Crux AI TPU Neocloud 2026: What UK Buyers Should Know

Crux AI launches with Google and Blackstone money behind it

The neocloud that Google and Blackstone quietly assembled to sell TPU capacity has officially named itself Crux AI. It's led by Benjamin Treynor Sloss, a former Google engineering VP and now Crux AI CEO, and it's backed by $5bn from Blackstone. That's a serious war chest for a company that has yet to bring its first planned 500MW of capacity online.

Treynor Sloss described the ambition as building 'an integrated platform designed end-to-end to operate reliably at enormous scale', with a goal of becoming, in his words, 'the infrastructure company trusted with the workloads that matter most.' For UK infrastructure buyers, the relevant question isn't the branding — it's how fast that capital turns into leasable racks, and who gets first access.

A 500MW start on the road to a 2GW ramp

Crux AI plans to bring 500MW of capacity online in 2027. On its site, the company says it plans to scale to 2GW of capacity overall. The company says its footprint will be a mix of powered shells, build-to-suit facilities, colo blocks, turnkey partner deals and first-party greenfield development — not a single delivery model but several running in parallel.

That mixed-delivery approach matters for buyers tracking capacity globally, including through resources like our UK AI data centre buildout tracker. A neocloud spreading risk across shells, build-to-suits and colo deals is effectively hedging against the exact bottleneck UK buyers already know intimately: power availability and construction lead times, a theme we've covered in depth around power as an AI bottleneck.

Hiring Meta's data centre chief signals delivery discipline

Crux AI's more telling move this week was hiring Alan Duong, who spent more than 12 years at Meta and led the team that redesigned Meta's paused data center projects for the AI era. He joins as chief development officer, responsible for what he called delivering 'multiple GW of data center capacity over the coming years.'

Duong's own framing is worth noting: he described data centre delivery as 'a chain that has to hold, end-to-end' — from site origination with real power, through community relations, efficient design, safe construction, and handover to operators who run facilities for decades. Crux AI also hired the CFO of Charter Communications earlier this year and currently lists 17 open roles, including a data centre construction manager, an energy manager, a head of operations, a lead supply negotiator and a VP of capital markets. That's an operator building an execution machine, not just a fundraising vehicle.

What this means for hyperscaler colocation pricing power in the UK

UK buyers negotiating colocation or GPU capacity in 2026 widely report tight power availability and long connection queues. A well-capitalised TPU neocloud scaling toward 2GW globally adds another large buyer chasing the same sites, contractors and grid connections that UK-based data centre developers depend on — even before Crux AI confirms any UK footprint specifically.

This is compounded by other capacity commitments already reshaping the market: Hypertec and 5C state they are targeting up to 2GW of AI-dedicated capacity for Together AI, including nearly 100,000 Nvidia Blackwell and future-generation GPUs. A July 2026 DCD report on Hypertec and 5C said 5C's European rollout includes 2GW of new capacity through 2029, with the UK named alongside France, Italy and Portugal as a priority market for first deployments. Buyers evaluating hosting high-density AI racks in UK colocation should treat these overlapping build-outs as evidence that pricing leverage is moving toward whoever secures power and shells first, not whoever negotiates hardest later.

Inference demand, not training, is now driving the numbers

JLL's 2026 Global Data Center Outlook projects AI inferencing will overtake training as the dominant AI data centre workload by 2027, and that nearly 100GW of new capacity will be added globally by 2030 as demand redistributes toward regional, lower-latency sites rather than only giant centralised training campuses. Nvidia's April 2026 materials claim the B200 reaches two cents per million tokens on GPT-OSS-120B, with a 15x reduction in cost versus the prior generation, and say hyperscalers are deploying nearly 1,000 NVL72 racks per week.

HPE has separately claimed its on-premises AI platform can cut token costs by up to 60% versus public cloud, and Sify is building smaller inference facilities across 20 secondary Indian markets rather than concentrating everything in megacampuses. The pattern for UK buyers is consistent: capacity decisions increasingly hinge on token economics and proximity to users, which is exactly the calculation our calculate GPU server needs for LLM inference tool is built to support.

How UK buyers should position for the ramp

HPE's internal analysis suggests its on‑premises AI platform can reduce token costs by up to 60% versus public cloud, and Sify plans smaller inference facilities across 20 secondary Indian markets rather than concentrating everything in megacampuses. The pattern for UK buyers is consistent: capacity decisions increasingly hinge on token economics and proximity to users, which is exactly the calculation our calculate GPU server needs for LLM inference tool is built to support.

• •Model inference workloads now against current token-cost benchmarks before committing to long colo terms
• •Review financing options for AI servers so capital isn't the reason you lose a queue slot to a better-funded rival
• •Cross-check any UK site quote against our AI servers data study and the wider UK AI data centre buildout tracker
• •Use the AI GPU sizing calculator to pressure-test capacity requests before signing multi-year colo commitments