Our roadmap to a decentralized inference network

Ghostral is a Public Benefit Corporation with a single objective of creating the first decentralized inference network (DIN). We believe this is the most effective way to move quickly and build the trust we need to reach our utility goal.

Once the network is working, our plan is to transition ownership from the company to the public.

Our roadmap

1. Build a private AI platform

We have launched ghostral.ai, a product that does not store or monetize your data. Think DuckDuckGo, but for AI. We do this by running open-source models instead of OpenAI's or Anthropic's, and we have open-sourced both our backend and our frontend.

2. Advance cryptographic AI and neutrality research

Right now, experts are a few years behind the AI wave when it comes to the best way to build a truly decentralized inference network. A few solutions exist, such as TEE hardware and homomorphic encryption, but each has its shortcomings. TEE is hardware specific and is currently available only under enterprise agreements from private clouds such as Azure, which makes it impossible to decentralize across multiple independent entities. Homomorphic encryption is very promising for making AI inference genuinely private, but it is significantly expensive to compute and needs further research before we understand how to run computation efficiently.

Our objective is to advance the field of AI cryptography by concentrating talent and raising awareness of its importance. We aim to refine existing technologies and build a truly auditable, decentralized inference network.

We also recognize that AI models inherit biases from their training data, which can reflect dominant political and cultural viewpoints while dismissing minorities. We want to advance research into neutral AI models as well, enabling a more equitable future in which enterprises can rely on unbiased systems when they interact with AI.

3. Create a decentralized inference network

Our inevitable goal is to create a decentralized inference network (DIN). What do we mean by that?

By a DIN, we mean a network with the following characteristics:

• Verifiable and encrypted inference
• Public ownership
• Anti-capture governance

Verifiable and encrypted (VE) inference is inference whose underlying model can be verified by the public, and whose content cannot be spied on by the inference provider. We want the public to agree on the best, safest, and most neutral model that should be served, and anyone should be able to verify that this exact model is being served, rather than a manipulated model acting in bad faith or steering opinion. We have been working on a protocol we internally call Proof of Inference, in which we hash the model weights and attach the resulting token to a public ledger. Inspired by the blockchain technology behind cryptocurrencies, we want inference to be verifiable.

We also want inference to be encrypted, so that an inference provider who joins the network cannot capture your information, download it onto their own machine, or even access it in the first place. The mainstream approach today is homomorphic encryption, which works by performing mathematical operations directly on ciphertext and returning an encrypted result. It is promising, but very slow. We want to accelerate innovation here to make it fast and viable, so we can achieve encrypted inference.

Public ownership comes second. Our objective is a complete liquidation of the company's shares on public markets, so that everyone can hold a stake in the inference network. Dead simple.

Anti-capture governance is closely tied to public ownership. We want to invest time and resources into understanding how to set up anti-capture governance for both the inference network and the company's ownership. We want to prevent large companies with datacenter facilities from owning the majority of the network and switching inference on and off wherever they please. We also want to prevent bad actors from seizing a majority of the company's voting power and abandoning our mission. We will invest resources into making our network resilient to these kinds of attacks, and into setting up a restricted ownership structure that guards against raids of all sorts and keeps the product democratic.

4. Full-dilution IPO

The moment we achieve a working DIN and have settled on a suitable anti-capture governance structure, we will IPO with full dilution, enabling the public to become a 100% stakeholder in the network.

Where we are today

Today is June 24th, 2026, and we are currently in Phase 1. We want to build an agentic platform that rivals OpenAI's and Anthropic's products. We are fortunate to benefit from a blooming open-source community that makes this possible.

We ask you to help our mission in any of the following ways:

• Switch your subscription from Claude or OpenAI to Ghostral
• Contribute to our open-source repository
• Spread awareness about the dangers of centralized AI
• Give us feedback and suggest features on Discord. We do not have your chat logs, so it is harder to learn what you want us to improve. 😉
• Join us: email us at mission@ghostral.ai and tell us how you want to help