Latest posts
-
LLM inference: inside ReadyServer’s shared-capacity API
ReadyServer’s shared LLM inference API has tight limits and a best-effort SLA. We report the queue, data-handling and model details the company disclosed.
-
Supermicro All-CPU Rack and the Shift to Execution Capacity
A rack with no GPUs, built for the CPU side of agent workloads. It also carries two quieter shifts: 800-volt DC power and agent governance in silicon.
-
AI-RAN: Supermicro, Nokia and NVIDIA Push Compute to the Network Edge
Supermicro, Nokia and NVIDIA bring AI-RAN hardware from the cell site to the core. Operators have measured the efficiency gains; the edge-services revenue is still unproven.
-
AI Factory Approach: Modelled, Not Measured
Vertiv and NVIDIA promise up to 30 per cent more token output by building AI factories faster. The numbers are models, and a competitor has already published measured data.
-
AI hardware: AXISCADES’ XIDA buys US supplier ITSJ
AXISCADES’ XIDA is buying San Jose supplier ITSJ for up to $7m to enter the AI hardware supply chain. We explain the deal, its terms and the tariff logic.
-
Private AI: what Accenture and Dell’s new group really signals
Accenture and Dell have formed a business group to sell private and sovereign AI. There is no disclosed revenue commitment, the economics hold only at high utilisation, and the reference customer is itself a partner.
-
AI Infrastructure: the utilisation gap behind the GPU bill
Info-Tech Research Group has published a blueprint telling enterprises to stop buying AI hardware and match architecture to workloads. Here is the measured utilisation data behind that advice, and why the fix is more operational than capital.
-
Floating Data Centers: The Bet to Put AI Compute on Ships
Floating data centers promise seawater cooling and on-site power without the years-long wait for a grid connection. The ocean solves the easy problems — and leaves the hard ones.
-
Sovereign AI Infrastructure: How the World’s Funds Are Buying the Build-Out
Sovereign wealth funds crossed $15 trillion and are moving from passive tech equity into owning AI’s data centres, power and chips. It’s a structural shift — with geopolitics and bubble risk attached.
-
Self-Hosted AI Coding: IBM Bob’s Air-Gapped Bet on AI Sovereignty
IBM’s self-hosted Bob offers a genuinely air-gapped path for agentic coding — a stricter claim than Copilot, Cursor or Amazon Q make, with caveats on models and productivity.