Decentralized AI

“There are a lot of people working, very, very smart people working on this stuff, that I have a lot of confidence that by mid-2025, when ICML happens, pay attention to the papers that are being submitted… and I think mid-25, you’re going to see decentralized AI go mainstream, like, no doubt.” — Greg Osuri, 00:38:57

“So we have to, you know, build nuclear reactors in the next two years, which is not going to be reality. So we’re going to have to explore very seriously about distributed training. Right? So, distributed training also means distributed grid. That means you can actually go, instead of concentrating too much compute in a single data center, you distribute that compute all over.” — Greg Osuri, 00:31:27

“I might make another prediction: with the advancement of DeepSeek, I think [the] market is going to realize funding private companies billions of dollars is not the way [to] AGI, versus having a shared ownership model that taps into global liquidity… I have a strong feeling that right now, with DeepSeek, AGI is actually possible on chain versus in a private setting.” — Greg Osuri, 00:43:25

“With DeepSeek it clearly proved that open source AI is going to overtake, and DeepSeek is going to even accelerate decentralization… I think the summer is actually accelerated — maybe here, maybe late winter or early spring.” — Greg Osuri, 00:37:42

“We’re not as fast as China to build out nuclear reactors — they can do it in three years. So I feel like we have a good five-year shot for decentralization to take mainstream.” — Greg Osuri, 00:19:56

“This paradigm shift is a game changer for training, and I think accelerated our AGI timeline significantly, because now we broke the data challenge.” — Greg Osuri, 00:12:50

“The real risk is us as a society going back to feudalism — I call this digital feudalism. There was a time in history that few people had most power in Europe and peasants had to depend on these people for accessing basic things like food and water, and that is a future we’re facing right now. The future is bleak where a corporation or a business dictates who gets to do AI, dictates a rationing of resources to use AI — and we’re kind of seeing that right now, Amazon and Google of the world rationing computing power… rationing, or we call it, the spice. But the spice must flow.” — Greg Osuri, 00:08:05

“Nvidia happens to be a biggest user of Akash — in fact, Akash is the only crypto protocol Nvidia uses. Akash is integrated into Nvidia products, and incredible companies like Nous Research, which happens to be a top tier AI team that’s working on decentralized training, to University of Texas to Rochester [Institute] of Technology — Akash has a wide ecosystem in the traditional AI world and also in the decentralized world.” — Greg Osuri, 00:12:22

“There’s no doubt in my mind that crypto is the only framework to have monetizable machine intelligence… it’s going to be multiple people, multiple organizations, multiple protocols interoperating with each other, permissionless.” — Greg Osuri, 00:46:54

“The major, major milestone — I think next 45 days — is going to be Akash AI… It is going to be the best AI developer product, and that’s the goal. It’s not about non-custodial nature, it’s not about crypto… AI developers are not going to use Akash because it’s decentralized; AI developers are going to use Akash because they get cheap GPUs and good developer experience.” — Greg Osuri, 00:44:16

“I love to see more work, more different approaches and more experimentation in the space… and really take the power away from the OpenAIs of the world. If you want to really disrupt them, you have to think of decentralization.” — Greg Osuri, 00:56:34

“Google wrote a paper called DiLoCo last year, and a company called Prime Intellect implemented [it]… they were able to train a 10 billion parameter model over the internet… another company called Nous Research is going to do a 15 billion parameter model soon… there are five other companies doing decentralized training.” — Greg Osuri, 00:20:58

“Training has been predominantly collocated… but that was until about 3 months ago… Prime Intellect, Nous Research… created by former Google DeepMind, big AI lab employees — they’ve solved distributed training fairly well. Now we have a 10 billion parameter model that’s fully distributed, training across the internet. There’s a 15 billion parameter model by Nous starting training in a couple of weeks — and they all use Akash. That’s how I know about these things.” — Greg Osuri, 00:13:25

“Now we’re getting into a bigger challenge for AI… it will be so obvious in a couple years that there’s no way we can scale AI the way we’re scaling AI. It’s obvious in the circles now, but I think it’ll be obvious to the world in a couple years that the only way to really build AI is through a decentralized system.” — Greg Osuri, 00:05:34

“It is perhaps the most important technological shift of our society… it is becoming to be a [feudalistic] society. People that have access to AI are going to win out over people that do not have access, and companies that control this access are going to amass so much power. If we don’t do something about it now, we’re going to lose out, because it’s happening very, very, very fast.” — Greg Osuri, 00:19:26

“I think Messari will eventually probably create a DML category sometime… DML is in a very, very early stages, but all of us are trying — I think we’re very close to catching up to the centralized AI.” — Greg Osuri, 00:40:14

“If you want to train advanced models, you need low latency, very minimal message-passing latency between two neurons. So if you have to do it in a decentralized way, that’s not possible… the trade-off is, yeah, you can get lower cost, but training time is going to be higher. So you’re never going to catch up to a GPT-4.” — Greg Osuri, 00:43:05

“It’s very important for machine learning to be decentralized and owned by the public. That’s established. And now the question is who is actually building this decentralized vision.” — Greg Osuri, 00:33:45

“Today GPT is trained on yours and my data, we gave for free essentially, so it’s taking our data and actually capturing quite a lot of value, and I think that’s wrong. So you should have control, sovereign, about your data and what you want to do with your data. If you want to get paid… I think that should be very primitive.” — Greg Osuri, 00:39:32

“This like new era of decentralized AI, which I believe will take off mid year, like June, July timeframe… there’s going to be an explosion of decentralized AI, but the beginning of the training, and then we’re going to go to inference… So every layer of AI will be decentralized.” — Greg Osuri, 00:47:57

“There’s enormous amount of research… Google really released a paper called [DiLoCo] and a bunch of companies actually implemented this paper and actually proved we can train a 32 billion parameter model fully distributed… you can have a cluster in New York… Singapore… San Francisco and actually contribute to a training run… That was not possible a few years ago.” — Greg Osuri, 00:26:13

“We’re still in the GPT-3 era of decentralized distributed training, but it’s progressing very fast, and only limiting factor is the amount of compute you can throw at it… how do we get the compute is through incentivization. So if you can create a viable incentive model, that’s where crypto plays a big major role — if you contribute compute, you somehow have rights in the future earnings of that model.” — Greg Osuri, 00:18:09

“If you can figure out how to do distributed training, I think that’s a big winner. So there’s a case to be made here, and that’s why I’m hyper bullish on decentralized AI, because the innovations that’s been happening over the last, let’s say, three to six months in distributed training is very promising.” — Greg Osuri, 00:17:25

“From an inference standpoint, which I believe is going to be much larger than training, as you start building more products, as AI gets more relevant into people’s lives…” — Greg Osuri, 00:12:24

“AGI — the ability to reason, plan, and learn broadly — could become the most important invention of the next 500 years. But it’s built on closed infrastructure. We risk bottlenecks, capture and fragility. We don’t just need open models. We need open infrastructure beneath them. We must decentralize the foundation before it ossifies.” — Greg Osuri, 04:26:47

“Nous pushed scale even further — a 405 billion parameter model trained in the wild, fully open source, fully decentralized. This isn’t the future. It’s the present. Together these breakthroughs prove… we can train powerful models without centralized servers, we can verify correctness without trust, and we can scale intelligence without monopolies.” — Greg Osuri, 04:31:19

“Centralized AI infrastructure creates single points of failure — physical, geopolitical and economic… We’re not just talking about uptime. We’re talking about national strategy. In a world where compute is power, open systems are the line of defense. This is how we preserve freedom in the age of digital acceleration, by decentralizing the power that drives it.” — Greg Osuri, 04:41:37

“Six months ago, decentralized AI training was a vision. Today, it’s a proven path… 405 billion parameters trained in the wild. This isn’t a science project. It is a movement… And just like Bitcoin proved for money, this proves for intelligence: we can build better systems by distributing them. Decentralization isn’t just viable, it’s inevitable.” — Greg Osuri, 00:06:46

“Until recently, decentralized AI was just considered impossible — too slow, too fragmented. But in 2025, that changed. Pluralis introduced async swarm training… Gensyn trained a 72 billion parameter model on a decentralized testnet — no campus, no cloud, just thousands of volunteers with GPUs. Bagel used ZK-LoRA… to verify decentralized fine-tuning without needing to rerun it. Nous is training a 405B parameter open model using distributed GPUs.” — Greg Osuri, 00:06:45

“When compute lives in a dozen hyperscale zones, it’s easy to surveil, easy to disrupt, easy to control. But when compute lives in 12,000 locations it’s resilient, it’s redundant, and it’s harder to censor, harder to kill, and harder to co-opt… We don’t win the AI age by concentrating compute. We win by distributing it. And that’s the American advantage.” — Greg Osuri, 00:15:43

“AGI — the ability to reason, plan, and learn broadly — could become the most important invention of the next 500 years. But it’s built on closed infrastructure. We risk bottlenecks, capture and fragility. We don’t just need open models. We need open infrastructure beneath them. We must decentralize the foundation before it ossifies.” — Greg Osuri, 00:02:12

“Nous pushed scale even further — a 405 billion parameter model trained in the wild, fully open source, fully decentralized. This isn’t the future. It’s the present. Together, these breakthroughs prove… we can train powerful models without centralized servers. We can verify correctness without trust. And we can scale intelligence without monopolies.” — Greg Osuri, 00:07:17

“We built AkashML, a distributed AI runtime designed to unlock real performance from globally scattered resources. It’s GPU native, it’s container first, and it’s lightning fast… We support inference today, fine-tuning tomorrow, and fully training soon after. Whether you’re running a star node in Austin or a micro grid in Nairobi, the experience is seamless.” — Unknown, 00:00:02

“Unlike other platforms that may remove unprofitable models without warning, AkashML surfaces real-time costs. If a model becomes expensive to maintain, users can choose to either attract more usage or shift to self-hosting. We empower communities to keep models alive… We’re not just building infrastructure — we’re restoring control to the people who create and rely on these models.” — Unknown, 00:01:29

“This shift does much more than fix our energy bottleneck. It reshapes access. Developers can build independently of big tech without begging for compute. These infrastructure policies would level the field for smaller players to build and deploy advanced AI models, decentralizing opportunity itself.” — Greg Osuri, 00:04:21

“The biggest model is 72 billion parameter model. For comparison, GPT-3 was 150 billion parameter model. So we’re very, very close to GPT-3… So if you were to ask me where distributed training is, it’s about pre-ChatGPT-3. And by end of the year, it will be ChatGPT-3 level.” — Greg Osuri, 00:19:37

“Can we actually reduce the amount of energy bandwidth that’s needed between nodes?… They were able to reduce communication or improve bandwidth requirements by 875x. So from, I think, like 80 gigabytes to, like, 70 megabytes.” — Greg Osuri, 00:14:16

“Why last year there was dismissal of this? No one even thought this was possible. This year, there are five papers that are all presenting at ICML. That’s a big shift because that is acceptance by traditional AI.” — Greg Osuri, 00:22:26

“In the meantime, I think decentralized AI is the only way. I looked at every possibility. Lay it out, and there is no magic solution to get more energy in America in the next four years.” — Greg Osuri, 00:40:30

“There is a lot of work that indicates distributed training will become a reality by end of the year, and by end of the year we will produce a model as good as GPT-3. By end of next year, or maybe going into 2027, there’s a good chance that we’ll be able to produce a state-of-the-art model fully trained distributed — or decentralized rather.” — Greg Osuri, 00:21:50

“I have high conviction over the next 3 to 5 years decentralized AI has the best shot to make a mark in the world. We got the best shot right now. If we don’t make progress in the next two years, I don’t think decentralized AI will take off… People are going to use a solution because there’s no other option — that’s really the best way to sell a product.” — Greg Osuri, 01:21:05

“We all know Bitcoin is the largest supercomputer in the world. Why? Because there’s an incentive structure to contribute compute to Bitcoin. Similarly, if there’s an incentive structure to contribute your compute to train a model, there is a possibility we may actually out-compete OpenAIs of the world, because now you have the public.” — Greg Osuri, 00:21:50

“We’re like under the radar, which I really like… because I think we’re about to blow up very soon in this decentralized AI space.” — Greg Osuri, 01:10:30

“There’s a lot of work that happened in the last few years on low communication algorithms, asynchronous algorithms, fault tolerance, all coming together. I believe in the next six [months] to one year you’ll actually see a truly decentralized model, or a model that’s trained on truly decentralized networks… it’s not decentralized yet, but it will be, very very very soon.” — Greg Osuri, 00:19:08

“Continuous non-disruptive verification is key. That means without affecting the workload that’s deployed on the GPU… you got to continuously verify, and no one has solved the problem, and I think we’re close to solving the problem and doing it in fully open source and decentralized manner.” — Greg Osuri, 00:29:58

“I feel like it’s going to be a new chapter post Mainnet 14 upgrade… that is going to set a new cadence in terms of how fast we’ll be shipping features. And that made me comfortable enough to commit to a public roadmap that I published on my Twitter… Verifiable compute I think is huge… virtual machines — again, 90% of AI uses virtual machines. So I’m really excited for these two capabilities, and I think the demand will catch up.” — Greg Osuri, 00:22:16

“Akash is like that very strong blockchain but very hard to make changes. It’s great if you’re Bitcoin — no changes to the blockchain — but we’re not Bitcoin, we’re AI. AI market demands a lot of change. We wanted new incentives, we wanted the burn mechanism… we want all these new capabilities that we said we’re going to do on the roadmap, but we discovered that we can’t really do much without upgrading the blockchain.” — Greg Osuri, 00:17:12

“A concrete decentralized training milestone is Nous Research trained a 15 billion parameter LLM over the public internet on volunteer machines and donated instances in about 10 days and about 11,000 steps, matching centralized run accuracy using WAN-aware optimizers. This demonstrates that with the right comms and async strategy, frontier class models can train beyond a single hyperscaler.” — Greg Osuri, 00:13:16

“If compute access is functionally gated by energy, we risk centralization by default — only a few players can pay the bill. So here’s the constraint in one line: AI demand is exponential; our energy response can’t be linear. We need architectures that use energy smarter — spreading load, lifting utilization, and collocating compute with abundant, cheaper renewables.” — Greg Osuri, 00:04:26

“An alternative is thousands of nodes — small data centers, campuses, labs, and individual providers cooperating over the internet to train one model. That is not science fiction. We’ve seen precedent at internet scale: Folding@home volunteers once peaked around 2.4 exaflops, outpacing the top supercomputer at that time. With the right algorithms and verification, we can turn the internet itself into a capable, resilient training fabric for PyTorch jobs.” — Greg Osuri, 00:06:39

“We’re not just reducing cost, we’re expanding who gets to build — that aligns with the open source ethos that made PyTorch what it is… We can turn idleness into intelligence… Make the internet a cluster, and we make clean energy a scheduling signal.” — Greg Osuri, 00:19:10

“Training is not there yet — I wouldn’t call training vaporware, that’s really promising… decentralized inference is complete vaporware. There hasn’t been any verifiable… inference that’s actually real.” — Greg Osuri, 00:49:21

“There have been 100 compute marketplaces that came after Akash. Most of them are just not decentralized in any manner — they’re just SaaS apps. So ignore anything that’s closed source, ignore [anything that’s not on-chain]… just because something has a token doesn’t mean it’s decentralized. So all that is junk.” — Greg Osuri, 00:45:43

“A big favorite for verifiability today is trusted execution environments, where it is vendor-attested workloads, essentially attested at the chip level, and with memory encryption it’s impossible to at least see inside a workload when it is encrypted at the memory level.” — Greg Osuri, 00:17:25

“The real interesting stuff I think in AI is distributed training… There hasn’t been a platform where you go and be like, ‘Hey, I’m going to contribute my computer to this training run, I want some tokens back.’ That’s the holy grail we’re working towards. No one has quite figured that out… nothing cohesive in terms of a framework that can bring all the primitives together… and make it as easy or as interesting as DeFi was during DeFi summer.” — Greg Osuri, 00:47:08

“My take is there is no other future that’s not decentralized, because the current model is not only failing, it cannot scale. There’s really no solution to the energy problem — it takes a Manhattan project level effort… we have competition from a nation state level, which is China, that’s building a gigawatt of solar capacity every 36 hours… The only way to solve that is to figure out how to distribute workloads.” — Greg Osuri, 00:08:01