—show that the bottleneck is no longer ideas or algorithms. It's power, silicon, and real estate.

Expect announcements to accelerate. Every major cloud provider is expanding AI-specific infrastructure. Regional players are racing to build before foreign competitors establish local operations. Startups with the capital to fund captive facilities are doing so to gain cost and performance advantages.

By end of 2027, the data center landscape will look fundamentally different: regional nodes, custom silicon, power-constrained expansion, and a much higher barrier to entry for new AI companies without billions to spend on infrastructure. The companies winning the AI race aren't just writing better algorithms. They're controlling the power grids and chip supply chains that make training those algorithms possible.