<\/p>\n
The cloud.<\/p>\n
It\u2019s a fabric that connects people in ways that were impossible just a few years ago \u2014 and we all pay for it.<\/p>\n
\u201cAll of us pay a \u2018cloud tax,\u2019\u201d Greg Osuri says. \u201cWe just don\u2019t realize it.\u201d According to Osuri, 50% of online subscription fees, for example, end up in the pockets of cloud providers.<\/p>\n
On the 0xResearch podcast (Spotify<\/a>\/Apple<\/a>), the founder and CEO of Overclock Labs and core contributor to Akash says that the cloud \u201ckeeps a modern society together.\u201d<\/p>\n Cloud technology brings incredible value by enabling \u201camazing services for us and storing our data\u201d but users have little to no control over this \u201cinvisible layer\u201d of society, he says.<\/p>\n Osuri explains that this lack of individual control results in \u201cuncontrollable costs, selective innovation, sometimes extremely anti-competitive practices and essentially, an oligopoly\u201d that persists behind the scenes.<\/p>\n \u201cThat\u2019s the current state of the cloud \u2014 an extremely important layer that is extremely opaque.\u201d<\/p>\n The goal of Akash, a decentralized computing network built on Cosmos<\/a>, is to \u201creverse that opacity,\u201d Osuri says. It seeks to create an \u201copen, transparent, permissionless layer [so] that all of us have a degree of ownership and a degree of sovereignty.\u201d<\/p>\n Akash acts as a marketplace that matches users with providers who might otherwise be sitting on idle computing power. Osuri explains that, at its core, it is a \u201csuper cloud,\u201d that aggregates computing resources across a range of public and private providers and offers \u201ca federated access point.\u201d<\/p>\n Users get access to providers by leasing services in an \u201copen setting,\u201d he says, typically using a reverse auction mechanism. \u201cYou specify your needs and you specify what you\u2019re willing to pay and providers bid on it.\u201d<\/p>\n One key advantage of Akash, Osuri points out, is that \u201cyou have a degree of concentration of computational power where [it] has been traditionally hyper-fragmented.\u201d<\/p>\n As of now, Akash is experiencing \u201cnear 100% utilization for high density GPUs<\/a>,\u201d Osuri says, \u201cand about 50% utilization cumulative for all resources, all types of GPUs, including the low end GPUs.\u201d\u00a0<\/p>\n That\u2019s quite a bit higher utilization than pretty much anything else out there in crypto, Osuri says. \u201cTraditionally, crypto has been good at oversupply and less demand. Here, we have the opposite problem: high demand and less supply, or supply that\u2019s not yet on the market.\u201d<\/p>\n Most demand for computing power right now is for high-end H100 chips and \u201cdata center grade homogenous hardware,\u201d Osuri says \u2014 equipment that powers AI<\/a> apps like ChatGPT.\u00a0<\/p>\n \u201cEvery time you hit a prompt in ChatGPT<\/a>, the time it takes to come back, it requires one H100 per concurrent user for that time,\u201d Osuri explains. \u201cNow, do the math. 100 million users. All of them at the same time.\u201d<\/p>\n \u201cThat\u2019s impossible,\u201d he says. \u201cThey have to prioritize the H100 time based on how many concurrent users they want to support and that dictates how fast the application works.\u201d<\/p>\n \u201cMore users, less chips means slower. More chips, less users means faster.\u201d<\/p>\n Osuri mentions one team that is working with a broader mix of chipsets for \u201cheterogeneous\u201d computing power. \u201cYou can train on a 4090 and H100 at the same time, leveraging the compute power from each of those chips,\u201d he says.<\/p>\n \u201cThere are lots of teams working on it,\u201d he says, adding that at some point in time in the near future, idle home gaming PCs and consoles might even be able to provide computing power to the marketplace.<\/p>\n As of now, Osuri says that most demand for the Akash service comes from machine learning developers looking for on-demand A100, H100 or A6000 chipsets.\u00a0<\/p>\n \u201cThere is no place right now on the cloud where you can get on-demand access to 100\u2019s,\u201d he says. \u201cAmazon, Google, Microsoft do not have A100s or H100s.\u201d<\/p>\n Read more:<\/strong> Don\u2019t give your life to Big Tech for free<\/a><\/p>\n Osuri reiterates that the Akash marketplace operates as a free market, with low demand resources costing around 80% less than they would via traditional services like AWS. On the flip side, high demand services like H100s are more expensive than what one would normally expect to pay.<\/p>\n \u201cWe don\u2019t have any control over price,\u201d he says. \u201cIt is a free market that sets the price.\u201d\u00a0<\/p>\n Osuri explains that Overclock Labs might provide guidance regarding pricing, \u201cbut we don\u2019t dictate what the price should be. And we never will. And that\u2019s critical for the protocol to function.\u201d<\/p>\n \u201cWe need to learn how to deploy these models on a decentralized network versus a centralized network,\u201d he says. \u201cIt\u2019s always been centralized because that was the only option<\/a>.\u201d<\/p>\n \u201cFor the first time, we actually have a viable decentralized option.\u201d<\/p>\n Don\u2019t miss the next big story \u2013 join our\u00a0free daily newsletter<\/a>.<\/strong><\/p>\nA computing power marketplace<\/h2>\n
Mixing things up<\/h2>\n
\n