May 22, 2026

Google is no longer just “building data centres” — it’s effectively assembling an AI supply chain that runs from silicon to megawatts. The clearest signal today: Google and Blackstone form AI TPU data center joint venture, with Blackstone putting in $5bn of equity and an explicit target of 500MW online by 2027. Stack that next to Google $15 Billion Missouri Expansion to Add Over 1 GW, and you can see the new playbook: capital markets + proprietary accelerators + power contracting, stitched into one programme.

The Big Stories

The Google–Blackstone joint venture is a big deal not because “$5bn” is a big number (it is), but because it formalises AI compute-as-a-service around Google’s TPU stack. Blackstone is contributing an initial $5bn in equity, while Google supplies TPUs, hardware, software and services, with 500MW expected online by 2027. If you’re a competitor, the uncomfortable part is that this isn’t just capacity; it’s capacity bundled to a specific accelerator ecosystem and distribution model.

Google also went very concrete on power in the US heartland. In Missouri, it announced a $15bn investment to build a new data centre in Montgomery County, contracting over 1GW of new generation capacity and supporting an additional 500MW with Ameren and Evergy. There’s also a $20m Energy Impact Fund aimed at household bill reductions, and the build will use advanced air-cooling to limit freshwater use. The subtext: the “social licence” and the megawatt story are now inseparable in US markets.

Amazon matched the long-duration commitment vibe in Southeast Asia. Amazon commits over $33bn to Southeast Asia cloud and AI across Indonesia, Malaysia, Singapore and Thailand through 2039, alongside projections of over $64bn GDP contribution and 56,300 FTE jobs in the local data centre supply chain. Amazon also pointed to renewable energy moves — including two projects in Singapore and a 210MW green tariff in Indonesia with PT PLN — which reads like a pre-emptive answer to the inevitable “where does the power come from?” question.

On the technology side, NVIDIA used COMPUTEX to push a very pointed message: rack-scale AI is now a systems-and-infrastructure product, not just a chip. NVIDIA wins COMPUTEX awards for AI infrastructure innovations, led by the Vera Rubin NVL72: a rack-scale, 100% liquid-cooled system (36 Vera CPUs and 72 Rubin GPUs) claiming up to 10x higher inference performance per watt and 10x lower cost per token, plus 6x more onboard energy storage for power smoothing. Whether every claim holds up is almost secondary; what matters is NVIDIA leaning hard into power, cooling, and grid-friendliness as first-class product features.

Energy storage also got a rare “data centre is the headline” moment in Europe. Invinity to deliver world’s largest flow battery for Swiss data centre at Flexbase’s site under construction in Laufenberg, Switzerland, now moving into the engineering phase with Equans Switzerland and Georg Fischer as local partners. Flow batteries aren’t the default choice for data centres today; seeing “world’s largest” attached to a live DC project is a reminder that longer-duration, cycling-friendly storage is creeping from theory into actual facility design.

Behind the Headlines

The most useful long-range reality check today came from modelling, not a deal announcement. Data center demand could drive 1,000 GW solar by 2050 says BloombergNEF sees nearly 1,000GW of utility-scale solar, 400GW of battery storage, and large additions of gas (370GW) and even coal (110GW) tied to data centre growth by 2050 — with thermal plants (~500GW) and fossil fuels providing about 51% of incremental generation for data centres. The uncomfortable takeaway for anyone selling a “clean AI” narrative is that the grid build-out implied here is brutally physical, and it doesn’t automatically resolve to renewables-only without policy and permitting doing a lot of heavy lifting.

KPMG put sharper numbers on the near-term version of the same problem. KPMG warns AI’s $15.7tn prize carries environmental costs projects data centre electricity use rising from ~415TWh (1.5% of global demand) in 2024 to ~945TWh by 2030, with water use jumping from ~560bn litres to ~1.2tn litres. This is why the Google Missouri note about advanced air-cooling isn’t just PR — water constraints are getting dragged into procurement decisions and site selection faster than many operators expected.

Connectivity can still be a bottleneck even when capital and silicon are abundant. Permitting, locates, and material costs slow BEAD fiber construction reports operators warning that permitting backlogs, locating failures, and higher materials costs are stretching BEAD timelines, with advice to prepare 6–12 months ahead using pre-drafted permits and one-year blanket permits. Duct pricing moving from about $0.25/ft pre-pandemic to $0.45–$0.50/ft now is the kind of mundane cost inflation that quietly breaks models. For data centres, it’s a reminder that “power and land” is only half the siting equation — fibre execution risk is real, and it’s showing up in public funding programmes first.