The decentralized cloud platform that wants to disrupt AWS
AI + DePin + Solana
Welcome to the alpha please newsletter.
Today I’m bringing you another interview with a project that I am immensely bullish on.
It spans a few verticals that are very hot this cycle: AI + DePin + Solana.
If you read my last piece on crypto x AI then you might be familiar with it.
io.net Cloud is a state-of-the-art decentralized computing network that allows machine learning engineers to access distributed Cloud clusters at a small fraction of the cost of comparable centralized services.
I chatted with COO Tory Green to get the alpha.
The IO token will be launching on Solana relatively soon, I strongly urge you to read this one. I will also include information on how to land the airdrop (at the end of the piece).
I am a private investor in io.net, and am a big believer in their platform, as their GPU clustering solution is truly unique.
io.net TL;DR
Decentralized AWS for ML training on GPUs.
Instant, permissionless access to a global network of GPUs and CPUs which is live right now.
They have 25,000 nodes.
Revolutionary tech that allows for the cloud clustering of GPUs together.
Can save large scale AI start ups 90% on their compute costs.
Integrated Render and Filecoin.
Built on Solana.
They just announced their $30M raise with a whole host of the biggest backers in the space.
Why should people pay attention to io.net?
We're not just competing with other crypto projects; we're competing with the cloud. One of our major offerings to customers is significantly lower prices, up to 90% cheaper. What we're truly offering is consumer choice, which is where it gets really interesting. Yes, you can get GPUs for 90% cheaper on our platform, which I highly recommend trying out. You can access cheap consumer GPUs, fully decentralized, with low speed, and at a significant discount. However, if you need high performance, you can recreate what you might get from AWS with top-of-the-line hardware like A100s, perhaps only 30% cheaper, but still less expensive than AWS. In some cases, we've even offered better performance than AWS, which could be crucial for specific industries like hedge funds.
For one of our main customers, we're offering a 70% discount compared to AWS, and about 40% off what they would get elsewhere. Our platform is user-friendly and permissionless, unlike AWS, which may require details like a business plan. Anyone can come on and spin up a cluster instantly, whereas with AWS, it could take days or weeks.
Comparing us to decentralized players, if you try to get a cluster on platforms like Akash, you'll find it's not instant. They're more like a travel agent, calling up their data centers to find available GPUs, which can take weeks. With us, it's instant, cheaper, and permissionless. We want to embody the Web3 ethos, and at the same time take down AWS and GCP.
What does your roadmap look like for 2024?
Well, there's the business roadmap and the technical roadmap. From a business perspective, TGE is approaching soon. We plan to have a summit this year where we'll announce a lot of product-related stuff. A lot of our focus is on continuing to build the network because, despite all the TGE excitement, we see ourselves as a real business and a legitimate competitor to AWS.
We're going to continue to heavily build out our sales team. We want to follow in the footsteps of the Chainlinks and Polygons of the world, and focus on hiring high-level sales executives from companies like Amazon and Google to form a world-class sales team. This will help us attract AI customers and get partnerships with entities like Hugging Face and Prettybase.
Our initial customer base is large AI startups who are facing significant AI compute costs. I'm part of a tech CFO group in the Bay Area, and one of the biggest concerns is the high cost of AI compute. There's a Series A SaaS startup spending $700,000 a month on AI compute, which is unsustainable. Our goal is to significantly reduce these costs for people like them.
Once we have proof of concept with these initial customers, we'll look into adjacent markets. With SOC two compliant GPUs on our network, we could target big tech companies or enterprises like JP Morgan or Procter & Gamble who surely have their own internal AI divisions. Our technology supports clusters of up to 500,000 GPUs, potentially allowing us to surpass AWS or GCP in capacity because they can’t physically have this many units in one location. This could attract major AI projects like OpenAI for future versions of GPT. However, building a marketplace requires balancing supply and demand; we currently have 25,000 GPUs in the network with a waitlist of 200,000. Over time, we aim to expand our network to meet the growing demand. So that's the business roadmap.
From a tech side, obviously, there's a lot to it. Right now, we support Ray, and we're also supporting Kubernetes, which we're actively working on. But as I mentioned, we're thinking about expanding our offerings. If you think about how AI works, for instance, when you use ChatGPT, that's the app, and ChatGPT is built on top of a model, which is GPT-3, and GPT-3 runs all its inferences on GPUs. We could eventually build that entire stack, starting with GPUs.
We're also partnered with Filecoin, and many of these partner data centers already have tons of CPUs and storage, so we could also start storing models. This would allow us to offer compute, model storage, and an SDK for building apps, creating a fully decentralized AI ecosystem, almost like a decentralized App Store.
What role will the token play in the network?
At a high level, it's a utility token that will be used to pay for compute on the network. That's the simplest explanation. I would also recommend checking out the website bc8.ai.
This is a proof of concept we built, a stable diffusion clone, and I believe it's the only fully on-chain AI dApp that exists right now. Users can create images with crypto payments, using Solana for microtransactions. Every transaction compensates the four key stakeholders involved in creating an image: the app creator, the model creator, us, and the GPUs used are all paid. Currently, we let people use it for free since we're both the app and model owner, but this is more a proof of concept than an actual business.
We plan to expand the network to allow others to host models and build fully decentralized AI apps. The IO token will power not just our models but any created model. Tokenomics are still being finalized and likely won't be announced until around April.
What made you choose Solana?
I think there's two reasons. One, we really liked the community, and two, quite frankly, it was the only blockchain that could support us. If you look at our cost analysis, every time someone does an inference, it's about five transactions. You have the inference, and then it's paying all the stakeholders. So when we did our cost analysis and had 60, 70, 100,000 transactions, it was necessary for all these transactions to be 1/100 of a cent or 1/10 of a cent. With the volume we do, Solana was really the only choice in our mind. Then the other thing too is, they've been so helpful as partners and the community is so strong. It was kind of a no-brainer.
What would you say is the size of your market?
I think it's intangible, you know. We've thrown around figures like a trillion, but even then, it's hard to grasp the full extent. For instance, there are forecasts by firms like Gartner suggesting that model training could account for 1% of GDP by 2030, which translates to about $300 billion. This statistic is relatively easy to find. However, when you consider statements from Nvidia's CEO, who mentioned that only 10% of the AI market would be for training, it shifts the perspective. If inferencing and training together represent a $300 billion market, then the entire AI GPU market, just for compute services, could be a $3 trillion market. And then you have projections like those from Kathy Wood, suggesting that AI as a whole could be an $80 trillion market. This shows the market's potential size is almost beyond comprehension.
What would you say are the biggest obstacles for io.net going forward?
Building a marketplace is tough, and while it might be a bit easier in crypto, it still comes with its challenges. For instance, most of our customers demand A100s, which are top-of-the-line, enterprise-grade GPUs costing around $30,000 each and are in short supply. They are very very hard to find right now. Our sales team is busting their ass to source these GPUs, which is a significant challenge due to their scarcity and high cost.
We also have a lot of 3090s that are more of a consumer product, and aren't in as high demand. This means we have to adjust our strategy and find customers specifically looking for these types of GPUs. You’re going to find this in any marketplace though, but we're tackling it by hiring the right people and implementing effective marketing strategies.
From a strategic point of view, the fear is, as I mentioned, we are currently the only platform capable of building decentralized clusters across different geographical locations. This right now is our competitive moat. In the short term, we have a significant competitive advantage, and I believe this extends into the medium term as well. For our collaborators like Render, it wouldn't make sense to try and replicate our model when they can just leverage our network and retain 95% of the value.
It took Ahmad and the team about two years to develop this capability. So it isn’t necessarily an easy thing to do. There's always the possibility though that someone else might figure this out in the future. By then, we hope to have built enough of a moat. We're already the largest decentralized GPU network by orders of magnitude, we have 25,000 GPUs compared to Akash's 300 and Render's couple of thousand.
Our goal is to reach a point where we have 100,000 GPUs and 500 customers, creating a network effect similar to Facebook's, where the question becomes, "Where else would you go?" We aim to become the go-to platform for anyone needing GPU compute, similar to how Uber dominates ride-sharing or Airbnb in accommodations. The key is to move quickly to secure our position in the market and become synonymous with decentralized GPU compute.
How to land the IO airdrop
There are two main ways to qualify for the IO drop:
They have a Galaxe campaign running called Ignition. You simply need to complete the tasks. It will require you to prove you are human by minting a galaxe passport, which is good because this can’t be sybilled.
Supply your GPU/CPU to io.net. Simply follow the instructions as stated in the docs. It can be done in about 10-15 minutes. It is quite simple even if you are non-technical.
And that’s your alpha.
Not financial or tax advice. This newsletter is strictly educational and is not investment advice or a solicitation to buy or sell any assets or to make any financial decisions. Crypto currencies are very risky assets and you can lose all of your money. Do your own research.
Thanks man, genuine alpha.