insights
Google, NVIDIA and Emerald AI Launch Flexible-Data-Center Alliance
Google, NVIDIA and Emerald AI launched AEMA to promote flexible AI data centers; its 100 GW figure is modeled potential, not installed capacity.
Insights
Everything we've published in insights.
-
-
insightsOracle and RWE Sign 433 MW West Texas Wind VPPA
Oracle’s 433 MW RWE wind VPPA covers three West Texas projects and supports its 2035 carbon-free AI data-center electricity goal.
-
insightsNVIDIA Vera Rubin NVL72 vs GB300 NVL72: MLPerf v6.1 Results
Vera Rubin's MLPerf preview shows up to 3.7× Qwen3-VL throughput over GB300, while four GB300 racks reached 99% scaling efficiency.
-
insightsAWS–Vistra 1.2 GW Comanche Peak PPA: 2027 Start, 2032 Ramp
Vistra’s SEC filing confirms a 20-year AWS nuclear PPA starting in Q4 2027 and reaching 1,200 MW by 2032—but not a dedicated data-center link.
-
insightsNVIDIA Dynamo Disaggregated Inference: Prefill, Decode and KV Routing
How NVIDIA Dynamo splits prefill and decode, moves KV cache with NIXL, and why one AKS test cut average TTFT 20.4× without proving universal gains.
-
insightsNVIDIA BESS Qualification: 12 Tests, 10 Core Requirements
What NVIDIA's BESS framework tests at the AC terminals—and why AI data-center developers still need site-specific electrical and fire-safety work.
-
insightsAlphabet Q2 2026 Capex: $44.9B Spent, $122.8B Not Yet in Service
Alphabet spent $44.9B on Q2 capex, held $122.8B of assets not yet in service and disclosed $85.2B in future uncommenced leases.
-
insightsMicrosoft AI Data Centre Capex: Cash vs Lease, FY2027 Accounting Shift
Microsoft’s $190B capex plan, the shift to $175B after lease reclassification, and what Q3/Q4 spending reveals about AI infrastructure.
-
insightsNVIDIA GB300 NVL72 vs HGX B300: GPU Memory and System Architecture Explained
A source-backed comparison of NVIDIA HGX B300 and GB300 NVL72: GPU count, HBM3e, memory bandwidth, NVLink scale, CPUs and rack power.
-
insightsNVIDIA GB300 NVL72 Networking Explained: NVLink, ConnectX‑8 and BlueField‑3
How GB300 NVL72 separates rack-scale NVLink, Spectrum-X GPU traffic, BlueField-3 CPU and storage traffic, and physical OOB management.
-
insightsGalaxy's 4.2 GW Texas Data Centers Get Conditional ERCOT Classifications
ERCOT conditionally classified five Galaxy data-center projects totaling about 4.2 GW. Here's what Base Load, Studied Load and verification mean.
-
insightsNVIDIA Invests $3.5B in MediaTek: What NVLink Fusion Adds
MediaTek will use NVLink Fusion for custom AI accelerators. What the $3.5 billion NVIDIA bond investment confirms, and what remains undisclosed.
-
insightsOpenAI's 1 GW Demand-Response Deal and Google Cloud's Utility AI
OpenAI's Georgia power deal, NextEra's reported AI savings and Santee Cooper's forecasting plans show different stages of grid flexibility.
-
insightsPrometheus Hyperscale Plans 1.5 GW Behind-the-Meter AI Campus in Texas
Prometheus Hyperscale and Istmo Energy plan a 1.5 GW behind-the-meter AI campus in Reeves County, with first power targeted for 2027.
-
insightsFloating Wind Meets AI: Exascale‑EnergyBank Norway Pilot
Exascale Labs and EnergyBank sign an MOU to test an off‑grid 800 kW AI data centre on Norway’s first floating wind turbine, with a roadmap to 10 MW per turbine.
-
insightsFlag Rebrands and Targets AI Data‑Center Connectivity
Flag drops GCX, invests $250 M and expands Pacific routes to serve AI‑driven data‑center traffic.
-
insightsVietnam's AI Data Centre Boom Bets Billions on One Fragile Subsea Cable
Vietnam's data centres are drawing billions in AI investment, but the buildout depends on a single submarine cable route and a tightening power supply.
-
insightsAmazon’s 199 MW Swedish PPA Deals Power AI‑Focused Data Centres
Amazon secures nearly 200 MW of wind power in Sweden, pushing its renewable portfolio toward 1 GW to fuel expanding AI‑intensive data centres.
-
insightsPJM Data‑Center Strategy: Grid Expansion vs. AI Load Surge
How PJM balances new generation, interconnection reforms and curtailment rules to accommodate AI data‑center growth.
-
insightsCerebras Overclocks WSE‑3 – Compute Power Density vs NVIDIA GB300
A data‑center‑focused look at Cerebras’s WSE‑3 Turbo over‑clock, its PFLOPS output, power delivery and how it changes rack architecture compared with NVIDIA’s GB300 NVL72.
-
insightsMicron $10B AI Memory Research Hub in Idaho Boosts Compute Power Density
Micron unveils a $10 billion AI memory research campus in Boise, Idaho, and what it means for AI compute power density.
-
insightsPJM’s $555/MW‑Day Backstop Auction and DOE’s Eddystone Extension: Inside the AI Power Crunch
PJM caps its Sept-Oct backstop auction at $555/MW-day to close a 6.8 GW gap, and DOE keeps Eddystone online through Nov - both driven by AI data-center demand.
-
insightsAI Infrastructure Capex Outlook: $3 Trillion by 2030 and the Drivers Behind It
We break down the $3‑4 trillion AI capex forecast, the $500B financing coalition, power demand spikes and revenue gaps shaping the AI build‑out.
-
insightsEdgeAI Foundations: Manchester Edge Data Center, Home‑Based GPU Networks, and EdgeCore Power Models
We break down the Manchester edge data‑center proposal, home‑based GPU networks, and EdgeCore's power and networking strategies for AI workloads.