Author
Editor · Founder and CEO, MJ Smart Solutions
AI engineer specialising in agentic systems and founder of MJ Smart Solutions in Bengaluru, building intelligent document processing, voice assistants and multi-agent platforms. Writes the Nexus on compute economics, model governance and agent security. Writing since March 2026. A published researcher and a product and UI/UX designer as well as an engineer, and studied at REVA University. That mix is the standard the Nexus holds itself to: sources opened and read rather than summarised second-hand, figures checked against the footnotes they come from, and every outbound link verified before a piece publishes.
Prompt injection tops the OWASP LLM risk list because instructions and data share one channel. Why filtering fails, which architectural defences hold, and what the research actually reports.
2,060 GW queued, a five-year median wait, and a 13 percent historical completion rate. Why the map of where large-scale compute exists in 2030 is being drawn by grid geography.
Every emergency stop ever specified assumes cutting power makes a machine safe. For a walking biped it is the hazard. The certification gap that explains the current deployment envelope.
Output costs five times input because decode is memory-bound. Batch costs half because utilisation is the provider's central problem. Beneath both sits a floor made of power contracts and grid queues.
Capital, energy times PUE, and overhead, divided by capacity times utilisation. The framework, the multipliers that matter, and the five places the arithmetic reliably breaks.
Decode reads the whole model to produce one token. Working from NVIDIA's own published rack figures, here is the arithmetic that explains why serving throughput ignores the number on the box.
Parameters times bytes per parameter decides feasibility before any licence review does. What 375B, 552B and 753B models actually cost to hold in HBM, and why active-parameter counts mislead.
A sandbox is four boundaries, and most teams draw two. Why credential scope and network egress decide what an incident costs, and why instructions have been measured failing as a control.
Attestation proves genuine hardware and a known measurement. It does not prove the code is safe, the data protected either side of processing, or that anyone checked the report — and 2026 research…
The Digital Omnibus moved the high-risk deadlines to December 2027 and August 2028. What binds an organisation deploying AI today, what was postponed, and why the delay is engineering time rather…
