From decentralised by Shlok Khemani

Hello!

On a sunny day in Memphis, Tennessee, a propeller spy plane repeatedly circled over an industrial building, its passengers frantically photographing the facilities below. This wasn’t a scene from Cold War espionage but from 2024. The target wasn’t a military installation or uranium enrichment site but a former appliance factory now housing one of the world’s most powerful supercomputers. The passengers weren’t foreign agents but employees of a rival data centre company.

Every few decades, a transformative technology emerges with the potential to unquestionably alter the trajectory of civilisation. What follows is a race between the world's most powerful entities to realise this technology first. The rewards are so immense, and the consequences of failure so devastating, that these entities rapidly mobilise their full arsenal of resources—human talent and capital—toward mastering the technology.

In the 20th century, two standout technologies fit this definition—nuclear weapons and space exploration. The race to harness these technologies involved the most powerful nation-states. The United States’ victories in both cemented its status as the world’s dominant superpower, ushering in an era of unparalleled prosperity. For the defeated—Nazi Germany and the Soviet Union—the consequences were devastating, even terminal.

The giant 44-acre K-25 plant in Oak Ridge, Tennessee, USA, where the uranium for the first atomic weapon was produced (source)

America's victory carried an enormous price tag. The Manhattan Project cost nearly $2 billion (approximately $30 billion adjusted for inflation) and employed over 120,000 people—one in every thousand Americans. The space race demanded even greater resources. The Apollo program cost $28 billion in the 1960s (roughly $300 billion in today's money) and involved over 400,000 people—one in 490 Americans. At its peak in 1966, NASA commanded 4.4% of the entire U.S. federal budget. 

The Apollo 11, just before liftoff on the mission to the moon (source)

The launch of ChatGPT in 2022 marked the dawn of a new race with civilization-altering proportions—the pursuit of artificial superintelligence (ASI). While AI is already woven into daily life—managing social media feeds, Netflix recommendations, and email spam filters—the emergence of large language models (LLMs) promises to transform everything: human productivity, media creation, scientific research, and innovation itself.

This time, the contenders aren’t nation-states (at least, not yet) but the world’s largest corporations (Microsoft, Google, Meta, Amazon), the hottest startups (OpenAI, Anthropic), and the wealthiest individual (Elon Musk). While Big Tech channels unprecedented capital into building the infrastructure for training ever-more-powerful models, startups are securing record-breaking venture capital funding. Elon is, well, doing Elon things (the data centre under surveillance belonged to his company, xAI).

And then there’s everyone else—enterprises, smaller companies, and startups—who may not aspire to build ASI but are eager to harness the cutting-edge capabilities unlocked by AI to optimise their business, disrupt an industry, or create entirely new ones. The potential rewards are so vast that everyone is scrambling to claim their share of this new machine-intelligence-driven economy.

At the heart of the AI revolution lies its most essential component: the graphics processing unit (GPU). Originally designed to power video games, this specialised computer chip has become the world’s hottest commodity. Demand for GPUs is so overwhelming that companies often endure months-long waiting lists just to acquire a few. This demand has catapulted NVIDIA, their primary manufacturer, into the position of the world’s most valuable company.

For businesses unable or unwilling to directly purchase GPUs, renting compute power has become the next best option. This has fueled the rise of AI cloud providers—companies operating sophisticated data centres tailored to meet the computational needs of the AI boom. However, the surge in demand and its unpredictable nature means that neither pricing nor availability is a guarantee.

I argued that crypto functions as a "Coasian" technology, designed to “grease the wheels, pave the roads, and strengthen the bridges” for other disruptive innovations to flourish. As AI emerges as the transformative force of our era, the scarcity and exorbitant cost of GPU access present a barrier to innovation. Several crypto companies are stepping in, aiming to break down these barriers with blockchain-based incentives.

In today’s article, we first step back from crypto to examine the fundamentals of modern AI infrastructure—how neural networks learn, why GPUs have become essential, and how today's data centres are evolving to meet unprecedented computational demands. Then, we dive into decentralised compute solutions, exploring where they can realistically compete with traditional providers, the unique advantages crypto networks offer, and why—although they won't give us AGI—they will still be essential for ensuring AI's benefits remain accessible to all.

Let’s start with why GPUs are such a big deal in the first place. 

GPUs

This is David, a 17-foot tall, 6-ton marble sculpture created by the genius Italian Renaissance master Michelangelo. It depicts the biblical hero from the story of David and Goliath and is considered a masterpiece for its flawless representation of human anatomy and masterful attention to perspective and detail. 

Like all marble sculptures, David began as an enormous, rough slab of Carrara marble. To get to its final, majestic form, Michelangelo had to methodically chip away at the stone. Starting with broad, bold strokes to establish the basic human form, he progressed to increasingly finer details—the curve of a muscle, the tension in a vein, the subtle expression of determination in the eyes. It took Michelangelo three years to free David from the stone.

But why discuss a 500-year-old marble figure in an article about AI?

Like David, every neural network starts as pure potential—a collection of nodes initialised with random numbers (weights), as formless as that massive block of Carrara marble.

This raw model is repeatedly fed training data—countless instances of inputs paired with their correct outputs. Each data point passing through the network triggers thousands of calculations. At every node (neuron), incoming connections multiply the input value by the connection's weight, sum these products, and transform the result through an "activation function" that determines the neuron's firing strength.

Just as Michelangelo would step back, assess his work, and course-correct, neural networks undergo a refinement process. After each forward pass, the network compares its output to the correct answer and calculates its margin of error. Through a process called backpropagation, it measures how much each connection contributed to the error and, like Michelangelo's chisel strikes, makes adjustments to its values. If a connection leads to an incorrect prediction, its influence decreases. If it helps reach the right answer, its influence strengthens.

When all data passes through the network (completing one forward and backward propagation step per data point), it marks the end of an "epoch." This process repeats multiple times, with each pass refining the network's understanding. During early epochs, the weight changes are dramatic as the network makes broad adjustments—like those first bold chisel strikes. In later epochs, the changes become more subtle, fine-tuning the connections for optimal performance—just as delicate final touches brought out David's details.

Finally, after thousands or millions of iterations, the trained model emerges. Like David standing proud in its finished form, the neural network transforms from random noise into a system capable of recognising patterns, making predictions, generating images of cats riding scooters, or enabling computers to understand and respond in human language.

Why GPUs?

Michelangelo, working alone on David, could make only one chisel strike at a time, each requiring precise calculations of angle, force, and position. This painstaking precision is why it took him three tireless years to complete his masterpiece. But imagine thousands of equally skilled sculptors working on David in perfect coordination—one team on the curls of hair, another on the muscles of the torso, and hundreds more on the intricate details of the face, hands, and feet. Such parallel effort would compress those three years into mere days.

Similarly, while CPUs are powerful and precise, they can perform only one calculation at a time. Training a neural network doesn’t require a single complex calculation but hundreds of millions of simple ones—primarily multiplications and additions at each node. For instance, the sample neural network mentioned earlier, with just 18 nodes and about 100 connections (parameters), can be trained on a CPU within a reasonable timeframe.

However, today's most powerful models, like OpenAI's GPT-4, have 1.8 trillion parameters! Even smaller modern models contain at least a billion parameters. Training these models one calculation at a time would take centuries. This is where GPUs excel: they can perform a large number of simple mathematical computations simultaneously, making them ideal for processing multiple neural network nodes in parallel. 

Modern GPUs are mind-bogglingly powerful. NVIDIA’s latest B200 GPU, for example, consists of over 200 billion transistors and supports 2,250 trillion parallel computations per second (2,250 TFLOPS). A single B200 GPU can handle models with up to 740 billion parameters. These machines represent feats of modern engineering, which explains why NVIDIA, selling each unit at $40,000, has seen its stock price surge over 2,500% in five years.

Jensen Huang presenting the NVIDIA B200

Yet even these formidable machines cannot train AI models alone. Recall that during training each data instance must pass through the model in a forward and backward cycle individually. Modern large language models (LLMs) are trained on datasets encompassing the entirety of the internet. GPT-4, for instance, processed an estimated 12 trillion tokens (approximately 9 trillion words), and the next generation of models is expected to handle up to 100 trillion tokens. Using a single GPU for such an immense volume of data would still take centuries.

The solution lies in adding another layer of parallelism—creating GPU clusters where training tasks are distributed among numerous GPUs working as a unified system. Model training workloads can be parallelised in three ways:

Data Parallelism: Multiple GPUs each maintain a complete copy of the neural network model while processing different portions of the training data. Each GPU processes its assigned data batch independently before periodically synchronising with all other GPUs. In this synchronisation period, the GPUs communicate with each other to find a collective average of their weights and then update their individual weights such that they are all identical. Consequently, they continue training on their batch of data individually before it’s time to synchronise again.

As models grow larger, a single copy can become too big to fit on one GPU. For example, the latest B200 GPU can hold only 740 billion parameters while GPT-4 is a 1.8 trillion parameter model. Data parallelism across individual GPUs doesn’t work in this case.

Tensor Parallelism: This approach addresses the memory constraint by distributing each model layer's work and weights across multiple GPUs. GPUs exchange intermediate calculations with the entire cluster during every forward and backward propagation step. These GPUs are typically grouped in servers of eight units, connected via NVLink—NVIDIA's high-speed direct GPU-to-GPU interconnect. This setup requires high-bandwidth (up to 400 Gb/s), and low-latency connections between GPUs. A tensor cluster effectively functions as a single massive GPU.

Pipeline Parallelism: This method splits the model across multiple GPUs, with each GPU handling specific layers. Data flows through these GPUs sequentially, like a relay race where each runner (GPU) manages their portion before passing the baton. Pipeline parallelism is particularly effective for connecting different 8-GPU servers within a data centre, using high-speed InfiniBand networks for inter-server communication. While its communication requirements exceed data parallelism, they remain lower than tensor parallelism's intensive GPU-to-GPU exchanges.

The scale of modern clusters is remarkable. GPT-4, with 1.8 trillion parameters and 120 layers, required 25,000 A100 GPUs for training. The process took three months and cost over $60 million. The A100 is two generations old; using today's B200 GPUs would require only about 8,000 units and 20 days of training. Just another demonstration of how fast AI is moving. 

But the GPT-4 class of models are old toys now. Training for the next generation of advanced models is underway in data centres housing clusters of 100,000 B100 or H100 GPUs (the latter being one generation older). These clusters, representing over $4 billion in GPU capital expenses alone, are humanity's most powerful supercomputers, delivering at least four times the raw computing power of government-owned ones.

Apart from securing raw compute, ASI aspirants run into another problem when trying to set up these clusters: electricity. Each of these GPUs consumes 700W of power. When you combine 100,000 of them, the entire cluster (including supporting hardware) consumes over 150MW of power. To put this in perspective, this consumption equals that of a city of 300,000 people—comparable to New Orleans or Zurich.

The madness doesn’t stop here. Most ASI aspirants believe that the LLM scaling laws—which suggest that model performance improves predictably with increases in model size, dataset size, and training compute—will continue to hold true. Plans are already in motion for training runs of even more powerful models. By 2025, the cost of each training cluster is projected to exceed $10 billion. By 2027, over $100 billion. As these figures approach the U.S. government's investment in the Apollo programs, it becomes clear why achieving ASI has emerged as the defining race of our era.

Metrics for models starting GPT-5 are estimations

As electricity consumption grows proportionally with cluster sizes, next year's training runs will demand over 1GW of power. The year after that, 10GW or more. With no indications of this expansion slowing, data centres are expected to consume roughly 4.5% of global generated by 2030. Existing power grids, already struggling with current model demands, cannot generate sufficient energy for future clusters. This raises a critical question: where will this power come from? Big Tech is taking a two-pronged approach.

In the long run, the only viable solution is for ASI aspirants to generate their own electricity. Given their climate commitments, this power must come from renewable sources. Nuclear energy stands out as the primary solution. Amazon recently purchased a data centre campus powered by a nuclear power plant for $650 million. Microsoft has hired a head of nuclear technologies and is reviving the historic Three Mile Island plant. Google has acquired multiple small nuclear reactors from California’s Kairos Power. OpenAI’s Sam Altman has backed energy startups like Helion, Exowatt, and Oklo. 

Microsoft is reopening the Three Mile Island Nuclear Plant (image source)

While the seeds of nuclear power are being sown now, the fruits (or power) will take several years to bear. What about energy requirements for the immediate generation of models? The interim solution involves distributed training across multiple data centres. Rather than concentrating massive power demands in one location, companies like Microsoft and Google are distributing their training clusters across multiple sites.

The challenge, of course, is getting these distributed systems to work together effectively. Even at the speed of light, data takes approximately 43ms for a round-trip journey from the U.S. East to West Coast—an eternity in computing terms. Additionally, if even one chip lags behind by, say, 10%, it causes the entire training run to be slowed by the same margin.

The solution lies in connecting data centres across multiple sites with high-speed fibre optic networks and applying a combination of the parallelism techniques discussed earlier to synchronise their operations. Tensor parallelism is applied to GPUs within each server, enabling them to function as a single unit. Pipeline parallelism, with its lower network demands, is employed to link servers within the same data centre. Finally, data centres in different locations (referred to as "islands") synchronise their information periodically using data parallelism.

Earlier, we noted that data parallelism proves ineffective for individual GPUs because they can't accommodate large models independently. However, this dynamic shifts when we're parallelising islands—each containing thousands of GPUs—rather than individual units. Training data is distributed across each island, and these islands synchronise periodically over the relatively slower (compared to NVLink and Infiniband) fibre optic connections.

Datacentres

Let's shift our focus from training and GPUs to the data centres themselves.

Twenty years ago, Amazon launched Amazon Web Services (AWS)—one of the most transformative businesses in history—and created a whole new industry known as cloud computing. Today’s cloud leaders (Amazon, Microsoft, Google, and Oracle) enjoy a comfortable dominance, making a combined annual revenue of close to $300 billion with margins of 30-40%.  Now, the emergence of AI has created new opportunities in a market that has remained largely oligopolistic for years.

The physical requirements, technical complexity, and economics of GPU-intensive AI data centres differ dramatically from their traditional counterparts.

We discussed earlier how energy-hungry GPUs are. This leads to AI data centres being much more power-dense and, consequently, producing more heat. While traditional data centres use giant fans (air cooling) to dissipate heat, this approach is neither sufficient nor financially viable for AI facilities. Instead, AI data centres are adopting liquid cooling systems where water blocks attach directly to GPUs and other hot components to dissipate heat more efficiently and quietly. (The B200 GPUs come with this architecture built-in). Supporting liquid cooling systems requires adding large cooling towers, a centralised water system facility, and piping to transport water to and from all GPUs—a fundamental modification to data centre infrastructure.

Beyond higher absolute energy consumption, AI data centres have distinct load requirements. While traditional data centres maintain predictable power consumption, AI workload power usage patterns are far more volatile. This volatility occurs because GPUs periodically alternate between running at 100% capacity and slowing to near-halt as training reaches checkpoints, where weights are either stored to memory or, as we saw earlier, synchronised with other islands. AI data centres require specialised power infrastructure to manage these load fluctuations.

Building GPU clusters is much harder than building regular computer clouds. GPUs need to talk to each other very quickly. To make this happen, they must be packed very close together. A typical AI facility needs more than 200,000 special cables called InfiniBand connections. These cables let the GPUs communicate. If just one cable stops working, the whole system shuts down. The training process can't continue until that cable is fixed.

These infrastructure requirements make it nearly impossible to retrofit traditional data centres with high-performance GPUs to make them AI-ready. Such an upgrade would require an almost complete structural overhaul. Instead, companies are building new data centres specifically designed for AI from the ground up, with different organisations pursuing this at varying scales.

At the forefront, leading tech companies are racing to build their own AI data centres. Meta is investing heavily in facilities solely for its own AI development, treating it as a direct capital investment since it doesn't offer cloud services. Microsoft is building similarly massive centres to power both its own AI projects and serve key clients like OpenAI. Oracle has also entered this space aggressively, securing OpenAI as a notable customer. Amazon continues to expand its infrastructure, particularly to support emerging AI companies like Anthropic. Elon Musk’s xAI, not wanting to rely on another company, chose to build its own 100,000 GPU cluster. 

Inside xAI’s 100,000 H100 GPU data centre (source)

Alongside the incumbents, "neoclouds" are emerging—specialised cloud providers focusing exclusively on GPU computing for AI workloads. These neoclouds are divided into two distinct categories based on scale.

Large neocloud providers, including CoreWeave, Crusoe, and LLama Labs, operate clusters of over 2,000 GPUs. They differentiate themselves from traditional cloud services in two ways: offering customised infrastructure solutions rather than standardised packages, and requiring long-term customer commitments instead of pay-per-use arrangements.

Their business model leverages these long-term agreements and customer creditworthiness to secure infrastructure financing. Revenue comes from premium rates charged for specialised services, and profits from the spread between low financing costs and customer payments. 

This is how such an arrangement typically works: a neocloud provider secures a three-year contract with a well-funded AI startup for 10,000 H100 GPUs at $40 million monthly. Using this guaranteed revenue stream of $1.44 billion, the provider secures favourable bank financing (at 6% interest) to purchase and install infrastructure worth $700 million. The monthly revenues of $40 million cover $10 million in operating costs and $20 million in loan payments, generating $10 million in monthly profits while the startup receives custom-built, dedicated computing power.

This model requires exceptionally careful customer selection. Providers typically look for companies with large cash reserves or strong venture backing—often valuations of $500 million or more. 

Small neoclouds offer GPU clusters of 2,000 or fewer and cater to a separate segment of the AI market—small and medium-sized start-ups. These companies either train smaller models (up to 70 billion parameters) or fine-tune open-source ones. (Fine tuning is the process of adapting a foundation model to specific use cases.) Both these workloads require moderate but dedicated compute for shorter periods. 

These providers offer on-demand computing with hourly rates for fixed-duration, uninterrupted cluster access. While this costs more than long-term contracts, it gives startups flexibility to experiment without committing to multi-million dollar agreements. 

Finally, apart from the cloud incumbents and neocloud providers, we have the middlemen of the AI infrastructure space: platforms and aggregators. These intermediaries don't own GPU infrastructure but instead connect compute resource owners with those who need them.

Platform providers like HydraHost and Fluidstack serve as the Shopify of GPU computing. Just as Shopify enables merchants to launch online stores without building e-commerce infrastructure, these platforms allow data centre operators and GPU owners to offer compute services without developing their own customer interfaces. They provide a complete technical package for running a GPU compute business, including infrastructure management tools, customer provisioning systems, and billing solutions.

Marketplace aggregators like Vast.ai function as the Amazon of the GPU world. They create a marketplace combining diverse compute offerings from various providers—ranging from consumer-grade RTX cards to professional H100 GPUs. GPU owners list their resources with detailed performance metrics and reliability ratings, while customers purchase compute time through a self-service platform. 

Inference

So far, our discussion has focused on training (or fine-tuning) models. However, once trained, a model must be deployed to serve end users—a process called inference. Every time you're chatting with ChatGPT, you're using GPUs running inference workloads that take your input and generate the model's response. Let’s go back to discussing marble statues for a minute.

This is also David—not Michelangelo's original, but a plaster cast commissioned by Queen Victoria in 1857 for London's Victoria and Albert Museum. While Michelangelo spent three gruelling years carefully chipping marble to create the original in Florence, this plaster cast was made from a direct mould of the statue—perfectly reproducing every curve, angle, and detail that Michelangelo had crafted. The intensive creative work happened once. Afterwards, it became a matter of faithfully replicating these features. Today, replicas of David appear everywhere from museum halls to Las Vegas casino courtyards.

This is exactly how inference works in AI. Training a large language model is like Michelangelo's original sculptural process—computationally intensive, time-consuming, and resource-heavy as the model gradually learns the right "shape" of language through millions of tiny adjustments. But using the trained model—inference—is more like creating a replica. When you chat with ChatGPT, you're not teaching it language from scratch but using a copy of a model whose parameters (like David's precise curves and angles) have already been perfected.

Inference workloads differ fundamentally from training. While training requires large, dense clusters of the latest GPUs like H100s to handle intensive computations, inference can run on single GPU servers using older hardware like A100s or even consumer-grade cards, making it significantly more cost-effective. That being said, inference workloads have their own unique demands:

These characteristics make inference workloads ideal for spot pricing models. Under spot pricing, GPU resources are available at significant discounts—often 30-50% below on-demand rates—with the understanding that service may pause when higher-priority customers need resources. This model suits inference because redundant deployment allows workloads to shift quickly to available GPUs if interrupted.

Against this backdrop of GPUs and AI cloud computing, we’re now in a position to start exploring where crypto fits into all of this. Let’s (finally) get to it. 

Where crypto fits in

Projects and reports frequently cite Peter Thiel's observation that "AI is centralising, crypto is decentralising" when discussing crypto's role in AI training. While Thiel’s statement is unquestionably true, we just saw ample evidence of Big Tech’s clear advantage in training powerful AI—it's often misappropriated to suggest that crypto and decentralised computers offer the primary solution to counterbalance Big Tech's influence.

Such claims echo previous overstatements about crypto's potential to revolutionise social media, gaming, and countless other industries. They are not only counterproductive but also, as I will argue shortly, unrealistic—at least in the short term.

Instead, I’m going to take a more pragmatic approach. I’m going to assume that an AI startup looking for compute doesn’t care about the tenets of decentralisation or mounting ideological opposition to Big Tech. Rather, they have a problem—they want access to reliable GPU compute at the lowest possible cost. If a crypto project can provide a better solution to this problem than non-crypto alternatives, they will use it. 

To that end, let’s first understand who crypto projects are competing with. Earlier, we discussed the different categories of AI cloud providers—Big Tech and hyperscalers, big neoclouds, small neoclouds, platforms providers, and marketplaces. 

The fundamental thesis behind decentralised compute (like all DePIN projects) is that the current compute market operates inefficiently. GPU demand remains exceptionally high, while supply sits fragmented and underutilised across global data centres and individual homes. Most projects in this sector compete squarely with marketplaces by aggregating this scattered supply to reduce inefficiencies.

With that established, let’s look at how these projects (and compute marketplaces in general) can aid with different AI workloads—training, fine-tuning and inference. 

Training

First things first. No, ASI is not going to be trained on a global network of decentralised GPUs. At least, not on the current trajectory of AI. Here’s why.

Earlier, we discussed just how big foundation model clusters are getting. You need 100,000 of the most powerful GPUs in the world to even begin competing. This number is only increasing with every passing year. By 2026, the cost of a training run is expected to cross $100 billion dollars, requiring perhaps a million GPUs or more.

Only Big Tech companies, backed by major neoclouds and direct Nvidia partnerships, can assemble clusters of this magnitude. Remember, we are in a race for ASI, and all the participants are both highly motivated and capitalised. If there is an additional supply of these many GPUs (there isn’t) then they will be the first to scoop them up.

Even if a crypto project somehow amassed the requisite compute, two fundamental obstacles prevent decentralised ASI development:

First, the GPUs still need to be connected in large clusters to function effectively. Even if these clusters are divided among islands in cities, they will have to be connected by dedicated fibre optic lines. Neither of these is possible in a decentralised setting. Beyond GPU procurement, establishing AI-ready data centres demands meticulous planning—typically a one-to-two-year process. (xAI did it in just 122 days but it’s unlikely Elon is going to be launching a token anytime soon.)

Second, just creating an AI data centre is not sufficient to birth a superintelligent AI. As Anthropic founder Dario Amodei recently explained, scaling in AI is analogous to a chemical reaction. Just as a chemical reaction requires multiple reagents in precise proportions to proceed, successful AI scaling depends on three essential ingredients growing in concert: bigger networks, longer training times, and larger datasets. If you scale up one component without the others, the process stalls.

Even if we do manage to somehow accumulate both the compute and get the clusters to work together, we still need terabytes of high-quality data for the trained model to be any good. Without Big Tech’s proprietary data sources, the capital to ink multi-million dollar deals with online forums and media outlets, or existing models to generate synthetic data, acquiring adequate training data is impossible. 

There has been some speculation of late that scaling laws may plateau, with LLMs potentially hitting performance ceilings. Some interpret this as an opening for decentralised AI development. However, this overlooks a crucial factor—talent concentration. Today's Big Tech firms and AI labs house the world's premier researchers. Any breakthrough alternative path to AGI will likely emerge from these centres. Given the competitive landscape, such discoveries would remain closely guarded.

Considering all of these arguments, I am 99.99% certain that the training of ASI—or even the world’s most powerful models—will not be trained on a decentralised compute project. In that case, what models could crypto actually help train? 

In order for models to be trained across separate GPU clusters placed in different geographic locations, we need to implement data parallelism between them. (Recall that data parallelism is how different islands of GPUs, each working on separate chunks of the training data, sync with each other). The bigger the model being trained, the greater the amount of data that needs to be exchanged between these islands. As we discussed, for frontier models with over a trillion parameters, the bandwidth needed is large enough to require dedicated fibre optics connections. 

However, for smaller models, bandwidth requirements decrease proportionally. Recent breakthroughs in low-communication training algorithms, particularly in delayed synchronisation, have created promising opportunities for training small-to-medium-sized models in a decentralised manner. Two teams are leading these experimental efforts.

Nous Research is an AI accelerator company and a leading player in open-source AI development. They're best known for their Hermes series of language models and innovative projects like World Sim. Earlier this year, they operated an LLM-ranking BitTensor subnet for a few months. They have dipped their toes into decentralised compute by releasing the DisTrO (Distributed Training Over the Internet) project, where they successfully trained a 1.2B parameter Llama-2 model while achieving an 857x reduction in inter-GPU bandwidth requirements.

The DisTrO report by Nous Research

Prime Intellect, a startup developing infrastructure for decentralised AI at scale, aims to aggregate global compute resources and enable collaborative training of state-of-the-art models through distributed systems. Their OpenDiLoCo framework (implementing DeepMind's Distributed Low-Communication method) successfully trained a billion-parameter model across two continents and three countries while maintaining 90-95% compute utilisation.

But how do these decentralised training runs work?

Traditional data parallelism requires GPUs to share and average their weights after every training step—impossible over internet connections. Instead, these projects let each "island" of GPUs train independently for hundreds of steps before synchronising. Think of it like independent research teams working on the same project: rather than constantly checking in with each other, they make significant progress independently before sharing their findings.

DisTrO and OpenDiLoCo only sync every 500 steps, using a dual optimiser approach:

When they do sync, rather than sharing all the weights, they share a "pseudo-gradient"—essentially the difference between their current weights and the weights from the last sync. This is remarkably efficient, like sharing only what's changed in a document rather than sending the entire document every time.

INTELLECT-1, a practical implementation of OpenDiLoCo by Prime Intellect,  is pushing this approach even further by training a 10B parameter model—the largest decentralised training effort to date. They've added key optimisations like:

INTELLECT-1, trained by over 20 GPU clusters distributed across the globe, recently completed pretraining and will soon be released as a fully open-source model. 

INTELLECT-1 training dashboard

Teams like Macrocosmos are using similar algorithms to train models in the Bittensor ecosystem. 

If these decentralised training algorithms continue to get better, they might be capable of supporting models of up to 100 billion parameters with the next generation of GPUs. Even models of this size can be very helpful for a wide variety of use cases:

Fine-tuning

Fine-tuning is the process of taking a pre-trained foundation model (usually an open-source one by Meta, Mistral, or Alibaba) and further training it on a specific dataset to adapt it for particular tasks or domains. This requires significantly less compute than training from scratch since the model has already learned general language patterns and only needs to adjust its weights for the new domain.

Compute requirements for fine-tuning scale with model size. Assuming training on an H100:

Given these specifications, fine-tuning doesn't demand the complex distributed training algorithms previously discussed. The on-demand model, where developers rent GPU clusters for short, concentrated periods, provides adequate support. Decentralised compute marketplaces with robust GPU availability are ideally positioned to handle these workloads.

Inference

Inference is where decentralised compute marketplaces have the clearest path to product-market fit. Ironically, this is the least discussed workflow in the context of decentralised training. This stems from two factors: inference lacks the appeal of 100,000 GPU "god model" training runs, and partly because of the current phase of the AI revolution. 

As of today, the majority of compute is indeed going towards training. The race to ASI is leading to massive upfront investments in training infrastructure. However, this balance inevitably shifts as AI applications move from research to production. For a business model around AI to be sustainable, the revenue generated from inference must exceed the costs of both training and inference combined. While training GPT-4 was enormously expensive, that was a one-time cost. The ongoing compute expenses—and OpenAI's path to profitability—are driven by serving billions of inference requests to paying customers.

Compute marketplace, decentralised or otherwise, by nature of aggregating a variety of models of GPU (old and new) from across the globe, find themselves in a unique position to serve inference workloads. 

Compute marketplaces, whether decentralised or traditional, naturally excel at inference workloads by aggregating diverse GPU models (both current and legacy) globally. Their inherent advantages align perfectly with inference requirements: broad geographic distribution, consistent uptime, system redundancy, and compatibility across GPU generations.

But why crypto?

We’ve discussed the different workflows decentralised compute can and cannot help with. Now, we need to answer another important question: why would a developer choose to secure compute from a decentralised provider over a centralised one? What compelling advantages do decentralised solutions offer?

Pricing and Range

Stablecoins achieved product-market fit by offering a superior alternative to traditional cross-border payments. A big factor is that stablecoins are simply much cheaper! Similarly, the single biggest factor that drives an AI developer’s choice of cloud provider is cost. For decentralised compute providers to compete effectively, they must first deliver superior pricing.

A compute marketplace, like all marketplaces, is a network effects business. The more the supply of GPUs on a platform, the greater the liquidity and availability for customers, which in turn attracts more demand. As demand grows, this incentivises more GPU owners to join the network, creating a virtuous cycle. Increased supply also enables more competitive pricing through better matching and reduced idle time. When customers can consistently find the compute they need at attractive rates, they're more likely to build lasting technical dependencies on the platform, which further strengthens the network effects.

This dynamic is particularly powerful in inference, where geographic distribution of supply can actually enhance the product offering by reducing latency for end users. The first marketplace to achieve this liquidity flywheel at scale will have a significant competitive advantage, as both suppliers and customers face switching costs once they've integrated with a platform's tooling and workflows.

The GPU marketplace network effects flywheel

In such winner-takes-all markets, bootstrapping the network and reaching escape velocity is the most critical phase. Here, crypto provides decentralised compute projects with a very powerful tool that their centralised competitors simply don’t possess: token incentives. 

The mechanics can be straightforward but powerful. The protocol would first launch a token that includes an inflationary rewards schedule, possibly distributing initial allocations to early contributors through airdrops. These token emissions would serve as the primary tool for bootstrapping both sides of the marketplace. 

For GPU providers, the reward structure should be carefully designed to shape supply-side behaviour. Providers would earn tokens proportional to their contributed compute and utilisation rates, but the system should go beyond simple linear rewards. The protocol could implement dynamic reward multipliers to address geographic or hardware-type imbalances—similar to how Uber uses surge pricing to incentivise drivers in high-demand areas. 

A provider might earn 1.5x rewards for offering compute in underserved regions or 2x rewards for providing temporarily scarce GPU types. Further tiering the reward system based on consistent utilisation rates would encourage providers to maintain stable availability rather than opportunistically switching between platforms.

On the demand side, customers would receive token rewards that effectively subsidise their usage. The protocol might offer increased rewards for longer compute commitments—incentivising users to build deeper technical dependencies on the platform. These rewards could be further structured to align with the platform's strategic priorities, such as capturing the demand in a particular geography.

The base rates for compute could be kept at or slightly below market rates, with protocols using zkTLS oracles to continuously monitor and match competitor pricing. The token rewards would then serve as an additional incentive layer on top of these competitive base rates. This dual pricing model would allow the platform to maintain price competitiveness while using token incentives to drive specific behaviours that strengthen the network.

By distributing token incentives, both providers and customers would start accumulating a stake in the network. While some, perhaps most, might sell these stakes, others would hold onto them, effectively becoming stakeholders and evangelists for the platform. These engaged participants would have a vested interest in the network's success, contributing to its growth and adoption beyond their direct usage or provision of compute resources.

Over time, as the network reaches escape velocity and establishes strong network effects, these token incentives can be gradually tapered off. The natural benefits of being the largest marketplace—better matching, higher utilisation, broader geographic coverage—would become self-sustaining drivers of growth.

How token incentives can supercharge the GPU marketplace flywheel

Censorship Resistance

While price and range are critical differentiators, decentralised compute networks address a growing concern: operational restrictions from centralised providers. Traditional cloud providers have already demonstrated their willingness to suspend or terminate services based on content policies and external pressures. These precedents raise legitimate questions about how similar policies might extend to AI model development and deployment.

As AI models become more sophisticated and tackle increasingly diverse use cases, there's a real possibility that cloud providers may implement restrictions on model training and serving, similar to their existing content moderation approaches. This could affect not just NSFW content and controversial topics, but also legitimate use cases in areas like medical imaging, scientific research, or creative arts that might trigger overly cautious automated filters.

A decentralised network offers an alternative by allowing market participants to make their own infrastructure decisions, potentially creating a more free and unrestricted environment for innovation.

The flip side of permissionless architecture is that privacy becomes more challenging. When compute is distributed across a network of providers rather than contained within a single trusted entity's data centres, developers need to be thoughtful about data security. While encryption and trusted execution environments can help, there's an inherent trade-off between censorship resistance and privacy that developers must navigate based on their specific requirements.

Trust and Contract Enforcement

Given the sky-high demand for AI compute, GPU providers can exploit their position to extract maximum profit from successful customers. In a post from last year, the famous solo developer Pieter Levels shared how he and other developers experienced their providers suddenly increasing prices by over 600% after sharing their AI app's revenue numbers publicly. 

Decentralised systems can offer a counter to this problem—trustless contract enforcement. When agreements are encoded on-chain rather than buried in terms of service, they become transparent and immutable. A provider can't arbitrarily increase prices or change terms mid-contract without the changes being explicitly agreed to through the protocol.

Beyond pricing, decentralised networks can leverage trusted execution environments (TEEs) to provide verifiable compute. This ensures developers are actually getting the GPU resources they're paying for—both in terms of hardware specifications and dedicated access. For example, when a developer pays for dedicated access to eight H100 GPUs for model training, cryptographic proofs can verify that their workloads are indeed running on H100s with the full 80GB of memory per GPU, rather than being silently downgraded to lower-end cards or having resources shared with other users. 

Permissionless 

Decentralised computer networks can provide developers with truly permissionless alternatives. Unlike traditional providers that require extensive KYC processes and credit checks, anyone can join these networks and start consuming or providing compute resources. This dramatically lowers the barrier to entry, particularly for developers in emerging markets or those working on experimental projects.

The importance of this permissionless nature becomes even more powerful when we consider the future of AI agents. AI agents have just started finding their footing, with vertically integrated agents expected to surpass the size of the SaaS industry. With the likes of Truth Terminal and Zerebro, we’re seeing the first signs of agents gaining autonomy and learning how to use external tools like social media and image generators. 

As these autonomous systems become more sophisticated, they may need to dynamically provision their own compute resources. A decentralised network where contracts can be executed trustlessly by code rather than human intermediaries is the natural infrastructure for this future. Agents could autonomously negotiate contracts, monitor performance, and adjust their compute usage based on demand—all without requiring human intervention or approval.

The Landscape

The concept of decentralised compute networks isn't new—projects have been trying to democratise access to scarce computing resources long before the current AI boom. Render Network has been operating since 2017, aggregating GPU resources for rendering computer graphics. Akash launched in 2020 to create an open marketplace for general compute. Both projects found moderate success in their niches but are now focussing on AI workloads.

Similarly, decentralised storage networks like Filecoin and Arweave are expanding into compute. They recognise that as AI becomes the primary consumer of both storage and compute, offering integrated solutions makes sense. 

Just as traditional data centres struggle to compete with purpose-built AI facilities, these established networks face an uphill battle against AI-native solutions. They lack the DNA to execute the complex orchestration required for AI workloads. Instead, they are finding their footing by becoming compute providers to other AI-specific networks. For instance, both Render and Akash now make their GPUs available on io.net's marketplace.

Who are these new AI-native marketplaces? io.net is one of the early leaders in aggregating enterprise-grade GPU supply, with over 300,000 verified GPUs on its network. They claim to offer 90% cost savings over centralised incumbents and have reached daily earnings of over $25,000 ($9m annualised). Similarly, Aethir aggregates over 40,000 GPUs (including 4,000+ H100s) to serve both AI and cloud computing use cases. 

Earlier, we discussed how Prime Intellect is creating the frameworks for decentralised training at scale. Apart from these efforts, they also provide a GPU marketplace where users can rent H100s on demand. Gensyn is another project betting big on decentralised training with a similar training framework plus a GPU marketplace approach. 

While these are all workload-agnostic marketplaces (they support both training and inference), a few projects are focussing only for inference—the decentralised compute workload we’re most excited about. Chief among these is Exo Labs, which enables users to run frontier-level LLMs on everyday devices. They have developed an open-source platform that allows for the distribution of AI inference tasks across multiple devices like iPhones, Androids, and Macs. They recently demonstrated running a 70-B model (scalable up to 400-B) distributed across four M4 Pro Mac Minis. 

Essential Infrastructure

When Satoshi launched Bitcoin in 2008, its benefits—digital gold with a hard supply and censorship-resistant money—were purely theoretical. The traditional financial system, despite its flaws, was functioning. Central banks hadn't yet embarked on unprecedented money printing. International sanctions weren't weaponised against entire economies. The need for an alternative seemed academic rather than urgent.

It took a decade of quantitative easing, culminating in COVID-era monetary expansion, for Bitcoin's theoretical benefits to crystallise into tangible value. Today, as inflation erodes savings and geopolitical tensions threaten dollar dominance, Bitcoin's role as "digital gold" has evolved from a cypherpunk dream to an asset adopted by institutions and nation-states.

This pattern repeated with stablecoins. As soon as a general-purpose blockchain in Ethereum was available, stablecoins immediately became one of the most promising use cases. Yet, it took years of gradual improvements in the technology and the economies of countries like Argentina and Turkey to be ravaged by inflation for stablecoins to evolve from a niche crypto innovation into critical financial infrastructure moving trillions of dollars in annual volume.

Crypto is by nature a defensive technology—innovations that seem unnecessary during good times but become essential during crises. The need for these solutions only becomes apparent when incumbent systems fail or reveal their true colours.

Today, we're living through AI's golden age. Venture capital flows freely, companies compete to offer the lowest prices, and restrictions, if any, are rare. In this environment, decentralised alternatives can seem unnecessary. Why deal with the complexities of token economics and proof systems when traditional providers work just fine?

But going by major technology waves of the past, this benevolence is temporary. We're barely two years into the AI revolution. As the technology matures and the winners of the AI race emerge, their true power will surface. The same companies that today offer generous access will eventually assert control—through pricing, through policies, through permissions. 

This isn't just another technology cycle at stake. AI is becoming civilisation's new substrate—the lens through which we'll process information, create art, make decisions, and ultimately evolve as a species. Compute is more than just a resource; it's the currency of intelligence itself. Those who control its flow will shape humanity's cognitive frontier.

Decentralised compute isn't about offering cheaper GPUs or more flexible deployment options (though it must deliver both to succeed). It's about ensuring that access to artificial intelligence—humanity's most transformative technology—remains uncensorable and sovereign. It's our shield against an inevitable future where a handful of companies dictate not just who can use AI, but how they can think with it.

We're building these systems today not because they're immediately necessary, but because they'll be essential tomorrow. When AI becomes as fundamental to society as money, permissionless compute won't just be an alternative—it will be as crucial to resisting digital hegemony as Bitcoin and stablecoins are to resisting financial control.

The race to artificial superintelligence might be beyond the reach of decentralised systems. But ensuring that the fruits of this intelligence remain accessible to all? That's a race worth running.

Touching grass in the Himalayas,Shlok Khemani