Gonka Daniil Liberman：打造去中心化 AI 全堆疊，算力也能 Web3

0:00AI 必須去中心化

0:00大家好，可以聽到我嗎？今天我要講的主題是 AI，不是加密貨幣，希望不會讓你們覺得無聊。我和我的兄弟還有我們的團隊都非常相信，AI 必須去中心化，理由非常多。Hi everyone. Yes, you can hear me. I'm here to speak about AI, not crypto. So I hope it won't be too boring for you. The thing is... We strongly believe, me and my brother and our team, we strongly believe that AI needs to be decentralized for many, many, many reasons.

0:24當我說去中心化時，我不是單純指加密貨幣，雖然加密貨幣確實是讓去中心化 AI 成真的關鍵。我今天想談的是算力。你們可能不知道，Bitcoin 從一開始只用家用電腦，到現在每小時消耗超過 20 GW 的電力，這比全世界目前所有 AI 算力消耗的電還多。And when I say decentralized, I don't mean it in a sense of cryptocurrency per se, even though crypto is going to be the way decentralized AI can even happen. What I want to talk about is compute. I don't know if some of you are aware, but Bitcoin itself, since the inception when it only started with our home computers, by now uses up to over 20 gigawatt of energy, 20 gigawatt per hour.

1:08Bitcoin 現在用的電比全世界 AI 算力還多。雖然未來 AI 算力一定會遠遠超過這個數字，但重點是：誰會擁有這些算力？誰能控制這些算力？這很重要，因為如果沒有一個能讓中心化供應商負責且透明的替代方案，讓大家公平取得 AI，我們可能會面臨前所未有的風險。Our energy is used by Bitcoin alone, which is more than all of the AI compute today worldwide. And even though AI compute is going to surpass this many, many, many times, the question is going to be who is going to own this compute and who is going to control this compute. And why it matters is because unless there is an alternative which actually makes the centralized providers accountable and transparent on the rules and the equality of the access to AI, we can end up with something really bad which we have not experienced before.

1:49AI 算力壟斷的危機

1:49這不會只是像銀行或社群媒體那樣的問題，也不只是像行動電信商那樣。AI 是生產力本身，會重塑整個經濟，甚至可能讓 GDP 翻四倍。問題是，什麼能阻止 AI 基礎設施的擁有者獨佔這些經濟成長？It's not going to be like banks, and it's not going to be like social media or mobile providers. It is going to be much worse because AI itself is means of production. AI itself will reshape economy, and AI itself will probably quadruple GDP, quadruple what economy is. And the only question is, what will stop the owners and the builders of AI infrastructure from owning all of the increase in the GDP and productivity which AI can bring.

2:28現在我們看到開放模型和封閉模型的競爭，但更重要的問題其實不是模型本身，而是算力的開放與否。目前全球超過 80% 的 AI 晶片算力都集中在美國，其中 70% 以上又被五家美國企業掌控。We can see right now that there is this battle between open model and closed models, but the more important question is even less about the open models and closed models, but open compute capacity versus closed compute capacities. Currently, globally, over 80% of all the AI-enabled chips compute sits in the United States. More over 70-plus percent of this AI compute sits in the hands of only five U.S. corporations.

3:06這看起來好像沒什麼，因為雲端運算也是這樣，AWS、GCP、Azure 幾乎壟斷全球 80% 的雲端算力。但雲端運算就像電力一樣，大家競爭讓它更便宜、更普及。但 AI 不一樣，AI 是勞動力本身，能取代我們很多工作，除非我們自己也參與其中。It seems to be like okay because it used to be before like that as well. We can see AWS, GCP, and Azure owning almost 80% of all of the cloud compute globally. But cloud compute is almost like electricity. Like the only way they own it is by making it cheaper and cheaper and more available for everyone, and they compete on this. With AI, it's a little bit different because AI is workforce. AI can replace us in many, many, many jobs unless we actually will be part of what's going on with AI.

3:43AI 的未來：壟斷還是全民參與？

3:43未來可能會變成大公司擁有所有 AI，也掌控 AI 創造的經濟；或者我們每個人都能平等取得 AI，大家依然在同一個經濟規則下互相提供服務，只是我們都被賦能了。我們發現，最終決戰點不在模型，而是在算力本身。So the future can look like corporations own all of the AI. And they own all of the economy AI generates, or in the future can be each of us have equal access to AI, and with this equal access to AI, we still play the same rules of the economy. We provide services to each other in the same way, except we are empowered. All of us are empowered. Because we realized. That this, the end game is going to be in, not in the models, but in the computer itself.

4:21我們預測未來會有很多開放模型，但主要市場和營收來源會是 API，不是模型本身。現在已經有很多強大的開放模型，未來只會更多。問題是，錢流去哪裡？目前大部分 AI 收入都被 OpenAI、Anthropic 和 Google 這三家公司吃下來了。We sort of predicted that there will be a lot of open models, and that the major market, the major revenue source is going to be API, not the models themselves. And that's how we can see a lot of really powerful open models on the market right now, and there will be even more. The only question is, where is the revenue? And currently, majority of the AI revenue stays with only two corporations, three corporations: OpenAI, Anthropic, and Google.

4:57他們靠 API 和訂閱服務掌控大部分算力，也主導 AI 基礎設施的建設。這就是為什麼幾年前我們意識到必須有一個去中心化協議，讓任何人都能無需許可地把 AI 算力接入一個去中心化網路，並且這個網路要有類似 Bitcoin 的工作量證明（Proof-of-Work）機制。They basically eat like 95% of all the AI revenues. There is API and subscription services, which allows them to acquire majority of the compute and be the one who actually owns the build-out of the AI infrastructure. That's how several years ago we ended up with understanding that there needs to be a protocol, a decentralized protocol. Which will allow permissionless, aiding the AI-enabled compute to a decentralized network, which anyone can participate, and this network should have a proof-of-work concept on the backbone of it, similar to Bitcoin.

5:41打造去中心化 AI 協議

5:41但這個網路的工作不該只是像 Bitcoin 那樣解數學謎題，而是要讓 API 執行開源模型的推論（inference），這才是 AI 現在和未來的主要收入來源。我們花了很多時間設計怎麼讓這種推論工作可以被驗證、可證明，並透過去中心化的驗證者網路來分配有限的代幣獎勵給真正提供算力的參與者。And the work in this network should not be multiplying the useless, in a sense, hashes. Just a mathematical puzzle of the Bitcoin hash, but it should be the API of open source, open models themselves, the inference, which is the future of their revenue, which is the now of the revenue in AI. That's how we spend time on building the principle of how to make API work. The inference, the inference work, I'm sorry, how to make it provable and verifiable through a decentralized network of validators, and then pay the limited supply of the coins to whoever is actually providing the compute to the end users through the decentralized API.

6:35我們就是這樣開始打造這個協議，並在一年前正式上線。協議推出的前四個月，網路規模每個月都成長四倍，證明了 AI 算力可以像工作量證明一樣運作。半年後我們才開放 API，現在每月已經有一兆顆代幣的用量，幾乎把現有幾千張 H100 等級 GPU 的算力都用滿了。And that's how we focused on building the protocol. We built this protocol and we launched this protocol a year ago, and in the very first days of this protocol, for the first four months, the network grew four times every month. That was the speed with which the network was growing, proving the whole concept that the AI compute can be perceived as a proof-of-work compute system. And then. We only launched the API half a year later after that, in June this year, and the usage of this network grew from zero to a trillion tokens a month by now. So as much as this sounds to be tiny, but this is almost 100% utilization of the current. A couple thousand of the H100 equivalents of the GPUs on the network.

7:30AI 基礎設施的龐大市場

7:30我們想強調，AI 的未來會有數兆美元的基礎設施投資。這其實很好理解，就像現在全球 50 億人平均每年花 220 美元在手機和行動網路上，這就是一兆美元的產業。美國人平均每年花 1,600 美元在電力上，其中只有 40% 是你家電費，60% 是藏在你用的商品和服務裡。What we want to explain is that the future of AI comprise of trillions of dollars of investment in infrastructure. And these trillions of dollars of investments are quite understandable. Like, if you will think how much money each of you spend right now on electricity or on mobile connectivity, like on average, 5 billion people worldwide spend like 220 bucks a year on the mobile phone and the mobile connectivity. Which comprise a trillion dollar worth of annual expenses on the mobile connection. If we talk about electricity itself, the average in the United States is going to be 1,600 per individual per year to be spent on electricity. And only 40% of this is your utility bill, and the 60% is how much electricity is hidden inside of every goods and services you use across a year.

8:47未來 AI 的規模可能會讓每個人每年都花上數千美元，這產業會非常巨大。預測到 2035 年，AI 基礎設施的總資本支出會到 10 到 20 兆美元。分攤到 20 億早期用戶身上，其實還比現在的電費低，但 AI 會滲透到經濟的每個角落。This is an entire generation of electricity divided by the U.S. citizens. So the massiveness of the future of AI can be that every single human being a year is going to be spending thousands of dollars. One to two thousand dollars at least on average, and this is how massive industry is going to be. The predictions are that by 2035, so in 10 years from now, the total capex, the total expenditure on the infrastructure of AI is going to be between 10 to 20 trillion dollars. Divided by two billion people of the early users, it is going to be less than what we spend on electricity right now. But AI is going to be crucial in all and every aspect of the economy.

9:33AI 晶片與產業瓶頸

9:33這就是為什麼我們會用 AI。現在我們遇到的瓶頸是 AI 晶片的產能。目前大家都知道 AI 幾乎都靠 Nvidia，已經有數兆美元投入製造 GPU 和資料中心，但這些都集中在美國少數幾家公司手上，這不應該這樣。That's why we're going to be using this. Currently, right now, we have a bottleneck. Like, this bottleneck is in AI chips and in production of AI chips. Currently. All we know about AI is primarily Nvidia, and trillions of dollars spent already on manufacturing and data centers, manufacturing GPUs and putting them in data centers. But this all right now concentrated in only few corporations in the States, which can't, should not be that way.

10:08我們需要打造一個人人都能平等參與、無需許可、不可竄改的基礎設施。不可竄改的意思是，如果我們讓 AI 存取銀行帳戶或個人資訊，就必須完全信任 AI 會為我們的利益行事。唯一能做到這點的方法，就是讓算力像 Ethereum 智能合約一樣不可竄改。And what we need is we need to build the infrastructure which is going to be equally available, permissionless, immutable. Immutable in the sense that we now need, if we will give access to the AI to our bank accounts or our private information. This will all require us complete trust that the AI is acting in our interest. And the only way how we can achieve this is if the compute is immutable the way it is immutable in terms of smart contracts of Ethereum.

10:45AI ASIC 與產業進化

10:45這就是我們打造這個協議的原因。未來會像 Bitcoin 一樣，最早是家用電腦，後來 2013 年變成 Nvidia GPU，2014 年就被 Bitcoin ASIC 主導。AI 也會走向類似的路，現在已經有數十家 AI ASIC 晶片製造商，還有多層軟體讓未來不只 Nvidia、AMD，還有很多廠商都能參與。That's how we build this protocol. And the future of this protocol is the similar future as it used to be in Bitcoin. When in Bitcoin, eventually it all started with home computers, then we moved to the GPUs of Nvidia by 2013, and then by 2014, it was already dominated by the Bitcoin ASICs. The similar future will be in AI as well. There are currently dozens of manufacturers of AI-powered AI chips, the AI ASICs, and there are currently several of the software layers which enables the future to be homogeneous. Not only the NVIDIA, not only the AMD, but tens of different producers of AI chips.

11:39比特幣的晶片在比特幣ASIC剛開始發展的時候，每兩年效能幾乎提升十倍。我們整個產業、我們所有人在打造比特幣的過程中，花了超過三千億美元，在十年內建構出這二十吉瓦的比特幣ASIC運算能力。這一次，規模將會更大。這一次，目前整個加密產業的總市值還不到三兆美元，過去曾經接近四兆，但未來有可能接近十兆。The chips of Bitcoin was increasing its output, its performance, almost ten times every two years in the beginning of the Bitcoin ASICs buildout. The industry, all of us, when we were building Bitcoin, we spent over $300 billion in the course of 10 years to build out this 20 gigawatt of Bitcoin ASIC compute. This time it's going to be much larger. This time, currently the entire crypto industry is less than $3 trillion combined, and it used to be close to 4, but it can be close to 10.

12:28如果我們將其視為一個 AI 基礎設施的建設，那麼我們所有人都需要把 AI 看作一個類似於 Bitcoin 的可能性。即使 AI 運算未來的建設中只有 10% 被切分出來，這也將帶來額外數兆美元的規模。If we embed what it is a build-out of AI infrastructure, and for that all of us need to look at AI as a possibility similar to Bitcoin, which will carve out piece, even if only 10% of the future build-out of AI compute, this is going to be additional trillions of dollars.

12:53早期參與者的機會

12:53在加密產業，早期投入去中心化 AI 網路算力的人，資產成長會像早期 Bitcoin 一樣。AI ASIC 的機會是真實存在的。像 Cerebras、Tenstorrent、Axelera 等公司，還有許多中國新創都在做 AI 晶片，這會大幅加速 AI 推論的創新速度。In the crypto industry, people who will invest in this early, people who will add up compute to the decentralized networks early, will see similar growth of their assets to what they have experienced with Bitcoin in the early years of Bitcoin. The ASIC opportunity with AI is real. Companies like Cerebras, companies like Tenstorrent, companies like Axelera and others, and there are dozens of the new Chinese companies producing chips right now that will empower and increase speed of innovation in AI inference.

13:39AI ASIC 的效能提升速度比摩爾定律還快，十年內提升 30 萬倍，這比電腦本身的進步還誇張。誰能生產更多算力，誰就能拿到最多代幣，這會驅動工程師和 IT 專家投入，把設備接上網路，像早期挖 Bitcoin 或 Ethereum 一樣開始賺錢。With a speed similar to Bitcoin, which was in just 10 years from the beginning of the ASIC revolution, the productivity of these chips increased 300,000 times, which is, like, unheard of. This is faster than Moore's law. This is faster than computers were developing themselves. As soon as it's specialized, and as soon as it's— Whoever produces more compute gets the most of the tokens of the network. This principle drives the innovation because now engineers can produce and IT specialists can buy, connect it to the network, and start printing money as it used to be with Bitcoin or Ethereum in the early days.

14:28去中心化 AI 的未來邀請

14:28我今天只有 30 分鐘，請大家多多關注我們，也思考你想參與 AI 基礎設施建設的哪一部分。我們團隊也在開發新型 AI 晶片，未來幾年會有更多廠商提供更便宜的推論晶片，讓大家都能分一杯 AI 建設的收入，這會是數兆美元的市場。I think I have only 30 minutes. Please follow us. Please look at what you actually want, what part of the actual future of AI infrastructure buildout you want to be. Our team, we are working on development of the new chips as well. And in the coming years, there will be dozens of chip manufacturers. Providing you with cheaper inference chips, which will generate revenue and carve out part of the revenue of current AI buildout, and it is going to be in trillions of dollars.

15:08這一切都會是完全相容、無需許可、去中心化，讓每個人都能平等取得 AI。歡迎大家加入我們！And it's going to be completely compatible and... Permissionless and decentralized and available, equally available for everyone to get access to AI. Join us.