The demand for AI content tools is exploding. But without fundamental change to how computing power is distributed and paid for, this tidal wave of innovation will take place behind closed doors.
Decentralization through open job marketplaces is a powerful alternative to the current AI bottleneck that can drastically lower prices, boost profitability for server providers and create a dynamic AI market through competition.
Let’s look at why AI compute is so expensive and what can be done about it.
Why AI compute is so expensive
The dawn of the AI era created a global gold rush for compute. While this is great for the handful of companies that enjoy a monopoly on computer clusters, it’s a disaster for innovation and accessibility of AI tools for small startups, independent developers, academic researchers, and other non-corporate actors . It’s also a recipe for sky high prices.
The world’s most famous chatbot, OpenAI’s ChatGPT, reportedly costs USD$700,000 per day to operate. And that’s just generating text. The most common generative AI media jobs, such as text-to-image, image-to-image and image-to-video require orders of magnitude more compute to generate. In fact, generating just 60 seconds of video could take up to 30 minutes of GPU time.
NVIDIA, a $3 trillion company, currently provides a staggering 92% of the GPU compute used in AI data centers, while Microsoft and OpenAI scoop up 69% of the foundational models and platforms market. With the necessary power for the world’s most transformative technology centralized and locked away behind closed doors, there’s no room for competition and little prospect of prices lowering.
Driving down prices with decentralization
Decentralized networks are a radical alternative to the current monopoly on AI-ready compute. These networks use blockchain-based coordination systems to connect geographically disparate GPUs and bring their compute power together in a vast, decentralized network.
Decentralized networks do not rely on servers from any one entity, which allows them to break the monopoly on pricing and access. They are most effectively able to outcompete centralized offerings through access to idle computing power. GPU providers then have extra opportunities to increase the profitability of their hardware, while lowering the cost.
Here’s how an AI video inference workflow could look on a decentralized network:
Select your decentralized GPU network provider of choice.
Scan their network for GPU resources based on location or capacity.
Select a server and pay for the required rental time.
Deploy your project and receive results.
Decentralization of GPU access is an enormous step forward, but a central component of this innovative model is also a potential blocker for lowering costs even further: having to rent a server.
Regardless of what you need done, you’ve still got to rent GPU resources, manage them, deploy software and pay for their running costs even when idle. A better solution must go even further to reduce responsibilities and costs.
Job-based marketplaces: reimagining competitive AI inference
Imagine a scenario in which you didn’t have to think about GPUs or rental responsibility. Instead, you simply pay for what you use, when you use it. No heavy hardware overheads, just the job you submit, on demand.
The good news is that this scenario actually exists. It’s called a decentralized open job marketplace and it’s one of the most attractive options for smaller developers, startups and most actors outside well funded companies. Decentralized open-job marketplaces are also networks of decentralized GPUs, but developers don’t have to reserve time on servers in order to process their AI tasks. Instead, they simply submit their individual AI inference tasks to the network and pay for them individually.
With a pay-per-task model, network users can immediately avoid the prohibitive costs of server rental. But open-job marketplaces also have other cost-saving initiatives baked in. When an AI task is submitted to the network, it doesn’t just go to the first server provider to see it. Hardware providers must then compete on price to complete the tasks as cheaply as possible. Because of this, network orchestrators are incentivised to complete tasks with the smallest possible charge on top of their own electricity costs.
Livepeer’s own AI inference offering is based on this very principle. Livepeer is a decentralized video processing network, dedicated to building the world’s open video infrastructure for the $100 billion streaming market. With ready access to idle compute power across thousands of GPUs in the network, Livepeer is uniquely placed to solve the AI video compute bottleneck and bring down costs for AI video.
Livepeer’s network is open access and the infrastructure providers compete to offer the services you need for the lowest price. All you have to do is submit a job to Livepeer’s network of GPUs and it will be done as cost effectively as possible, scaling automatically to meet your needs.
Here’s how an AI video inference job could look on Livepeer:
A user of an AI enabled application built on Livepeer takes an action that requires an image-to-image, image-to-video, video upscaling or subtitle generation job. The app submits the job to the network.
The job is then routed through the Livepeer AI gateway and picked up by a network orchestrator running a GPU for the lowest possible price.
The job is executed, resubmitted and the result is delivered back to the user..
With a pay-per-task model, the prohibitive overheads of GPU reservations for AI video generation are completely removed. Server providers are able to maximise the productivity and profitability of GPUs already running on the network, while users can make the most of idle servers to generate AI video at a fraction of the usual cost. It’s a win-win situation.
That’s not to say there aren’t a few tradeoffs that come along with a pay-per-job model rather than a rent-a-server model. Decentralized open job marketplaces favor popular AI models, which means that the network needs a lot of providers for the most common job types for it to be cost-effective.
AI models need to be open so everyone can access, run and compete on them. Proprietary models could be run once contractual agreements are set up with operators, but that’s not so different from using cloud-based services. Niche, custom or proprietary workflows may still need to use a server-rental marketplace.
Competition must prevail for an affordable, accessible AI
Without fundamental change, the development of AI will take place behind closed doors by the world’s most powerful and unaccountable companies. AI’s potential for transformation is hard to understate, yet its successful implementation requires open access and affordability. This will not happen without a systematic dissembling of the structures that allow them to lock out competition and keep prices high.
Decentralized, open marketplaces can break the AI computing bottleneck and incentivise GPU providers to compete on price. Livepeer’s network is already creating the world’s open video infrastructure that will support the next generation of video, cheaper and more efficiently than anywhere else. If you want to see a better, more affordable AI for everyone, come and join us.
Click here to boost your GPU profitability with Livepeer.
If you’re a video developer, check out Livepeer Studio for a full suite of streaming and AI services.
If you’re a developer tinkering with generative AI media, start experimenting Livepeer AI now.
Livepeer is a video infrastructure network for live and on-demand streaming. It has integrated AI Video Compute capabilities (Livepeer AI) by harnessing its massive GPU network and plans to become a key infrastructure player in video-focused AI computing.Twitter | Discord | Website