Databricks spends $3B abroad as power, not money, limits AI buildout

The spending

Databricks is investing approximately $3 billion across Europe and the Middle East to accelerate enterprise AI adoption, the company says. The announcement also covers more than $500 million in the Netherlands and more than $350 million in Singapore, the company says, which shows the spending is not a single-country bet but a broad push to put capacity and staff near enterprise customers. For a company selling data and AI platforms to large organisations, being close to the customer matters for data rules, latency and sales. m:…3895

At home, Databricks opened a new 235,000-square-foot headquarters at One Sansome Street in San Francisco, the company says, and is investing over $1 billion into its San Francisco operations by early 2028. The company has also committed to hosting its Data + AI Summit in San Francisco through 2030, the company says, and the 2026 summit drew over 30,000 attendees, according to the company. A headquarters of that size is a statement of intent: Databricks wants its research and product teams in one place, where much of its hiring happens. m:…5218

The spending is backed by growth figures the company reports itself. Databricks grew over 80 percent year over year and surpassed a $7 billion revenue run-rate as of August 2026, the company says, and it reports more than 12,000 employees worldwide. Those numbers are company figures, not independently verified results, but they set the scale of the bet: spending about $3 billion across Europe and the Middle East means planning for demand to keep growing, not for a pause. m:…3895 m:…5218

The power problem

The constraint on that growth is not money. The UK government estimates the country will require 6 GW of AI data centre capacity by 2030, according to GridBeyond's account of the new Data Centre Covenant, and reaching that target requires a threefold increase in the current estate. Data centres currently account for 73 GW of the UK's electricity demand connection queue, the same account says, while the UK's peak electricity demand in 2025 is approximately 45 GW. In plain terms, the queue for grid connections is larger than the entire country's peak demand, so most projects in the queue will wait, shrink or never connect. m:…4126

The UK government's answer is the Data Centre Covenant, which sets five tests for strategic data centre projects covering energy costs, communities, environment, jobs and security, according to GridBeyond. Strategic data centres are required to design for a power usage effectiveness of 1.3 or better, the account says, and Ofgem is proposing a Data Centre Commitment Fee of 237,500 to 712,500 pounds per megawatt for projects above 40 MW. The fee and the tests are designed to push speculative projects out of the queue. m:…4126

Flexibility is now part of the deal. The government is exploring the power to direct large users to temporarily curtail demand as a last resort, according to GridBeyond, and flexible connections, including non-firm supply and hybrid arrangements, are considered a necessary intervention for grid access. A new ministerial centre will use powers under the Planning and Infrastructure Act to speed up strategic data centre connections, the account says. The message to developers is direct: a fast connection is available, but only if the project can bend when the grid is under strain. m:…4126

Working around the grid

WhiteFiber's answer to the same problem is to stop treating a data centre as one building. Power availability is the primary constraint for AI infrastructure scaling, according to the company's account, because data centre power budgets are fixed by facility, utility connection and cooling capacity. WhiteFiber Continuum links the QTS ATL1 and QTS SUW1 data centres into a single GPU cluster, the company says, pooling the power capacity of two sites instead of waiting for one site to get a bigger connection. m:…0223

The two facilities are 83 kilometres apart, according to WhiteFiber, and the inter-site link provides 136 terabits per second of aggregate bandwidth with round-trip latency of 0.9 milliseconds across the span. The system currently runs 512 Nvidia H200 GPUs in an active-active configuration, the company says. The design uses 12 dark fibre strands leased from Zayo and a network fabric built on DriveNets AI Fabric, according to the account, with WEKA NeuralMesh replicating data intelligently to the site where the GPU job lands. The point is that the cluster behaves as one machine even though its power comes from two separate grid connections. m:…0223

What this means for buyers

For executives, the two halves of this story pull in the same direction. Databricks is spending about $3 billion across Europe and the Middle East because enterprise demand for AI is there, and the UK's estimate of 6 GW of AI data centre capacity needed by 2030 shows governments expect that demand to keep rising. But the 73 GW queue shows that announcing capacity is easier than connecting it. A vendor's investment plan is not the same as a live data centre, and the gap between the two is measured in grid connections, not in dollars. m:…3895 m:…4126

The practical questions are therefore about power, not product. Can a supplier show a grid connection with a date, or only a plan? Can a project accept a flexible connection and curtailment, or does it need firm power from day one? And can the work be spread across sites, as WhiteFiber does with its 512 GPUs, so that no single connection becomes the bottleneck? Databricks' spending answers the demand question. The grid queue answers the harder one. m:…0223 m:…4126