bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 1쪽
0G: Towards Fully Decentralized AI Operating System
Zero Gravity Labs
0g@0g.ai
September 4, 2025
Abstract
The advent of large language models, coupled with the rapid de-
velopment of parallel computing hardware (GPUs, TPUs, and NPUs),
has significantly accelerated the progress toward AGI. Recent advanced
models, trained on vast amounts of data from across the internet, have
demonstrated significant superiority over humans in many aspects and
real-world scenarios. However, the current centralized and monopolis-
tic model training process has raised many concerns regarding the secu-
rity, privacy, and fairness of applying these powerful models. Therefore,
there recently appears a strong demand for transparency and democra-
tization in the production and deployment process of AI products. De-
centralization powered by blockchain technologies has been recognized
as a promising path to achieve this. Unfortunately, the infrastruc-
tures in current blockchain industry still face tremendous scalability
challenges in achieving the internet-level scale required for real-world
AI systems and application scenarios. 0G makes a first step towards
the ultimate solution by providing a decentralized AI operating sys-
tem through a modular and layered architecture design. It consists
of a storage network, a data availability network, and a data serving
network, managing the computing and storage resources in all these
networks in a permissionless way, and orchestrating them organically
with the separate 0G consensus network. All of these network com-
ponents utilize carefully designed and innovative sharding mechanisms
to achieve infinite horizontal scalability, thereby removing obstacles to
the true democratization of AI.
1 Introduction
Artificial Intelligence (AI) has rapidly evolved into a powerful tool that is
transforming industries and reshaping our everyday lives. However, the
journey of AI development and deployment is often confined within central-
ized platforms, leading to significant concerns around privacy, fairness, and
alignment. As we continue to witness the proliferation of AI, a critical trend
is emerging: the democratization of AI through decentralization.
Centralized AI platforms, while convenient and powerful, introduce sev-
eral critical issues. When a handful of entities control the AI infrastructure,
they also control the data and the algorithms. This centralization creates
a scenario where users’ privacy can be compromised, either through data
breaches or through misuse of personal information. Moreover, the decision-
making processes of AI systems can become opaque, making it difficult to
ensure fairness and accountability. Alignment, or ensuring that AI systems
act in accordance with human values and intentions, becomes particularly
challenging in centralized environments. If AI development is monopolized
by a few organizations, the risk of misalignment with the broader popula-
tion’s values increases, potentially leading to outcomes that are beneficial
to some but detrimental to others.
Decentralization offers a promising alternative. By moving AI workflows
into a decentralized environment, the entire process becomes more transpar-
ent. Every step of the AI lifecycle—from data collection and processing to
model training and deployment—can be tracked, audited, and verified. This
transparency helps in addressing concerns about provenance (where the data
comes from) and attribution (who owns or has contributed to the data and
models). Furthermore, a decentralized approach makes it easier to ensure
that AI systems are aligned with a broader and more diverse set of values,
as it is less likely to be controlled by a single entity with a narrow agenda.
However, while the benefits of decentralizing AI are clear, the challenges
are equally daunting. The AI workflow is inherently complex, involving mas-
sive amounts of data processing and neural network computing. Replicating
this workflow in a decentralized environment is non-trivial. It requires well-
designed abstraction and super scalable infrastructure that can handle the
demands of AI workloads while maintaining the benefits of decentralization.
0G makes a first step towards the ultimate solution by providing a de-
2
centralized AI operating system through a modular and layered architecture
design. It consists of a storage network, a data availability (DA) network,
and a data serving network, managing the computing and storage resources
in all these networks in a permissionless way, and orchestrating them or-
ganically with the separate 0G consensus network. All of these network
components utilize carefully designed and innovative sharding mechanisms
to achieve infinite horizontal scalability, thereby removing obstacles to the
true democratization of AI.
The 0G storage network consists of a bunch of storage nodes connected
through a peer-to-peer gossip network. Each data block written into the
storage network accompanies a transaction on the consensus network to
record the commitment and the ordering of the data. To scale the data
throughput infinitely, the storage network is organized in a partitioned way
and connects to an arbitrary number of consensus networks that run in
parallel and independently. The data requests from different independent
applications may be written into different storage partitions and their data
commitments can be recorded into different consensus networks simultane-
ously. All the consensus networks share the same set of validators with the
same staking status so that they keep the same level of security. Each stor-
age node actively participates in a mining process by submitting proof of
accessibility for a specific piece of data to a smart contract deployed on a
consensus network. Once the proof is validated by the smart contract, the
storage node gets rewarded accordingly. This incentive-based mechanism
rewards the nodes for contributions rather than punishing them for misbe-
haviors, so it can better encourage nodes to participate in the maintenance
of the network, and hence can promote the network to achieve better scal-
ability in practice. 0G storage is also designed as a general storage system
with multiple stacks of abstractions and structures including an append-only
log layer for archiving unstructured data and a key-value layer for manag-
ing mutable and structured data. This allows 0G to support reliable data
indexing and a greater variety of data types for AI processing.
The 0G DA network comes from the demand for off-chain verification
of executed states in Layer2s or decentralized AI networks with optimistic
mechanism. Specifically, the off-chain verifier, usually the light client, needs
to be able to access the entire transaction history data to verify the execution
of the transactions. In Layer2 scenarios, the blocks containing executed
3
transactions in Layer 2 networks need to be published and stored somewhere
for light client to conduct further verification. Similar requirements also
exist in decentralized AI infrastructures where the results of training or
inference tasks on devices in the networks need to be further verified due to
the demands of users or system incentives. For example, in ORA/OpML [3]
scenarios, it requires participants to provide fraud proofs for specific AI
tasks during a challenge window and the proofs may need to contain the
data and models used in those tasks. Some incentive mechanisms may also
require randomly choosing and verifying old historical tasks to give rewards
accordingly, and hence to achieve a good trade-off between verification cost
and effectiveness.
In contrast to other alternative DA solutions, e.g., Celestia [5] and
EigenDA [6], 0G DA targets to the ultimate scalability and security. It
embraces the idea of separating the workflow of data availability into both
the Data Publishing Lane and the Data Storage Lane. The large volume of
data transfers happen on the Data Storage Lane that is supported by the DA
nodes with horizontal scalability through erasure-coding based data slicing,
while the Data Publishing Lane guarantees the data availability property
by checking the aggregated signatures of the corresponding DA nodes on
the consensus network, which only requires tiny data flowing through the
consensus protocol to avoid the broadcasting bottleneck.
The DA nodes in the network also need to be stakers on 0G consensus
and form a series of quorums with majority honesty assumption. This is
the security foundation of the data availability. The quorums of 0G are
constructed randomly by the consensus system through verifiable random
function (VRF) which theoretically guarantees the same distribution of the
honest participants as in the validator set of the entire consensus network, so
that the data availability client cannot collude with them. In other words, as
subsets of all the validators, the quorums share the same security property
as the entire consensus network. The aggregated signatures of the quorum
will be submitted to the consensus of 0G for data availability confirmation,
which can be orders of magnitude faster and more efficient than Ethereum.
Combined with the multi-consensus design of0G with infinite scalability, the
consensus protocol will not become the bottleneck of the data throughput.
In addition, as with EigenDA, the data availability client needs to conduct
erasure coding for the data to have it split into chunks. In practice, this
4
process is very costly and can be the major bottleneck for the throughput
of a single client. 0G provides a GPU-accelerated solution for the erasure
coding process to significantly speed it up.
Further, in order to support efficient and scalable AI computation and
data access, 0G proposes a decentralized serving network. It is a general
framework that can support all kinds of serving workloads including data re-
trieval, AI inference, and even AI training tasks. It consists of a set of smart
contracts and distributed software to allow any type of services to integrate
into the decentralized network. Attached with 0G serving framework SDK,
a service provider can register and publish the service type, pricing, and
verification method into the smart contract. A user can discover a service
that she is willing to use through our platform and pre-pay certain amount
of 0G tokens to the smart contract. She then can directly interact with
the service, sending request, receiving and verifying the response, and ac-
knowledging the response with signature. The service provider maintains
the request/response sequence with user acknowledgment and send it to the
smart contract for settlement at any time to get the corresponding reward
from the user paid fees. Through this way, the service workload can be
perfectly partitioned for horizontal scalability and the blockchain overhead
is minimized. Since all the service process traces are sent to the smart
contract for settlement, these information can also be used to evaluate the
contribution and quality of the services in the entire network, which enables
the appropriate and fair incentives for better service providers.
2 Overview of the System Design
0G system consists of a data availability layer ( 0G DA) on top of a decen-
tralized storage system ( 0G Storage). There are one or multiple separate
consensus networks that are part of both the 0G DA and the 0G Storage.
For 0G Storage, each consensus network is responsible for determining the
ordering of the uploaded data blocks, realizing the storage mining verifica-
tion and the corresponding incentive mechanism through smart contract.
For 0G DA, the consensus is in charge of guaranteeing the data availability
property of each data block via verifying the aggregated signatures from the
corresponding data availability quorum.
Figure 1 illustrates the architecture of the 0G system when deploying a
5
single consensus network. When a data block enters the 0G DA, it is first
erasure coded and organized into multiple consecutive chunks through era-
sure coding. The merkle root as a commitment of the encoded data block
is then submitted to the consensus layer to keep the order of the data en-
tering the system. The chunks are then dispersed to different storage nodes
in a quorum in 0G Storage. The storage nodes periodically participate the
mining process by interacting with the consensus network to accrue rewards
from the system. The DA client then collects the aggregated signatures from
the quorum and submits the signatures to the consensus for verification.
0G Consensus
Smart
Contract
0G Storage
Network
Storage
node
Validator
node
Disperse
server
DataBlob
Chunk1
Chunk2
DataBlob
Chunk2Chunk1
Erasure
coding data
to chunks
Data root&
commitment
Disperse chunks
to storage nodes
1
2
3
4 Aggregated
signatures of quorum
DA quorum
Figure 1: The architecture of the 0G system.
When deployed with multiple POS-based consensus networks, each val-
idator participates in the maintenance of all the consensus networks by us-
ing a shared staking status. The shared staking status can be recorded in
a smart contract on one of the multiple consensus networks or a blockchain
network outside of the 0G consensus network set, e.g. Ethereum. Assuming
C0 is the consensus network that maintains the shared staking status, the
token T0 on C0 would be staked for POS voting. When token Ti is produced
on network Ci (i ̸= 0) via incentives, it can be mapped to T0 on C0 by burn-
ing Ti on Ci through a secure cross-chain bridge channel. When a validator
maintains the POS protocol in any network Ci, it always uses the amount
of staked T0 on C0 as its voting power in Ci. Therefore, all the network
6
Ci share the same level of POS security. By using Ethereum as network
C0, it brings the benefit of incarnating the token T0 as a standard ERC20
token and also enables the easy integration with existing popular restaking
framework like EigenLayer [7]. Note that, the sources of the shared staking
status may also come from the high value tokens (e.g., BTC and ETH) in
restaking frameworks like EigenLayer and Babylon [8], so that to make 0G
consensus reach the Bitcoin or Ethereum level of security. As illustrated in
Figure 2, this shared staking design introduces the unlimited scalability to
0G system.
0G Storage Network
0G Chain
0
0G Chain
1
0G
Chain
2
$0G0
$0G1
$0G2
$0GERC20
V0 (0x5124...1c34)
V’0 (0x5124...1c34)
V’’0 (0x5124...1c34)
burn
mint
stake
fetch
stake
Figure 2: The infinite scalability of the 0G system.
3 Design of the 0G Storage
0G Storage employs layered design targeting to support different types of
decentralized applications. Figure 3 shows the overview of the full stack
layers of 0G Storage.
The lowest is a log layer that is a decentralized system consisting of
multiple storage nodes that form a storage network with a built-in incentive
mechanism that rewards data storage. The ordering of the uploaded data
is guaranteed by a sequencing mechanism that provides log-based seman-
tics and abstraction. This layer is used to store unstructured raw data for
7
permanent persistency.
Log System
Key-Value Store
Transaction Processing ACID for applications with concurrent update, e.g., Google doc with collaborative editing
Enables the mutability of the data with access
control and allows applications like
decentralized social network
Archive system which provides permanent
persistency and append-only property
Figure 3: Full stack solution of 0G Storage.
On top of the log layer, 0G Storage provides a Key-Value store runtime
to manage structured data with mutability. Multiple key-value store nodes
share the underlying log system, putting the structured key-value data into
the log entry and appending it to the log system. They play the log entries
in the shared log to construct the consistent state snapshot of the key-
value store. The throughput and latency of the key-value store are bounded
by the log system, so that the efficiency of the log layer is critical to the
performance of the entire system. The key-value store can associate access
control information with the keys to manage the update permission for the
data. This enables the applications like social network, e.g., decentralized
Twitter, which requires the maintenance for the ownership of the messages
created by the users.
0G Storage further provides transactional semantics on the key-value
store runtime to enable concurrent updates for the keys from multiple users
who have the write access permission. The total order of the log entries
guaranteed by the underlying log system provides the foundation for the
concurrency control of the transactional executions on top of the key-value
store. With this capability, 0G Storage can support decentralized applica-
tions like collaborative editing and even database workloads.
4 Log System Protocol
The log layer of 0G Storage provides decentralized storage service via a
permissionless network of storage nodes. These storage nodes collaboratively
8
serve archived data, where each node optionally specifies which portion of
data it keeps in local storage.
…
Host Blockchain
create new data entry
read mining status claim reward
Raw data Storage request
Users
hash commitment
Storage
Storage nodes
Mining proof
Status
Mining
random query Queried chunks
Figure 4: Overview of 0G Storage protocol. Data creation and reward distribution
are fully decoupled: users directly submit storage requests to the 0G Storage contract
deployed on host blockchain; storage nodes claim rewards by proving to the contract that
they have the ability to answer random queries to archived data. 0G tokens are handled
in another isolated contract, which enables non-blocking token transfer.
4.1 Protocol Overview
The storage state of 0G Storage network is maintained in a smart contract
deployed on 0G blockchain. This contract is called the 0G Storage contract
and the underlying blockchain is called the host blockchain or host chain for
short.
The design of 0G Storage network fully decouples data creation, reward
distribution, and token circulation.
The 0G Storage contract is responsible for data storage request process-
ing, data entry creation, and reward distribution. Data storage requests
are submitted by users who wish to store data in the 0G Storage network,
where each request includes necessary metadata such as data size and com-
9
mitments, and it comes along with the payment for storage service. Data
entries are created for accepted data requests, keeping record of stored data
while reward distribution is handled independently through a mining pro-
cess. storage nodes submit mining proofs to the 0G Storage contract to
claim rewards for maintaining the 0G Storage network.
4.2 Storage Granularity
The log layer of 0G Storage is updated (append-only) at the granularity
of log entries, where every entry is created by a storage-request transaction
sent to the 0G Storage contract. The log layer is akin to a filesystem, with
every log entry corresponding to a file.
The log system operates at the level of fixed-sizesectors, with each sector
storing 256 B of data. To avoid one sector being shared by distinct log
entries, every log entry must be padded to multiple sectors.
The mining process of 0G Storage requires proving data accessibility to
random challenge queries. To maximize the competitive edge of SSD storage,
the challenge queries are set to the level of 256 KB chunks, i.e. 1024 sectors.
As every challenge query requires the miner to prove accessibility to a whole
chunk of data, storage nodes would maintain data at the granularity of
chunks.
4.3 Data Flow
In 0G Storage, committed data are organized sequentially. Such a sequence
of data is called a data flow, which can be interpreted as a list of data entries
or equivalently a sequence of fixed-size data sectors. Thus, every piece of
data in 0G can be indexed conveniently with a universal offset. This offset
will be used to sample challenges in the mining 프로젝트로 돌아가기