The emergence of ChatGPT has brought deep learning and GenAI to the forefront of public discourse. Text generation, code analysis, and problem-solving services offered by large language models (LLMs) have provided significant value to both analysts and creatives. This trend has not been lost on VCs — AI ventures received a staggering $42.5 billion in aggregate equity investment in 2023 alone. AI/ML startups have over 40% representation in YCombinator’s latest batch. Unsurprisingly, decade growth rates of 37.3% and 23% for the AI market and AI job market respectively suggest that demand is here to stay.

The Case for Decentralization 

At its core, AI is powered by an increasingly scarce resource: GPU. The ability of GPUs to rapidly process matrix operations in parallel makes them incredibly powerful for ML applications. Since 1950, GPU transistor counts have doubled bi-annualy, following a trend known as Moore’s Law. However, as transistors approach 5 nanometers in diameter, physical constraints are proving this trend impossible to continue. The demand for compute power is far exceeding current supply, forcing costs upward for AI-powered startups and demanding creative solutions. 

There are currently three approaches to handling this scarcity: quantum computing, custom architectures, and parallelization. Unfortunately, quantum computing is in its infancy and the development of custom chip architectures is a game only available to the major players (think Google, Meta, etc.). Therefore, parallelization — the distribution of compute jobs across many chips — is the only feasible solution for small-scale operations. However, as accessible GPU supply continues to decline, chip aggregation is becoming unaffordable. 

Decentralized GPU networks are the logical solution to this problem. Underutilized data centers, desktops, and crypto mining rigs store untapped compute power that can be distributed to AI developers for mutual benefit. This technology creates an efficient market, where developers can access the computing power they need and nothing more, without facing upfront hardware costs. Additionally, it allows hardware owners to passively monetize their unused resources. Clearly, this technology will augment GPU supply, decrease compute costs, and fuel the next wave of AI generation. 

The Emerging Leaders 

Decentralized GPU is an emerging field with no dominant player. Refining parallel cloud-tech and amassing sufficient chip supply has been a recurring problem for all early ventures. Currently, 3 major names are attempting to obtain addressable market share: Render Network, HyperLink, and IO.net. In this report, we will address the varied approaches and unique value propositions offered by each of them. 

Render Network 

Like many of its competitors, Render Network uses blockchain technology to support GPU AI and rendering jobs on the cloud. The firm began targeting artists and media professionals to provide high-speed graphics rendering for mobile media. Render allows studios to produce large-scale 3D effects that would typically require prohibitively costly capital expenditure, as well as to render scenes at maximum resolution for further post-processing and to mint NFTs. However, they have recently made a strategic pivot towards addressing the augmented / virtual reality and ML development markets. Render Network technologies allow for real-time VR/AR streaming as well as prompt-based Image and VideoGen with blockchain-powered digital rights management (DRM). 

Render’s decentralized architecture consists of two layers: an off-chain rendering network and an on-chain payment system backed by proprietary RENDER contracts. The rendering network handles the provision of GPU nodes, graphics production, and transmission back to the client. Currently, GPU nodes are delegated based on a reputation scoring system, which scores the dependability of “Node Operators” to complete jobs, and the specific hardware requirements of each render task. The blockchain network uses public ledgers and the RNDR token to verify that all transactions are processed correctly and allows for easy correction in the case of network failure. 

Render uses a tiered-pricing model. Tier 1 Service, the most premium subscription, uses a separate pool of centralized nodes with the highest VRAM and priority. Tier 2 Service is the most costly decentralized subscription which gives user access to the most powerful parallel nodes. These customers receive a discount to Tier 1 dependent on network conditions. Lastly, Tier 3 users receive the lowest priority nodes with the slowest rate of job completion.

Historically, Render has been regarded as the industry leader and even dubbed the “Web3 NVIDIA.” A recent 387% YoY surge in RNDR token value and their integration with Apple’s M2 processing chip for mobile rendering would suggest that their dominance is continuing. However, it is clear that their lethargic pivot towards prioritizing AI has caused them to suffer. The firm seems to hold an irrational allegiance to a rendering market that is not growing at anywhere near the pace of AI. It is unclear whether this is due to stubborn management, risk aversion, or difficulties in changing their infrastructure, but one thing is certain: it is undoubtedly stiffening their growth and VC interest. The firm has failed to fundraise since 2021 and is quickly falling out of favor with the market. While their user-base has allowed them to hold on to relevance, it is clear that they must shift their operations toward AI compute for them to have any hopes of future success. 

HyperLink 

HyperLink is an emerging player in decentralized computing that has taken a unique marketing approach: targeting the prosumer rather than the developer. HyperLink is building a cloud computing platform marketed as a passive investment strategy as well as a source of GPU supply. Starting with HyperLink Phase 2, the firm boasts potential 500% ROI per annum for compute suppliers. The platform is attractive to prosumers who can purchase GPU resources, download their software, and host their processing power on the cloud to the tune of thousands of dollars per year. Their UX/UI experience is easy, intuitive, and accessible to anybody with an underutilized personal computer. HyperLink successfully targets both sides of the market, avoiding sophisticated jargon and complex onboarding for the supplier and providing raw GPU compute for those in demand. On the supply side, they adeptly market their service as a portfolio diversification tool, connoting the platform as a fixed income investment akin to treasury bonds and high-yield savings. On the demand side, they offer AI-ready GPU, without falling into the Render-pitfall of overly-niche customer concentration. 

However, potentially the most compelling reason to be bullish on HyperLink is in their unit economics. From supplier, to cloud, to consumer, the firm has been able to keep costs shockingly low, offering far lower prices than centralized industry leaders, like Amazon AWS and Microsoft Azure. This has allowed them to maintain impressive margins while still returning market-beating value to passive investors. The main reason for their cost supremacy is their in-house development of proprietary GenAI. HyperLink has been able to produce custom ImageGen, VideoGen, TextGen, SpeechGen, and CodeGen LLMs that are offered to all consumers of their compute power. This has been monumental, as it allows the firm to evade the OpenAI API costs burdening most AI startups leveraging GPT-powered technologies. The firm faces no unit costs for providing GenAI to their clients, providing them with a fortified economic moat. 

For this reason, HyperLink is an emerging power that all investors should monitor in the coming years. Their proprietary technology in combination with their shrewd marketing campaign makes them a powerful force — and they have millions in annual recurring revenue to show for it. 

IO.net 

IO.net is a decentralized GPU network that shares many of the characteristics of its peers. It takes advantage of underutilized compute power from all available sources and uses the Solana-backed IO Token to power its transactions. It is similar to Render in that it targets AI professionals on both the supply side and demand side. IO Tokens are gained by offering GPU uptime, and then used to purchase GPU nodes for personal development. Being Solana-backed, IO token transactions are immensely efficient and carbon-neutral, which will undoubtedly have its advantages in an ESG-focused economy. Its software is sophisticated and technical, and this is reflected by its state-of-the-art network infrastructure. 

Compared to its competitors, IO.net provides the most comprehensive suite of development services. Their decentralized architecture allows for inference workflows to be exported and distributed among parallel nodes, vastly improving classification efficiency via proprietary task delegation algorithms. Additionally, IO offers parallelized hyperparameter tuning, a feature unaddressed by any of their competitors. Hyperparameters refer to AI model “settings”, such as learning rate, train-test split, and batch size, which are manually chosen by developers and remain unchanged throughout training. The IO infrastructure leverages proprietary training libraries which clone training jobs with randomized hyperparameters across parallel GPUs, allowing developers to fine-tune these criteria with ease. Lastly, IO offers built-in support for reinforcement learning with distributed workloads, optimized for parallelization on the cloud. All of these features are analyzed in depth in their thorough and publicly available documentation. 

IO.net claims to be 10x to 20x more efficient than any other current cloud offerings, and this comes at no surprise in the context of their experienced development team and impressive docs. For this reason, IO.net is likely to be the most highly-touted decentralized GPU service by seasoned professionals and AI enthusiasts. It is marketed accordingly as a network-backed programming library rather than an ease-of-use tool for quick jobs. 

When benchmarked against HyperLink and Render, IO.net is likely to hold firm in second place. It is a more comprehensive tool suite than Render Network and is specifically designed to address the budding AI development market. Render will not be able to compete without a significant operational detour and IO.net will likely consume its market share in the coming years. However, its complexity and failure to address the everyday potential supplier will make it difficult for the firm to keep up with HyperLink’s growth. HyperLink is the only major player that has successfully addressed both sides of the market, and this is why they will inevitably emerge as a market leader. While IO’s targeted niche is lucrative and receptive, it is unlikely to achieve mainstream success without the adoption of the common man with an idle NVIDIA GPU.

Summary 

In conclusion, the decentralized GPU market is an essential vertical to monitor from a VC perspective. It is an obvious and prudent solution to a macro supply shock in the processing chip market that is destined for continued growth. Parallelization is the future of AI development, and there is no other technology that can offer such cost-effective and accessible parallel computing. AI ventures will continue to dominate the world of VC and angel investment, and there is no reason to believe that decentralized GPU will not lay the foundations of their resource provision.