Good morning 👋
From a $3.5B GPU computing pact to subsea cable deals, this week’s Grid Boiler explores how AI’s thirst for power, data and connectivity is reshaping energy and telecom infrastructure.
🔥 The Big Boil
AI infrastructure ecosystem is accelerating on all fronts.
Hyperscalers and partners are contracting vast new capacity (computing, cooling, power and networking) even as regulators scramble to keep up.
For example, U.S. regulators are proposing to eliminate mandatory public review of new data‐center air permits, a signal of the pressure to speed projects despite grid and community impacts.
At the same time, industry surveys and insurance data show that global data center buildout is exploding: Verisk calculates U.S. data center counts have topped 2,500 and expects data-center insurance premiums to more than double from ~$10B in 2026 to $23B by 2030.
Corporations and financiers are pouring billions into every layer – from multi-gigawatt power infrastructure (Flex/Power and Cummins deals) to multi-Tbps cables (SubCo’s APX East) to high-density cooling technologies (Liteon/DCX investment) – often with formal financing programs or MOUs.
In summary, the AI boom is now a physical-industrial buildout: massive new sites (e.g. Meta’s $1.2B Idaho DC), enabling power and network “skeletons,” and novel finance vehicles (NVIDIA’s compute-lease platforms and private data-center funds) are all converging. This issue’s stories quantify that surge and its hotspots, while highlighting bottlenecks (power, permitting, water) and novel solutions (microgrids, edge deployment).
⚡ What’s Boiling
Nscale secures 100,000 NVIDIA GPUs in $3.5B AI contract
BNeocloud provider Nscale signed a multi-year deal with robotics company Figure to deploy up to 100,000 NVIDIA “Vera Rubin” GPUs at Nscale’s Barstow, TX data center (a 234 MW DC site previously leased from Bitcoin miner Ionic Digital). The deal’s initial value is $3.5 billion, with plans to scale above $6 billion as more GPU capacity comes online. This make Nscale a key compute provider for Figure (and already for Microsoft at Barstow) on a preferred basis.
Why it matters: It showcases unprecedented scale (100k GPUs, 234 MW) of contracted AI compute and signals the rise of specialist “compute factories” aligned with AI-first customers.
Flex to buy EPC Power for $4.4B, boosting AI power systems:
Singapore-based Flex Ltd. agreed to acquire EPC Power – a North American developer of high-voltage DC power conversion and solid-state transformer technology – for a reported $4.4 billion. The deal (closing expected Q4’26) folds EPC’s 800 V power architecture into Flex’s data-center power segment. EPC’s systems are tailored for ultra-dense AI loads, allowing higher-power data halls (e.g. 20–50 MW racks with DC distribution).
Why it matters: This vertical integration into “grid-to-chip” power could overcome one of AI’s main bottlenecks – delivering reliable high-density power – accelerating buildout of large AI data centers.
Evolution DC pre-leases 40 MW first phase in Thailand:
Evolution Data Centres announced that its first Thai campus (Bangkok “TH01” project) has a binding contract with a “tier-1 global hyperscaler” for Phase 1 capacity. The second phase – tailored for AI workloads – will add an additional 40 MW of IT load. The campus (in Bang Na, Bangkok) is scheduled to come online later this year, making it Evolution’s first facility outside Australia.
Why it matters: Hyperscalers are expanding AI-optimized capacity into Southeast Asia. A new 40 MW phase in Bangkok indicates rapid regional growth and reflects infrastructure alignment (both power and fiber) to support low-latency AI inference.
Telxius/Cable landing: new Cancún edge DC for Tikal cable:
Telefónica’s Telxius is constructing a carrier-neutral cable landing station and edge data center in Cancún, Mexico. The facility will host the Tikal subsea system (linking Puerto Barrios, Guatemala to Boca Raton, FL) scheduled for H1 2027. It consolidates international fiber, terrestrial backbone and local compute in one hub.
Why it matters: This bridges Central America to North America for high-speed connectivity and offers a new edge compute point near the Caribbean. It signals growing attention to distributed data capacity in emerging markets to reduce latency for regional AI services.
🧠 Boiled Down
Today’s Term: 800V DC Power Architecture
In plain English: a way of moving electricity through a data center at a much higher DC voltage, reducing current and electrical losses before power reaches AI servers.
Why you’re hearing about it: AI racks are demanding far more power, pushing operators to rethink traditional data-center electrical systems. Flex’s planned acquisition of EPC Power puts this technology directly into the spotlight.
Why it matters: As GPU density rises, getting electricity efficiently from the grid to the rack becomes a design constraint of its own. Better power architecture can mean less wasted energy, smaller electrical infrastructure and more compute supported inside the same facility.
🤯 Did You Know? Australian AI-cloud firm Firmus is committing $300 million for up to 150 Tbps of capacity over 25 years on the planned APX East subsea cable linking Australia and the U.S.
Till next time,
Grid Boiler Team