Infrastructure & Hardware

NVIDIA details Transformer Engine optimizations for dropless MoE training in JAX: grouped GEMM for ragged expert shapes, MXFP8 quantization, and fused dispatch/combine via NCCL EP. DeepSeek-V3 671B throughput rose from 103 to 1,068 TFLOPS/GPU (10.4x), with 97% scaling efficiency at 1,024 GPUs.

Brookfield Asset Management agreed to acquire a minority interest in American Real Estate Partners, parent of PowerHouse Data Centers, for an undisclosed sum, closing in Q4 2026. PowerHouse reportedly has an 8GW-plus pipeline; Brookfield calls it early days of a multi-decade AI infrastructure buildout.

Damac Digital and Vodafone Türkiye inaugurated the first phase of a data center in Izmir, launching at 4MW with plans to reach 20MW. Total investment has grown to $300 million; the seismically isolated site holds 650 racks. Damac Digital targets over 700MW operational by Q1 2027.

Edged announced that OMA01-1 and OMA01-2 in Council Bluffs, Iowa have been topped out, with Turner Construction and Jacobs on the build. Started in October 2025, the Project Lola campus comprises two roughly 285,000 sq ft buildings and is expected to total around 200MW.

A 36-year-old Chinese construction worker died after being hit by a reversing crane lorry at an unnamed data center site in Rayong, Thailand, on September 3. The driver and supervisor were questioned but not charged. A local official called for stricter safety standards and inspections at data center builds.

APAC operator DayOne broke ground on Phase I of its Kodaira Data Center in Tokyo, delivering 15MW initially and 42MW at full build. It follows an 80MW Fuchu campus being developed with Gaw Capital, whose 18MW first phase is due core-and-shell by mid-2027. DayOne was spun out of GDS in 2025.

Marsh's Michael Mathews argues that the scale of data center development is changing how risk is understood and capacity structured. A $500 million project no longer stands out, city-scale builds are common, and a $10 billion 2020 facility could now cost $14-15 billion, with generation assets extending exposure beyond property losses.

Rupesh Mainali of Reliability Engine argues that the first coolant problem in an AI data center may surface as odd rack behavior or slowed training jobs rather than a cooling failure. Liquid cooling introduces a fluid system with a history; if small changes in coolant condition are not preserved, they become hard to interpret later.

Thor Equities submitted plans for a 150MW data center in Kingston, Bartow County, Georgia. The Kingston Technology Hub is estimated at $1 billion, targets completion in 2030, and would span 350,000 sq ft, generating about $2 million a year in local tax revenue and around 50 full-time jobs.

IEEE Spectrum details how OpenAI designed Jalapeño, its first in-house AI accelerator unveiled Aug 25 (13.4 PFLOPS 4-bit, 232 GB memory, up to 3.6x lower latency than Nvidia's GB300). A team of under 100 paired internal LLMs with the XLS high-level synthesis flow, going from concept to first silicon in under 20 months and RTL to tapeout in nine. After first silicon, AI-written kernels rose from 0.31% to 88.94% of theoretical peak in about 40 hours; Broadcom handled physical design.

London-based Teravolt forecasts AI demand of 410GW by 2036 against 170GW available from the grid, a 240GW shortfall. Since grid buildout takes 5-15 years versus 1-3 for data centers, it argues repurposing Bitcoin farms, aluminum smelters, and old thermal plants, with permits and infrastructure in place, is the more lucrative path.

Nigel Jakeman of Turbo Power Systems says AI data centers face a speed-to-power wall as transmission upgrades take 5-10 years. Rather than waiting in utility queues, operators should design sites with battery storage, power electronics and microgrid principles that stabilize the grid, citing EPRI's Flex MOSAIC work. An 800V HVDC backbone cuts conversion stages from four-plus to two, yielding over $1.2M a year in opex savings at 10MW scale.

Schneider Electric's Steven Carlini says rack densities jumping from 20kW to over 140kW have made air cooling obsolete for chips like NVIDIA Blackwell. Moving direct-to-chip liquid cooling from pilot to production brings eight challenges, including galvanic corrosion from CDU material mismatches, fluid chemistry and pressure deviations, shared-asset warranties and chiller plant trade-offs, requiring IT and facility systems to be engineered as one appliance.