AI Infrastructure & LLMs
Large language models, AI assistants, data centres, AI chips, cooling, optical networks and the power systems behind modern artificial intelligence.
-
Sparse AI Models Still Need Dense Infrastructure
Mixture-of-experts models activate only part of their parameters for each token. The unused parts still shape memory requirements, and routing creates new work.
-
Data-Centre Backup Power Carries an Air-Quality Cost Beyond the Electricity Meter
Backup engines protect digital services, but operating hours and emissions controls determine their local footprint. Clean-grid claims do not answer every question.
-
AI Training Needs a Reliable Save Button as Much as Faster Chips
A training run must survive failures and save progress without holding thousands of GPUs idle. Checkpointing connects performance to storage and recovery.
-
A Cloud Server’s Second Life Requires More Than a Recycling Rate
Microsoft’s Circular Centers recover retired cloud hardware. A combined reuse-and-recycling rate cannot show how much equipment keeps working or what impact is avoided.
-
NVIDIA’s $1 Billion Science Pledge Faces a Test Beyond More Compute
NVIDIA has pledged $1 billion over five years for US science. New Genesis Mission projects show why computing access must connect to experiments and validation.
-
The Yandex Data-Centre Strike Turns Physical Resilience Into an AI Issue
A reported drone strike forced Yandex to suspend its Sasovo data centre. Core services continued, but the incident exposes the physical limits of cloud redundancy.
-
A National Compute Grid Wants to Turn Idle GPUs Into Shared Infrastructure
A new US consortium says it will pool underused AI capacity through a common scheduler. Its 760 MW pipeline and utilisation claims still need independent proof.
-
OpenAI’s 8 GW Ohio Campus Begins With an 800 MW Reality Check
OpenAI has joined an Ohio data-centre plan targeting about 8 GW of IT capacity. The first 800 MW is expected in 2028; permits, finance and power remain ahead.
-
AI Inference Is Becoming a Scheduling Problem Between Two Different Jobs
Reading a prompt and generating a response stress AI infrastructure differently. Separating the jobs can help, but it adds a memory-transfer and scheduling problem.
-
Osaka’s New AI Data Centre Starts With a Mix of Air and Liquid Cooling
Digital Realty has started building a 24 MW Osaka facility planned for late 2028. Hybrid cooling and connectivity are central to its AI infrastructure design.
-
Kolibri’s Open Weights Make Sovereign AI a Concrete Infrastructure Choice
Aleph Alpha’s new German-English model can run on infrastructure its users control. Open weights expand that choice while leaving hardware and validation work.
-
Japan’s Chiba AI Plan Connects Computing Expansion to the LNG Supply Chain
JERA, Dell and RHAELM propose a repeatable power-to-compute model, starting at Chiba. The up-to-400 MW project targets operations around 2028.