The AI power crisis

A survey of what AI workloads are doing to data centres: cooling past the limits of air, water use, and what to do with the waste heat.

Co-author with Jenil Saliya and Tirth Acharya. Later extended as a solo draft, “Power, Heat, and Water”.

Per-chip power now exceeded: 1,000 W (source: The AI Power Crisis survey, 2026)
Rack power typical for AI: 30–50 kW (source: The AI Power Crisis survey, 2026)
Cooling approaches compared: 5 (source: The AI Power Crisis survey, 2026)
Sections: 9 (source: The AI Power Crisis survey, 2026)

Why air stops working

The physics of air as a coolant, the chip-power threshold, and how efficiency degrades as fans fight rising temperatures.

The liquid options, compared

Air-assisted liquid cooling, direct-to-chip cold plates, and single- and two-phase immersion, on performance, ten-year cost of ownership and operational risk.

The water problem

Direct evaporation on site, the larger indirect footprint of generating the electricity, and what both do to stressed watersheds.

Cooling without water

Dry cooling architectures, what operators have committed to, and the efficiency price of going water-free.

Using the heat

District heating, greenhouse agriculture and industrial reuse of what is today thrown away.

Rules and optimisation

The ASHRAE thermal guidelines, overcooling, and AI control of the cooling plant itself.

Open problems

The research challenges the field has not solved, as a map for future work.

My work at Pravi Research: AI Builder at Pravi Research. Over these months I have designed, built and shipped six systems that are in real use on public-sector programmes and inside the company. Clients, places, data and figures stay private; the engineering is described on its own page for each system.
Multi-camera video analytics platform: A platform that takes in many live camera streams, analyses them on GPU for vehicles and number plates, records them, and shows operators a wall of cameras that stays responsive under load.
Field reporting and approval platform: A mobile app for staff in the field and a web portal for the offices behind them. What is recorded on site moves up a chain of reviewers, each seeing only their own jurisdiction, until an action is taken and recorded.
Large-scale public records portal: A portal that lets officials search, filter and export a very large table of public records without waiting, with access decided by role and every use logged.
Asset registry portal: A registry of physical assets with a stable identity for each one, loaded from legacy spreadsheets, where offices at each level see and maintain the assets under them.
HR and recruitment platform: An internal platform for the company’s people operations: employee records and roles on one side, and a recruitment pipeline that takes a candidate from application to decision on the other.
Company website and content system: The company’s public website, rebuilt so that the team can edit its pages, careers and articles themselves through a content management system, without a developer.
Reform.ai: Takes a research paper as PDF and returns it as a compilable LaTeX manuscript in a target journal’s format.
Two-Box Audio Deepfake Detection: A small anti-spoofing model that says whether speech is synthetic, and a glass-box layer that says why, in plain language.
Defend the Kingdom: Winning entry at the Infineon Technologies hackathon: a resource-dispatch simulation over a changing road graph, written and hardened in six hours.
Mixed-precision LLM quantization: Deciding, layer by layer, which parts of a language model can drop to 4 bits and which must stay at 8.
Structure or shortcuts? Cross-domain transfer in scientific figure classification: Four vision architectures trained to classify scientific figures from one field, then tested on another, to see what they actually learned.
Voice calling agent: A multi-tenant agent that answers inbound calls and runs outbound campaigns to book appointments, in English and Hindi.
Traffic analytics pipeline: A desktop system for traffic monitoring, enforcement and emergency response: it detects, tracks and reads vehicles across camera feeds, measures speed and flags anomalies, built and tested against the CityFlow benchmark.
CampusRecruit AI: Screens a campus drive’s resumes in two stages: hard rules first, then an LLM scoring each candidate on five dimensions.
Swarm Rescue: Five simulated robots search an unknown arena for victims, coordinating only through a shared pheromone map.
The AI power crisis: A survey of what AI workloads are doing to data centres: cooling past the limits of air, water use, and what to do with the waste heat.
GenX: An end-to-end business analytics dashboard generator. Give it the data, a goal and the objectives; a multi-agent system works out the KPIs and builds the dashboard.
AetherPrompt: A prompt engineering platform: it turns a rough idea into a structured, high-performance prompt, scores it on eight criteria, generates improved variants and tests them across models.
Battery life prediction: Predicting how many cycles a lithium-ion cell has left from its charge curves, with a model small enough for an embedded board.
Intrusion-detection replication: A from-scratch replication of two papers on the UNSW-NB15 network-intrusion dataset: ten classifiers across seven feature scalings.
Multi-Agent Audiobook Generator: A fully automated multi-agent system that turns any topic into a narrated audiobook, running entirely on local models.
AI Research Assistant Network: An autonomous multi-agent system that takes an open research question all the way from query to finished report.
Multi-Agent Content Creator: A pipeline of CrewAI agents that automates social media and blog content, from trend research to a scheduled post, with a person approving before anything is published.
Multi-Agent Film Making System: A generative pipeline that turns a one or two sentence synopsis into a complete animated short film of three to four minutes.
AI Newsroom Team: A multi-agent system that runs the editorial process of a professional newsroom: agents gather the news, write it, verify the facts and format the article.
Smart Data Analyst Bot: A data assistant that takes a question in plain language and returns the SQL result with a chart and an explanation.
Intelli-Gen: My first startup, funded under Gujarat’s Student Startup and Innovation Policy. It turns 5–15 reference photos into a full training dataset for machine vision, and the research behind it became a paper.
FlavorForge: Photograph your ingredients; a vision-language model identifies them and writes a recipe that fits your kitchen, your time and your skill.
God’s Eye: An image processing automation tool: enhancement, restoration, detection, segmentation, classification and captioning in one place.
VisioBrain: A one-click data analysis platform: upload your data, choose an analysis, and get charts and a report without writing code.
AI Genie: An image generation program on Stable Diffusion: text to image, image plus text to image, and image to image.
Proof-of-Deepfake: Detects deepfake videos and proves it: the system finds the original video on social media and shows the two side by side, with the upload date and channel.
OmniSeeker: One search box over Google, YouTube and Wikipedia, with results curated and personalised using natural language processing.

One-page resume