Snapshot: Gensyn (opens new window) is a machine learning compute network. It allows developers to train deep learning models over a network of connected devices, from data centers with excess capacity to personal laptops with latent GPUs. Gensyn connects these devices into a single, virtual cluster that developers can access on-demand and peer-to-peer. This dramatically expands the global supply of ML compute, bringing down costs for developers and offering greater capacity than traditional cloud alternatives.
The impact: Access to compute is the biggest bottleneck in AI today. Developers are unable to access GPUs to train deep learning models, either directly or from large cloud providers. Meanwhile, spending on compute for training is estimated to reach 1% of US GDP by 2031 (opens new window). Gensyn solves this problem by unlocking latent compute sources and uniting them into a single cluster that can run ML workloads at scale. This reduces costs and offers greater scale in model training at comparable levels of performance. And because Gensyn is an open network - not reliant on any single intermediary - it remains accessible to anyone who wants to use it, including those building open source AI models and applications.
On joining PLN: In June 2023, Gensyn raised $43 million in a Series A funding round led by a16z and a group of investors, including Protocol Labs. The startup joined Protocol Labs network later that month.
What’s on the radar: Gensyn has been primarily research-focused up until Q2 of 2023, so the latest funding round will go towards accelerating the deployment of its network – including its first testnet release later this year.
# How it started: Expanding low-cost access to compute
How can we expand access to high-performance compute? This is the key question that brought the co-founders of Gensyn, Harry Grieve (opens new window) and Ben Fielding (opens new window), together in the accelerator program Entrepreneur First (opens new window) in early 2020. By summer, Gensyn was officially born. The co-founders had a very specific, shared vision: an open marketplace that could connect developers with low-cost compute to train their models.
In designing this system, Gensyn decided to focus on machine learning training, which involves repeatedly exposing the model to relevant data (and making slight updates to its internal “weights”) until it produces acceptable outputs. This process is computationally intensive, and requires a massive amount of specialized computing power in order to produce a properly trained model. Research (opens new window) shows the GPU rental market is dominated by the $677 billion cloud compute market that charges high margins. This makes it difficult for startups and smaller players, typically with three or four GPUs, to successfully enter the space – let alone compete.
Gensyn determined that they could increase the supply of ML compute by tapping into various latent sources, including smaller data centers, personal laptops and even smart phones. If these could be linked together and utilized as a single cluster (from the developer’s perspective), they could offer a real alternative to the existing cloud-based model.
“Machine learning is revolutionizing the way we work and the way we live, but the biggest roadblock to continuing this revolution is access to compute power,” said Grieve. “It's very expensive to train models, so training has become a bottleneck to advancements in the AI and machine learning space.”
# From cost effective compute to accurate machine learning models
Gensyn allows any type of hardware to supply compute, so long as they can perform the relevant workloads. While this allows the network to achieve maximal scale, it introduces novel challenges, including how to verify that computations were performed correctly. To solve this, Gensyn is designing a mechanism that includes both cryptographic and game theoretic components to ensure proper verification.
In June 2023, Gensyn raised a $43 million Series A funding round from a group of investors, including a16z, Protocol Labs, CoinFund, Canonical Crypto, Eden Block, Maven 11, and angel investors. This brings the total amount of funds raised by Gensyn to over $50 million to date, which will go towards expanding its workforce, introducing its protocol and finding more ways to make AI infrastructure more accessible.
“We’re in the middle of a machine learning gold rush with new entrants totally rethinking the way we do model training – and incentivised to do so,” said Grieve.
# The future: A new, accessible machine learning training protocol
Through knowledge-sharing with Protocol Labs, Gensyn has found inspiration in the classic architecture of IPFS’s Dedicated Gateways (opens new window) as a method of whitelisting content in a permissionless network. Grieve believes that value accrues easily via decentralized networks, due to natural supply and demand. This method will also boost computing power by leveraging underutilized hardware globally.
Gensyn is preparing to launch its first testnest by the end of 2023.
“We are in a new industrial revolution, this time driven by intelligence and powered by compute. By cutting out the margins and the middlemen, we can build larger models at lower prices. This means the world will look very different in the next five years,” said Grieve.