Research
When a measurement, a model or a market can be trusted enough to act on
Each paper takes one place where capital, clinical decisions or automated systems rely on a number nobody has checked, and checks it. The subjects so far are a Raman model after its instrument changes, a city's industry after a recession, an imaging chip measured against the diagnosis it serves, the GPUs, cooling and power behind AI, from the collateral value of a chip to the water a rack needs, and a self-improving AI system that can be graded only by events that had not happened when it was built.
Areas
Where the work sits
AI in science and machine learning
Whether a model's uncertainty guarantee survives the conditions it is deployed under, and what scientific workloads need from the hardware they run on. Spectral Passport tests conformal prediction across eight Raman spectrometers and 168 transfer routes. Biology Does Not Run on One Chip matches eight classes of life-science computation to GPU precision, memory and energy. The Weismann Machine proposes self-improving AI that inherits only a compact recipe for learning, kept behind a barrier from everything a lifetime learns, and is selected on forecasts of events that resolve after a sealed cutoff.
Biophotonics
Vibrational spectroscopy as evidence a buyer or regulator can rely on, and fluorescence instruments run by an agent. See Spectral Passport and The Photon Scheduler.
Optoacoustics, thermodynamics and multiphysics
The photoacoustic signal chain from laser exposure limits through thermoelastic conversion (the Grüneisen parameter), acoustic attenuation and thermal noise, used to derive a chip specification from a clinical decision. See Outcome-Defined Silicon.
Industrial organization
How industries regrow after shocks, tested on 25 years of BLS employment data with a registered, hashed forecast for 2026. See Where an Industry Keeps Its Memory and The Pedicle.
Capital and AI infrastructure
What GPU-backed credit facilities disclose, and when a GPU stops paying its way. See Lending Against Machines Nobody Appraises and The Economic Life of an Accelerator.
GPU hardware, energy and cooling
How precision, memory and power differ across six NVIDIA data-centre GPUs and what biology needs from each, the thermodynamics and water cost of 130 kW racks, and what grid flexibility is worth to an AI campus. See Biology Does Not Run on One Chip, Hot Water for Hot Chips and Speed to Power.
Product strategy
Specifying deep-technology products from the outcome they must deliver, and measuring what AI is worth against an explicit no-AI comparator. See Outcome-Defined Silicon and The AI Counterfactual.
Papers
Research papers, newest first
Research paper · 17 November 2026
The Weismann Machine
Abhay Chakra Sadineni
Self-improving AI systems now edit their own code, archive their variants and draw parents from the archive largely by benchmark score. The agent being improved writes its own successor from its own traces, and each child inherits its parent’s code, edits included. In 1892 August Weismann argued that…
Read the paperResearch paper · November 2026
Where an Industry Keeps Its Memory
Abhay Chakra Sadineni
When a shock removes part of an industry from a place, does the industry grow back to where it was? Three answers compete. Locational fundamentals predict full regrowth. Agglomeration predicts that the loss persists because the stock of firms, workers and suppliers is itself the advantage. Institutional commitments…
Read the paperResearch paper · November 2026
Spectral Passport
Abhay Chakra Sadineni
A Raman calibration model that comes with a conformal prediction interval carries a promise that, on new samples, the true concentration falls inside the interval at a stated rate. The promise rests on the new samples being exchangeable with the calibration samples. This paper tests what happens to…
Read the paperResearch paper · November 2026
Lending Against Machines Nobody Appraises
Abhay Chakra Sadineni
Between July 2025 and October 2026, AI cloud operators borrowed tens of billions of dollars in facilities described as secured by graphics processors. This paper reads the public record for eleven of them, plus one corporate secured facility included as a comparator, and codes what each discloses. The…
Read the paperResearch paper · November 2026
Outcome-Defined Silicon
Abhay Chakra Sadineni
The chips inside a medical imaging system are usually specified the way all chips are, by channel count, resolution, noise, bandwidth and power, chosen by engineers from what the previous generation offered and what the bill of materials allows. The clinical outcome the system exists to support enters…
Read the paperResearch paper · November 2026
The Economic Life of an Accelerator
Abhay Chakra Sadineni
The largest buyers of data-centre GPUs depreciate them over five to six years. A common critique holds that a new chip every year makes three years more realistic. Both sides argue from the release calendar. This paper argues from prices. It sets public rental indices for the NVIDIA…
Read the paperResearch paper · November 2026
Biology Does Not Run on One Chip
Abhay Chakra Sadineni
Data-centre GPUs are now designed for training and serving large language models, and each generation's headline gain is in low-precision tensor arithmetic. Life-science computing is not one workload. Molecular dynamics runs in single precision with fixed-point accumulation. Electronic-structure codes need double precision. Structure prediction is limited by memory…
Read the paperResearch paper · November 2026
Hot Water for Hot Chips
Abhay Chakra Sadineni
A rack of 72 Blackwell GPUs is provisioned for about 132 kW. Most of that heat leaves through cold plates in water that may enter at up to 45°C and return at up to 65°C, and the rest leaves through air. This paper works through the thermodynamics of…
Read the paperResearch paper · November 2026
Speed to Power
Abhay Chakra Sadineni
Grid operators now ask large AI campuses for three kinds of flexibility. On a scale of seconds, they want synchronised training jobs to stop swinging tens of megawatts. On a scale of hours, they want the right to curtail the campus during the few hundred hours a year…
Read the paperCurrent interests
Questions under way, not yet published.
- Foundation models for spectroscopy and imaging, and how their uncertainty should be certified when the instrument changes.
- Deep learning for calibration transfer between spectrometers, combined with standardisation methods from chemometrics.
- Independent verification of the physical assets behind AI infrastructure credit.