The “experts” keep watching which AI model is number one. I’m watching who controls the compute keeping those models alive.
GPUs, inference, retrieval, verification, developer access, private compute and energy are all becoming one massive battlefield. Bittensor already has ten subnets fighting to control a different piece of that infrastructure.
These projects aren’t all building the same product. Some are selling access to hardware, some are making inference easier, some are protecting private workloads and others are proving the computation actually happened.
The Compute Wars isn’t coming. Its already here.
SN64CHUTES
The Inference War
Chutes is fighting for one of the most important layers in AI: inference.
It doesn’t simply connect developers with available GPUs. Chutes turns decentralized compute into something people can actually build a real product on.
A developer can deploy an open-source model, connect through an API and let Chutes handle the GPUs, scaling and infrastructure underneath it. You don’t need to rent a machine, install everything yourself and hope the server stays alive.
That is a much bigger opportunity than people realize.
Most developers don’t care where the GPU is sitting. They care if the API is fast, reliable, affordable and easy to use. The easier Chutes makes decentralized infrastructure feel, the more likely developers are to use it instead of a traditional cloud provider.
Owning GPUs is one thing.
Turning them into a product developers actually want to use is another.
That is the layer Chutes is trying to own.
SN95ACTUAL
The Distributed Compute War
AI companies are spending billions building larger data centers, but Actual is asking a completely different question.
What if a lot of the compute already exists?
Gaming PCs, Macs, workstations and consumer GPUs sit idle for a huge part of every day. Instead of waiting years for more data centers to come online, Actual wants to connect that unused hardware into one distributed inference network.
The machines are already built and already paid for. The problem is they are scattered across different homes, offices and countries doing nothing most of the time.
Actual is trying to make thousands of independent machines feel like one usable pool of compute.
That is obviously not easy. Different hardware, internet speeds and locations all need to work together without the user seeing the mess underneath.
But if Actual can make that network reliable, the opportunity is massive.
They aren’t trying to build the world’s biggest data center.
They are trying to prove we already have one, its just disconnected.
SN51LIUM
The GPU Marketplace War
Lium is probably the most direct compute play on this list.
People with available GPU machines list their hardware, and people who need compute can rent it for training, inference, rendering, data processing or other heavy workloads.
Simple idea, massive market.
Today, anyone needing serious compute normally ends up dealing with AWS, Azure or another centralized cloud provider. That can get expensive fast and you are still limited by whatever hardware those companies have available.
Lium gives independent GPU owners a way to earn from machines that would otherwise sit idle. At the same time, developers get another place to find the hardware they need.
The hard part is making sure everything works like its supposed to.
The machine has to be real. The performance has to match what was advertised and it has to stay available when someone is paying to use it.
Lium doesn’t need to invent a better GPU.
It needs to become the marketplace where compute gets bought and sold.
SN4TARGON
The Private Compute War
The more powerful AI becomes, the more valuable the information going into these models becomes.
Financial records, medical data, company IP, customer information and private models are not things serious businesses will send through infrastructure they can’t trust.
That is where Targon comes in.
Targon provides confidential GPU and CPU compute using hardware-protected environments. The goal is allowing private workloads to run without exposing the data or model to the person operating the machine.
Instead of blindly trusting whoever owns the server, the hardware helps prove the environment is secure.
I think this part of compute is going to matter a lot more than people realize.
Right now, most people still think about AI as a chatbot answering random questions. Companies are starting to connect AI to internal systems, private documents and some of the most valuable information they own.
Once that happens, privacy isn’t some cool extra feature.
It becomes a requirement.
Targon is trying to make decentralized compute secure enough for workloads businesses would never risk running on an open machine.
SN28GM
The Developer Access War
The best infrastructure in the world is worthless if developers don’t use it.
That is the problem GM is attacking.
GM gives developers a familiar way to access AI models without needing to understand every miner, provider or piece of infrastructure behind the request.
Developers can use code that works with the same type of setup they already know from OpenAI, Anthropic and Gemini. GM handles the provider access and routing underneath it.
That matters because developers are not going to rebuild their entire application just to experiment with decentralized AI.
The easier it is to switch over, the more likely people are to actually try it.
GM also uses Trusted Execution Environments to help protect requests while they are being processed. That adds another layer for companies that don’t want private prompts moving through exposed infrastructure.
GM isn’t fighting to own the GPUs.
It is fighting to own the relationship with the developers using them.
And history has shown the platform with the easiest developer experience normally gets the builders.
SN29HOτFLOAτ
The Optimization War
When most people talk about solving the AI compute shortage, their answer is simple.
Buy more GPUs.
Hoτfloaτ is taking a different approach. Instead of only adding more hardware, it is trying to make the hardware we already have perform better.
Miners compete to build faster and more efficient inference systems using better batching, caching, quantization and model-serving techniques.
The goal is not only producing the correct output. It is producing that output faster while using less time, memory and compute.
That might not sound as exciting as another giant GPU data center, but the value could be just as big.
If better software allows one GPU to serve twice as many users, you basically doubled its useful compute without buying another machine.
Every AI company is going to care about that.
Hardware is expensive. Power is expensive. Memory is limited and users still expect the answer immediately.
Hoτfloaτ is not trying to win by owning more GPUs.
It is trying to make every GPU count.
SN31REC4LL
The Retrieval War
The smartest AI model in the world is still useless if it can’t find the right information.
That is the problem Rec4ll is focused on.
Rec4ll works on the retrieval side of AI. Before a model produces an answer, the system needs to search documents, databases or knowledge bases and pull the right information into the response.
Most companies don’t need a model that memorized the entire internet.
They need a model that can find the correct answer inside their own data.
That could be customer records, product information, internal research, company documents or anything else the public model was never trained on.
A model can sound extremely confident while using the wrong source or completely missing the information it needed. Better retrieval makes the final response more useful, more accurate and easier to trust.
This is going to become a huge piece of enterprise AI.
The biggest model doesn’t always create the smartest system.
Sometimes the winner is the one that knows where to look.
SN53ENGY
The Verification War
Running an AI model is not enough in a decentralized network.
You also need to know the miner actually performed the work they claimed.
A miner could say it used an expensive model while secretly using a smaller one. It could skip layers, cut down the computation or return something that looks close enough while using far fewer resources.
The user might never know the difference.
Engy is building a verification layer designed to catch that behavior. The network uses commitments and audits to check whether miners honestly performed the inference they were paid to do.
That is what makes Engy different from a normal inference network.
It isn’t only asking who answered the request.
It is asking if the answer was produced honestly.
As decentralized AI grows, trust becomes just as valuable as the compute itself. Cheap inference doesn’t mean much if the miner can fake the work and keep the difference.
Compute you can’t verify is compute you can’t fully trust.
SN96VERATHOS
The Proof War
Engy is working on auditing the computation.
Verathos is taking the next step and trying to prove it with cryptography.
The goal is creating mathematical proof that the correct model, weights and computation were actually used to produce an output.
Not a company promise.
Not a dashboard saying the job completed.
Actual proof that can be independently checked.
That might sound like overkill right now, but imagine AI handling financial transactions, healthcare decisions, private company systems or critical infrastructure.
In those environments, “trust me” will not be good enough.
Verified inference is the starting point, but verified training and fine-tuning could become an even bigger opportunity. Those jobs are expensive, take a long time and can be extremely difficult for a customer to independently confirm.
When someone pays for a massive training job, they should be able to prove the work really happened.
Engy is checking the work.
Verathos wants the math to prove it.
SN110GREEN COMPUTE
The Sustainability War
AI has another compute problem that still doesn’t get enough attention.
Power.
Every new model requires more GPUs, more electricity, more cooling and more infrastructure. As AI demand continues growing, energy could become one of the biggest bottlenecks facing the entire industry.
Green Compute is approaching that problem from a different direction.
Instead of competing only to deliver the fastest compute, it focuses on where the energy powering that compute comes from. The goal is rewarding infrastructure powered by cleaner energy sources like hydro, solar and wind.
Today, that might sound like a niche.
Tomorrow, it could become a requirement.
Enterprises are under pressure to lower emissions, governments continue increasing sustainability requirements and AI companies are already being questioned about how much electricity their models consume.
Green Compute isn’t only asking how we power AI.
It is asking where that power should come from in the first place.
THESE ARE NOT TEN VERSIONS OF THE SAME PRODUCT
The more I researched these projects, the more obvious it became.
They are solving completely different problems across the same compute stack.
Chutes is turning decentralized inference into a product developers can use.
Actual is unlocking hardware that is already sitting idle.
Lium is creating a marketplace for GPU access.
Targon is protecting private workloads.
GM is making decentralized AI easier for developers to access.
Hoτfloaτ is making every GPU more efficient.
Rec4ll is improving how AI finds information.
Engy is checking whether the work was actually performed.
Verathos is building cryptographic proof for AI computation.
Green Compute is asking how all of that infrastructure should be powered.
That is why I don’t think Bittensor has one compute narrative.
There are multiple battles happening at the same time.
Some subnets are competing to own infrastructure. Some are competing to own developer adoption. Others are competing to own privacy, efficiency, verification or energy.
Together, they are building something much bigger than ten individual subnets.
They are building the decentralized AI compute stack.
THE GAME IM WATCHING
The Compute Wars are not about one subnet winning everything.
They are about who controls each layer of the future AI stack.
Bittensor does not need one subnet to dominate the entire market. It needs each layer to become useful enough that developers can build the next layer on top of it.
Some of these projects won’t make it.
Some will completely change direction.
But a few could become infrastructure used by thousands of applications without the people using those applications ever knowing Bittensor is underneath them.
People spend too much time debating which AI model wins.
I’m more interested in who is building the infrastructure those models will not be able to live without.
That is the game I’m watching.
— Shizzy