Traditional DC: 10 kW per rack.
NVIDIA GB300 NVL72: 132 kW per rack.
13x density increase. Air cooling hits a physics wall at 30 kW. Every AI rack above that needs liquid cooling.
Liquid cooling market: $2.84B in 2025, projected $21-44B by 2032. 33% CAGR driven by thermodynamics.
Facilities that cannot adapt become stranded assets.
The hyperscaler CAPEX race in 2026:
Amazon: $200B
Alphabet: $185B
Microsoft: $120B+
Meta: $100B+
Combined: $600B+. ~75% tied to AI infrastructure.
Meta reinvesting 57% of revenue into infra. No precedent in tech history.
Jensen Huang: "The largest industrial buildout in human history."
They are not building data centers. They are building AI Factories.
https://resistancezero.com/article-18.html
#HyperscalerCapex #AIFactory #BigTech #CloudComputing #Infrastructure
The network war nobody talks about: Ultra Ethernet vs InfiniBand.
UEC 1.0 launched June 2025. Open Ethernet now approaches InfiniBand performance.
Ultra Ethernet wins inference scale-out. InfiniBand holds frontier training (sub-1us latency, proven at 100K+ GPUs).
AI hardware refreshes under 5 years. Facilities that cannot support 3-ton racks or liquid cooling face obsolescence.
The network war nobody talks about: Ultra Ethernet vs InfiniBand.
UEC 1.0 launched June 2025. Open Ethernet now approaches InfiniBand performance.
Ultra Ethernet wins inference scale-out. InfiniBand holds frontier training (sub-1us latency, proven at 100K+ GPUs).
AI hardware refreshes under 5 years. Facilities that cannot support 3-ton racks or liquid cooling face obsolescence.
The hyperscaler CAPEX race in 2026:
Amazon: $200B
Alphabet: $185B
Microsoft: $120B+
Meta: $100B+
Combined: $600B+. ~75% tied to AI infrastructure.
Meta reinvesting 57% of revenue into infra. No precedent in tech history.
Jensen Huang: "The largest industrial buildout in human history."
They are not building data centers. They are building AI Factories.
https://resistancezero.com/article-18.html
#HyperscalerCapex #AIFactory #BigTech #CloudComputing #Infrastructure
Traditional DC: 10 kW per rack.
NVIDIA GB300 NVL72: 132 kW per rack.
13x density increase. Air cooling hits a physics wall at 30 kW. Every AI rack above that needs liquid cooling.
Liquid cooling market: $2.84B in 2025, projected $21-44B by 2032. 33% CAGR driven by thermodynamics.
Facilities that cannot adapt become stranded assets.
Singapore and Batam are 20 km apart. Batam electricity costs 60% less. Land costs 40-100x less. Labor is 4-5x cheaper.
Yet most hosting companies still pick Singapore. Why?
I wrote an objective analysis with a decision matrix for six workload types. The answer depends entirely on what you are running.
https://resistancezero.com/article-19.html
#DataCenter #Singapore #Batam #Infrastructure #SoutheastAsia #CloudComputing
The Singapore-Batam Digital Corridor is already being built.
Front-end in Singapore (peering, low-latency apps). Back-end in Batam (AI training, storage, DR). A dedicated 24-fiber-pair submarine cable connects them at under 2ms latency.
This is what happens when one location runs out of space and the other has cost advantages too large to ignore.
Full breakdown: https://resistancezero.com/article-19.html
#DataCenter #DigitalInfrastructure #SubmarineCable #SiteSelection #TechInfrastructure
Decision matrix for data center site selection in Southeast Asia:
Latency-sensitive front-end → Singapore (5/5)
AI training / batch → Batam (5/5)
Enterprise compliance → Singapore (5/5)
Cost-sensitive compute → Batam (5/5)
Multi-site resilience → Both
Singapore has 26+ submarine cables and 260+ peering members. Batam has 60% cheaper electricity and 40-100x cheaper land. The right answer depends on the workload.
Decision matrix for data center site selection in Southeast Asia:
Latency-sensitive front-end → Singapore (5/5)
AI training / batch → Batam (5/5)
Enterprise compliance → Singapore (5/5)
Cost-sensitive compute → Batam (5/5)
Multi-site resilience → Both
Singapore has 26+ submarine cables and 260+ peering members. Batam has 60% cheaper electricity and 40-100x cheaper land. The right answer depends on the workload.
