What Is DePIN? Decentralized GPU Compute vs. Centralized Cloud, and Why VOLT Leads

Three companies control roughly 65% of all global cloud infrastructure. If you need a GPU today, you are almost certainly renting it from either Amazon, Google, or Microsoft. You’ll pay Big Cloud’s prices, operate under their terms, and be at the mercy of availability they allocate first to their largest customers. That is the status quo with centralized cloud providers. DePIN is the only alternative.

DePIN, which stands for Decentralized Physical Infrastructure Networks, turns the hyperscaler model inside out. In a DePIN cloud computing marketplace, no single company owns all of the hardware and reselling access. Instead, many thousands of independent hardware operators contribute their resources to a shared, permissionless marketplace. Buyers and sellers transact directly, governed by transparent on-chain coordination rather than opaque enterprise contracts.

For AI teams, the most consequential flavor of DePIN is decentralized GPU compute. In this realm, VOLT is DePIN’s largest and most mature implementation, with thousands of GPUs across 138+ countries, all available for instant deployment at 50–70% below hyperscaler pricing.

This article is your primer. You’ll learn how DePIN works, why it has structural advantages over centralized cloud, and how VOLT's network delivers those advantages in practice.

Centralized Cloud’s problems exist by design 

Big Cloud’s GPU compute landscape is faulty on purpose. It’s designed to be a natural monopoly, where anyone who wants to build an AI app, train their model, or conduct LLM research is subject to predatory contracts that make it harder, not easier to function. 

Pricing is ruled by oligopoly, not free market competition

AWS, Azure, and GCP don't meaningfully compete on GPU pricing. In fact, they don’t really want to. 

H100 80GB instances cost $4.99–6.98/hr on AWS, $4.50–6.20/hr on Azure, and $4.20–5.50/hr on GCP. Why is this spread so narrow? These providers don't need to undercut each other because switching costs are prohibitively high once a team is embedded in one ecosystem's tooling.

Real GPU price discovery does not exist. In the centralized cloud compute model, the open market clears GPU supply against GPU demand. The hyperscaler quotes the price, and the customer either pays it or goes elsewhere, which usually means confronting another hyperscaler’s remarkably similar prices.

Availability is rationed, not elastic

If you’ve ever tried to provisions H100 and A100 instances on AWS, chances are you’ve faced 3–6 month waitlists. GCP GPU quotas require 24–72 hour approval processes, and even then requests could very well be denied outright. Azure H100 inventory sells out continuously. The official narrative is "high demand," but the structural reality is something else entirely: these providers have little incentive to overbuild capacity because unoccupied GPUs are simply CapEx without returns.

For startups and researchers, this means compute is unavailable when you need it most, regardless of the ability to pay.

Vendor lock-in is the business model

AWS SageMaker, GCP Vertex AI, and Azure ML are purpose-built to deter teams from leaving by making it prohibitively expensive to do so. Hyperscalers’ proprietary APIs, data formats, and orchestration are all pitched to AI teams as conveniences, but there is nothing convenient about that stack being added to your eventual migration bill as a switching cost. So, projects that have been on AWS for the last three might not not leave but instead negotiate a renewal. 

These egress fees practically guarantee the lock-in. Moving 100TB of training data out of AWS or GCP costs $8,000–12,000. Data sovereignty is just not a principle in the hyperscaler model, unless of course you want to pay for it as a premium feature when you sign the contract. 

Access is gated and unequal

Not everyone can get a GPU from AWS. If you’re a new AI startup or just kickstarting some LLM research, you will need a billing account, a business entity, quota approval, and often a minimum spend commitment. Not every project or research team can manage this; and, really, they shouldn’t have to.

Beyond that, teams in countries without formal corporate structures, with thin credit histories, or in regions outside the hyperscalers' primary markets, are effectively excluded.

All of this has consequences that are anti-competitive and, in some ways, monopolistic. AI capability is now being concentrated in a handful of well-funded teams with deep AWS/GCP relationships. Everyone else? Left waiting.

What DePIN changes, and its structural advantages

DePIN doesn't just build more data centers to solve the GPU shortages. Instead, it makes available GPU clusters that are sitting idle, like gaming rigs, research clusters, and crypto mining operations, and underutilized enterprise colocations. 

Here’s the reality: the world has far more GPU compute than the hyperscalers' inventory implies, or wants anyone to know. The only problem was that it just hadn’t aggregated into a market. That’s what DePIN in general, and VOLT in particular, solved. 

Price discovery through actual free market competition

In a DePIN compute marketplace, supply-side participants (GPU operators) compete for user demand. A provider offering RTX 4090s in Singapore competes against one in Frankfurt. The price is set by market clearing, not by a corporate pricing committee. This produces meaningful, sustained discounts relative to hyperscaler rates.

VOLT's marketplace delivers 50–70% below hyperscaler pricing as a direct consequence of this mechanism. H100 80GB instances run $1.49–2.20/hr. A100 80GB instances run $2.30/hr. RTX 4090 (unavailable on any hyperscaler) runs $0.28/hr. These prices reflect what independent operators need to cover their costs and earn margin in a competitive market.

Elastic supply without artificial rationing

A decentralized network can add supply continuously as new operators onboard hardware. What this means for AI startups and LLM researchers is they won’t have to experience a procurement cycle, CapEx committee approval, or inventory constraint from a company's balance sheet. Quite simply, when GPU demand spikes, new operators spin up; and supply naturally follows demand rather than lagging behind by quarters.

VOLT maintains thousands of GPUs with high availability — no waitlists, no quota requests, no approval process. Deployment time is under 2 minutes via CLI or dashboard. An H100 cluster that would take 4 months to provision on AWS is live on VOLT in the time it takes to drink a coffee.

Permissionless, instant access

A permissionless network sets programmatic, rule-based conditions for participation. No discretionary gatekeeping exists. Any team that can pay can access VOLT's network without a business entity requirement, minimum commitment, or account manager relationship. A researcher in Lagos or Karachi or Warsaw with a crypto wallet and a training script can deploy GPUs today.

The teams and countries currently excluded from hyperscaler access represent enormous untapped AI capacity. 

No vendor lock-in by design

DePIN networks are built on open standards because they have to be. A decentralized marketplace can't force proprietary APIs on a fragmented supply side. VOLT instances use standard SSH access, Docker containers, PyTorch, TensorFlow, Kubernetes. Every tool works because it's the industry standard, not a proprietary SDK.

This means you can deploy on VOLT today, then move to a different provider tomorrow. We’ve made the workload portable because nothing about VOLT's infrastructure is proprietary. That's the structural counterpoint to the hyperscaler model.

Geographic distribution

VOLT's thousands of independent providers reaches over 138 countries, which places GPU compute closer to end users than AWS's 33 regions or GCP's 39 regions. For inference workloads where latency matters, a distributed DePIN network can frequently outperform a centralized data center on the other side of the globe.

By the numbers: DePIN vs. Centralized Cloud

Example at scale: A team running 10x A100 80GB GPUs continuously:

• AWS: 10 × $4.55/hr × 730 hrs/month = $33,215/month
• VOLT: 10 × $2.30/hr × 730 hrs/month = $16,790/month
• Annual savings: $197,100

Those kinds of savings can be reinvested, either with more infra or a new team member. 

How VOLT implements DePIN for GPU compute

VOLT is DePIN's largest GPU network by deployed capacity. Understanding how the network functions matters because the implementation details determine whether the structural advantages above actually hold in practice.

Supply: independent operators

VOLT aggregates GPU supply from independent operators who provide hardware. This could be data centers, colocations, bare-metal servers, or even properly vetted consumer hardware. Operators earn revenue by fulfilling compute jobs. The network uses on-chain coordination to handle job matching, payment settlement, and operator reputation.

The supply side's diversity is the feature. No single data center outage takes down the whole network. When AWS us-east-1 has a bad day, every AWS customer has a bad day. When one VOLT provider cluster goes offline, the job routes around it.

Demand: bare-metal access

Buyers deploy via VOLT's CLI, dashboard, or API. Provisioned instances offer full bare-metal SSH access without virtualization overhead, hypervisor tax, or restricted root access. CUDA drivers, system libraries, kernel parameters are all fully configurable.

This matters for performance. Virtualized GPU passthrough on hyperscaler instances carries 3–10% overhead. VOLT's bare-metal access eliminates that entirely, delivering the GPU's rated TFLOPS without a hypervisor taking a cut.

GPU selection: The full stack

VOLT offers the widest GPU selection in the market:

• H100 SXM / PCIe 80GB: $1.49–2.20/hr — frontier training and inference
• A100 80GB: $2.30/hr — production training workloads
• A100 40GB: $1.40/hr — fine-tuning and mid-scale inference
• RTX 4090 (24GB): $0.28/hr — fine-tuning up to 13B parameters, inference, development
• L40S / A6000: available for professional rendering and mixed compute

The RTX 4090 option, powerful consumer GPU, is worth re-emphasizing because it simply does not exist on any hyperscaler offering. At $0.28/hr, a 4x RTX 4090 cluster running LLaMA 13B inference full-time costs $817/month vs. $8,935/month for 4x A100 40GB on AWS.

No waitlists, quotas, or approvals

This deserves its own section because the friction of the hyperscaler approval process is invisible until you encounter it just once. On AWS, you can submit a quota increase request, wait up to 6 months, potentially get denied, appeal, and wait around some more. 

On VOLT, you can type a deploy command, and your GPU is up and running in under 2 minutes.

For a startup with a 2-week runway to a demo, those 6 months are the entire company. DePIN removes the gatekeepers.

Who should use DePIN compute

DePIN GPU compute is not the right answer for every team in every situation. Let’s be precise instead of making any claims that just aren’t true. 

DePIN is the right choice when:

• Cost matters more than managed services – If you're running PyTorch training or vLLM inference and comfortable managing your own stack, VOLT saves 50–70% immediately.
• Speed of access matters – No other platform provisions H100s in under 2 minutes with no approval required.
• You're building on open-source models (LLaMA, Mistral, Qwen, Stable Diffusion) – These workloads are NVIDIA GPU-native and run identically on VOLT hardware.
• You're a global team or an international researcher without a corporate AWS relationship – VOLT's permissionless access removes the bureaucratic barrier.
• You want portability – No proprietary dependencies means you can leave whenever you want.

Centralized cloud may still win when:

• You need formal enterprise compliance certifications (HIPAA, FedRAMP) for regulated production workloads – AWS and GCP have deep cert portfolios; VOLT's SOC 2 is in progress.
• You're deeply integrated with a managed ML platform (AWS SageMaker, GCP Vertex AI) and the migration cost exceeds the compute savings.
• Your workload is JAX/TPU-native – GCP's TPU v5p offers price-performance that GPU networks can't match for that specific case.

The honest answer for most AI teams is you should use VOLT for training and development, evaluate hyperscalers for compliance-gated production inference, and run the numbers on whether managed services justify the 2–3x cost premium for your specific workflow.

Getting Started on VOLT

Deploying on VOLT takes under 2 minutes:

# Install the CLI

curl -fsSL https://buildonvolt.com/install.sh | bash

# Deploy an H100 cluster

VOLT deploy \

  --gpu h100-80gb \

  --count 4 \

  --image nvcr.io/nvidia/pytorch:24.01-py3

# Or an RTX 4090 cluster for development

VOLT deploy \

  --gpu rtx-4090 \

  --count 4 \

  --image nvcr.io/nvidia/pytorch:24.01-py3

New accounts receive $100 in credits. And again, there won’t be a waitlist, a quota approval, or an account manager to deal with just to provision the compute. 

Want to calculate your savings? Compare your current AWS/GCP bill against VOLT pricing at VOLT/pricing. 

For most teams running 4+ GPUs continuously, the annual savings exceed $100,000

Why DePIN matters for AI’s future

The centralization of GPU compute means that a handful of companies in a handful of countries control who gets to train AI models, and dictate costs, terms, and any access restrictions. Over time, that type of concentration compounds, as teams without hyperscaler access might not even get expensive compute; they will get no compute at all.

DePIN's structural answer is to turn idle hardware everywhere into accessible infrastructure for everyone. The network's value grows as more operators contribute supply and more teams contribute demand. This is a model that hyperscalers simply cannot replicate without undermining their own pricing power.

The compute oligopoly is not inevitable. DePIN is what happens when it isn't.

Ready to deploy? Start building on VOLT →