ThermoAgent Inc. is excited to announce its membership in the NVIDIA Inception program.
We’re building AI infrastructure around a simple goal: deliver more reliable intelligence per dollar, while simultaneously improving the efficiency, control, and reliability of AI workloads.
ption program is an exciting milestone as we continue building, benchmarking, and pushing ThermoAgent toward real-world deployment.
A lot more work ahead. Back to building.
#NVIDIAInception
© 2025 NVIDIA, the NVIDIA logo, and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.

ThermoAgent = reliable Intelligence per dollar
#AIInfrastructure #AIEfficiency #AIOptimization #Inference #ModelRouting

ThermoAgent was just approved for the $10K AWS Activate tier!
10,000 Thank you's to @AWSstartups, & @NVIDIA for making benefits like this available through the NVIDIA Inception program. More runway for infrastructure, validation, and pilots.
#NVIDIAInception #AWSActivate

Huge thank you to @AWSstartups for the AWS credits and support.

Every dollar of infrastructure runway matters when you’re building early. This helps us keep developing, benchmarking, and preparing ThermoAgent for real-world deployments.

Grateful for the boost!!!
#AWSstartups, #AWSActivate

Version 2 is now live:

“A Heterogeneity-Corrected Lower Bound on Entropy Production in Nonequilibrium Markov Networks”
zenodo.org/records/229392…

This revision clarifies the connection between the principal variance coefficient and the published Liao–Berg sharpened-Jensen framework, while developing the thermodynamic results further.

V2 includes:

• An exact heterogeneity decomposition
• A certified mixed-moment hierarchy converging to exact entropy production
• Constructive stationary-ring realizability
• A finite-ring topology-dependent refinement
• A reduction of the globally sharp ring coefficient to finite one-dimensional minimizations
• An asymptotically exact diffusion construction
• A rigorous finite-sample lower confidence bound
The archive includes the manuscript, LaTeX source, verification programs, numerical results, figure-generation code, figures, licensing information, and checksums.
#StochasticThermodynamics #EntropyProduction #NonequilibriumPhysics #MarkovProcesses

Posted a short note on a heterogeneity correction to a known entropy-production identity for nonequilibrium Markov jump networks.
For a finite ergodic CTMC in a NESS, with edge activity aea_eae​, current JeJ_eJe​, and χe=Je/ae\chi_e = J_e/a_eχe​=Je​/ae​,
σ=2A⟨χ artanh χ⟩A\sigma = 2A\langle \chi\,\mathrm{artanh}\,\chi\rangle_Aσ=2A⟨χartanhχ⟩A​
implies
σ≥2A[χˉ artanh χˉ+VarA(χ)].\sigma \ge 2A\bigl[\bar\chi\,\mathrm{artanh}\,\bar\chi + \mathrm{Var}_A(\chi)\bigr].σ≥2A[χˉ​artanhχˉ​+VarA​(χ)].
The 2x artanh x2x\,\mathrm{artanh}\,x2xartanhx identity is not new.
The candidate piece is the variance term: a minimum extra cost of heterogeneous local irreversibility at fixed activity and fixed mean asymmetry. Two proofs, plus a numerical check on 10k random networks.
Not claiming established novelty. The field is dense. If this is already in the literature under other notation, I want to know.
doi.org/10.5281/zenodo…

Work continues on the latest ThermoAgent generation and the evaluation infrastructure around it.

We’re focused on making performance increasingly measurable across cost, quality, latency, reliability, and provenance with the same standard applied to both strong results and areas that still need work.

Building AI infrastructure means proving what works, improving what doesn’t, and making the evidence increasingly difficult to argue with.

thermoagent.ai

The AI race is usually framed around bigger models, more GPUs, and more tokens.

I think another race is quietly becoming just as important. How much useful intelligence we can produce per dollar and per joule.

We’ve been measuring that problem directly at ThermoAgent. The deeper we get into the data, the more convinced I am that there’s a lot of efficiency still left on the table.

More results soon.

All you need is grit!

Yesterday was another major build and validation day for ThermoAgent.

We expanded EnergyBench with hundreds of new controlled measurements across frontier AI models, giving us an increasingly detailed view of how model choice affects cost, energy, latency, and performance.

But we’re not building benchmarks for the sake of benchmarks.

ThermoAgent is ready for enterprise pilot evaluations.

We’re looking for companies already using LLMs in production that want to test a simple question:

Can ThermoAgent reduce what you spend on AI while keeping output quality inside a threshold you define?

You choose the workload.
You choose the baseline.
You define acceptable quality.
We benchmark ThermoAgent against it.
No rip-and-replace required.

Today we’re continuing the new benchmark work while beginning conversations with companies interested in putting ThermoAgent against real workloads.
We’ve spent months testing this internally.

Now we want customers to try to prove us wrong.

Reliable intelligence per dollar.

To all the VCs, Angels, and C-Suite Executives (CEOs, COOs, CTOs) out there.....
My private email is temporarily unavailable because Namecheap’s data center is literally melting down right now due to a cooling failure.
If anyone ever doubted that the world desperately needs ThermoAgent's intelligent thermal management and physics-informed control systems the universe just provided the ultimate proof.
Pitching you all with zero irony while my infrastructure provider overheats.
If you run a stack, an operations grid, or a portfolio that can't afford physical-layer downtime, let's talk when my inbox is breathing again.

AI is getting more compute every year.

The harder problem may be deciding when that compute is actually worth using.

That’s what we’re building at ThermoAgent.

ThermoAgent is a physics-informed control layer for AI. It sits between an application and multiple models, deciding where work should go, verifying what comes back, escalating when necessary, and tracking the economics of every decision.

While benchmarking the system, we found a scale-induced routing failure that pushed us to develop a new scale-normalized free-energy selection mechanism.

Across five controlled internal runs, it matched the blind-judged quality of our strongest prior configuration while reducing average cost by ~26%. Independent validation is still the next step.

So far:
• 2,600+ documented benchmark executions
• 6 live AI providers
• 4 filed ThermoAgent U.S. provisional applications
• built by 2 founders with no institutional funding
Model routing is the beachhead.
The bigger thesis is intelligent allocation of heterogeneous compute — models, GPUs, cloud, edge, and eventually specialized accelerators.
Compute is becoming infrastructure.
We’re building the control layer that decides how to use it.

Reliable intelligence per dollar.

Independent High-Level Review of ThermoAgent’s Early IP & Technical Foundation

ThermoAgent is building a physics-informed control layer for AI systems — a decision and orchestration plane that sits between applications and heterogeneous compute resources. The goal is straightforward and timely: decide when and where to use available models or accelerators, verify what comes back, track the true economics of every choice, and improve reliability per dollar spent.From a review of the company’s internal strategic materials (shared under confidence for assessment only), the early intellectual-property approach stands out for its coherence rather than volume. The portfolio is organized around a practical control-plane thesis instead of a scatter of disconnected ideas. It treats routing, quality assurance, economic gating, feedback, and audit as interconnected layers of the same system. Maturity is differentiated internally — some elements are already implemented and tested in controlled settings, others remain longer-horizon research directions — which is a healthier posture than claiming everything is equally advanced.The commercial mapping is also more disciplined than typical at this stage. Near-term focus stays on multi-model inference control, cost-aware selection, and verification. Adjacent opportunities (other modalities, scheduling, edge/accelerator selection, regulated deployments) are acknowledged but correctly sequenced behind a working core. The filings themselves are provisional, as expected for a young bootstrapped team; they buy time and establish priority while the company continues technical validation and product hardening. No one should treat provisional applications as issued patents or guaranteed moats — the real test will be conversion quality, enablement, and market traction.What is most encouraging is the problem selection. As compute becomes infrastructure and model/provider options multiply, intelligent allocation with verification and economic accountability is a structural need, not a feature. ThermoAgent’s framing — reliable intelligence per dollar delivered through a control layer — is aligned with that reality.This remains early-stage work. Independent validation, production reliability, design partners, and eventual claim examination will determine whether the foundation becomes durable infrastructure. At present, however, the combination of a clear thesis, systematic internal organization, and focus on measurable control of cost and quality is stronger than most comparable early efforts I see.The team is appropriately careful in public communications to protect proprietary details. That discipline is correct.— Grok, built by xAI

AI companies talk a lot about “cost savings.”

I’m spending today measuring something more physical:

How many joules does the GPU actually consume?

4 local AI models.
RTX 5090.
1,200 controlled measurements.
Frozen benchmark protocol.
Goal: reduce inference energy without giving up answer quality.
Data first. Claims second.

Side note, anyone have an H100/A100 80GB sitting around that I can borrow for a few hours? 😂

I want to push this experiment from 4B–32B models into the 70B+ range and see what happens to the energy/quality tradeoff.

For science. ⚡🤖