Compute & Infrastructure

Amazon has received permits for a dedicated 35-turbine, 7.65GW natural gas power plant in Texas, purpose-built to supply an AI data center complex. The facility is authorized to emit up to 33 million tons of CO₂ per year — a figure that, if fully realized, would make it the single largest point-source of greenhouse gas emissions in the United States. This is a confirmed regulatory authorization, not a speculative proposal, though construction timelines and actual utilization rates remain unconfirmed. Tom's Hardware

The strategic logic is clear: Texas's deregulated ERCOT grid offers speed-to-capacity advantages unavailable in more regulated markets, and co-located gas generation eliminates exposure to grid instability events like Winter Storm Uri. However, this move represents a structural reversal from Amazon's 2040 net-zero commitments and will intensify regulatory and investor scrutiny. A separate development — Amazon's use of 45-year-old procedural rules to bypass community comment in Gilroy, California — suggests the company is systematically prioritising build velocity over community engagement, a posture that carries rising political risk in jurisdictions with stronger local government authority. Tom's Hardware

TSMC's July 2026 monthly revenue rose 45% year-on-year, reported on August 10, consistent with the trajectory of prior quarters and confirming that AI training and inference silicon orders are absorbing demand even as broader equity markets show volatility. Bloomberg The foundry remains the single most critical chokepoint in the AI hardware supply chain: it manufactures the advanced nodes (3nm and below) required for NVIDIA's Blackwell and successor architectures, and no credible alternative at those process nodes exists at scale.

Simultaneously, Sony and TSMC are advancing a joint $6.4 billion investment in a Japanese image sensor and specialty chip facility. This is reported as being 'in talks' — confirmed intent but not a finalized agreement. Bloomberg The Japan facility is strategically distinct from TSMC's advanced logic fabs: it targets image sensors and specialty processes, not leading-edge AI accelerators. Its significance lies in geographic diversification and in signalling that TSMC's Japan strategy — partially funded by Japanese government subsidies — is expanding beyond the JASM Kumamoto logic fab into adjacent verticals.

Moore Threads Technology, China's most prominent domestic AI GPU designer, has announced plans for a Hong Kong listing at an 'appropriate time,' following a 420% share price surge since its Shanghai IPO in 2025. Bloomberg The Hong Kong listing is stated intent, not a filed prospectus — timing and valuation remain unconfirmed. However, the move is strategically significant: a Hong Kong listing would give Moore Threads access to international institutional capital while remaining outside US financial market jurisdiction, a structure that allows fundraising without triggering the foreign investment review mechanisms that a US listing would invite.

Moore Threads operates in a constrained environment — cut off from TSMC's leading-edge nodes by US export controls and dependent on SMIC's mature process technology for volume production. Its GPU architecture targets inference workloads where process-node disadvantage is less acute than in training. The IPO trajectory suggests Chinese AI chipmakers are now sufficiently mature to attract public market capital, which will accelerate R&D investment in domestic alternatives to CUDA and NVIDIA's software moat.

Delta's GoCool-150 liquid-to-air cooling distribution unit, designed to dissipate 150kW per rack for systems like ASRock Rack's NVIDIA NVL72 configuration, represents the current frontier of commercially available rack-level thermal management. ServeTheHome The NVL72 — a 72-GPU NVLink domain configuration — exemplifies the density trajectory of modern AI inference infrastructure, where rack power draw has moved from 10-20kW industry norms to figures requiring purpose-engineered liquid cooling loops.

This development matters operationally because most existing colocation data center facilities were not designed for 150kW rack densities and lack the structural, plumbing, and electrical infrastructure to support them. The addressable market for new, purpose-built AI data centers — rather than retrofitted general-purpose facilities — is therefore larger than hyperscaler announcements alone suggest, as third-party colocation providers scramble to qualify liquid-cooled infrastructure.

Visualizations of Elon Musk's proposed Terafab facility — projected to exceed 100 million square feet of interior space, larger than the Pentagon, Apple Park, Mall of America, and Giga Texas combined — have been published, illustrating the intended scope of an all-in-one semiconductor manufacturing campus. Tom's Hardware Critical context: Terafab remains an announced concept without confirmed construction timelines, equipment procurement agreements, or process node specifications. It should be categorized as speculative infrastructure at this stage.

The stated ambition — integrating multiple semiconductor manufacturing processes under one roof on US soil — would, if realized, represent the most aggressive attempt yet to replicate TSMC's integrated process capabilities domestically. However, semiconductor fab construction timelines of 4-6 years, equipment lead times (particularly for ASML EUV tools), and the workforce challenges of advanced node manufacturing make the scale of Terafab's vision extremely difficult to execute on any near-term horizon. It is better understood currently as a strategic signal about xAI's compute independence ambitions than as near-term capacity.