In this piece, we explore the key pitfalls of centralized compute for generative AI applications and app developers. Just as importantly, we outline a path for businesses to avoid those risks and secure their products through a novel decentralized model that leads to more reliable, cost effective, and scalable compute.

The risks of centralization became all too clear in July, when a faulty Crowdstrike update sparked a worldwide outage for Microsoft clients worldwide. The cost to enterprises was stunning: for example, Delta reported at least a half-billion dollars in losses after having to cancel more than 6,000 flights and reroute hundreds of thousands of passengers — and that doesn’t even take into account the major hits to its reputation and the pending liabilities it faces from angry customers and federal regulatory groups in the coming days.

That failure crippled more than just major companies.  Small businesses from locksmiths in Australia to restaurant owners in New York City were left struggling to access critical files or pay employees. And centralization risk will only get worse if businesses continue to rely on traditional solutions in the age of generative AI.

“We’re reaching an age where a lot of functions that were typically done by humans … are going to be delegated by AI agents. And so these types of issues are going to become more prevalent,” Ifeoma Ajunwa, an Emory University law professor, recently told reporters.

The danger of exposing your business to a single point of failure is just one of the many risks of centralization. The reliance on just a few centralized cloud providers for compute has led to other significant downsides for businesses, including artificially high costs, restrictive lock-in requirements, and crippling bottlenecks — all concerns that are especially pertinent to generative AI companies and developers as they build the applications of the future.

The drawbacks of the web2 cloud for generative AI

Centralized cloud systems, in which just a few companies control access to key compute tools or capacity, are prone to significant drawbacks including …

Centralization risk

• Currently, NVIDIA holds 92% of the market share in the GPU compute used in AI data centers space, while Microsoft and OpenAI hold 69% of the market for foundational AI models and platforms.

• Industry monopolization doesn’t just expose companies to a single point of technological failure, as Microsoft users faced in the wake of the Crowdstrike outage. It also poses significant regulatory and business risks as well: If just one of these major companies fails, or is closed down by a government, all those who rely on it will collapse too.

Extractive agreements

• Because they face little competition, the leaders of centralized industries are able to force vendors to accept unfavorable terms that may lock them into long service agreements with fixed capacity, or may require them to rent servers and reserve space even if they don’t use them.

• The combination of high costs and unideal terms results in significantly less access for startups, academics, or individuals who do not have the buying power of more established competitors.

High costs

• Regulations favor incumbents, limiting access for competitors that become even worse amid supply chain shortages that favor established cloud operators (like AWS or Microsoft Azure) and hardware providers (including NVIDIA and AMD). This leads to consolidated compute in the hands of a few industry actors that can then set higher prices as a result.

Restricted access

• The closed loop systems created by centralized compute can prohibit certain people from accessing generative AI and other resources.

• These permissioned systems mean that legacy cloud giants like Microsoft or market-controlling hardware providers like NVIDIA decide who can use these transformative technologies.

These drawbacks were clear in the past. However, the advent of generative AI has exacerbated them. With such high demand and limited supply, centralized gatekeepers are able to charge even more exorbitant fees while having even less incentive to improve compute quality or agreement terms. And that effect is particularly felt by non-legacy clients and smaller generative AI companies, developers, and researchers.

How decentralization can help generative AI applications

The pitfalls of centralized compute demands a novel solution. A decentralized compute model, one in which generative AI needs are distributed across a broad variety of cloud compute providers, provides a new possibility for open access inference that is more cost effective and reliable, as well as infinitely scalable on demand.

Less risk with decentralized compute providers

• There is no single point of failure, with cloud compute capacity spread out between a large network of providers. This makes it harder for businesses to lose access to services due to a cyber attack, software crash, regulatory crackdown, or any other unforeseen event that can damage a centralized operator’s integrity.

Less fee extraction with more competitive services

• Providers that charge exorbitant fees or prices will be outbid by others, leading to a more efficient market for cloud compute services. A decentralized system gets rid of monopolized services and extractive middlemen, leading to less costly overhead and more savings that can be passed onto users.

Reduced costs with pay-per-task marketplaces

• While there is still a huge demand for generative AI services, there is also an ocean of small, medium, and large-scale cloud providers who are capable of meeting it in a decentralized system.

• Rather than pay for compute capacity they don’t need, companies can operate on a pay-per-task basis, greatly increasing cost efficiency and reducing risk by allowing them to scale or downsize production as they see fit.

More equitable access and contract terms

• Because there is no centralized provider, a decentralized system can offer permissionless cloud compute services on demand, without any long-term service agreement or any single entity dictating who can access services and who cannot.

• That greater flexibility helps level the playing field and expand access for less-established companies, including individuals, startups, academics, and other non-enterprise institutions looking to access generative AI compute services.

Case Study: Decentralized compute for generative AI videos

Many businesses believed cloud computing prices would drop over the years, as new technologies emerged to lower costs. Instead, cloud inflation has been the norm since the arrival of generative AI, with a reported year-over-year increase of 3.7% from early 2023 to 2024 and an upward trend since September 2022.

A number of industries are facing serious ramifications: consider video creation, a sector that AI has transformed by taking a process that used to require dedicated crews of workers and hours of editing and shifting it into tasks that can be performed with just a few text prompts. Although OpenAI’s Sora demos and the top open-source AI video model Stable Diffusion have showcased the industry’s potential, it remains inaccessible to many due to the high cost and limited compute offered by centralized market leaders like Microsoft Azue and Amazon Web Services.

Using a decentralized compute network similar to what Livepeer has built, video developers can start addressing these challenges by leveraging permissionless networks to access low-cost, high-performance processing that is globally available and affordable.

On Livepeer AI, video app developers are able to integrate gen AI features — including text-to-image, image-to-image, and image-to-video conversion — on a per-task basis rather than having to reserve expensive compute capacity they might not use, as they currently have to do with centralized cloud providers.

It is able to do this by tapping into Livepeer’s already-established network transcoding millions of minutes of video weekly, making it one of the largest and most reliable decentralized sources of affordable AI-ready compute.

Developers can scale these AI efforts infinitely, since all cloud compute capacities are performed through a permissionless decentralized system: there’s no need to negotiate new agreements with a single centralized provider, since compute is spread out among as many competing providers that are needed to meet their demand.

Emerging decentralized compute technologies and networks like these will be critical to producers going forward — and if adopted, they could enable a future where AI generative video is the de facto norm for content creators across the globe.

Learn more about Livepeer AI here.

Sign up for the waitlist to stay updated.

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