Technology

Gonka Co-Creator Gleb Morgachev on AI Model Verification, GPU Costs and Decentralized Compute

Written by
Read time
9 min
Published
Filed under
Add to Preferred Sources
AI processor chip on a circuit board

When developers send a request to an AI provider, they generally have to trust that the provider runs the model they asked for. The developer may never notice if the provider uses a smaller, cheaper model that produces a plausible response.

Gonka uses a decentralized network of independent GPU operators to address that problem. Developers access AI models through an API, while other network participants randomly check inference results to detect operators that use the wrong models. The protocol also tests the computing capacity of participating hosts and gives operators voting rights over supported models and network rules.

Gleb Morgachev, Head of Engineering at Product Science and a co-creator of the Gonka protocol, spoke with TechBullion about the economics of decentralized AI inference. The interview also covered the challenges of coordinating GPUs across different locations and the technical difficulties of distributed model training. He also discussed how independent compute networks could support personal AI agents in the years ahead.

1. Can you explain how your decentralized compute network works and what differentiates it from traditional cloud-based GPU infrastructure?

My work on language models made me skeptical of the idea that releasing model weights would make AI widely accessible. Most developers don’t want to manage GPUs themselves. They want to call an AI API and get a response. But they still depend on whoever runs the service.

Say your application calls an API for a particular open model. You get an answer back, but you usually have to trust the provider that it ran the model you asked for. A smaller, cheaper model might produce a convincing response, and you wouldn’t necessarily know the difference. With Gonka, developers can use a familiar API through a broker, and independent GPU operators handle the requests. Other participants randomly check inference results to detect operators that aren’t running the models they’re supposed to.

There’s also the question of who makes the decisions. With a traditional AI API, one company decides which models to offer and what to charge. In Gonka, the hardware operators vote on which models the network supports and the rules behind its pricing. The prices themselves respond to demand. And these operators have a reason to keep both developers and hardware providers satisfied. If the prices are too high or the model selection is too limited, developers can go elsewhere. I want developers to have the convenience of an AI API without depending on a single company for everything that happens behind it.

2. How do you coordinate distributed GPUs and ensure that AI workloads can be executed efficiently across the network?

Imagine you have one server in Warsaw and another in Singapore. You can’t expect them to communicate like two machines sitting next to each other in the same data center. Distance introduces latency, and if your system needs those servers to constantly exchange data, you’re going to have performance problems.

When a developer sends an inference request to Gonka, a host acting as a Transfer Agent passes it to another host, the Executor, which runs the model. That machine handles the computation and sends the result back. Meanwhile, the relevant information gets recorded on the blockchain. The model doesn’t have to wait for a new block before it can start answering.

I tend to look at how long those requests actually take and how often they complete successfully. Having thousands of GPUs connected to a network is useful, of course, but the number alone doesn’t tell me much about how well the system performs.

3. AI compute has become one of the most expensive and strategically important resources in the technology industry. What economic inefficiencies in the current GPU market is Gonka trying to address?

You can have an expensive GPU sitting idle in one place while a developer somewhere else is struggling to find enough compute to run a model. That frustrates me.

Larger companies usually have a better idea of how much computing power they’ll need, so they can reserve capacity ahead of time. But a smaller team might need several GPUs to test a model today and hardly any next week. Cloud providers do offer on-demand access, of course, but the price and availability of the specific hardware you need can still be a problem.

I understand why cloud providers prefer predictable contracts. They have expensive data centers to finance, and guaranteed revenue helps with that. Independent GPU operators face a different problem. They have expensive hardware to keep busy, but finding enough customers isn’t always easy. Gonka lets them offer that computing power through a shared network. Developers get another way to run AI models through an API, and operators can earn money from hardware that would otherwise sit idle.

4. What makes decentralized AI compute economically viable compared with centralized providers, particularly when it comes to cost, utilization, and access to high-performance GPUs?

I wouldn’t assume decentralized compute is always going to be cheaper. Someone still has to buy the hardware, pay for electricity and keep those machines running. And when the hardware is spread across different locations, you have additional costs to deal with.

But you don’t need to build a huge data center to become a host. An operator can connect a small cluster to Gonka, provided the hardware meets the network’s requirements, and earn fees for the inference requests it handles. They can start with what they can afford and add more capacity as demand grows. Developers, meanwhile, can access AI inference without having to operate that hardware themselves. Of course, the machines still need enough memory and processing power to run the requested models.

A fair comparison looks at what a developer pays for each successful request and how long it takes. A low price means little if requests keep failing or response times are unpredictable. Gonka has its own pricing at the protocol level, and independent brokers may charge differently.

5. How do you ensure that participants contributing compute are rewarded for useful, reliable resources rather than simply for contributing raw GPU capacity?

If an operator tells me they have eight powerful GPUs, I want to know what those machines can actually deliver. A list of hardware specifications doesn’t tell you whether that computing power will be available when the network needs it.

Gonka uses Proof of Compute to assess that capacity. During short tests called Sprints, hosts run Transformer-based calculations, and the verified results determine their computational weight. We also use Confirmation Proof of Compute to check whether they can maintain that capacity. Serious failures can cost hosts their rewards.

Hosts earn fees for processing AI requests and receive additional rewards in GNK, the network’s native token, based on their verified computing power. The tests give us a measure of each host’s computing capacity, but actual usage depends on how many requests developers send.

Finally, for us it’s really important to align those tests with the real, useful work these hosts can deliver. Otherwise, it creates an incentive to optimize the hardware for those tests rather than for fast, reliable inference.

6. Training and inference have very different technical and economic requirements. How is Gonka approaching each, and where do you see the greatest opportunity for decentralized infrastructure?

Inference is easier to handle across a network like ours. If a model is already running on a host with enough computing power, it can answer requests without constantly exchanging data with servers elsewhere. Developers can already use Gonka for inference today.

Distributed training is much harder, and it gets expensive quickly. When GPUs in different locations train a model together, they have to exchange updates. If one connection slows down, other machines may end up waiting. You’re still paying to run that hardware during those delays.

We’re exploring approaches such as DiLoCo, which allows workers to go longer between synchronization steps. We also need to work out checkpointing and recovery. If a host goes offline halfway through training, we should be able to recover the work rather than start the whole job again.

I think training has to remain a major goal for any decentralized AI network that’s thinking long term. We can’t assume frontier AI labs will keep releasing model weights forever. They could decide to stop, or governments could restrict those releases. But I don’t want to make big promises until we’ve demonstrated what the network can do with real training runs.

7. Decentralization introduces questions around governance, verification, security, and reliability. How are you thinking about these challenges as the network scales?

A decentralized network should let people join without asking anyone’s permission or revealing their identity. We’ve designed Gonka to work that way wherever possible. But when you don’t know who’s operating the hardware, you can’t rely on personal trust to keep the network running.

The rules can’t come from one company, either. Gonka’s hosts vote on major protocol changes, including which AI models the network supports. They also decide how community funds are spent, whether it’s a marketing initiative or rewarding developers who contribute code. Voting power comes from verified computing work, so the people providing the hardware have a say in how the network operates.

Every host also has to prove its computing capacity. That’s where Proof of Compute comes in. We can check what participants contribute without having to take their word for it.

Then there’s the question of which models hosts choose to run. You could have plenty of computing power across the network but still struggle to serve a particular model if too few hosts support it. We recently introduced Dynamic Coefficients to address that. Governance can set a target share of computing power for each model based on expected demand, and the protocol gradually adjusts reward coefficients to encourage hosts to meet those targets. Hosts still choose which models to run, but now they have a financial incentive to make capacity available where it’s needed.

8. Looking beyond compute itself, what new applications or businesses could become possible if developers had access to an open, globally distributed layer of AI infrastructure?

I’d really like to see personal AI agents built on open, decentralized infrastructure. These agents are already consuming enormous amounts of compute. According to OpenRouter, agents used nearly eight times as many tokens as human users on its platform in early October. Agent token consumption was also 34 times higher than in early February. And we’re only beginning to see what happens when millions of people have agents working for them throughout the day.

Look at OpenAI’s Dots or Meta’s Muse. These agents can keep working on tasks even when you’re not actively talking to them. But they rely on their providers’ models and infrastructure. Those companies decide where the services are available, what resources users can access and how much they’ll pay.

I’d like developers to be able to build similar personal agents on a globally distributed network of independent compute providers. They could choose which models to use without being tied to a single AI lab. And if the network supports regional routing, they could also control which countries or regions process their users’ data. I think that could make personal AI agents available to far more people, including those in markets that major providers don’t currently serve.

9. Looking five years ahead, what do you believe the AI compute landscape will look like, and what role do you want Gonka to play in creating that future?

I expect the big cloud providers to remain major players five years from now. They have the money to build enormous data centers, and tightly connected GPU clusters will still be important for certain workloads.

Look at what happened with Bitcoin mining. It started on ordinary computers and eventually moved to machines built specifically for mining. AI is more complicated because different models have different hardware requirements. Still, I expect more specialized accelerators to emerge, particularly for inference.

Developers will also have more reasons to think carefully about where they run their models. A cloud provider can change its prices or restrict access to certain services. If you’ve built your entire application around that provider, moving elsewhere can be expensive and difficult. I expect more developers to look for infrastructure that gives them greater control over those decisions.

I’d like Gonka to become a regular option for developers who need to run open AI models. I want to see independent operators across different countries serving applications that people use every day. And five years from now, I’d be much more interested in how many developers actually rely on the network than how many GPUs we’ve managed to connect to it.

Spotted an error? Request a correction
Never miss a storyThe TechBullion Weekly newsletter and your preferred channels.Subscribe
Share this
Comment

More executive interviews

Founders, CEOs and technology leaders talk about what they build, how they fund it, and where AI, finance and data are taking their industries next.

All interviews