.png)
Humanity stands at an AI Oppenheimer moment.
Elon Musk pointed out, “As we advance technologically, it’s crucial to ensure that AI serves the interests of the people, not just those in power. People-owned AI offers a path forward.”
At the intersection with crypto, is where AI can democratize itself. Starting with open source models, then with AI for the people, by the people, of the people. While the goal of web3xAI is noble, its practical adoption depends on its usability and compatibility with the existing AI software stack. This is where IONET’s unique approach and tech stack come in.
IONET’s decentralized Ray Framework is the trojan horse to launch a permissionless AI Compute Marketplace to web3 and beyond.
IONET is leading the charge in bringing GPU abundance. Unlike other generalized compute aggregators, IONET is bridging decentralized compute with the industry-leading AI stack, by rewriting the Ray Framework. This approach paves the way for wider adoption within and beyond web3.
Competition for resources across the AI stack rages on. AI models proliferated over the past few years. Within hours of Llama 3 launch, Mistral and OpenAI released new versions of their frontier AI models [1].
The three layers in the AI stack undergoing resource competition are: 1) Training data, 2) Advanced algorithms, and 3) Compute units. Compute power allows AI models to advance performance by scaling their training data and model size. According to an OpenAI empirical study of transformer-based language models, performance improves smoothly as we increase the amount of compute used for training.
In the past 20 years, compute usage has surged. Epoch.ai’s analysis of 140 models reveals a 4.2x annual increase in training compute for landmark systems since 2010. The latest OpenAI model, GPT-4, needs 66 times more compute than GPT-3 and about 1.2 million times more than GPT.
AI nationalism is evident with significant investments from the US, China, and other countries totaling around $40bn each. Most funds will focus on producing GPUs and AI chip factories. OpenAI's CEO, Sam Altman, aims to raise up to $7 trillion for a venture to enhance global AI chip manufacturing, emphasizing that “compute is going to be the currency of the future.”
Aggregating long-tail compute resources could significantly disrupt the market. Challenges with centralized cloud providers such as AWS, Azure, and GCP include long wait times, limited GPU flexibility, and burdensome long-term contracts, especially for smaller entities and startups.
Underutilized hardware from data centers, crypto miners, and consumer GPUs can meet the demand. A 2022 DeepMind study found training smaller models on more data is often more effective than using the latest, most powerful GPUs, signaling a move towards more efficient AI training using accessible GPUs.
IONET structurally disrupts the global AI compute market. IONET’s globally distributed end-to-end platform for AI training, inference, and fine-tuning aggregates the long tail of GPUs to unlock cheap high-performance training.
IONET aggregates GPUs from data centers, miners and consumers across the world. AI startups can deploy decentralized GPU clusters in a few minutes by specifying the cluster location, hardware type, ML stack (Tensorflow, PyTorch, Kubernetes), and instantly pay on Solana.
GPUs without proper parallelism infrastructure are akin to a reactor without power lines — present but unusable. As emphasized in the OpenAI blog, the limitations on parallelism, both hardware and algorithmic, significantly affect the compute efficiency per model, constraining the model's size and usefulness during training
IONET leverages the Ray framework to transform clusters of thousands of GPUs into a unified instance. This innovation enables IONET to form GPU clusters regardless of their geographical dispersion, addressing a major hurdle in the compute marketplace.
Ray stands out as an open-source unified compute framework, streamlining the scaling of AI and Python workloads. Embraced by industry leaders like Uber, Spotify, LinkedIn, and Netflix, Ray facilitates the integration of AI into their products and services. Microsoft offers clients the opportunity to deploy Ray on Azure, while Google Kubernetes Engine (GKE) simplifies the deployment of OSS ML software by supporting Kubeflow and Ray on its platform.
Ahmad presenting his work on decentralized Ray framework at Ray Summit 2023
We first met Tory, CEO of IONET, while he was the COO of high-growth startups in fintech, and knew that he was a next-level operator with decades-long experience scaling startups to meaningful impact. Speaking with Ahmad and Tory, we immediately realized this was the dream team to bring Decentralized AI Compute to web3 and beyond.
Ahmad's brainchild, IONET, was born from an eureka moment during practical application. Developing Dark Tick, an algorithm for ultra-low-latency high-frequency trading, demanded substantial GPU resources. To navigate costs, Ahmad developed a decentralized version of Ray framework to cluster GPUs from crypto miners, inadvertently forging a resilient infrastructure that tackled broader AI compute challenges.
By leveraging token incentives, as of mid-2024, IONET has onboarded over 100k GPUs and 20K cluster-ready GPUs, including a substantial number of NVIDIA H100 and A100s. Krea.ai is already leveraging io.net’s decentralized cloud services, IO Cloud, to power their AI model inferences. IONET has recently announced partnerships with projects like NavyAI, Synesis One, RapidNode, Ultiverse, Aethir, Flock.io, LeonardoAI, Synthetic AI and many others.
By relying on a globally distributed network of GPUs, IONET can:
IONET stands at the forefront of innovation by decentralizing Ray. Harnessing Ray Core and Ray Serve, their distributed GPU clusters efficiently orchestrate tasks across decentralized GPUs.
The push for open-source AI models is a nod to the collaborative spirit of the original internet, where folks can permissionlessly plug into HTTP and SMTP.
The advent of crowd-sourced GPU networks takes the torch of permissionless spirit in its natural evolution. By crowd-sourcing the long tail of GPUs, IONET is opening the floodgate to valuable compute resources, creating a fair and transparent market that prevents power concentration within a few hands.
We believe in IONET’s vision of AI Compute-as-a-currency via the decentralized Ray clustering technology. In a world increasingly composed of the ‘haves’ and ‘have-nots’, IONET will ultimately “Make Internet Open Again”. 🚀
_________________
LongHash is a crypto-native venture capital firm since 2017. We invest in and build alongside visionary founders forging the next evolution of the open economy. Follow Roy (x.com/0xroylu) and Raghav (x.com/@0xRaghav) for cutting-edge research on data availability and AI.
_________________
[1] “AI race heats up as OpenAI, Google, and Mistral release new models.” The Guardian, https://www.theguardian.com/technology/2024/apr/10/ai-race-heats-up-as-openai-google-and-mistral-release-new-models. Accessed April 19, 2024
[2] “Welcome to the Era of AI Nationalism” Economist https://www.economist.com/business/2024/01/01/welcome-to-the-era-of-ai-nationalism. Accessed April 19, 2024