Theta launched in 2019 as a decentralized content delivery network (CDN) for video, but the company has evolved considerably since then.
Today Theta Labs builds Theta EdgeCloud, decentralized physical infrastructure network (DePIN) that supplies GPU computing power for AI, along with enterprise AI products such as customer service AI agents. This is a short account of how that transition came about and what Theta does now.
Theta Labs builds Theta EdgeCloud, a decentralized network that supplies GPU computing power for AI, along with enterprise AI products such as customer service AI agents.
From video delivery to GPU computing
Theta's original decentralized network was designed to make video streaming cheaper and more efficient. Rather than relying entirely on central servers, viewers could share their spare bandwidth to relay streams to one another and earn token rewards for doing so.
In 2021 Theta launched the Theta Edge Network, which extended this model from bandwidth to processing power. Individuals could run edge nodes that used their CPUs and GPUs to encode and transcode video.
The same year, Theta filed a patent for an edge computing platform coordinated by a blockchain, which was granted in 2023.
As demand for AI computing accelerated and GPUs became scarce and expensive, it became clear that this capacity was far more valuable for AI in addition to video delivery, and in May 2024 Theta launched EdgeCloud. Theta now operates its video delivery network as part of EdgeCloud, and it has expanded that original infrastructure to include compute, AI, and the products built on it.
How Theta EdgeCloud works
EdgeCloud combines GPU capacity from several sources and manages it as a single platform. The distributed edge network, which has grown from around 10,000 nodes in early 2024 to more than 30,000 today, is made up of machines run by individuals and organisations who contribute spare capacity in return for rewards. This is combined with partner data centres and enterprise cloud infrastructure from Google Cloud and AWS, which provide high-end hardware such as NVIDIA H200 and A100 GPUs.
When a customer submits a job, the platform routes it to the hardware best suited to its requirements and budget, so large training runs can be placed on data centre GPUs while inference, fine-tuning and other tasks that can be divided into smaller pieces run on consumer cards such as the RTX 4090 and 5090. Because much of this hardware would otherwise sit idle, EdgeCloud can offer compute at 50 to 70 percent below the rates of the major cloud providers. Universities including Stanford, Imperial College London, KAIST and NTU Singapore use it for AI research.
- 30,000+edge network nodes today, up from around 10,000 in early 2024
- 50–70%below the rates of the major cloud providers, which EdgeCloud can offer because much of its hardware would otherwise sit idle
- 30+universities, including Stanford, Imperial College London, KAIST and NTU Singapore, use it for AI research
The role of the blockchain
Coordinating thousands of independently owned machines raises questions that a conventional cloud provider never has to answer, such as how to confirm that a node has actually completed the work it was given and how to pay a global community of contributors fairly.
The Theta blockchain was built to handle this initially for video segments and now for AI compute jobs, rendering and more. Each compute task can be registered onchain and assigned to a node, the result can be verified by a smart contract, and the contributing node is then rewarded automatically.
This gives contributors a transparent record of their work and earnings, and it allows EdgeCloud to draw on hardware around the world without a separate agreement with every owner. For customers, the blockchain simply runs in the background.
Customer service AI agents for enterprises
The same infrastructure powers Theta's enterprise AI solutions, the most prominent of which are its customer service AI agents. These give organisations a conversational AI assistant, embedded in their website or app, that answers questions from customers and fans at any hour. Each agent is built on the organisation's own information and proprietary data from NHL, MLS APIs, supplied as text, documents, web pages or live data, and its knowledge base stays continuously synchronised so that answers about tickets, schedules, products or policies remain current. Because the agents draw on this verified material rather than relying only on the general knowledge of the underlying language model, they are more dependable in a customer-facing role.
The agents are designed to work alongside human support teams. Conversations can be handed to a member of staff in real time during scheduled hours, and when the team is offline the agent collects contact details for follow-up. Through Theta Intelligence, organisations can also turn these conversations into dashboards showing what customers ask about most and where they run into difficulty.
AI Gaming Services
In Summer 2026, Theta launched its AI Gaming vertical, beginning with its AI Characters product, reflecting how quickly generative AI has moved from experimentation into everyday game development.
AI Characters gives developers a way to add AI-powered non-playable-characters (NPCs) to games and other interactive experiences, and it runs on the same EdgeCloud infrastructure as Theta's other products.
Developers describe the game world, including its setting, rules and atmosphere, and then define each character's personality, goals, speaking style and knowledge, with preset templates available for common archetypes. Characters respond to player dialogue, actions and in-game events. They can also return structured decisions that feed directly into a game's logic, which allows them to function inside an actual game loop rather than as chatbots.
Theta has published a public developer repository containing a working reference implementation and the full API specification. It has also released a poker demo in which players face AI opponents whose personalities stay consistent throughout a game.
Theta today
Theta now serves universities, AI startups, enterprises and sports organisations across the US, Europe and Asia. Its early work on distributing video across a peer-to-peer network turned out to be the foundation for something broader, and the company's focus today is on giving organisations affordable, flexible access to the computing power that AI depends on.
Key terms in plain English
About Theta
- Theta Network
- A blockchain-powered decentralized cloud for AI, media and entertainment. It works as a dual network: Theta EdgeCloud supplies GPU computing power, and the Theta blockchain handles payments, rewards and smart contracts.
- Theta Labs
- The company that builds Theta EdgeCloud and its enterprise AI products, such as customer service AI agents.
- Theta EdgeCloud
- Theta’s hybrid cloud-edge platform, launched in May 2024, giving developers, researchers and enterprises on-demand access to GPU computing power for AI, video, rendering and any containerized task.
- Theta Edge Network
- The decentralized network of community-run Theta Edge Nodes that powers EdgeCloud. It has grown from around 10,000 nodes in early 2024 to more than 30,000 today.
- Hybrid cloud-edge
- An architecture that combines community-run edge nodes with partner data centres and enterprise cloud infrastructure from Google Cloud and AWS, managed together as a single platform.
- Decentralized GPU cloud
- A cloud whose GPU capacity comes from many independent owners rather than one company’s data centres, coordinated as a single platform.
- DePIN
- Decentralized physical infrastructure network: real-world hardware, such as GPUs, contributed by many independent owners and coordinated as one platform.
- CDN
- Content delivery network: a network of servers that delivers video and other content to viewers quickly.
- Neocloud
- A cloud provider that specialises in renting GPU compute for AI training and inference, rather than offering the broad range of services of a traditional cloud. Theta EdgeCloud is a decentralized example. Read more: What is a neocloud?
- GPU-as-a-service
- Renting GPU computing power on demand, by the hour or per job, instead of buying and running the hardware yourself.
Tokens and the blockchain
- THETA
- Theta’s governance and security token. Holders stake it to run Validator or Guardian nodes, and the protocol distributes newly generated TFUEL to stakers in proportion to their stake. Read more: Why EdgeCloud drives TFUEL and THETA value
- TFUEL
- Theta’s operational token. Every compute job, node payment and on-chain transaction runs through it, and node operators are always paid in TFUEL, whatever currency the customer paid in. Read more: Why EdgeCloud drives TFUEL and THETA value
- TDROP
- A token on the Theta blockchain that rewards activity on the ThetaDrop NFT marketplace, gives holders governance rights over ThetaDrop and rewards stakers. Read more: Intro to TDROP
- Theta blockchain
- Theta’s EVM-compatible, multi-chain blockchain, which supports Turing-complete smart contracts. For EdgeCloud it handles payments and rewards, and can register compute tasks and verify their results.
- Smart contract
- Code that runs on a blockchain and executes automatically when its conditions are met. On EdgeCloud, a smart contract can verify the result of a task and trigger the node’s reward.
- Staking
- Locking up tokens to help secure a blockchain network, in return for rewards.
- Theta Metachain
- A “chain of chains”: the Theta mainnet as the main chain, with any number of purpose-specific subchains above it. All are EVM-compatible and use TFUEL as gas. It launched in December 2022 to let the network scale horizontally. Read more: Theta Metachain overview
- TNT20
- Theta’s token standard, the counterpart of Ethereum’s ERC-20. TDROP is a TNT20 token.
- EVM-compatible
- Compatible with the Ethereum Virtual Machine, so Ethereum-style smart contracts and developer tools work on Theta.
Nodes and the network
- Edge node
- A computer run by an individual or organisation that contributes spare capacity to the network in return for rewards. Read more: Theta Edge Node
- Elite Edge Node
- An Edge Node whose operator also stakes TFUEL on the Theta blockchain. Read more: Elite Edge Node staking
- EdgeCloud Client
- Lightweight software that lets community members share idle GPU capacity with EdgeCloud customers. Operators set their own hourly rental rate, and rewards are calculated in USD and paid in TFUEL. Read more: EdgeCloud Client guide
- Community node
- A GPU run by an individual or organisation in the Theta community and rented to EdgeCloud customers, for example for inference, fine-tuning or media generation.
- Node operator
- Anyone who contributes hardware to the network. Operators are paid in TFUEL for the work their machines complete.
- Validator Node
- A node that proposes, votes for and finalizes new blocks on the Theta blockchain. Anyone can run one without approval from Theta Labs. Read more: Running a Theta Validator
- Guardian Node
- A node that seals blocks and acts as a check on malicious or non-functional Validator Nodes, forming a second layer of defence. Community members can run one by staking THETA. Read more: Guardian Node guide
- Enterprise Validator
- A large organisation that runs a validator node, staking THETA, producing blocks and taking part in on-chain governance. Theta’s enterprise validators include Google, Samsung, Sony, CAA and Binance.
EdgeCloud products and infrastructure
- On-demand Model APIs
- Serverless access to AI models through a simple API, paid for as you use them, with no servers to provision. Behind the scenes, part of the work runs on community-operated nodes. Read more: On-demand Model APIs
- Dedicated model deployment
- Running your own deployment of a model on EdgeCloud, instead of the shared on-demand service, to tune performance and latency. Templates cover models such as Llama 3, Stable Diffusion and Whisper. Read more: Dedicated AI Model Serving
- GPU Node
- A rentable GPU machine on EdgeCloud with SSH access, used for model training or any other computation. It works much like a virtual machine: you can stop, restart and reconfigure it. Read more: GPU Node docs
- GPU cluster
- Several GPU Nodes of the same type in the same region that can communicate directly with each other, which makes distributed AI model training possible. Read more: GPU cluster docs
- Persistent storage
- Storage volumes that keep their data after the GPU Node they are attached to is stopped or deleted, and that can be shared between nodes. Read more: Persistent Storage docs
- Containerized workload
- Any job packaged as a container image, such as a custom AI model, a data analysis job or a scientific simulation, which EdgeCloud can run on GPU hardware. Read more: Containerized workloads
- Jupyter Notebook
- An open-source web application for documents that combine live code, equations, visualizations and text, widely used to prototype AI models. EdgeCloud can launch one for you. Read more: Jupyter Notebook on EdgeCloud
- EdgeCloud API key
- A credential for managing EdgeCloud deployments programmatically. Each key is scoped to a single project, so it can only reach that project’s resources. Read more: EdgeCloud API keys
- EdgeCloud Video Services
- Theta’s video-on-demand and livestreaming services, formerly known as the Theta Video API and now part of EdgeCloud. Read more: EdgeCloud Video Services
- Peer-to-peer (P2P) delivery
- Delivering video by letting viewers relay streams to one another using spare bandwidth, rather than relying only on central servers. It is where Theta began.
- NFT-based DRM
- Digital rights management for video in which viewers must own an NFT from a chosen collection to receive the decryption keys. Read more: NFT-based DRM
AI fundamentals
- Artificial intelligence (AI)
- Software that performs tasks normally associated with human intelligence, such as understanding language, recognising images or making decisions.
- Generative AI
- AI models that create new content, such as text, images, audio or video, rather than only classifying or predicting.
- Large language model (LLM)
- An AI model trained on vast amounts of text that can read, write and reason about language. It is the technology behind chatbots such as ChatGPT.
- Multimodal AI
- AI that works with more than one type of data, such as text, images and audio, at the same time.
- Training
- Building an AI model by feeding it large amounts of data. It is a finite project with an end date, after which the model is fixed. Read more: Why EdgeCloud drives TFUEL and THETA value
- Inference
- Running an already-trained AI model to produce answers or other outputs. Read more: Why EdgeCloud drives TFUEL and THETA value
- Fine-tuning
- Further training an existing AI model on specific data so that it performs better at a particular task.
- Token (in AI)
- A small chunk of text that a language model reads or writes. Many model APIs are billed per token. Not to be confused with THETA or TFUEL.
- Context window
- How much text a model can consider at once. A long context window lets a model reason across several documents together.
- Mixture of experts (MoE)
- A model design that activates only part of its parameters for each request, which keeps inference cost and latency manageable at large scale.
- Quantization
- Storing a model’s weights in lower-precision formats, such as 8-bit or 4-bit instead of 16-bit, to reduce its memory footprint. Read more: Distributed inference explained
- Latency and throughput
- Latency is how long a single request takes. Throughput is how many requests the system can handle at once. Tuning for one often costs the other.
GPUs and AI chips
- GPU
- Graphics processing unit: the chip that does the heavy parallel maths behind AI training and inference. Read more: Find the GPU built for your workload
- VRAM
- A GPU’s own memory. It largely determines how large a model a card can hold.
- NVIDIA RTX GPUs
- Consumer-grade GPUs, such as the RTX 4090 and 5090, that handle inference, fine-tuning, media generation and prototyping well. Many community nodes run them. Read more: Renting an Nvidia RTX GPU for AI workloads
- Enterprise GPUs (A100, H100, H200)
- Data-centre accelerators used for the most demanding work, such as pretraining frontier models. EdgeCloud pairs them with community GPUs. Read more: Find the GPU built for your workload
- AWS Trainium
- AWS’s own chip for cost-efficient model training at scale, available through Theta EdgeCloud Hybrid. Read more: Find the GPU built for your workload
- AWS Inferentia
- AWS’s own chip for serving models efficiently at production scale, available through Theta EdgeCloud Hybrid. Read more: Find the GPU built for your workload
Running AI models across many GPUs
- Distributed inference
- Running a single AI model across multiple GPUs instead of one, by splitting its layers, the computations inside those layers, or the incoming requests. Read more: Distributed inference explained
- Tensor parallelism
- Splitting the operations inside each layer of a model across GPUs, so several devices compute a share of the same layer at once. It reduces the latency of a single request but needs fast links between GPUs. Read more: Distributed inference explained
- Pipeline parallelism
- Assigning different layers of a model to different GPUs, like an assembly line. It needs less communication between GPUs, which suits slower or more distant links. Read more: Distributed inference explained
- Prefill and decode
- The two phases of answering a prompt. Prefill processes the whole input and builds the working memory; decode then generates the response one token at a time. Read more: Prefill/decode disaggregation benchmark
- Prefill/decode disaggregation
- Running prefill and decode on separate sets of GPUs so each suits its own workload. Theta EdgeCloud benchmarked the approach on a 235-billion-parameter model. Read more: Prefill/decode disaggregation benchmark
- KV cache
- The working memory a model builds from a prompt and reuses as it generates a response. It grows with context length.
- Parallax
- A distributed serving framework built by Gradient. EdgeCloud uses it to serve the Qwen3-32B model across community GPUs using pipeline parallelism. Read more: Distributed inference explained
AI agents and Theta products
- AI agent
- A branded conversational assistant, built for an organisation, that answers customer questions from its approved sources on websites, apps and platforms such as Discord. Read more: AI customer service agents for sports teams
- Agentic AI
- AI that can plan and act towards a goal, using tools, rather than only answering questions. Read more: Custom Agentic Tools
- RAG (retrieval-augmented generation)
- A technique in which an agent searches a curated knowledge base and uses the verified information it finds to generate its answer, rather than relying on memory alone. Read more: AI Agent docs
- Knowledge base
- The documents, web pages, APIs and live data an AI agent draws on. Kept up to date, it keeps the agent’s answers current. Read more: Updating the knowledge base
- Agentic tools
- External capabilities, such as web search, API calls and text-to-SQL, that let an AI agent act beyond retrieval. Read more: Custom Agentic Tools
- Live agent escalation
- Handing a conversation from an AI agent to a member of staff in real time. Read more: Live Agent Escalation
- Digital mascot
- An AI agent given its own name and character, so it speaks in the voice of the organisation it serves.
- Theta Intelligence
- Theta’s premium analytics tier for AI agents. It turns anonymised conversations into real-time dashboards, alerts and reports on what customers are asking. Read more: Unlock customer insights with Theta Intelligence · Theta Intelligence FAQ
- AI Insights report
- A written summary of how an AI agent is used over a period, covering volume, sentiment and the topics that come up most. Read more: AI Insights reports
- AI Characters
- Theta’s product for adding AI-powered non-playable characters to games and other interactive experiences. Read more: Build AI Characters into your game
- NPC
- Non-playable character: a character in a game that the player does not control.
- MCP (Model Context Protocol)
- An open standard that lets AI assistants connect to external tools and services. EdgeCloud has an official MCP server, so an assistant can run models, manage GPU Nodes and check spending. Read more: The official MCP server · MCP server docs
- Agent-native
- Built so AI agents can use a platform directly: discovering, deploying and managing resources through APIs and MCP rather than a dashboard. Read more: Agents can now deploy EdgeCloud GPUs
For updates on what we're building and working on, follow our blog at blog.thetatoken.org.
