Introduction
Artificial Intelligence (AI) continues to revolutionize the world — from generative models like GPTs to complex scientific simulations. Yet, beneath the breakthroughs lies a growing crisis: the energy cost of intelligence. Training and deploying large AI models consume massive amounts of power, pushing the limits of existing data centre infrastructure.
Enter Extropic AI, a Silicon Valley startup that believes the future of AI cannot be sustained by incremental GPU optimizations alone. Instead, they propose a radical rethinking of how computers work — inspired not by digital logic, but by thermodynamics and the physics of the universe.
Extropic is developing a new class of processors — thermodynamic computing units — that use the natural randomness of physical systems to perform intelligent computation. Their goal: to build AI processors that are both incredibly powerful and orders of magnitude more energy-efficient than current hardware.
This blog explores the full story behind Extropic AI — their mission, technology, roadmap, and how they aim to build the ultimate substrate for generative intelligence.
Company Overview
| Aspect | Details |
|---|---|
| Company Name | Extropic AI |
| Founded | 2022 |
| Founders | Guillaume Verdon (ex-Google X, physicist) and Trevor McCourt |
| Headquarters | Palo Alto, California |
| Funding | ~$14.1 million Seed Round (Kindred Ventures, 2024) |
| Website | https://www.extropic.ai |
| Mission | To merge the physics of information with artificial intelligence, creating the world’s most efficient computing platform. |
Extropic’s founders believe that AI computation should mirror nature’s own intelligence — distributed, energy-efficient, and probabilistic. Rather than fighting the randomness of thermal noise in semiconductors, their processors embrace it — transforming chaos into computation.
The Vision: From Deterministic Logic to Thermodynamic Intelligence
Traditional computers rely on binary logic: bits that are either 0 or 1, flipping deterministically according to instructions. This works well for classic computing tasks, but not for the inherently probabilistic nature of AI — which involves uncertainty, randomness, and high-dimensional sampling.
Extropic’s vision is to rebuild computing from the laws of thermodynamics, creating hardware that behaves more like nature itself: efficient, adaptive, and noisy — yet powerful.
Their tagline says it all:
“The physics of intelligence.”
In Extropic’s world, computation isn’t about pushing electrons to rigidly obey logic — it’s about harnessing the natural statistical behavior of particles to perform useful work for AI.
Core Technology: Thermodynamic Computing Explained
1. From Bits to P-Bits
At the heart of Extropic’s innovation are probabilistic bits, or p-bits. Unlike traditional bits (which hold a fixed 0 or 1), a p-bit fluctuates between states according to a controlled probability distribution.
By connecting networks of p-bits, Extropic processors can natively sample from complex probability distributions — a task central to modern AI models (e.g., diffusion models, generative networks, reinforcement learning).
2. Thermodynamic Sampling Units (TSUs)
Extropic’s hardware architecture introduces Thermodynamic Sampling Units (TSUs) — circuits that exploit natural thermal fluctuations to perform probabilistic sampling directly in silicon.
Each TSU operates using standard CMOS processes — no cryogenics or exotic quantum hardware needed. These TSUs could serve as building blocks for a new kind of AI accelerator that’s:
- Massively parallel
- Energy-efficient (claimed up to 10,000× improvements over GPUs)
- Noise-tolerant and self-adaptive
3. Physics Meets Machine Learning
Most AI models — particularly generative ones — rely on random sampling during inference (e.g., diffusion, stochastic gradient descent). Today’s GPUs simulate randomness via software, wasting energy. Extropic’s chips could perform these probabilistic operations in hardware, vastly reducing energy use and latency.
In essence, Extropic’s chips are hardware-accelerated samplers, bridging physics and information theory.
The Hardware Roadmap
Extropic’s development roadmap (as revealed in their public materials) progresses through three key phases:
| Stage | Codename | Timeline | Description |
|---|---|---|---|
| Prototype | X0 | Q1 2025 | Silicon prototype proving core thermodynamic circuits |
| Research Platform | XTR-0 | Q3 2025 | Development platform for AI researchers and early partners |
| Production Chip | Z1 | Early 2026 | Full-scale chip with hundreds of thousands of probabilistic units |
By 2026, Extropic aims to demonstrate a commercial-grade thermodynamic processor ready for integration into AI supercomputers and data centres.
Why It Matters: The AI Energy Crisis
AI growth is accelerating faster than Moore’s Law. Data centres powering AI models consume enormous electricity — estimated at 1–2% of global energy use, projected to rise sharply by 2030.
Every new GPT-like model requires hundreds of megawatt-hours of energy to train. At this scale, energy efficiency is not just a cost issue — it’s a sustainability crisis.
Extropic AI directly targets this bottleneck. Their chips are designed to perform AI computations with radically lower energy per operation, potentially making large-scale AI sustainable again.
“We built Extropic because we saw the future: energy, not compute, will be the ultimate bottleneck.” — Extropic Team Statement
If successful, their processors could redefine how hyperscale data centres — including AI clusters — are designed, cooled, and powered.
Applications
1. Generative AI and Diffusion Models
Generative models like Stable Diffusion or ChatGPT rely heavily on sampling. Extropic’s chips can accelerate these probabilistic operations directly in hardware, boosting performance and cutting power draw dramatically.
2. Probabilistic and Bayesian Inference
Fields like finance, physics, and weather forecasting depend on Monte Carlo simulations. Thermodynamic processors could make these workloads exponentially faster and more efficient.
3. Data Centre Acceleration
AI data centres could integrate Extropic chips as co-processors for generative workloads, reducing GPU load and energy consumption.
4. Edge AI and Embedded Systems
Energy-efficient probabilistic computing could bring powerful AI inference to low-power edge devices, expanding real-world AI applications.
Potential Impact
If Extropic succeeds, the implications extend far beyond chip design:
| Impact Area | Description |
|---|---|
| AI Scalability | Enables future large models without exponential energy growth |
| Sustainability | Massive reduction in energy and water use for data centres |
| Economic Shift | Lowers cost per AI inference, democratizing access |
| Hardware Industry | Challenges GPU/TPU dominance with a new compute paradigm |
| Scientific Research | Unlocks new frontiers in physics-inspired computation |
In short, Extropic could redefine what it means to “compute.”
Challenges and Risks
While promising, Extropic faces significant challenges ahead:
- Proof of Concept – Their technology remains in prototype stage; no large-scale public benchmarks yet.
- Hardware Ecosystem – Software stacks (PyTorch, TensorFlow) must adapt to use thermodynamic accelerators.
- Adoption Barrier – Data centres are heavily invested in GPU infrastructure; migration may be slow.
- Engineering Complexity – Controlling noise and variability in hardware requires precise design.
- Market Timing – Competing architectures (neuromorphic, analog AI) may emerge simultaneously.
As with any frontier technology, real-world validation will separate hype from history.
Extropic vs Traditional AI Hardware
| Feature | GPUs/TPUs | Extropic Thermodynamic Processors |
|---|---|---|
| Architecture | Digital / deterministic | Probabilistic / thermodynamic |
| Core Operation | Matrix multiplications | Hardware-level probabilistic sampling |
| Power Efficiency | Moderate (~15–30 TFLOPS/kW) | Claimed 1,000–10,000× higher |
| Manufacturing | Advanced node CMOS | Standard CMOS (room temperature) |
| Cooling | Intensive (liquid/air) | Minimal due to lower power draw |
| Scalability | Energy-limited | Physics-limited (potentially higher) |
Global Context: Why This Matters Now
AI has reached a stage where hardware innovation is as critical as algorithmic breakthroughs. Every leap in model capability now depends on finding new ways to scale compute sustainably.
With the rise of AI data centres, space-based compute infrastructure, and sustainability mandates, energy-efficient AI hardware is not optional — it’s essential.
Extropic’s “physics of intelligence” approach could align perfectly with this global trend — enabling AI to grow without draining the planet’s energy grid.
Future Outlook
Extropic’s upcoming milestones will determine whether thermodynamic computing becomes a footnote or the next revolution. By 2026, if their Z1 chip delivers measurable gains in energy and performance, the AI industry could face its most profound hardware shift since the invention of the GPU.
A future where AI models train and infer using nature’s own randomness is no longer science fiction — it’s being built in silicon.
“Extropic doesn’t just want faster chips — it wants to build the intelligence substrate of the universe.” — Founder Guillaume Verdon
Final Thoughts
Extropic AI isn’t another AI startup — it’s a philosophical and engineering moonshot. By uniting thermodynamics and machine learning, they’re pioneering a new physics of computation, where energy, noise, and probability become features, not flaws.
If successful, their work could redefine the foundation of AI infrastructure — making the next generation of intelligence not only faster, but thermodynamically intelligent.
The world has built machines that think. Now, perhaps, we’re learning to build machines that behave like nature itself.
on a state space 
(finite or countable, sometimes continuous) that satisfies the Markov property:
define the transition matrix 
with entries 
.
for all 
.
is finite with size 
is an 
row-stochastic matrix.
-step transition probabilities are entries of the matrix power 
: 
,
.
.
. Let 
.
satisfies 
and 
. Write 
.
.
so 
. Divide both sides by 
(digit-by-digit): 
, therefore 
.
gives 
so 
. Then 
.
.
e.g. first component 
with 
and 
. If 
the stationary distribution is uniform: 
.
if 
for some 
.
if both 
. Communication partitions 
:![\mathbb{E} [\tau_i]<$\infty$](https://elasticstrain.com/wp-content/ql-cache/quicklatex.com-0993cc7d96152089fcb7c00344d42c2a_l3.png "Rendered by QuickLaTeX.com")
.
.If 
the state is aperiodic.
(row vector) is stationary if 
.
for all 
.
, 
,
.
is the smallest 
.
has real eigenvalues 
. Roughly,
.
time to hit set 
be the hitting time of set ![h(i) = \mathbb{E}_i[T_A]](https://elasticstrain.com/wp-content/ql-cache/quicklatex.com-dc2bd060e348c5c8281361883c0a2f4b_l3.png "Rendered by QuickLaTeX.com")
you can solve linear equations: 
.
satisfying, for 
, 
, and
.
.
and death rates 
, the stationary distribution (if it exists) has product form: 
.
edge. Stationary 
.
for Google matrix 
.
(Metropolis–Hastings, Gibbs).
on large state space. Strategy: construct an ergodic chain with stationary distribution 
.
:
.
, propose 
.
move to 
; otherwise stay at 
with ![{\mathbb{E}_\pi[|f|] < \infty}](https://elasticstrain.com/wp-content/ql-cache/quicklatex.com-72fa532587c974f9c3b79ce3660165a1_l3.png "Rendered by QuickLaTeX.com")
,![\frac{1}{n}\sum_{t=0}^{n-1} f(X_t) \xrightarrow{a.s.} \mathbb{E}_\pi[f]](https://elasticstrain.com/wp-content/ql-cache/quicklatex.com-37be0a40083e9fec176ffb3517453ae8_l3.png "Rendered by QuickLaTeX.com")
,![\sqrt{n} (\overline{f_n} - \mathbb{E}_{\pi}[f])](https://elasticstrain.com/wp-content/ql-cache/quicklatex.com-6062cce9a064b8988528948ede27f6a0_l3.png "Rendered by QuickLaTeX.com")
converges in distribution to a normal with asymptotic variance expressible via the Green–Kubo formula (autocovariance sum).
,
.
to the spectral gap.
with normalization 
=1.
by power iteration: repeatedly multiply an initial vector by 
is recurrent in 1D and 2D (returns w.p. 1) but null recurrent in 1D/2D (no finite stationary distribution); in 3D it’s transient.
.
vs 
.
.
.
.
microstates compatible with a given macrostate, Boltzmann gave the famous formula:
,
is Boltzmann’s constant (
).
, Gibbs generalized Boltzmann’s formula:
) ensemble: 
and partition function 
, one obtains thermodynamic relations like:
.
J·K⁻¹.This equivalence underlies deep results connecting information processing to thermodynamics (see Landauer’s principle below).
and cold reservoir at 
.For a reversible cycle:
work.The catch: resetting the demon’s memory (erasure) costs at least 
.
.
:
Conversion to electronvolts (eV):1 eV = 
, the von Neumann entropy generalizes Gibbs:
, giving black holes thermodynamic behavior and posing the information-paradox puzzles that drive modern research.
quantifies irreversibility.
to equilibrium free-energy differences ΔF:
,
and ⟨⋅⟩ denotes average over realizations. The Jarzynski equality has been experimentally verified in molecular pulling experiments (optical tweezers etc.) and is a powerful tool in small-system thermodynamics.
.
and latent heats; statistical estimations use measured distributions 
to compute:
. Entropy 
.
→ engine → cold reservoir 
into engine, 
out, 
rejected; annotate entropy transfers 
and 
.
, and memory erasure cost 
and annotate
.
