March 13, 2026

Bridge Data Centers just put a big number on the board in one of the world’s most power-constrained markets: a planned S$3–5bn build-out for more than 2GW of “AI-ready” capacity in Singapore. That scale isn’t just a flex — it’s a reminder that the AI infrastructure race is now as much about power strategy as it is about land, fibre, or GPUs. And in the background, US grid modelers are increasingly blunt about what happens when demand lands faster than generation.

The Big Stories

Bridge Data Centers plans S$3-5bn AI-ready data centre expansion is the day’s headline because it combines capital, capacity, and geography in a way that forces competitors to react. Bridge says it will invest S$3–5bn with global partners to develop more than 2GW in Singapore, and it’s explicitly floating hydrogen-based power and other alternative approaches to meet higher power needs and sustainability goals. The “AI-ready” label is doing real work here: it’s shorthand for high-density power and cooling, and it telegraphs that Singapore demand is being chased with designs that look more like compute plants than traditional colo.

On the power-systems side, the warning shot comes from Faster-than-expected data center growth could increase fossil generation. The U.S. EIA modeled a scenario where data-center-heavy regions see 50% higher growth in 2026–27, and found it could drive natural gas generation up 7.3% (123 BkWh across 2025–27) and lift wholesale prices — with ERCOT 2027 prices up to $37/MWh higher under the high-demand assumption. The key caveat is the point: this scenario assumes no additional generating capacity beyond the February STEO, which is exactly why it matters to investors — the price impact isn’t “AI needs gas,” it’s “AI arrives before the supply-side paperwork clears.”

Networking vendors are also trying to re-architect around AI density, and Arista came with a concrete claim. Arista launches XPO liquid-cooled optics for AI datacenters introduces a liquid-cooled, cold-plate pluggable optics form factor aimed at replacing OSFP in high-density AI builds, with 204.8T per open rack unit and support for up to 400W per module. Arista’s own example is telling: in a 400MW datacenter, it says XPO could reduce switch rack counts from ~1,400 to 352. Whether XPO becomes the default or not, the direction is clear: optics and thermals are becoming first-class constraints, not afterthoughts.

Cooling and flexibility R&D is getting federal dollars, too. DOE funds Berkeley Lab projects on load flexibility, cooling sits inside a $155m DOE investment across 16 projects, including three led by Lawrence Berkeley National Laboratory — REFLEX for load flexibility, a data center cooling testbed network, and an industrial R&D platform (FOOD Center). The near-term implication is less about a single breakthrough and more about an agenda shift: DOE is explicitly treating load-shaping and thermal performance as scalable levers for cost and grid impact.

In the deal file, RadiusDC to acquire phoenixNAP Phoenix colocation, expand to 26MW is a clean illustration of how second-tier markets are being assembled into real platforms. RadiusDC agreed to buy phoenixNAP’s Phoenix, Arizona data center and colo business (closing expected Q2 2026), expand DC1 to 8MW, and develop DC2 up to 18MW with initial phases online beginning H1 2028 — taking the Phoenix I campus to ~26MW. phoenixNAP keeps ~80% of its global business and stays on as a tenant, which reads like a liquidity event that still preserves operational continuity for existing customers.

Behind the Headlines

Local politics is increasingly a gating factor for large powered-land proposals, even when the developer’s pitch is “jobs and investment.” Gibraltar enacts moratorium on proposed 100MW Raeden data center shows the pattern: a one-year moratorium now covers Raeden’s proposed 100MW project at the former McLouth Steel site, despite Raeden estimating up to $2bn of investment and 60–100 permanent jobs. What moved the room wasn’t abstract skepticism — it was specific anxieties: air quality, diesel generator counts (30–35), potential environmental impacts (Humbug Marsh), and even electricity rates as Michigan regulators prepare a data center tariff filing. The lesson isn’t that projects won’t get built; it’s that “permit risk” is turning into “social license risk,” and timelines should be underwritten accordingly.

Geopolitics is also being reframed as a resilience design requirement, not just an insurance line item. Gulf conflict exposes data centre and connectivity resilience gaps is triggered by a 28 February escalation and Iranian strikes that it says affected three AWS data centres, and it translates that into four practical imperatives: data embassies, satellite/NGSO connectivity, cyber preparedness, and supply-chain scenario planning. The details worth dwelling on are the physical chokepoints — undersea cables like AAE-1, EIG, FALCON, IMEWE, PEACE — and the operational constraint that regulatory lead times are “measured in months.” If you’re building multi-region availability, the uncomfortable takeaway is that redundancy isn’t just another availability zone; it may require a different connectivity stack.

And the fiscal bargain between data centers and host states is starting to fray in the US. Environmental group urges end to Virginia data center tax exemption flags a live budget fight over Virginia’s sales tax exemption, with the Piedmont Environmental Council targeting a projected $1.9bn tax break as the House and Senate reconcile budgets. The Senate version would remove the exemption; the House would keep it; and Governor Abigail Spanberger is non-committal while floating a possible consumption tax on data center energy use. That mix — budget pressure plus energy politics — is exactly how “incentives” become “penalties,” and it’s something operators will need to price into their next wave of capacity commitments.