Why Open-Source AI Needs Incentives

Most people experience artificial intelligence through a remarkably simple object: a text box.

You type a question, wait a few seconds and receive an answer. The interaction feels almost weightless. Somewhere on the internet, intelligence appears on demand.

The physical reality behind that text box is rather less elegant.

AI models depend on GPUs, data centres, electricity, cooling, networking equipment, storage and enormous amounts of capital. Even after a model has been trained, serving it to millions of people remains expensive. Someone has to buy the machines, keep them online and pay the electricity bill while users casually ask the model to rewrite an email for the fourth time.

This physical requirement is one of the main reasons AI is becoming so centralized. A growing number of open-source models may be available to everyone, but only a relatively small number of companies can operate the infrastructure needed to train and serve the largest models at global scale.

Open source has opened the model.

It has not automatically opened the factory.

This is where I think Bittensor adds something genuinely different. It introduces an economic reason for people around the world to contribute compute, models, storage, data and other useful resources. Someone who owns suitable hardware does not have to donate it to an open-source project out of goodwill. He can connect it to a subnet, compete with other providers and earn rewards when his work is useful.

The idea sounds simple once it is stated.

Open-source AI gives people access to the technology. Bittensor tries to give them a reason to provide the infrastructure on which that technology runs.

That missing incentive layer may prove extremely important.

Open source solved only part of the problem

Open-source AI has already achieved something remarkable.

Developers can download powerful models, inspect their architecture, modify them, fine-tune them and run them outside the platforms of the companies that originally trained them. A startup does not necessarily have to begin by asking OpenAI or Anthropic for permission. It can build with open weights and retain far more control over the resulting product.

But a model file sitting in a repository is not yet an AI service.

The model still needs a machine capable of running it. If many people want to use it, the service needs many machines, reliable networking, monitoring, load balancing and engineers who prevent the whole system from disappearing during the night. Training a new model requires even more coordination and compute.

This is the uncomfortable gap inside many open-source visions.

The code may be open, but the infrastructure remains expensive.

A community can publish software, share research and improve models together. What it cannot easily do is persuade thousands of strangers to buy GPUs, pay their electricity bills and remain available whenever a customer submits a request.

Enthusiasm helps.

Money tends to be more dependable.

Bittensor turns the contribution of digital resources into an economic activity. Each subnet defines a commodity it wants miners to produce. Miners compete to provide it, validators measure their performance and the network rewards the participants judged to be creating the most value. Bittensor subnets currently coordinate commodities including inference, training, compute and storage.

This changes the character of open participation.

A person running a model is no longer merely helping a project. He may be operating a small business. A developer who discovers a faster way to serve a model can compete for rewards. A team with several GPUs can join an existing market without first building its own consumer brand.

The subnet provides the rules, the competition and the payment system.

Whether it provides a good product depends on how well those rules are designed.

The hidden resource sitting in homes and offices

There is an enormous amount of computing hardware in the world that is not used continuously.

Gaming computers sit idle while their owners sleep or work. Companies own workstations that are busy for several hours and quiet for the rest of the day. Universities, developers and small AI teams operate machines that rarely achieve perfect utilisation.

None of this means that an ordinary laptop can suddenly train the next frontier model. Different AI tasks require different types of hardware, memory and network connections. A MacBook is useful for local AI, but it is not secretly an H100 cluster wearing thinner clothes.

Still, the larger observation remains valid: computing capacity is widely distributed, while the systems capable of coordinating it remain limited.

Most spare capacity is economically invisible. The machines are scattered across different countries, owned by people who do not know one another and connected through networks of uneven quality. A conventional company cannot easily build a dependable AI service from this collection.

Bittensor asks whether an incentive market can help organise it.

A subnet can specify the work it needs, establish how that work will be measured and reward the machines that perform best. If joining the market is profitable, more hardware providers may participate. Poor providers can lose rewards while better ones gain a larger share.

The network does not have to care whether the machine sits inside a hyperscale data centre or a small server room. It cares whether the machine produces the result the subnet rewards.

The current response to growing AI demand is generally to build more data centres and concentrate more compute inside facilities owned by large companies. Those facilities will remain essential. Some workloads require tightly connected clusters and cannot be distributed casually across household internet connections.

But before assuming every new unit of intelligence requires another enormous centralized facility, it seems reasonable to ask whether some of the hardware already in existence can be used more effectively.

Bittensor provides an economic structure for attempting exactly that.

Chutes made the idea real for inference

Chutes is perhaps the clearest current example because it already offers developers access to open-source models through familiar APIs.

Inference is what happens when a trained model actually performs a task. When you send a prompt to a language model and receive an answer, the model is running inference. A single request may not require much time, but serving large numbers of users requires reliable and scalable GPU infrastructure.

Chutes coordinates that infrastructure through a decentralized network of compute providers. Developers can access hosted open-source models or deploy their own workloads without managing every underlying machine themselves. Chutes describes itself as a decentralized compute platform for deploying and running open-source AI models in production.

This is where the incentive layer becomes concrete.

Miners have a financial reason to supply GPUs. Chutes can coordinate those machines into a service. Developers pay to use the inference, and the application built on top may never reveal that Bittensor is involved.

The end user simply receives an answer.

I find this important because it shows that decentralized compute does not have to remain a theoretical network used mainly to discuss decentralization. It can look like an ordinary AI service from the outside.

A developer uses an API.

A miner earns from providing hardware.

The customer may have no idea either exists.

That is already much closer to infrastructure than to a crypto experiment.

Training is much harder

Inference is only one part of AI. Training is the process through which the model learns in the first place, and distributing that work is considerably more difficult.

During the training of a large model, many GPUs need to exchange information and remain synchronized. These machines are usually placed inside the same data centre and connected through extremely fast networks. Even small communication delays can slow the process considerably.

A collection of gaming computers spread across France, Germany and Canada does not naturally behave like one well-designed training cluster.

The machines may have different GPUs, different amounts of memory and wildly different internet connections. One participant may disconnect. Another may submit poor work. A malicious miner may attempt to earn rewards without performing the promised computation.

This is one reason decentralized training matters so much. It is also why I would be suspicious of anyone presenting it as a solved problem.

If a Bittensor subnet can help coordinate useful training across independent machines, the network begins to decentralize more than access to existing models. It begins to decentralize the creation of new ones.

That is a much more ambitious objective.

IOTA and the attempt to train together

IOTA is one of the strongest examples of this direction.

Earlier decentralized-training systems often asked each miner to train a complete model independently. This allowed competition, but it also meant the size of the model was limited by the hardware available to each individual miner.

IOTA takes a different approach. Its architecture is designed to divide training across a network of heterogeneous and unreliable devices, allowing miners to contribute to one shared training process rather than each building an isolated model. Different machines can work on different parts of the model or different portions of the training data, while the system attempts to measure and reward their individual contributions.

This is technically difficult for reasons that are easy to understand.

The network must keep the training process moving when some machines are slow or disappear. It must compress the information travelling between participants so internet bandwidth does not become an impossible bottleneck. It must work out whether each miner genuinely improved the model and reward that contribution fairly.

IOTA’s published architecture includes data and pipeline parallelism, activation compression and mechanisms intended to estimate the value contributed by different miners. Early work suggests that model layers can be distributed across participants so the total model is not limited to the memory of one machine.

This does not mean that someone can connect an old laptop tonight and begin training the world’s largest model.

It means the architecture is exploring how many different machines could cooperate without being owned and operated by one company.

That is the part I find exciting.

A distributed-training algorithm alone is interesting research. A distributed-training algorithm connected to an open reward system may be able to attract the hardware and contributors required to become an actual network.

Bittensor supplies the economic coordination that the technical architecture needs.

Where Parallax fits

Parallax belongs to the same wider effort to make heterogeneous hardware useful, although it should not be described as a distributed-training project.

Parallax focuses on decentralized inference. It explores how a large model can be divided and served across a pool of different volunteer GPUs, even when no individual machine can hold the complete model. Its scheduler decides how model layers should be placed across available hardware and which sequence of GPUs should handle each incoming request. Tests described in the Parallax paper used real volunteer nodes and showed improvements over other decentralized-inference approaches.

IOTA asks how independent machines can help train a model together.

Parallax asks how independent machines can cooperate to run an existing model efficiently.

Both projects confront the same physical reality: the world contains many machines, but they differ in capacity, speed and reliability. Turning them into one usable service requires intelligent coordination rather than simply connecting them to the internet and hoping for a sudden outbreak of cooperation.

This is another place where Bittensor’s incentives may help.

A scheduler can decide where work should go. The economic network can give hardware owners a reason to remain available and perform that work well.

Technology coordinates the machines.

Incentives coordinate the owners.

Why privacy becomes the decisive issue

There is an obvious problem with sending AI workloads to computers operated by strangers.

The stranger owns the machine.

A business may be comfortable asking an unknown GPU to generate a public image or answer a general question. It will be much less enthusiastic about sending customer records, legal documents, source code or medical information.

Without a credible privacy layer, decentralized compute may remain limited to workloads that contain nothing sensitive.

This is where trusted execution environments—usually shortened to TEEs—become important.

A TEE creates a hardware-isolated environment inside a computer. Data and code can be processed inside that protected area while the machine’s normal operating system and administrator are restricted from inspecting or modifying what happens there. The aim is to allow someone to use hardware he does not own without granting the hardware operator ordinary access to the workload.

A miner can supply the machine.

The user supplies the private data.

The hardware attempts to prevent the miner from reading it.

That changes what decentralized compute may eventually be used for.

Chutes has developed a confidential-compute path using Intel TDX and NVIDIA confidential-computing protections, with remote-attestation mechanisms intended to prove that workloads are running inside the expected protected environment. Chutes also publishes its TEE infrastructure as open source and explicitly discusses the remaining trust boundaries rather than pretending the technology makes every risk disappear.

TEEs are not magic privacy boxes. They depend on hardware manufacturers, secure configuration and software that does not contain vulnerabilities. Attestation can prove that particular code is running, but the code itself still needs to be understood and trusted.

A TEE reduces the amount of trust placed in the machine operator.

It does not remove trust from the system completely.

Still, for many real-world workloads, this may be enough to make decentralized infrastructure practical.

Spare capacity becomes more valuable when it can be private

Spare hardware without incentives is difficult to coordinate.

Incentives without privacy can only attract a limited range of workloads.

Privacy technology without a market may remain an expensive feature used by a few specialist providers.

The combination is much more interesting.

Bittensor can reward people for supplying hardware. Systems such as Parallax and IOTA explore how very different machines can cooperate. TEEs can protect sensitive information while it is being processed on hardware the user does not control.

Together, these pieces point toward an AI infrastructure market that looks very different from the conventional cloud.

A small business could submit a private workload without building its own GPU cluster. Independent hardware providers could perform the computation and receive payment. The business would not need to know the miners, and the miners would not need to see the business’s data.

Imagine a legal firm that wants to search thousands of confidential documents with an open-source model. It does not want to send those documents to a normal consumer chatbot, but buying and operating several high-end GPUs makes little sense for a small firm.

A mature Bittensor service could offer private model inference through a network of TEE-equipped miners. The firm pays for the compute, receives the results and avoids owning the entire infrastructure.

The same idea could eventually serve researchers, hospitals, journalists, local governments and small AI companies.

Not every workload will fit.

But there are many organisations that want the advantages of AI without handing their most sensitive data to one dominant platform.

An open market for confidential compute could become enormously valuable to them.

Why centralized providers cannot offer exactly the same structure

Centralized cloud and AI companies can also provide strong security and confidential-computing products. Some may offer better guarantees than young decentralized networks for quite some time.

The difference lies in the structure rather than the mere existence of privacy technology.

With a centralized provider, one company controls the infrastructure, account, billing system and rules of access. It can change its prices, restrict users or discontinue a service. The customer ultimately depends on that company continuing to behave as expected.

A decentralized market can potentially allow several independent providers to compete underneath one service. If one provider becomes unavailable or expensive, workloads can move elsewhere. No individual miner needs to become the permanent trusted intermediary.

That does not automatically make the decentralized service safer.

It makes the supply side more open and replaceable.

I think this distinction will become increasingly important as AI moves deeper into business and personal life. People will want privacy, but they may also want freedom from complete dependence on one model company or cloud provider.

Bittensor can connect those two desires: more private execution and more competitive infrastructure.

Open-source AI needs an economy

The deepest point is not really about spare gaming computers.

It is about whether open AI can develop an economy capable of supporting itself.

Closed AI companies have a clear economic system. Investors provide capital. The company hires employees, purchases hardware and sells access to the resulting models. Control and revenue accumulate inside the company.

Open-source projects often distribute the technology much more widely, but the economic structure underneath them is weaker. A model can be copied by everyone while the expensive work of training, hosting and improving it remains dependent on a small group of companies, researchers or sponsors.

Bittensor offers another possibility.

The model can remain open while the people supplying useful work are paid through competitive markets. Compute providers earn for compute. Model builders earn for performance. Storage providers earn for reliable storage. Validators earn for judging the network accurately.

The open system does not have to survive entirely on donations and goodwill.

It can pay its contributors.

This is why I think the incentives matter more than the token itself. TAO and alpha tokens are the accounting instruments, but the larger idea is to create markets around useful machine work.

Open source gave people permission to build.

Bittensor tries to make building economically sustainable.

The honest limitations

It is easy to become carried away by this vision.

I do, occasionally.

But several enormous problems remain.

Decentralized training is still early. Communication between distant machines is expensive and slow. Measuring a miner’s contribution to one shared training process is difficult. Hardware may disappear unexpectedly. Incentive systems can be gamed.

TEEs introduce hardware-based privacy, but also new trust assumptions and possible security failures. Consumer hardware varies enormously, and much of it will never be suitable for demanding AI work. A distributed network may also be less energy-efficient than a well-optimised data centre for certain workloads.

Most importantly, protocol rewards can attract hardware before paying customers exist.

A subnet may appear to have a thriving supply network because miners earn tokens for participating. The stronger test is whether outside customers are willing to pay for the resulting service. At some point, real revenue needs to replace at least part of the subsidy.

Otherwise the network is not yet producing sustainable infrastructure.

It is paying people to rehearse it.

That rehearsal can still be valuable. New industries often need subsidised experimentation before customers understand the product. But we should not confuse the bootstrapping phase with completion.

Why I still think this is groundbreaking

After all those qualifications, I remain extremely enthusiastic about the basic idea.

The world is moving toward more AI, more compute demand and probably more dependence on a small number of companies that control the largest infrastructure.

Open-source models provide one response by making intelligence itself more accessible. But they do not solve the question of who owns the machines, who pays the contributors and who turns scattered hardware into a reliable service.

Bittensor adds an economic layer.

A person with useful hardware can become a miner and earn rewards. Independent machines can be coordinated into inference or training systems. Open-source models can be served without one company owning every GPU. Confidential compute can make at least some private workloads possible across machines operated by strangers.

Perhaps Chutes becomes the dominant open-inference service. Perhaps IOTA succeeds in coordinating large-scale distributed training. Perhaps Parallax or another architecture makes heterogeneous GPUs surprisingly useful. More likely, some of today’s projects will fail and a later subnet will discover the model that actually works.

The winner matters less to me than the fact that Bittensor creates a place where these experiments can compete.

Centralized AI labs experiment behind closed doors and decide internally which approaches survive.

Bittensor allows separate teams to create markets, attract hardware and prove their ideas in public.

That is a different method of building infrastructure.

It may turn out to be less efficient than centralized coordination in many areas. It only needs to become better in a few important ones.

If it does, the future of AI infrastructure will not belong exclusively to the companies capable of constructing the largest warehouses.

Some of it may belong to people who connect their machines, protect the users’ data and earn money for contributing useful work to an open network.

That is the possibility Bittensor creates.

And I think it is fantastic.