What the telemetry actually predicts
Moving from a small number of point probes to a continuous spatial thermal field, and establishing which signals genuinely precede thermal events rather than merely accompanying them.
Research
Infrastructure intelligence and autonomous operations
This is CoralDC's most important research domain, and the company says so directly. The evidence indicates that automated thermal control has a larger measured effect on delivered compute than any hardware change available to a modular infrastructure company. Published results credit predictive control with approximately 10 percent more compute per megawatt, which is roughly an order of magnitude more valuable than the energy it saves.
The Constraint
The constraint
Cooling systems are conventionally tuned to hold a setpoint and react to deviation. Reacting to coolant temperature means responding after heat has already been produced, which forces conservative operation and thermal headroom that is paid for in capacity that is provisioned and not used.
Relevance
Why it matters for AI infrastructure
Because the effect is measurable and large. Published figures for automated cooling control in a rack-scale AI reference design credit approximately 10 percent more compute per megawatt through reduced thermal spiking. Independently, A/B tested production results on rack-scale AI hardware report a 75 to 80 percent reduction in thermal spike magnitude relative to a tuned conventional control loop.
Set against CoralDC's own modelling, this reframes the entire economic argument. Energy and water savings from better thermal architecture are real. The compute released by better thermal control is roughly an order of magnitude larger in value. For an infrastructure buyer, the argument is not that Coral saves electricity. It is that Coral delivers more usable compute from the same electrical envelope.
Under Study
What CoralDC is studying
A telemetry schema that holds across a fleet
What is measured, at what rate, in what units, in every pod. Unglamorous, foundational, and the prerequisite for everything above it.
Whether distributed fleet data is different in kind
CoralDC's structural position is unusual: many small sites across varied climates, buildings and workloads rather than one large campus. Whether that produces a materially more useful dataset than a single facility is the company's central research question in this domain, and it is not yet answered.
A customer-inspectable digital twin
And whether a thermal performance guarantee can be underwritten against it rather than against a datasheet.
Under Evaluation
What CoralDC is evaluating
Reinforcement learning and model-predictive cooling control. Distributed fibre optic thermal sensing. Continuous dielectric fluid condition monitoring. Digital twin and simulation platforms. Predictive maintenance approaches built on thermal and fluid data.
Unresolved
What remains uncertain
How much generalises across heterogeneous sites. Ten deployments across different climates, buildings, hardware generations and workloads is not a dataset in the statistical sense. It is a set of anecdotes with high dimensionality. CoralDC believes the fleet approach compounds and has not yet demonstrated that it does.
Data sovereignty constrains the fleet-learning premise directly. The regulated institutions CoralDC serves will contractually restrict pooling of their operational data. Any fleet intelligence capability has to be built within that constraint, and CoralDC regards resolving this properly as a precondition rather than an afterthought.
CoralDC does not claim an AI operating system for infrastructure. It claims disciplined instrumentation and an intention to earn the intelligence layer on top of it.
Partnership
Where partnership could accelerate this
Control and optimisation software companies. Sensing and instrumentation companies. Digital twin and simulation platforms. University controls, thermal and machine learning groups, particularly on the question of learning across small heterogeneous fleets under data-sharing constraints, which is an interesting research problem independent of CoralDC's commercial interest in it.
Who Should Talk to Us
Who should talk to us
Infrastructure control and optimisation companies. Sensing and telemetry companies. Simulation and digital twin platforms. Academic groups in controls, thermal engineering and applied machine learning.
Next Step
Bring us the problem, not the pitch.
CoralDC evaluates technologies continuously across its research domains. Submissions go to engineering, not to a marketing queue.
FAQ
Direct answers for infrastructure buyers.
What is AI-driven cooling control?
It is control software that predicts thermal behaviour from system telemetry and adjusts cooling in advance, rather than reacting to measured coolant temperature after heat has been produced. Published results in rack-scale AI reference designs credit this approach with approximately 10 percent more compute per megawatt through reduced thermal spiking.
Why does thermal control affect how much compute you get?
Because accelerators reduce their own performance when they approach thermal limits. Suppressing thermal spikes keeps hardware in its full performance envelope for more of the time, so the same electrical supply and the same silicon deliver more work.
Does CoralDC build its own control software?
CoralDC builds the telemetry schema, the instrumentation and the integration. The control algorithm itself is a domain where specialised companies are considerably further ahead, and CoralDC's position is to integrate rather than to reproduce that work.