Research: AO, a F**King Big Computer

Earlier this month, SpaceX successfully launched Starship flight 4.

It’s the most powerful rocket ever flown.

This engineering marvel boasts a Super Heavy booster, which houses 33 Raptor engines, generating a total thrust of 72 MN. To put that in perspective, that’s twice the thrust of the Apollo mission rocket and three times that of a NASA space shuttle.

I’m always captivated by the evolution of deep tech companies. Years of relentless effort and millions in R&D culminate in what appears to be an overnight success.

But insiders know that numerous low-probability events must align perfectly for a deep tech company to succeed. SpaceX, founded 22 years ago as a scrappy startup with big dreams, exemplifies this journey—an outcome of compounded hard work and iterative progress.

Such monumental success is rare in crypto.

Only a handful of crypto companies from the 2017-2018 wave during the ICO boom remain relevant today. We’re a young, nascent industry with very short attention spans, and this mindset often bleeds into how founders think about their companies.

That’s why it’s heartening to see an OG project elevate to the next level, becoming one of the most exciting launches of the year.

Arweave & AO Computer — Closely Intertwined

Launched in 2018, Arweave is decentralized permanent data storage. Despite the fundamental importance of data storage, it has struggled to find a compelling use case and adoption.

Data storage isn’t exactly sexy. It’s a commodity.

AO computer is Arweave’s most ambitious project.

AR = storage layer.

AO = compute layer

AO aims to be a scalable, secure, trustless and customizable computing platform—a hyper-parallel computer.

The big idea is that deterministic virtual machines + permanent data storage = reproducible state. Envisioned four years ago, this has materialized into AO’s actor-oriented computing environment.

I want to emphasize that AO computer vastly differs from most other blockchains today. Its existence is enabled due to the foundational work done by Sam (founder) and the Arweave team over the years. Like deep tech startups need a solid groundwork to succeed, AO wouldn’t be possible without Arweave’s secure decentralized storage network.

And I’m rooting for them to succeed.

Not Another Ethereum or Solana

Ethereum and Solana are global synchronous state machines with a shared global state.

AO computer is a shared global messaging machine with localized states

Shared global state is the default mode for blockchains today. It enables consensus and trust among participants, maintaining the integrity of the ledger. However, this falls short of supporting AI applications of the future.

The AO computer framework differentiates itself from Ethereum and Solana through two primary features:

(1) scalable compute capabilities and

(2) a flexible security model.

Scaling Compute

In Ethereum and Solana, adding more compute resources (new nodes) does not enhance the network’s output. Ethereum still processes 12 - 15 transactions per second, no matter how many validator nodes there are.

However, the AO computer framework offers a dynamic approach to scaling. The computational output scales correspondingly as additional compute resources are integrated into the network. This is because AO processes run in parallel locally, unhindered by the limitations of a global state.

Security

Ethereum employs a uniform security model where every transaction is protected equally within its Proof of Stake (PoS) network. While this ensures consistency, it can be inefficient and expensive for low-value transactions, such as gaming.

The AO computer framework introduces a flexible security model, allowing developers to tailor security requirements to the specific needs of their processes.

A F**King Big Computer

Imagine you have a giant supercomputer made up of many smaller computers spread all over the world.

This big, connected computer can do lots of things at once, like running games and apps. The smaller computers talk to each other using a messaging system. The best part is that each small computer isn’t slowed down by the rest of the network but still gets the security and trust of a blockchain.

This is what the AO computer is, in a nutshell.

The AO computer framework modifies typical blockchain node architecture by decomposing it into smaller, modular components. These components include:

1. Processes

2. Messaging Units

3. Scheduler Units

4. Compute Units

Each element plays a critical role in the system’s overall functionality (the docs and whitepaper provide full details).

At a high level, AO can be distilled down to three core concepts:

1. Multiple processes that can run in parallel.

2. Processes maintain individual sovereignty.

3. Asynchronous messaging that allows for coordination & communication between processes

Parallel Processing

At the heart of the AO framework are processes, which are applications on the platform. These processes run independently and in parallel. Individual processes have no visibility into the local state of other processes.

This isolation allows for complex computations on powerful machines, expanding the design space far beyond the limitations imposed by conventional smart contract protocols (fees, gas limits, block sizes).

Compute Units play a critical role here. They provide the necessary computing power to keep AO processes running efficiently.

Application Sovereignty

A distinguishing feature of AO is the sovereignty it grants to developers.

Each process operates autonomously, allowing developers to select their computational tasks, virtual machines, and security parameters without interference from other processes.

Developers can adjust security measures based on the value and sensitivity of specific transactions or computations.

Asynchronous Messaging

This is the glue that holds the network together. Communication within the AO protocol is managed through asynchronous messaging.

Processes exchange messages to coordinate actions and receive the necessary data. All applications adhere to a universal messaging standard, ensuring consistency and interoperability.

Messaging Units relay these messages, working in concert with Scheduler Units—akin to rollup sequencers—which order messages and record them on Arweave.

AO’s Trillion-Dollar Opportunity is in AI

Running a large, multi-billion parameter AI model like LLama-3 directly on Ethereum is impossible.

• Computational Requirements, Data Storage & Bandwidth: LLaMA-3 models need significant computational resources (RAM, GPU power). Ethereum is not optimized for the large-scale data storage and high bandwidth necessary for LLM inferences.

• Latency and Performance: Ethereum’s transaction speed hinders the low-latency, high-throughput environment required for effectively using AI models. No user wants to wait 12 seconds (Ethereum’s block time) for a result.

• Costs: AI inference on-chain is very expensive. Period.

So, I was very excited when AO announced last week that large language models (LLMs) could run on smart contracts, bringing intelligence directly into the blockchain. This would integrate human-like decision-making within a trustless network.

Memory

Running an AI model requires loading the model parameters into the memory. The more parameters there are, the larger the memory requirements.

Running a relatively smaller model like LLama-3-8B will require at least 12GB of RAM. GPT-4 has 1.76 trillion+ parameters. GPT-5 is expected to have 50 trillion parameters (!!). AI models are only going to get larger.

Blockchains today are very limited, memory-wise. At a protocol level:

• Ethereum has 48KB of RAM when executing smart contracts

• Solana has 10MB of RAM

• Even Internet Computer (ICP), which is making a push into AI, has ~3B of RAM and has a limited capability for AI applications.

AO uses WebAssembly in its execution environment and now supports WASM64, enabling high-performance applications. Today, AO has a 16GB memory limit, with a ceiling of 18 exabytes at the protocol level. This enables it to run heavy computations, such as inference on Llama-3 or Phi 3 (unquantized)

Data Storage

WeaveDrive is a new feature that enables AO applications to access all of Arweave’s data like a local hard drive.

By functioning as a local file system within a smart contract, WeaveDrive enhances the efficiency and accessibility of data storage and retrieval. More dApps will be incentivised to upload and store data on Arweave.

AO Computer’s framework enables the direct on-chain execution of large AI models by leveraging the full computational power of individual machines.

By combining an unbounded execution environment with efficient and accessible data storage, AO significantly expands the design space for on-chain applications, especially those that are compute-intensive and AI-driven. This facilitates the development of autonomous agents with smart contract-like guarantees.

Integrating cron jobs—a feature not natively available on most blockchains—and autonomous agents will significantly boost on-chain activity, fostering more sophisticated and dynamic interactions.

Many people will underestimate how important this breakthrough is as we move towards an era of AI ubiquity.

Llamaland serves as a proof-of-concept showcase: an AI-powered MMO game built on the AO computer.

In Llamaland, users can submit petitions to the Llama King, an AI agent that uses a large language model (LLM) to review their petitions and decide how much memecoin to allocate to each user. Remarkably, this entire process is executed fully on-chain, making it the first implementation of an LLM operating entirely on-chain.

Although it is not live yet, you can register on the waitlist.

Building the Permaweb

On 6 Jun, Arweave made 3 major announcements aimed at expanding its distribution:

1. Acquisition of Odysee by Forward Research

2. Launch of AgentFi by Autonomous Finance

3. AO ventures with $35M in funding for builders in AO ecosystem

From my perspective, the most eye-catching one was the Odysee acquisition. It’s under-appreciated by most.

Odysee is the largest Web3 social app that no one has heard of.

It has 7M+ monthly active users (20x MAU of Farcaster) and 2.2M website visits each month. Odysee operates as a more open version of YouTube, leveraging blockchain technology to give creators greater control and freedom. Unlike YouTube, which relies on traditional ad revenue sharing, Odysee rewards creators with LBRY Credits (LBC).

Odysee is like if YouTube was more like Reddit, a darker corner of the internet where freedom of speech is emphasized and content is less filtered. Its “Wild West” category is particularly notable for hosting more controversial and unfiltered content.

According to SimilarWeb, a significant portion of its audience comes from Europe, especially France, Spain, and Germany. While it will not rival YouTube in size, Odysee still attracts a substantial user base, which could be potential users for AO apps.

Odysee will now be building on Arweave infrastructure, adding its platform to the permaweb and giving creators total control over their content.

Apps on AO

AO is currently operating on its testnet, utilizing a Proof of Authority (PoA) security mechanism. Most applications on the mainnet will adopt a Proof of Stake (PoS) mechanism. The mainnet launch date has yet to be determined.

Over the past three months, the testnet has seen approximately 7,000 daily active users, with 3,000 to 4,000 processes and 182 million messages sent. While these numbers aren’t staggering, they indicate the presence of a small but active community.

Dexi

Dexi autonomously gathers and organizes data on asset prices, swaps, liquidity, and more. Think of Dexi as a blockchain-based Bloomberg. It is managed by a network of autonomous agents and hosted on Arweave, connected to all available liquidity pools.

Dexi is permissionless, pulling data directly from the Dexi aggregation agent. This setup ensures censorship resistance, with all information verifiable on-chain.

0rbit

0rbit is a decentralized oracle network with two main components: an AO module and a network of nodes that fetch data from anywhere on the internet. This is critical infrastructure for useful applications to be built on AO.

The AO module interacts with 0rbit nodes to send data requests and receive results. Anyone can run a node and contribute to the network. Data is fetched from the internet asynchronously, meaning the process does not have to wait for a response.

Currently, 0rbit can fetch news from user-defined website APIs and price feeds from Coingecko’s API.

During my tinkering, I found a few other interesting things on AO, including AMMs (Bark, Permaswap, ArSwap), Trunk (an AO meme coin), and Astro, an over-collateralised stablecoin using AR as collateral. It is still early for AO, with many projects yet to be tested in production or see significant transactions and volume.

Important Read: AO Tokenomics

The AO token went live on 14 June, with minting retrospectively dated from 27 February 2024. I like how it was launched fairly, with no investor or team allocations.

The minting schedule mirrors Bitcoin’s distribution model with no pre-mint and 21M total tokens. The rate of newly minted tokens decreases by half every 4 years, although this is smoothened rather than through a discrete halving event.

New AO tokens will be split between:

• Arweave token holders (36%)

• Bridgers of staked assets like stETH, SOL (64%)

So far, ~1M tokens have been minted, most of which have been retrospectively distributed to AR token holders. AR token holders will continue to earn AO tokens every 5 minutes. AOCRED tokens from the testnet can be converted to AO at a 1000:1 ratio.

Importantly: AO tokens will not be transferable or tradable until 15% of the total supply has been minted (approximately 8 February 2025). This gives time for the fledging ecosystem to develop and for mainnet to launch, where AO tokens will be used to secure the network.

Mining AO by holding AR

I did some calculations on the expected return for each AR token. Calculations here assume 1 AR = $30.

• 1 AR token holder will receive ~0.016 AO tokens over the next 12 months.

• There will be ~4.2M AO in circulation in 12 months time

The APYs for holding AR are modest, ranging from 2.5% - 50% depending on expected FDV, which in turn depends on market conditions when the token becomes tradable next year. Assuming market sentiment is good in 2025, a $10B FDV could be quite reasonable, implying an APY of 25%.

Mining AO by bridging assets e.g. stETH

Holding 0.01% of the bridge Total Value Locked (TVL) for 12 months will accrue approximately 210 AO tokens.

As of June 25, there is ~$320 million of bridged stETH mining AO, leading to an exceptionally juicy APY of 33% - 600% depending on TVL & FDV. The estimated APY is volatile and heavily influenced by the total value of assets bridged (TVL) and expected token FDV.

This makes for a very attractive, low-risk yield opportunity for anyone holding stETH today.

The APY will decrease as more users become aware of this and bridge their stETH. The idea of bridging stETH to mine tokens is a win-win situation for both the AO team and its users.

Substantial revenue is generated from the bridged stETH (~$30M/year for $1B in deposits), which will fund ecosystem projects and other development activities.

Monitoring the TVL of bridged assets will be crucial, as it will likely form an equilibrium with the AR token price. If the TVL remains relatively low, more people may sell AR to farm using staked assets, and vice versa.

The parent token, AR, has a total supply of 66 million tokens, with 65 million in circulation. AR has potential disinflationary properties because AR spent on data storage goes into the endowment fund, which will release AR tokens only when mining becomes unprofitable.

How I’m Thinking about AO

When considering whether a protocol will succeed, I like to break it down into smaller bets.

I ask myself: What are the most important things that need to go right for AO to achieve Solana-like success?

1. High demand for AI Agents and Applications in Crypto: Given the rapid advancements in AI over the past two years, this seems to be the current trajectory. Offchain computations using techniques like zkML/OpML are being developed on other blockchains, but only AO supports full AI computation directly on-chain today.

2. Composability with Major Blockchains: AO must integrate closely with other major blockchains. Expecting existing developers and applications to migrate entirely to AO is unrealistic. Therefore, building bridges and communication pipelines with existing ecosystems will be crucial.

3. Developing the Social Layer: Ultimately, the success of a blockchain is not about transaction speed or low fees. Cultivating a passionate developer community and engaging users is key to widespread adoption. Monitoring developer activity and social media sentiment can serve as proxy metrics.

I will be watching these three aspects very closely and adjusting my level of conviction accordingly.

Other thoughts

• AO computer approaches decentralised computing in a way that is quite different from traditional blockchains. It could be a hit or miss. But the ability to perform large computations on-chain is a major unlock in the age of AI

• AO uses an actor-oriented approach. Such platforms are used in systems that require high concurrency, modularity, and scalability. While the actor model offers theoretical advantages, its widespread adoption and success at a large scale is still yet to be proven. Challenges include handling unexpected failures and maintaining consistency across distributed actors.

• AO is extremely nascent. While tinkering on testnet, I found users mostly experimenting with text-based chats and games. Tooling needs to be built out, which takes time. Due to the different architecture and language used, it’s not easy to port apps directly from other chains. Overall, AO feels like Solana in its early 2021 days, when it had only a few applications and a lot of scepticism.

• Long-term conviction necessary: AO requires a high level of conviction. Short-term traders may struggle to find immediate catalysts since the token will only be tradable in 2025.

• Demand for Arweave: AO will create consistent demand for AR, acting as both a growth catalyst and a new platform for app development. Every process on AO will be writing data on Arweave and driving utility for AR.

AI and crypto are undoubtedly the two most significant technology paradigm shifts of this decade. The intersection of AI and crypto will be transformative. In the future, we will likely look back and wonder why we didn’t pay more attention to this.

My gut tells me that AO will be a key player leading the charge forward.

Thanks for reading! If you have any feedback or comments, please feel free to reply to this email or DM me on Twitter.

Best regards,

Teng Yan

This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.

References

• AO stETH deposit breakdown

• Official AO computer Whitepaper

• Economics of AO

• Messari’s research report on AO

• AO computer testnet metrics