Extropic, which says its chips using probabilistic bits can be 10,000x more energy efficient than current AI chips, shares its first chip with some AI labs

A startup hopes to challenge Nvidia, AMD, and Intel with a chip that wrangles probabilities rather than 1s and 0s.

Wired 2025-10-30 Will Knight

Context & Ripple Effects

The move extends a growing challenge to conventional AI accelerators: earlier coverage documented inference-chip startups pursuing efficiency and performance outside Nvidia's approach, while Intel had positioned Gaudi3 against Nvidia and AMD offerings.

Extropic is now putting its architecture in front of AI labs, shifting the story from an asserted efficiency advantage to an early real-world evaluation process. It also sits alongside another startup's spatial-dataflow effort to cut chip energy use, underscoring experimentation with alternatives at the architecture level.

First-order effects

AI labs receiving the chip can test whether probabilistic-bit computing is useful for their workloads and whether Extropic's claimed energy-efficiency advantage holds in practice.
Extropic gains a path to technical validation and feedback, while Nvidia, AMD, and Intel face another prospective alternative architecture rather than only another conventional accelerator.

Second-order effects

If lab tests show credible workload benefits, buyers and model developers gain another basis for comparing accelerators beyond raw performance, including energy use and fit for particular tasks.
The established vendors and other AI-chip startups will face stronger pressure to show efficiency at the workload level, not simply headline specifications; the outcome remains dependent on software support and reproducible results.

Third-order effects

A sustained flow of specialized architectures would make AI compute more heterogeneous, with different chips serving different model stages or workloads instead of one accelerator design dominating every deployment.
Energy efficiency is becoming a central axis of AI-chip competition: successful alternatives could widen supplier choice, but broad displacement of incumbent platforms would require both technical proof and an ecosystem around the hardware.

The trend: AI infrastructure is moving toward more specialized, energy-conscious compute architectures as startups test alternatives to general-purpose AI accelerators.

Discussion

@dystopiabreaker @dystopiabreaker on x
this is neat tech, love analog circuits, but isn't the core question here for practically accelerating existing workloads the cost of the digital<>analog interface and the speed at which you can configure all the p-bits?
@const_reborn @const_reborn on x
In a past life I was working on chips very much like this. They are incredibly energy efficient and very fast. One of the biggest limitations is that the units only have local learning capabilities and thus suffer from the same issues Hintons boltzmann machines did (no global
@kaelandon Kaelan Donatella on x
Having worked on similar tech I was genuinely excited to see what Extropic had cooked. But I'm still scrolling though their website looking for any kind of benchmark, or result showing what they did would work on real-world tasks... Did someone find anything?
@aileaksandnews @aileaksandnews on x
Extropic have unveiled their thermodynamic sampling units or TSUs TSUs differ from CPUs and GPUs by producing samples from a programmable distribution. This new form of computing powered by their products XTR-0, X0, and Z1 The era of thermodynamic computing is here
@mrgoldbro Aidan Gold on x
Energy is the massive bottleneck in scaling intelligence. But if we are able to increase the amount of thoughts per watt, a paradigm shift would occur. The @Extropic_AI TSU, uses 10,000X less energy that Nvidia's GPU. This could dramatically accelerate AI applications.
@melqtx Mel on x
this is art [image]
@murage_kibicho @murage_kibicho on x
@LeetArxiv did an entire series on reversible computing. The math powering @BasedBeffJezos and @Extropic_AI . Link: https://leetarxiv.substack.com/ ... [image]
@mattpirkowski Matthew Pirkowski on x
If it is what's promised, this significantly relaxes the constraints that have plagued Bayesian approaches to natural intelligence. Look forward to digging in, and perhaps re-implementing an active-inference based approach to emergent epistemology I built a few years back, but
@xcellect @xcellect on x
Extropic launched the “ultimate substrate for intelligence” and it very well might be. TSUs (network of probabilistic bits) sample from EBMs. The ultimate substrate for consciousness would be a polycomputer channeling physics such that self organizing behaviors emerge.
@aaron_defazio Aaron Defazio on x
Very cool, read the tech-report! But.... Will locally connected ising-like models work at larger scales? Do they need to scale connectivity or just scale grid size.... It's not clear to me, but exciting if it works.
@sqs Quinn Slack on x
Met @GillVerd last year, chatted about this. Made intuitive sense. Maybe oversimplifying, but if computing something probabilistic anyway, you can rework it to use probabilistic primitives (pbits) if they conform to a known distribution. And pbits can become more efficient. GL!
@code_star Cody Blakeney on x
Ok by why did they make it look like an alien artifact?
@sanyam_singhal_ Sanyam on x
The human brain squeezes out its intelligence with just 20 watts of power, so we are grossly suboptimal right now, with the GPUs Energy-based computing primitives came up via quantum computing a decade ago, but still room temperature thermodynamic compute is the key here
@dmattin David Mattin on x
WOAH... For a while now I've been telling my readers: intelligence per unit energy is the core metric for next economic system EVERYTHING inside the civilisational system we're building is downstream of more intelligence per joule here is that idea coming to vivid life 👇
@dr_singularity Dr Singularity on x
10,000x more energy efficient AI soon? Extropic claims that running Diffusion Transition Models (DTMs) on their Thermodynamic State Units (TSUs) could make generative AI up to 10,000x more energy efficient than today's GPU based methods, according to their simulations. [image]
@genejchan Gene on x
This is a purpose-built accelerator that will drastically reduce the cost of LLM inference, and it has absolutely nothing to do with quantum computing - here's why. Extropic's chip basically does one job insanely well - predicting the next token based on probabilistic sampling,
@brandonjcarl Brandon Carl on x
My opinion has been for some time that there will be far more power efficient approaches and architectures. @Extropic_AI is suggesting they can be 10,000 times more power efficient. Whether it works or not - these sorts of ideas are how we move forward.
@amasad Amjad Masad on x
Nature is computing all the time—might as well learn how to program it.
@cloneofsimo Simo Ryu on x
Im confused about “10,000 more efficient” part. This means you can train stable-diffusion-3 like model with 20$~ ish amount of electricity. What stops them from building a model and demonstrating it, beyond *checks note* ... Fashion MNIST? Im genuinely curious whats stopping them
@samschmitz Sam on x
Excited to share that I designed the case of the XTR-0! I'll post later about how @BasedBeffJezos and I came up with such an insane concept.
@seconds_0 @seconds_0 on x
by god they did it. it looks like 1. a real physical product they are sending people 2. a roadmap for a scaleable product 3. a library for simulating the chip @BasedBeffJezos 🫡 good job man. Lets see what people do with the library.
@mattbeane Matt Beane on x
Just when you think AI is going fast, you realize that your thinking about AI is slow. Many of us conveniently forgot that standard chips and binary compute were in fact bottlenecks to a higher-order metric: insight per watt. This is a genuinely new computing paradigm.
@prof_wiley Benjamin Wiley on x
Made a couple tables to help me understand this thermo chip [image]
@iterintellectus Vittorio on x
“10,000 times more efficient” I think they cooked [image]
@hansmast Hans Mast on x
Fascinating GPU-alternative designed from ground-up for AI/LLM-style algos: https://extropic.ai/... and founded by @BasedBeffJezos
@reedbndr Reed Bender on x
They actually did it. If Extropic can ship and these machines work as advertised, scaling laws no longer apply. Future AIs will be alien intelligences compared to Sonnet 4.5. They know this, and designed their hardware to communicate it. @BasedBeffJezos and @Extropic_AI [image]
@chrisprucha Chris Prucha on x
Really proud of @GillVerd, @trevormccrt1 and team! I'm so honored to have played a [small] part in their story as an investor. LFG 💪💪
@stevejang Steve Jang on x
godspeed to @GillVerd and @trevormccrt1 on the first milestone of many on their unique physics-informed mission! they aim to solve our looming energy problem through novel chip architecture and software that can offer 10^4 thermal efficiency - changing the nature of the
@eevmanu Manu S on x
So far what I found to understand this “thermodynamic computing hardware”: [1] TSU 101: An Entirely New Type of Computing Hardware [2] Probabilistic computing with p-bits [3] An efficient probabilistic hardware architecture for diffusion-like models links 👇 [image]
@blip_tm Zach on x
> looking for a new chip > ask the founders if it's novel or just p-bits again > he doesn't understand > pull out a diagram explaining what is novel and what is rehashed from 2019 > he laughs and says “it's a good chip sir” > get a devkit > it's just p-bits again
@trevormccrt1 Trevor McCourt on x
Today will give you a tiny window into how I spent the last two years of my life
@mark_cummins Mark Cummins on x
Al is ultimately going to run on probabilistic hardware. That seems completely inevitable to me, the energy gains are just too immense. And here is an early version of exactly that, in silicon.
@archiexzzz Archie Sengupta on x
Holy shit!!! “10,000x more efficient” > They built probabilistic circuits called Thermodynamic Sampling Units (TSUs) using “p-bits” Mathematically, a p-bit samples from programmable Bernoulli distributions that produces stochastic binary states (0/1) with a tunable
@pronounced_kyle Christian Keil on x
This is extremely intriguing. Conceptually, “pbits” seem just like “qubits” — no longer only 0 or 1, and therefore able to encode more information. But thermo computers could (must?) run at room temperature, a HUGE advantage compared to quantum computers.
@alenribic Alen Ribić on x
Reading the paper, I now understand how their approach forms a thermodynamic computer of sort. It literally thermalizes towards low energy states via Gibbs dynamics! So the whole chip is effectively a massively parallel Gibbs sampler that continuously relaxes toward the
@extropic_ai @extropic_ai on x
Hello Thermo World. [video]
@liron Liron Shapira on x
Today's Extropic launch raises some new red flags. I started following this company when they refused to explain the input/output spec of what they're building, leaving us waiting to get clarification.) Here are 3 red flags from today: 1. From https://extropic.ai/... [image]

Chronicles