Decentralized GPU networks are pitching themselves as a lower-cost layer for working AI workloads, whereas coaching the most recent fashions stays concentrated inside hyperscale knowledge facilities.
Frontier AI coaching entails constructing the most important and most superior programs, a course of that requires 1000’s of GPUs to function in tight synchronization.
That stage of coordination makes decentralized networks impractical for top-end AI coaching, the place web latency and reliability can’t match the tightly coupled {hardware} in centralized knowledge facilities.
Most AI workloads in manufacturing don’t resemble large-scale mannequin coaching, opening house for decentralized networks to deal with inference and on a regular basis duties.
“What we’re starting to see is that many open-source and different fashions have gotten compact sufficient and sufficiently optimized to run very effectively on client GPUs,” Mitch Liu, co-founder and CEO of Theta Community, informed Cointelegraph. “That is making a shift towards open-source, extra environment friendly fashions and extra economical processing approaches.”
From frontier AI coaching to on a regular basis inference
Frontier coaching is concentrated amongst just a few hyperscale operators, as working massive coaching jobs is pricey and sophisticated. The most recent AI {hardware}, like Nvidia’s Vera Rubin, is designed to optimize efficiency inside built-in knowledge heart environments.
“You’ll be able to consider frontier AI mannequin coaching like constructing a skyscraper,” Nökkvi Dan Ellidason, CEO of infrastructure firm Ovia Programs (previously Gaimin), informed Cointelegraph. “In a centralized knowledge heart, all the employees are on the identical scaffold, passing bricks by hand.”
That stage of integration leaves little room for the free coordination and variable latency typical of distributed networks.
“To construct the identical skyscraper [in a decentralized network], they must mail every brick to 1 one other over the open web, which is very inefficient,” Ellidason continued.
Meta educated its Llama 4 AI mannequin utilizing a cluster of greater than 100,000 Nvidia H100 GPUs. OpenAI doesn’t disclose the scale of the GPU clusters used to coach its fashions, however infrastructure lead Anuj Saharan mentioned GPT-5 was launched with help from greater than 200,000 GPUs, with out specifying how a lot of that capability was used for coaching versus inference or different workloads.
Inference refers to working educated fashions to generate responses for customers and purposes. Ellidason mentioned the AI market has reached an “inference tipping level.” Whereas coaching dominated GPU demand as just lately as 2024, he estimated that as a lot as 70% of demand is pushed by inference, brokers and prediction workloads in 2026.
“This has turned compute from a analysis value right into a steady, scaling utility value,” Ellidason mentioned. “Thus, the demand multiplier by inside loops makes decentralized computing a viable choice within the hybrid compute dialog.”
Associated: Why crypto’s infrastructure hasn’t caught up with its beliefs
The place decentralized GPU networks really match
Decentralized GPU networks are finest suited to workloads that may be cut up, routed and executed independently, with out requiring fixed synchronization between machines.
“Inference is the quantity enterprise, and it scales with each deployed mannequin and agent loop,” Evgeny Ponomarev, co-founder of decentralized computing platform Fluence, informed Cointelegraph. “That’s the place value, elasticity and geographic unfold matter greater than good interconnects.”
In observe, that makes decentralized and gaming-grade GPUs in client environments a greater match for manufacturing workloads that prioritize throughput and adaptability over tight coordination.
“Shopper GPUs, with decrease VRAM and residential web connections, don’t make sense for coaching or workloads which are extremely delicate to latency,” Bob Miles, CEO of Salad Applied sciences — an aggregator for idle client GPUs — informed Cointelegraph.
“At present, they’re extra suited to AI drug discovery, text-to-image/video and huge scale knowledge processing pipelines — any workload that’s value delicate, client GPUs excel on value efficiency.”
Decentralized GPU networks are additionally well-suited to duties equivalent to accumulating, cleansing and making ready knowledge for mannequin coaching. Such duties usually require broad entry to the open net and may be run in parallel with out tight coordination.
One of these work is tough to run effectively inside hyperscale knowledge facilities with out in depth proxy infrastructure, Miles mentioned.
When serving customers all world wide, a decentralized mannequin can have a geographic benefit, as it will probably cut back the distances requests must journey and a number of community hops earlier than reaching an information heart, which may enhance latency.
“In a decentralized mannequin, GPUs are distributed throughout many places globally, usually a lot nearer to finish customers. In consequence, the latency between the person and the GPU may be considerably decrease in comparison with routing visitors to a centralized knowledge heart,” mentioned Liu of Theta Community.
Theta Community is going through a lawsuit filed in Los Angeles in December 2025 by two former staff alleging fraud and token manipulation. Liu mentioned he couldn’t touch upon the matter as a result of it’s pending litigation. Theta has beforehand denied the allegations.
Associated: How AI crypto buying and selling will make and break human roles
A complementary layer in AI computing
Frontier AI coaching will stay centralized for the foreseeable future, however AI computing is shifting away to inference, brokers and manufacturing workloads that require looser coordination. These workloads reward value effectivity, geographic distribution and elasticity.
“This cycle has seen the rise of many open-source fashions that aren’t on the scale of programs like ChatGPT, however are nonetheless succesful sufficient to run on private computer systems geared up with GPUs such because the RTX 4090 or 5090,” Liu’s co-founder and Theta tech chief Jieyi Lengthy, informed Cointelegraph.
With that stage of {hardware}, customers can run diffusion fashions, 3D reconstruction fashions and different significant workloads domestically, creating a chance for retail customers to share their GPU assets, in accordance with Lengthy.
Decentralized GPU networks will not be a alternative for hyperscalers, however they’re turning into a complementary layer.
As client {hardware} grows extra succesful and open-source fashions change into extra environment friendly, a widening class of AI duties can transfer exterior centralized knowledge facilities, permitting decentralized fashions to slot in the AI stack.
Journal: 6 weirdest units folks have used to mine Bitcoin and crypto
