Skip to content

The CCS Interview

Aethir's Rydon: Web3 won't overtake AWS, but stability is its edge

Mark Rydon is co-founder of Aethir, a decentralized cloud network that supplies enterprise-grade GPU compute for AI and gaming. Recorded in July 2024, shortly after the platform and token launched, the conversation covers the GPU shortage, why Aethir excludes consumer hardware for now, and how Web3 can compete with AWS. Figures and plans are as described by Rydon at the time.

Interview by Ashton Addison Recorded July 2024 5 min read

Key takeaways

  • Aethir addresses GPU scarcity by creating an enterprise-focused decentralized marketplace connecting underutilized data center GPUs with AI companies.
  • Enterprise data center GPUs operate at only 15-18% utilization rates, representing significant latent supply that Aethir aims to unlock.
  • Aethir's network consists of 43,000 enterprise-grade GPUs in data centers, differentiating it from consumer-focused DePIN projects.
  • Large language model training costs are increasing exponentially while GPU supply and chip production cannot keep pace with demand.
  • Aethir prioritizes enterprise service requirements like uptime and reliability over consumer participation, unlike traditional compute DePIN projects.

The GPU squeeze

Ashton Addison: What does Aethir do, and what problem is it solving?▶ 00:26

Mark Rydon: We're an infrastructure project, so it's best to start with the problem. Elon Musk has said GPUs are harder to buy than drugs, and Meta bought just under $11 billion worth of Nvidia's top GPU chip for its Llama 3 ambitions. Startups and small to medium-sized enterprises find it increasingly hard to access this resource. We broke the problem into four pieces. First, buy pressure: OpenAI, Google, Microsoft, Amazon and Anthropic all have an insatiable appetite for GPUs, because more GPUs means training AI quicker. Second, Nvidia can only produce so many chips per year, and that isn't yet enough to meet global demand. Third, LLM complexity: later-generation models are exponentially more expensive to train, in compute, data and cost, but we're not seeing a similar ramp-up in chip supply or performance. Finally, these resources are poorly managed. Enterprise GPUs sitting in US data centers run at roughly 15 to 18% utilization. So we built an enterprise-focused GPU compute marketplace. Data centers, tech companies, enterprises and even crypto mining organizations with latent GPUs can connect to our decentralized cloud network, earn rewards for doing work, and supply the global demand market. Buyers can then access compute that is cheaper and more available than with the current big providers.

Ashton: Is Aethir making a meaningful dent in the gap between compute supply and demand?▶ 15:13

Mark: Short answer: no, not right now. The gap between global compute demand and supply is so immense that unless you're Amazon or Google, you're not making a dent. I'd have to find the exact quote, but Microsoft is undertaking what is the biggest infrastructure project in modern history, deploying more capital to stand up compute data centers than has ever been deployed on any infrastructure project. Estimates are that, with current infrastructure, the world won't be able to produce enough power to satisfy compute demand just five years from now. But I can guarantee that nothing scales quicker in a compute ecosystem than a decentralized compute network. As more compute comes online and we refine our onboarding, I do think we can make a very meaningful dent. Many tech companies buy compute that sits in their own data center. If they get their forecasting wrong, over-buy or delay production, that expensive compute burns a massive hole in their pocket. That's where Aethir has an incredible opportunity to de-risk compute for companies that don't have the business development teams or infrastructure to on-sell it through the cloud.

The biggest thing you need to be optimizing for as a Web3 company is stability, because that's where Web2 companies rip us to shreds on credibility.

Why Aethir is enterprise-only

Ashton: Can individuals with consumer GPUs contribute to the network and earn?▶ 06:02

Mark: Traditionally, compute DePINs have focused on latent consumer compute, such as gaming PCs and idle MacBooks. You can build a network of GPUs and a strong community quickly, because that hardware is a sunk cost to its owners. But it's very hard to sell to enterprise customers. They'll ask how you can guarantee the GPUs aren't switched off at night, or that a home's bandwidth isn't throttled by a Zoom call or Netflix. You struggle to meet enterprise requirements for service, performance and uptime on consumer equipment. So can consumers contribute to Aethir's network? Not today. We've built a 100% enterprise-grade DePIN. We have 42,000 to 43,000 GPUs, all of them in data centers or on enterprise-grade network infrastructure. That has let us onboard the biggest AI companies. We have the largest LLM training contract on the back of a DePIN, with FedML training its own open-source model on our H100s, and the biggest telecommunications company in the world is on our infrastructure. You couldn't do that with consumer GPUs. Do we intend to open the door to consumer compute in future? In principle, yes, maybe within the next six months, we'll see. But it has to be for a meaningful product we can actually sell. Aethir is very much about real business.

Web3 versus the hyperscalers

Ashton: Can Web3 cloud infrastructure take real market share from AWS and traditional cloud?▶ 19:23

Mark: Anyone thinking Web3 will overtake Web2 in any major sector is probably a little bit delusional today. But can we have a very meaningful position? One hundred percent. Compared with AWS, the GPUs are the same, the data centers are the same, and we're linked to the web through the same subsea cables. We don't have special Web3 GPUs. What's different is two levers. First, we can onboard other people's compute, which is an advantage in unit economics and speed to market. Second, we have the token, which can incentivize behavior, subsidize, and move currency across borders into regions where you don't have a presence. The trap many projects fall into is using the token only as a discount, or to heavily subsidize network growth. You attract people with a short-term interest in your network. Then an enterprise buyer asks for a three-year contract and whether all those GPUs will still be there. The biggest thing you need to be optimizing for as a Web3 company is stability, because that's where Web2 companies rip us to shreds on credibility. Use the token to incentivize stability first, then for price and compute incentives. Unless you solve for stability, you won't be around long enough for someone like Amazon to care about you.

Ashton: Crypto markets skew short-term, but infrastructure needs a long horizon. How do you see that?▶ 25:09

Mark: Infrastructure takes time, and we're definitely in this for the long run. That's why we chose to be enterprise-first, even though we knew it would take us longer to get to market. It's why we focus so strongly on real business and on onboarding real revenue and real customers. Five years is going to be a blink of an eye for us. We're going to be here for decades, and I think we'll be part of building some really significant pieces of the AI economy.

Nothing scales quicker in a compute ecosystem than a decentralized compute network.

How to take part

Ashton: What are the ways to get involved if you don't own an enterprise GPU fleet?▶ 27:40

Mark: The node sale is now closed. We sold just over 72,000 Checker nodes and raised around 42,000 ETH, which was anywhere from $125 million to $150 million depending on the ETH price. That's the biggest node sale in Web3 history, and a strong indicator of the power of community. The checkers are effectively the police force of the network. They keep it safe and free from bad actors, and a significant amount of the token supply is distributed to those node operators every year. For people who missed out, there's a staking product. If you want to earn rewards but don't have hardware to contribute, you can contribute your tokens to the staking pools. For those who want the hardware side, there's Aethir Edge, a premium physical device that looks like a short, fat PlayStation 5. You plug it in at home and it passively operates as an enterprise-grade network node. We partnered with Qualcomm, and it uses their latest AI chip. Big tech cares about the edge, meaning compute closer to users, but it's very hard to unify with varied consumer hardware. By selling the same box everywhere, we homogenize that layer and can deliver enterprise workloads to a unified edge network, while owners earn rewards.

This interview has been edited for length and clarity. Recorded July 2024.

Mark Rydon
Mark RydonMark Rydon is co-founder and Chief Strategy Officer of Aethir, a decentralized GPU cloud network for AI and gaming workloads, and is based in Singapore. He co-founded Aethir in August 2021 with Daniel Wang, who serves as CEO. He began his career in civil engineering… Full profile →
Share this interviewXLinkedIn
▶ Watching the interview