Akash Founder Sees Energy and Policy Bottlenecks Threatening AI Growth

The Four Horsemen of AI With Akash Network — Ai4

Ai4

Centralized AI faces a crisis driven by energy grids, regulation, and privacy risks that only decentralized compute networks can resolve.

AI-assisted summary of The Four Horsemen of AI With Akash Network — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background This talk was delivered by Greg Osuri, founder of Akash Network, published on YouTube by the Ai4 channel on August 14, 2026. According to the page description, the presentation addresses how regulation, compute limits, energy constraints, and privacy threats jeopardize AI progress, while proposing solutions like adaptive policy and decentralized compute.

## TL;DR Greg Osuri argues that AI progress in 2026 is hindered not by bad models, but by four structural bottlenecks: fragmented regulation, compute concentration, physical energy constraints, and privacy failures. He illustrates how these 'Four Horsemen' form a self-reinforcing loop driven by centralization that inflates costs and triggers public backlash. To counter this, Osuri showcases production-ready decentralized solutions, including internet-scale distributed training, tapping into idle consumer GPUs via Akash Home Nodes, and integrating compute with residential solar power.

## Atmosphere & style The presentation carries an urgent, analytical, and pragmatically optimistic tone, blending high-stakes technological diagnosis with concrete engineering solutions. Osuri speaks with authoritative clarity and sharp wit, using striking metaphors and stark statistical charts to construct a compelling narrative arc. Watching the talk feels like a lucid wake-up call that shifts from sobering structural warnings to an empowering, actionable blueprint for a resilient AI ecosystem.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) The bottleneck to AI progress in 2026 has shifted from models, data, and talent to law, chips, electrons, and trust—the Four Horsemen. - [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) The Four Horsemen operate in a self-reinforcing feedback loop driven fundamentally by centralization. - [00:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=29s) Fragmented regulation acts as a fixed-cost trap, giving hyper-capitalized incumbents an unfair advantage over smaller startups. - [00:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=29s) Model weights have become a key battleground in geopolitical regulatory disputes and copyright litigation. - [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Soaring frontier model training costs are turning high-end compute into a defensive moat dominated by a few private companies. - [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Open compute marketplaces significantly improve GPU utilization compared to the low efficiency of traditional cloud providers. - [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Millions of idle consumer gaming GPUs represent a massive untapped reservoir for running open-source inference models. - [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Pluralis demonstrated that pre-training frontier models over standard consumer internet connections is technically feasible. - [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) AI energy demands are surging rapidly, creating a severe mismatch with long physical infrastructure lead times. - [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) Soaring data center interconnection requests are straining regional power grids, driving up energy prices and triggering legislative moratoria. - [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) Combining home rooftop solar with localized compute nodes offers a way to bypass grid power bottlenecks. - [01:42](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=102s) Centralized cloud architectures expose corporate conversations to legal subpoenas and data breaches. - [01:42](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=102s) Local and edge inference using open-source models provides the strongest guarantee for data privacy and compliance. - [01:42](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=102s) Deploying edge workloads on decentralized networks offers substantial cost savings over traditional cloud infrastructure. - [02:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=124s) Organizations can build resilience by evaluating risks across all four domains, diversifying models, and leveraging idle hardware.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the talk, preserving all concrete facts, arguments, numbers, names, and the speaker's texture and tone.

---

### [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) The Paradox of Free Intelligence and the Four Horsemen Greg Osuri, founder of Akash Network, opens his talk by pointing to a striking contrast. Just nine days prior, a Beijing startup released the largest model ever trained for free to the public—an open model functionally distributed like a torrent that anyone can download anytime. This brings up a fundamental question: if frontier intelligence is freely accessible, why does AI feel like it is controlled by only five permission-gated companies where developers must wait in line for API access or energy? The answer lies in one core issue: **centralization**.

While the last decade framed AI bottlenecks around models, data, and talent, the reality in 2026 is that AI progress is limited by **law, chips, electrons, and trust**—or regulation, compute, energy, and privacy. Osuri clarifies that these "Four Horsemen" are not the harms caused by AI, but rather the structural threats to AI's ongoing progress. Crucially, they do not operate in isolation; they are tightly bound together in a self-reinforcing loop driven by centralization.

``` +-------------------------------------------------+ | CENTRALIZATION | +-------------------------------------------------+ | ^ v | Privacy Failures ---> Heavy Regulation | | | | | v | | Compute Control | | | | v v | Backlash & Moratoria <-- Energy Stress (Gigawatts) --+ ```

- **Loop Dynamics**: Privacy failures trigger regulatory crackdowns; regulation dictates who owns and accesses compute; compute concentration drives massive gigawatt-scale data center deployments; extreme energy stress provokes public and economic backlash, which in turn leads back to further regulation.

---

### [00:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=29s) The First Horseman: Fragmented Regulation and the Fixed-Cost Trap The first horseman, regulation, is rapidly intensifying across the globe. The European Union recently activated its comprehensive AI Act, carrying maximum penalties of €35 million or 7% of global turnover—a stick even bigger than GDPR. Osuri quips that some might argue the EU generates more revenue from penalties than from taxes. The timing is almost theatrical: while attendees gather at an AI conference in Las Vegas, Brussels has officially activated strict transparency and high-risk compliance rules.

Meanwhile, the landscape in the United States is a chaotic "patchwork." All 50 state legislatures have introduced AI proposals, around 38 states have enacted laws, and nearly 100 new AI-related statutes have been created. Highlighting the lack of federal preemption, the U.S. Senate voted 99 to 1 against freezing state-level AI laws. Federal executive orders lack teeth according to legal experts, and even local attempts at regulation—such as California's governor vetoing a bill designed to cap the amount of compute used to train models—show that while everyone wants to regulate, nobody agrees on what or how to regulate.

- **Copyright & Model Weights**: High-stakes legal battles over training data resulted in the largest copyright infringement settlement to date involving Anthropic. Furthermore, model weights themselves have become the new front line in global regulatory disputes: Washington is currently investigating how the latest Qwen3 weights were derived, while Beijing contemplates export controls on model weights. - **The Uncomfortable Math**: Regulation represents a fixed compliance cost that scales sub-linearly. A 50-person startup faces roughly the same regulatory burden as a 50,000-person enterprise, ensuring that hyper-capitalized incumbents with the biggest balance sheets inevitably win under heavy regulatory regimes. - **Distributive Solutions**: To counter this centralization, Osuri proposes open-weight distribution to prevent single-company API control, jurisdiction-agnostic permissionless compute marketplaces, and **protocol learning**—a mechanism where a model is collectively trained across decentralized internet nodes without any single company owning all the weights.

---

### [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) The Second Horseman: Compute Concentration and Internet-Scale Clusters Compute costs are compounding at roughly 3.5 times per year, doubling approximately every five months. While training GPT-4 cost around $80 million, training top-tier models in 2026 costs between $200 million and $500 million, with estimates reaching $1 billion to $3 billion for next-generation frontier models. At these price points, the list of organizations capable of competing narrows drastically. Furthermore, 75% of frontier models originate from the U.S., and 90% come from private industry rather than academia, turning compute into a defensive moat rather than a mere operational input.

To disrupt this concentration, decentralized open marketplaces like Akash have provided the compute infrastructure to bootstrap projects such as Prime Intellect, Nous Research, Brev (prior to its acquisition by Nvidia), and Venice AI. Unlike traditional cloud providers where utilization lingers in the single digits due to opaque capacity management, open marketplaces operate at 60% to 70% utilization because open inventory allows for far better planning.

``` +-----------------------------------------------------------------+ | COMPUTE BOTTLENECK | +-----------------------------------------------------------------+ | GPT-4 Training Cost: ~$80 Million | | 2026 Frontier Models: $200 Million - $500 Million | | Next-Gen Projections: $1 Billion - $3 Billion | +-----------------------------------------------------------------+ | DECENTRALIZED RESOURCE ALTERNATIVE | | 150M Steam Gamers --> 3/4 Own Nvidia RTX 4090/5090 GPUs | | ~20 Hours Downtime/Day --> Capably Runs 70B Quantized Models | +-----------------------------------------------------------------+ ```

- **The Steam GPU Reservoir**: Approximately 150 million gamers are active on Steam at any given time, and three out of four hold modern Nvidia GPUs (such as RTX 4090s or 5090s). These consumer cards can easily run quantized 70-billion parameter inference models and sit idle for roughly 20 hours a day. Mobilizing even a single-digit percentage of these GPUs via tools like the Akash Home Node (a simple USB boot drive) expands global inference capacity exponentially without building new data centers. - **Distributed Training Proof-of-Concept**: Addressing the common objection that heavy model training requires co-located data center compute, Osuri points to a breakthrough by frontier lab Pluralis. They pre-trained an 8-billion parameter model from scratch across 330 internet-connected nodes, requiring only an RTX 4090 with under 200 Mbps bandwidth per node. The run achieved 63% of the training efficiency of an interconnected H100 cluster, proving that the cluster can indeed be the internet itself.

---

### [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) The Third Horseman: The Unforgiving Physics of Energy Constraints Unlike policy or software, energy is governed strictly by the laws of physics and grid infrastructure limits. According to International Energy Agency (IEA) figures, global data center energy consumption was roughly 215 TWh in 2024 and is projected to reach 945 TWh by 2030 (about 3% of all global electricity usage) before climbing toward 1,200 TWh by 2035, with AI identified as the primary catalyst. In the United States—home to nearly half of global data center capacity—Berkeley Lab estimates that AI will consume nearly 12% of total U.S. electricity production by 2030.

This rapid surge creates a severe mismatch between AI software scaling speed and physical infrastructure lead times. In 2025, U.S. electric utilities received 700 GW of data center interconnection requests—representing approximately 700 nuclear reactors worth of capacity and far exceeding the entire country's 2023 peak capacity of 477 GW. Building new transmission lines takes 4 to 8 years, while power transformer lead times now stretch up to 4 years. Regional grids are feeling the strain: ERCOT in Texas received 143 GW in requests against an 85 GW peak, while PJM Interconnection (serving 65 million people) saw wholesale energy prices surge by 75.5% according to its independent monitor, sparking public backlash and pushing at least 12 state legislatures to consider data center moratoria.

- **Residential Solar and Compute Integration**: Osuri emphasizes that the nation's largest energy producers are residential homes equipped with rooftop solar, which can easily generate 10 kW—more than enough to power localized inference nodes. - **Span + Akash Home Node**: Smart electrical panel manufacturer Span projected it could reach 6,000 to 8,000 homes in 6 months to deploy workstation-grade GPUs (such as RTX 4090s and RTX 6000 Ada GPUs). These home nodes consume excess solar energy during peak generation hours, solving localized power constraints while generating income for homeowners.

---

### [01:42](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=102s) The Fourth Horseman: Privacy Failures and the Distributed Solution Privacy has evolved from a secondary feature into a hard deployment gate for AI adoption. Osuri points to the legal battle between *The New York Times* and OpenAI, where a court order subpoenaed approximately 20,000 ChatGPT user conversations, including temporary and deleted chats. The issue is not inherent malice, but an architectural flaw: any data transmitted to a centralized cloud service is vulnerable to legal discovery and breach.

A study analyzing one million AI conversations revealed that 4% of corporate user chats and 22% of uploaded files contained sensitive corporate data—leaving enterprises "one subpoena away from being disclosed to the public." Regulatory enforcement against entities like Clearview AI demonstrates that strict compliance regimes make cloud-hosted AI non-viable for sensitive sectors like healthcare and legal services.

- **Local and Edge Inference**: The strongest privacy guarantee is data that never leaves the local network. With open-source models now rivaling proprietary alternatives, companies can run models locally and offload surplus compute needs onto permissionless networks like Akash. - **Enterprise Proof Point (Razer)**: Gaming hardware manufacturer Razer ran an April Fools' AI campaign that took user pet photos and rendered interactive holographic images. Operating on consumer GPUs through decentralized compute, they processed 11,000 unique image generations over 5 days at a cost of just $0.01 per generation—compared to $0.03 to $0.15 per generation on traditional cloud platforms.

---

### [02:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=124s) Actionable Framework for confronting the Four Horsemen Osuri concludes by stressing that decentralized AI is an active, production-ready reality rather than a speculative science project. Open frontier models can be executed anywhere without reliance on centralized providers, distributed training across the internet is proven, and decentralized home nodes are operating in the wild.

He leaves the audience with a concrete action checklist for their organizations:

- **Evaluate Risk Metrics**: Assess enterprise vulnerability across all four domains—regulation, compute costs, energy dependence, and privacy guarantees. - **Diversify Models and Infrastructure**: Move away from single-provider reliance on OpenAI or Anthropic; host open models locally and test edge deployments. - **Monetize Idle Hardware**: Deploy unused corporate or consumer GPUs onto Akash Home Nodes to earn a measurable return on investment. - **Advocate for Interoperable Regulation & Grid Innovation**: Back clear, unified regulatory standards to avoid a fractured 50-state legal environment, and invest in distributed, solar-friendly edge compute architectures to prevent an "AI apocalypse."

## Notable quotes > [00:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=29s) The EU recently activated its AI Act, carrying maximum penalties of €35 million or 7% of global turnover—a stick even bigger than GDPR. — Greg Osuri

> [00:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=29s) A 50-person startup faces roughly the same regulatory burden as a 50,000-person enterprise, ensuring that hyper-capitalized incumbents with the biggest balance sheets inevitably win under heavy regulatory regimes. — Greg Osuri

> [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Compute costs are compounding at roughly 3.5 times per year, doubling approximately every five months. — Greg Osuri

> [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Unlike traditional cloud providers where utilization lingers in the single digits due to opaque capacity management, open marketplaces operate at 60% to 70% utilization because open inventory allows for far better planning. — Greg Osuri

> [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) Mobilizing even a single-digit percentage of these GPUs via tools like the Akash Home Node expands global inference capacity exponentially without building new data centers. — Greg Osuri

> [00:36](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=36s) The run achieved 63% of the training efficiency of an interconnected H100 cluster, proving that the cluster can indeed be the internet itself. — Greg Osuri

> [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) In 2025, U.S. electric utilities received 700 GW of data center interconnection requests—representing approximately 700 nuclear reactors worth of capacity and far exceeding the entire country's 2023 peak capacity of 477 GW. — Greg Osuri

> [01:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=98s) These home nodes consume excess solar energy during peak generation hours, solving localized power constraints while generating income for homeowners. — Greg Osuri

> [01:42](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=102s) A study analyzing one million AI conversations revealed that 4% of corporate user chats and 22% of uploaded files contained sensitive corporate data—leaving enterprises 'one subpoena away from being disclosed to the public.' — Greg Osuri

> [02:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=124s) Decentralized AI is an active, production-ready reality rather than a speculative science project. — Greg Osuri

## Mentioned Greg Osuri, Akash Network, EU AI Act, GDPR, Anthropic, Qwen3, GPT-4, Prime Intellect, Nous Research, Brev, Nvidia, Venice AI, Steam, Pluralis, International Energy Agency (IEA), Berkeley Lab, ERCOT, PJM Interconnection, Span, The New York Times, OpenAI, ChatGPT, Clearview AI, Razer

## Video description Greg Osuri - Founder, Akash Network

AI’s greatest threats may not come from rogue models, but from the systems required to build, power, govern, and trust them. Regulation is fragmenting across borders, advanced compute is concentrating in fewer hands, energy demand is colliding with infrastructure limits, and privacy concerns are eroding public confidence. Together, these four forces could slow innovation, raise barriers to entry, and determine who gets to shape the future of AI. This session will examine how these threats reinforce one another and explore practical solutions, including adaptive policy, decentralized compute, new energy strategies, and privacy-preserving architectures. Attend to gain a clear framework for confronting the Four Horsemen and building an AI ecosystem resilient enough to survive them.

[Source](https://www.youtube.com/watch?v=Cc4rGyh2XM4)

中文简报

Akash创始人称能源与政策瓶颈正威胁AI发展中心化AI正面临由电网、监管和隐私风险引发的危机，唯有去中心化算力网络才能化解。 Greg Osuri指出，2026年AI的发展阻碍并非源于模型本身，而是来自四大结构性瓶颈：碎片化监管、算力集中、物理能源限制和隐私漏洞。他阐述了由中心化驱动的这“四骑士”如何形成自我强化的循环，从而推高成本并引发公众反弹。为应对这一挑战，Osuri展示了具备生产就绪能力的去中心化解决方案，包括互联网级分布式训练、通过Akash Home Nodes利用闲置消费级GPU，以及将算力与住宅太阳能发电相结合。 2026年AI进步的核心瓶颈已从模型、数据与人才转向法律、算力、电力与隐私这“四骑士”，中心化结构正将这四大挑战锁定在相互加剧的恶性循环中。 (0:04) 碎片化监管正沦为合规固定成本陷阱，使资本雄厚的科技巨头形成天然护城河，同时模型权重也已成为地缘政治与版权诉讼的全新博弈点。 (0:29) 前沿模型动辄数亿美元的训练成本加剧了算力集中，但开放算力市场与数千万张闲置消费级GPU正展现出巨大潜力，且已有团队验证了基于普通民用网络进行分布式预训练的可行性。 (0:36) 数据中心激增的用电需求正撞上电网扩容的物理极限，结合住宅屋顶光伏的边缘计算节点不仅能缓解用能危机，还能从架构底层解决集中式云端的隐私与诉讼风险。 (1:38)

Reading summary

Akash Network founder Greg Osuri argues that artificial intelligence is structurally threatened not by model limitations, but by a self-reinforcing centralization loop spanning regulation, compute access, energy supply, and privacy failures. Regulatory Fragmentation and Incumbent Advantage (0:29) Osuri asserts that intensifying global regulation functions as a regressive fixed cost that favors deep-pocketed incumbents over smaller developers. The European Union has enacted its AI Act with potential fines up to €35 million or 7% of global turnover, while the United States operates under an uncoordinated patchwork where all 50 states have introduced AI bills, roughly 38 have enacted laws, and the U.S. Senate voted 99 to 1 against preempting state statutes. Meanwhile, model weights have become regulatory battlegrounds, illustrated by U.S. scrutiny of Qwen3 weights and proposed export controls in Beijing. Because compliance burdens scale sub-linearly, Osuri warns that heavily regulated environments entrench central cloud providers. To mitigate this barrier, he advocates for open-weight models, jurisdiction-agnostic compute markets, and protocol learning, where models are collectively trained across decentralized nodes rather than controlled by a single corporate entity. Compute Cost Escalation and Distributed Consumer Hardware (0:36) Frontier AI model training costs are compounding at roughly 3.5 times annually, jumping from approximately $80 million for GPT-4 to between $200 million and $500 million in 2026, with next-generation models projected at $1 billion to $3 billion. Osuri notes that 75% of frontier models originate in the U.S. and 90% stem from private industry, transforming advanced compute into an exclusive corporate moat while traditional cloud infrastructure runs at single-digit utilization rates. As an alternative to hyper-centralized clusters, Osuri highlights decentralized compute pools operating at 60% to 70% utilization. He points out that roughly 75% of the 150 million active Steam gamers own modern Nvidia GPUs—such as RTX 4090s and 5090s capable of running quantized 70-billion-parameter models—which sit idle around 20 hours per day. Demonstrating feasibility, research lab Pluralis trained an 8-billion-parameter model from scratch across 330 distributed internet nodes running consumer RTX 4090s, achieving 63% of the training efficiency of an interconnected H100 data center cluster. Grid Physical Limits and Distributed Residential Energy (1:38) Energy constraints are dictated by physical infrastructure timelines rather than software release cycles. Citing International Energy Agency projections, Osuri notes data center electricity consumption will rise from 215 TWh in 2024 to 945 TWh by 2030 and roughly 1,200 TWh by 2035, with Berkeley Lab estimating AI could absorb 12% of total U.S. generation by 2030. In 2025, U.S. utilities received 700 GW in data center interconnection requests—exceeding the nation's 477 GW 2023 peak capacity—while equipment lead times reached up to four years for transformers and eight years for transmission lines. This grid stress has produced localized friction: ERCOT faced 143 GW in requests against an 85 GW peak, PJM Interconnection wholesale power prices spiked 75.5%, and at least 12 state legislatures began weighing data center moratoria. Osuri proposes pairing decentralized inference hardware with residential rooftop solar installations, which generate up to 10 kW. For example, a deployment initiative by smart electrical panel maker Span targeted 6,000 to 8,000 homes to power local GPU nodes during peak generation periods. Privacy Failures and Edge Inference Deployments (1:42) Centralized cloud architectures create inherent legal and security liabilities for enterprise AI adoption. Osuri references the copyright dispute between The New York Times and OpenAI, which resulted in a court subpoena for approximately 20,000 ChatGPT conversations, including deleted sessions. He adds that an analysis of one million AI interactions revealed that 4% of enterprise user prompts and 22% of uploaded files contained confidential corporate information. Because data transmitted to centralized APIs remains subject to legal discovery and breach, Osuri contends that local and edge processing offers the only strict privacy guarantee. In a commercial deployment example, gaming hardware firm Razer processed 11,000 image generations over five days using decentralized consumer GPUs for an interactive campaign, reducing costs to $0.01 per generation compared to an estimated $0.03 to $0.15 on conventional cloud platforms. Osuri concludes that distributed compute and open-weight models have progressed from experimental concepts into production-ready alternatives capable of bypassing centralized bottlenecks. However, realizing this decentralized architecture at scale requires enterprises to systematically audit their vendor reliance, monetize idle consumer and corporate hardware, and actively support interoperable standards over fragmented state-level regulations.

精炼阅读摘要

Akash Network 创始人 Greg Osuri 指出，尽管开源模型让前沿智能变得可自由获取，但AI发展正被中心化模式衍生的“四大骑士”——监管、算力、能源与隐私所阻碍，四者相互强化形成系统性瓶颈，唯有转向去中心化与分布式架构才能打破垄断僵局。自增强闭环与碎片化监管合规成本 (0:29) Osuri 认为，制约2026年AI发展的核心瓶颈已从传统的数据与人才转向法律、芯片、电能与信任。这四项要素在中心化体系下构成了相互强化的闭环：隐私泄漏引发严苛监管，监管限定了算力的归属与访问权限，算力高度集中导致吉瓦级大型数据中心落地，进而引发严重的区域电网压力与公众抵制，最终倒逼政府出台更多限制措施。在全球监管层面，合规环境正急剧碎片化且偏向头部机构。欧盟《人工智能法案》规定了高达3500万欧元或全球营业额7%的违规罚款；美国联邦层面缺乏统一法案，50个州议会均提出了AI提案，其中约38个州已通过法案并累计产生近百部新规，参议院更以99比1投票反对冻结各州立法。Osuri 强调，监管属于次线性扩展的固定成本，50人的初创团队与5万人的大型企业面临着相似的合规负担，这种机制客观上将模型权重和合规门槛转变为资本巨头的专属护城河。为此，他提倡通过开放权重分发、与司法管辖区解耦的无许可算力市场，以及由多节点协同训练的“协议学习”来规避单点控制。算力成本指数增长与闲置硬件再利用 (0:36) 前沿模型训练成本呈现几何级跃升，每年复合增长约3.5倍，平均每5个月翻一番。训练GPT-4耗资约8000万美元，而2026年顶尖模型的训练成本已达2亿至5亿美元，下一代前沿系统预计将攀升至10亿至30亿美元。当前全球75%的前沿模型来自美国，90%源于私营企业，算力已完全演变为防御性商业壁垒。传统公有云服务商由于容量管理机制不透明，算力利用率通常仅停留在个位数，而开放算力市场的动态调度可将利用率提升至60%至70%。针对基础设施集中化的挑战，去中心化方案尝试激活沉睡的消费级硬件。Steam平台拥有约1.5亿活跃游戏玩家，其中四分之三配备了现代Nvidia GPU（如RTX 4090或5090），这些显卡每天闲置约20小时且足以运行700亿参数的量化推理模型；通过简单的引导盘启动家庭节点，即可在不新建数据中心的情况下大幅扩充全球推理能力。在预训练层面，前沿实验室Pluralis进行了验证：其利用330个互联网节点（单节点仅需单张RTX 4090与200 Mbps以下带宽）从零预训练一个80亿参数模型，取得了相当于互连H100集群63%的训练效率，证明了利用开放互联网节点执行分布式训练的可行性。电网物理瓶颈与户用光伏算力整合 (1:38) 与软件迭代不同，能源供应受制于电力物理规律和基础设施建设周期。国际能源署（IEA）数据显示，全球数据中心耗电量在2024年约为215太瓦时（TWh），预计到2030年将达945 TWh（占全球用电总量的3%），并在2035年逼近1200 TWh，其中AI是主要增长动力。伯克利实验室估计，到2030年AI将消耗美国总发电量的近12%。算力扩张速度与电网承载力之间出现了严重脱节。2025年美国公用事业机构收到的数据中心并网申请容量达700吉瓦（GW），相当于700座核反应堆的发电量，远超全美2023年477 GW的峰值负荷。然而，建设输电线路需4至8年，变压器交货周期长达4年。得州ERCOT电网收到了143 GW的申请（其峰值负荷仅85 GW），覆盖6500万人口的PJM电网批发电力价格暴涨75.5%，导致至少12个州议会开始考虑对数据中心实施建设冻结。对此，Osuri 指出家庭屋顶光伏平均可产生10千瓦电力，智能电气面板制造商Span正推动与Akash合作，计划在数千个家庭部署搭载工作站级GPU的节点，利用日间冗余太阳能运行本地计算并为户主创造收益，缓解集中供电压力。云端数据隐私脆弱性与本地边缘推理 (1:42) 隐私保障正在成为AI企业级落地的硬性准入门槛。Osuri 指出，任何上传至中心化云服务的数据在架构上都存在泄露与法律传票风险。在《纽约时报》诉OpenAI案中，法院通过传票调取了约2万名ChatGPT用户的对话记录，包含临时与已删除的对话内容。另有一项针对100万次企业AI对话的研究显示，4%的用户对话和22%的上传文件中包含企业敏感数据，使企业面临暴露风险。针对医疗与法律等高敏感行业，数据不离开本地局域网是确立隐私合规的前提。随着开源模型性能追赶专有模型，企业可在本地部署推理，并将溢出的非敏感算力需求分流至去中心化网络。以硬件制造商Razer为例，其在愚人节互动推广中通过去中心化消费级GPU集群，在5天内完成了1.1万次独立图像渲染，单次生成成本仅为0.01美元，相较于传统云端0.03至0.15美元的报价，验证了低成本边缘计算在实际商业场景中的可用性。 Osuri 总结强调去中心化AI已从概念验证走向生产落地，建议企业评估自身在监管、算力开支、电力依赖和隐私合规上的脆弱性，降低对单一闭源API的依赖并盘活闲置硬件；然而，大规模分布式网络在处理超大模型长距离同步时的网络延迟、节点高流失率以及跨司法管辖区的最终法律责任界定，仍是该模式全面替代中心化架构前有待检验的工程与制度难题。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台上，演讲者为 Akash Network 创始人 Greg Osuri。视频简介指出，当前 AI 面临的最大威胁并非来自失控的算法模型，而是源自构建、供电、治理和信任系统层面的多重瓶颈，本场演讲旨在剖析这些威胁如何相互强化，并提出韧性破局方案。 ## TL;DR 演讲围绕阻碍 AI 演进的“四骑士”——监管、算力、隐私与能源展开，指出这四大危机均由底层高度中心化的架构所驱动并形成恶性闭环。巨额合规成本与各州碎片化法律加剧了算力垄断，模型训练成本的指数级飙升与吉瓦级数据中心的电力负荷正逼近电网物理极限，而云端集中式架构更让企业隐私在司法传票面前荡然无存。面对这一困局，Osuri 展示了开源权重、分布式协议学习（如 Pluralis 实验）、家用光伏余电消纳以及消费级 GPU 算力池等去中心化实践，呼吁行业摆脱对单一云巨头的依赖，重构抗风险能力更强的 AI 基础设施。 ## Atmosphere & style 整场演讲节奏紧凑、警示意味浓厚，Osuri 以极其锐利且充满危机感的叙事开场，直接将台下听众拉入 AI 基础设施撞墙的严峻现实。演讲风格兼具宏观政策洞察与硬核技术底色，通过援引触目惊心的电价涨幅、司法传票案例与实测数据，在层层剖析系统性死局的同时，迅速转向务实且富有极客精神的分布式解法，给观众带来一种从技术压迫感走向开源破局的振奋与紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) 阻碍 AI 演进的上限不再是算法或人才，而是由中心化驱动的“四骑士”：监管、算力、隐私与能源。 - [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) 前沿开源模型性能已逼近顶尖商业模型并可免费下载，但开发者依然受制于五大科技巨头的 API 与电力许可，形成了中心化悖论。 - [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) AI 四骑士相互强化形成恶性闭环：隐私泄露引发严苛监管，监管加剧算力垄断，算力集中引发能源短缺与电价暴涨，能源危机进一步倒逼反弹与停建禁令。 - [02:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=149s) 欧盟《人工智能法案》罚金高达 3500 万欧元或全球营收的 7%，加剧了合规壁垒；而美国各州碎片化立法令跨区域合规成本呈次线性缩放，极度不利于初创企业。 - [06:30](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=390s) 前沿模型训练成本每 5 个月翻倍，明年单次训练成本预计将达到 10 亿至 30 亿美元，导致高校与初创公司被挤出前沿竞争，算力异化为垄断护城河。 - [06:30](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=390s) 全球游戏玩家闲置显卡构成了庞大未开发算力矿藏，哪怕仅激活个位数的百分比，也能在不新建数据中心的情况下提供爆发式推理算力。 - [06:30](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=390s) 前沿实验室 Pluralis 借助 330 个无许可家用节点成功跨网预训练出 8B 模型，打破了前沿模型训练必须依托高速局域机房的物理神话。 - [11:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=698s) AI 发展速度与电网基建周期根本错位：美国公用事业面临 700 GW 的数据中心并网申请，而输电线路与变压器扩建需耗费 4 到 8 年。 - [11:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=698s) 家庭屋顶光伏拥有高达 10 kW 的输出能力，通过结合 Akash Home Node 可就地将过剩太阳能转化为本地算力，规避电网过载风险。 - [16:26](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=986s) 司法传票可依法强制调取云端大模型的历史与已删除对话，绝对隐私的唯一标准是“数据永远不离开本地内网”。 - [19:07](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=1147s) 雷蛇商业实测证明去中心化消费级 GPU 在图像生成任务中将单张成本压至 1 美分，相比传统云服务的 3 至 15 美分大幅降低。 - [19:07](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=1147s) 企业应当推进基础设施多元化以摆脱对闭源 API 的依赖，通过本地私有部署开源模型与分布式算力规避中心化系统的系统性风险。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) 开场：最强开源模型问世与中心化悖论就在演讲开始的 9 天前，一家位于北京的初创公司刚刚向公众免费发布了有史以来规模最大的训练模型。演讲者 Greg Osuri 站在舞台上直言，这个最前沿、性能最强悍的模型就在手边，任何人随时都可以免费下载，“从本质上来说，它就像一个种子文件（Torrent）”。这引出了一个极其尖锐的核心问题：如果人类最前沿的智能已经是免费且人人唾手可得的，为什么我们感知到的 AI 世界，却依然牢牢被区区五家科技巨头所把持？为什么开发者必须卑微地排队申请 API 访问权限、排队等待电力配额，做任何事情都需要征得某些巨头的许可？Osuri 指出，这个悖论的根源可以用一个词来概括——“中心化”（Centralization）。过去十年里，整个行业普遍认为 AI 的发展上限取决于算法模型、数据量和顶尖人才；但到了 2026 年，现实规则被彻底改写，AI 正在被法律、芯片、电子以及信任这四道枷锁死死钳制。这正是阻碍 AI 发展的“四骑士”（The Four Horsemen）：监管（Regulation）、算力（Compute）、隐私（Privacy）与能源（Energy）。Osuri 特别澄清，这四个要素并不是指 AI 对人类社会造成的潜在危害，而是威胁 AI 自身技术演进的实质性阻碍。更关键的是，这四匹马绝非各自孤立行动，而是由“中心化”这一核心引擎死死咬合在一起的恶性循环：隐私泄漏引发严苛监管，监管法则决定算力归属并加剧算力垄断，算力的高度集中迫使巨头建设吉瓦（GW）级的超大规模数据中心，而极端的能源压力最终导致社会反弹，进一步倒逼更沉重的监管。 ```mermaid flowchart TD accTitle: AI 四骑士的中心化恶性闭环 subgraph Engine["核心驱动：中心化 (Centralization)"] direction TB P["隐私失效 (Privacy Failures)"] -->|引发严苛合规与执法| R["监管壁垒 (Regulation)"] R -->|推高门槛与管控，掌控算力所有权| C["算力集中 (Compute Concentration)"] C -->|建设吉瓦级园区，电网负荷承压| E["能源危机 (Energy Bottleneck)"] E -->|电价上涨、社会反弹与停建禁令| R end ``` ### [02:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=149s) 第一骑士：监管——合规高墙与碎片化泥潭谈及第一匹骑士“监管”，欧洲刚亮出的重锤足以令所有人倒吸一口冷气：欧盟最新落地的《人工智能法案》（AI Act）最高罚金达到了 3500 万欧元，或者企业全球年营业额的 7%。这一惩罚力度甚至超过了以往令人闻风丧胆的 GDPR，“有人甚至调侃，欧盟靠开出违规罚单赚的钱，恐怕都要超过他们的税收收入了”。其时间点更是充满戏剧性：就在全球从业者汇聚拉斯维加斯参加 AI 大会时，布鲁塞尔在 8 月 2 日正式激活了高风险与透明度等核心合规条款。Osuri 坦言自己并非宣扬“监管即邪恶”，人类确实需要合理规范，但监管带来的真正威胁往往更加隐蔽和致命。在大西洋彼岸的美国，情况则演变成一场混乱的法律拼贴实验：全美 50 个州全部提出了各自的 AI 立案法案，其中约 38 个州已正式签署立法，累计产生了约 100 部新法律；参议院更是以 99 比 1 的悬殊比例投票反对冻结各州的 AI 立法权。白宫虽然出台了强力行政命令，但在法律界看来根本缺乏实质约束力。与此同时，训练数据的版权之争日趋昂贵，Anthropic 刚刚遭遇了史无前例的天价版权侵权和解案。更惊人的是，“模型权重”（Model Weights）本身已经成为全球地缘竞争的崭新前线——华盛顿正在严审最新模型权重的合法来源，北京在酝酿对权重实施出口管制，而加州州长此前甚至需要动用否决权去阻止一项试图限制模型训练算力上限的法案。这种“人人想管，却没人知道该怎么管”的局面让行业陷入泥潭。监管背后隐藏着极其残酷的不对等数学逻辑：合规成本属于固定成本，具有次线性缩放（sub-linear scaling）特征。这意味着一个 50 人的初创团队，在合规审计上需要承担与 50,000 人科技巨头几乎相同的法务开支，这必然导致唯有资产负债表雄厚万分的巨无霸企业才能在这场耐力赛中胜出。打破这一死局的唯一护城河是“分布式”： - **开源权重**：将尖端能力下放至大众手中，摆脱单一商业公司对 API 接口的绝对垄断。 - **无许可算力市场**：以跨司法管辖区的架构规避单一地域的政策阻断。 - **协议学习（Protocol Learning）**：模型被拆解并分布在互联网上的不同独立节点协同训练，没有任何个体或机构能够独占模型与权重。 ### [06:30](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=390s) 第二骑士：算力——天文数字的成本与闲置显卡突围第二匹骑士是让所有开发者倍感窒息的“算力成本”。业内的共识是，算力训练开销大约每 5 个月翻一番，年复合增长率高达 3.5 倍。回顾当年，训练 GPT-4 大约花费了 8000 万美元；而 2026 年训练最新模型的成本已经暴涨至 2 亿到 5 亿美元之间；预计到了明年，打造一个顶尖的前沿模型将一举迈入 10 亿至 30 亿美元的天价门槛。在这样的资本绞肉机面前，具备入场资格的机构名单正急剧萎缩。这种集中不仅体现在资本上，还体现在地理和机构分布上——全球约四分之三的前沿模型诞生于美国，约 90% 源于大型商业公司而非高校科研院所；学术界已被彻底挤出前沿竞争舞台，算力不再仅仅是基础设施投入，而已异化为牢不可破的商业护城河。为了打破这种垄断，Osuri 介绍了自己打造了 6 年的去中心化开放算力市场 Akash Network。与传统云服务商个位数的资源利用率相比，去中心化算力市场凭借公开透明的库存管理，利用率通常可达 60% 至 70%。除了数据中心资源，更庞大的算力金矿隐藏在千家万户： - 全球主流游戏平台（Steam）上在任意时刻都有约 1.5 亿玩家在线，其中四分之三的人配置了 NVIDIA RTX 4090 或 5090 显卡。 - 这些消费级显卡性能惊人，已能流畅运行 700 亿参数（70B）的量化模型，而它们每天平均闲置长达 20 个小时。 - 只要调动这些闲置消费级显卡中个位数的百分比接入网络，就能在无需新建任何数据中心的前提下，实现全球推理算力的指数级爆发；用户只需通过 Akash Home Node 使用一个 U 盘启动设备，游戏主机即可摇身一变为收益节点。在面对“分布式节点只能做简单推理，无法承担需要极高带宽互联的复杂训练”的传统质疑时，Osuri 用事实做出了强力反击：2026 年 7 月 14 日，前沿实验室 Pluralis 在 Arweave 上公布了一项重大突破——他们成功通过分布在互联网各处的 330 个无许可节点，从零预训练出了一个 80 亿参数（8B）的完整模型。参与节点的门槛仅仅是一张 4090 显卡与不足 200 Mbps 的家用网络带宽，其协同训练效率竟达到了高速互联 H100 集群的 63%。“互联网本身，如今已经进化成了一整台超级计算机集群。” ### [11:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=698s) 第三骑士：能源——电网物理极限与分布式屋顶光伏第三匹骑士是无法讨价还价的客观物理法则——能源。根据国际能源署（IEA）发布的数据，全球数据中心的用电量在 2024 年约为 215 太瓦时（TWh），到 2030 年预计将飙升至 945 TWh，仅新建数据中心的电力消耗就将吞噬全球总用电量的 3% 左右；这一数字在 2035 年更将逼近 1200 TWh，其中 AI 是最主要的电量吞噬者。聚焦到承载了全球近半数数据中心的美国，伯克利实验室（Berkeley Lab）给出的预测中值显示，到 2030 年，美国近 12% 的发电量将被 AI 吞噬，这在电网规划层面上几乎是不可承受之重。更深层的危机在于开发节奏的根本错位：2025 年，美国公用事业公司收到的数据中心并网申请总量竟高达约 700 吉瓦（GW），这相当于约 700 座核反应堆的发电容量，甚至直接超越了美国在 2023 年达到的 477 GW 全国峰值负荷用电需求。一边是 AI 以日新月异的软件速度进行指数级扩张，另一边却是电网传输线路长达 4 至 8 年的漫长建设周期，甚至连重型变压器的供货排期都排到了 4 年之后。“AI 正以软件的速度狂飙，而能源只能以基建的龟速前行，两者存在着不可调和的底层矛盾。” - **得州电网（ERCOT）危机**：面临 143 GW 的并网申请负荷，而其历史峰值需求仅为 85 GW，申请量接近翻倍。 - **民生与电价反噬**：涵盖 6500 万美国居民的最大区域电网运营商 PJM，其电力批发市场价格暴涨了 75.5%，直接导致民众用电账单飞涨；已有约 12 个州因能源危机紧急出台了针对新建数据中心的暂停禁令（moratoriums）。 - **分布式能源解法**：全美最大的发电单元正是普通家庭。家庭屋顶光伏系统能够轻松输出约 10 千瓦（kW）的充沛电力，足以驱动一套高水准推理节点。通过与智能电表面板公司 Span 合作，Akash Home Node 计划在 6 个月内铺开 6,000 到 8,000 个家庭节点，配备 RTX 4090 或 RTX 6000 Pro 等专业显卡，直接就地消纳过剩太阳能，将原本造成电网拥堵的能源转化为实打实的本地算力。 ### [16:26](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=986s) 第四骑士：隐私——云端暴露的风险与本地化最后防线第四匹骑士是关乎公众信任的基石——隐私。在《纽约时报》起诉 OpenAI 侵权的诉讼案中，法院向 OpenAI 下发传票，强制要求调取并公开多达 2 万条用户历史对话记录。尽管 OpenAI 自身也抗议这是一场“隐私噩梦”，但在法律机器面前却根本无力抗拒，甚至连用户标记为临时对话以及早已点击删除的私密记录也被迫全盘上交。Osuri 强调，这并不是因为云服务厂商心怀恶意去偷看隐私，而是因为现有的中心化云架构在底层设计上就注定了“一旦数据被发送至云端，隐私便不复存在”。相关调研数据同样触目惊心：在一项针对 100 万条企业交互对话的抽样研究中，约 4% 的日常员工对话包含了企业的绝密核心数据；而在用户上传至 ChatGPT 等云端平台的各类文档中，竟有 22% 包含企业敏感机密。对于医疗、法律等强监管行业而言，面对潜在的司法传票风险，使用中心化云端大模型几乎等同于行业自杀。隐私问题已然从单纯的技术考量，变成了决定 AI 落地生死的关卡（deployment gate）： - 科技巨头兜售的加密算法、机器遗忘（Machine Unlearning）等概念无法提供终极安全保障。 - 真正无可辩驳的绝对隐私，其唯一金标准就是“数据永远不离开本地内网”。 - 伴随开源小钢炮模型能力的全面井喷，企业完全可以在本地设备或通过 Akash 这类去中心化计算网络闭环调度算力，在根本上抹除数据外泄与法律传票调取的系统性风险。 ### [19:07](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=1147s) 闭环系统与破局实践：从科学实验到现实落地在演讲的最后阶段，Osuri 再次强化了中心化系统的恶性闭环逻辑：隐私失效招致强力监管，监管手段加剧算力垄断，集中算力催生吉瓦级超大园区并引发能源危机，能源短缺诱发高昂电价与社会停建禁令，这反过来又刺激监管出台更苛刻的限制措施。但打破这套枷锁的分布式工具已经齐备： - 像北京团队开源的模型已达前沿水平，任何开发者都可以直接下载并在本地部署，彻底打破了由 OpenAI 或 Anthropic 筑起的护城河。 - Pluralis 于 7 月 14 日证实的无许可跨网训练技术，击碎了“大模型训练必须扎堆部署在同一物理机房超高速局域网内”的神话。 - 硬件大厂雷蛇（Razer）的商业化实战成果打破了去中心化算力属于“科幻玩具”的偏见：在 2026 年 4 月 1 日发起的一场全网宠物全息图像生成裂变营销活动中，雷蛇在 5 天内利用去中心化消费级 GPU 生成了约 11,000 张独一无二的图像，将单张生成成本暴力压缩至 1 美分，相比传统云厂商 3 到 15 美分的报价缩减了数倍。 ```mermaid flowchart LR accTitle: 中心化云与去中心化架构的性能与成本对比 subgraph Centralized["传统中心化云端模式"] direction TB C1["云端超大集群 API"] --> C2["数据中心电力告急<br/>PJM 批发电价 +75.5%"] C2 --> C3["企业机密存在传票风险<br/>雷蛇测试成本: 3-15 美分/图"] end subgraph Decentralized["去中心化与本地分布式模式"] direction TB D1["开源权重 + 家用/消费级 GPU"] --> D2["消纳家庭光伏 10kW 余电<br/>算力利用率 60-70%"] D2 --> D3["数据不离网绝对隐私<br/>雷蛇测试成本: 1 美分/图"] end ``` Osuri 呼吁参会者拿出手机拍下演讲总结的风险评估框架：周一回到工作岗位后，请立刻对照“监管、算力、能源、隐私”四大维度对自身业务进行风险体检。企业今年应加速基础设施的多元化进程，不再一味盲目依赖外部闭源 API，尝试在本地私有化部署开源大模型；如果有闲置的闲散显卡，不妨将其接入 Akash Home Node，赚取切实的闲置收益；在宏观层面上，他恳请全行业共同推动具备互操作性的统一法律规则，拒绝 50 个州各自为政的破碎法网，并积极拥抱对电网友好的分布式光伏与算力设施，唯有如此，全行业才能齐心协力规避这场看似近在咫尺的“AI 启示录大危机”。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=4s) 从本质上来说，它就像一个种子文件（Torrent）。 — Greg Osuri > [02:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=149s) 有人甚至调侃，欧盟靠开出违规罚单赚的钱，恐怕都要超过他们的税收收入了。 — Greg Osuri > [02:29](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=149s) 人人想管，却没人知道该怎么管 — Greg Osuri > [06:30](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=390s) 互联网本身，如今已经进化成了一整台超级计算机集群。 — Greg Osuri > [11:38](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=698s) AI 正以软件的速度狂飙，而能源只能以基建的龟速前行，两者存在着不可调和的底层矛盾。 — Greg Osuri > [16:26](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=986s) 一旦数据被发送至云端，隐私便不复存在 — Greg Osuri > [16:26](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=986s) 数据永远不离开本地内网 — Greg Osuri > [19:07](https://www.youtube.com/watch?v=Cc4rGyh2XM4&t=1147s) 大模型训练必须扎堆部署在同一物理机房超高速局域网内 — Greg Osuri ## Mentioned Greg Osuri, Akash Network, Akash Home Node, OpenAI, Anthropic, Pluralis, Arweave, 雷蛇（Razer）, 欧盟（EU）, 《人工智能法案》（EU AI Act）, GDPR, 《纽约时报》（The New York Times）, 国际能源署（IEA）, 伯克利实验室（Berkeley Lab）, 得州电网（ERCOT）, PJM, Span, NVIDIA, Steam, GPT-4, ChatGPT ## Video description Greg Osuri - Founder, Akash Network AI’s greatest threats may not come from rogue models, but from the systems required to build, power, govern, and trust them. Regulation is fragmenting across borders, advanced compute is concentrating in fewer hands, energy demand is colliding with infrastructure limits, and privacy concerns are eroding public confidence. Together, these four forces could slow innovation, raise barriers to entry, and determine who gets to shape the future of AI. This session will examine how these threats reinforce one another and explore practical solutions, including adaptive policy, decentralized compute, new energy strategies, and privacy-preserving architectures. Attend to gain a clear framework for confronting the Four Horsemen and building an AI ecosystem resilient enough to survive them. [Source](https://www.youtube.com/watch?v=Cc4rGyh2XM4)

Source quotes

[0:04] Greg Osuri: 从本质上来说，它就像一个种子文件（Torrent）。 [2:29] Greg Osuri: 有人甚至调侃，欧盟靠开出违规罚单赚的钱，恐怕都要超过他们的税收收入了。 [2:29] Greg Osuri: 人人想管，却没人知道该怎么管 [6:30] Greg Osuri: 互联网本身，如今已经进化成了一整台超级计算机集群。 [11:38] Greg Osuri: AI 正以软件的速度狂飙，而能源只能以基建的龟速前行，两者存在着不可调和的底层矛盾。 [16:26] Greg Osuri: 一旦数据被发送至云端，隐私便不复存在 [16:26] Greg Osuri: 数据永远不离开本地内网 [19:07] Greg Osuri: 大模型训练必须扎堆部署在同一物理机房超高速局域网内

Source quotes

[0:29] Greg Osuri: The EU recently activated its AI Act, carrying maximum penalties of €35 million or 7% of global turnover—a stick even bigger than GDPR. [0:29] Greg Osuri: A 50-person startup faces roughly the same regulatory burden as a 50,000-person enterprise, ensuring that hyper-capitalized incumbents with the biggest balance sheets inevitably win under heavy regulatory regimes. [0:36] Greg Osuri: Compute costs are compounding at roughly 3.5 times per year, doubling approximately every five months. [0:36] Greg Osuri: Unlike traditional cloud providers where utilization lingers in the single digits due to opaque capacity management, open marketplaces operate at 60% to 70% utilization because open inventory allows for far better planning. [0:36] Greg Osuri: Mobilizing even a single-digit percentage of these GPUs via tools like the Akash Home Node expands global inference capacity exponentially without building new data centers. [0:36] Greg Osuri: The run achieved 63% of the training efficiency of an interconnected H100 cluster, proving that the cluster can indeed be the internet itself. [1:38] Greg Osuri: In 2025, U.S. electric utilities received 700 GW of data center interconnection requests—representing approximately 700 nuclear reactors worth of capacity and far exceeding the entire country's 2023 peak capacity of 477 GW. [1:38] Greg Osuri: These home nodes consume excess solar energy during peak generation hours, solving localized power constraints while generating income for homeowners. [1:42] Greg Osuri: A study analyzing one million AI conversations revealed that 4% of corporate user chats and 22% of uploaded files contained sensitive corporate data—leaving enterprises 'one subpoena away from being disclosed to the public.' [2:04] Greg Osuri: Decentralized AI is an active, production-ready reality rather than a speculative science project.

Original transcript

[0:04] So, 9 days ago a startup in Beijing released the largest ever model trained for free to the public. [0:16] In fact, I have it right here. This is the latest and the greatest model ever to be trained [0:23] available for anyone to download anytime. And it is functionally a torrent. [0:34] Now, that begs the question. If the frontier intelligence is free for anyone to use [0:41] why does it feel like there are only five companies that control AI? [0:46] And you have to wait in line for API access or for energy and you need someone's permission. That's [0:54] the question for this talk. And the answer I'll I'll argue comes down to one word. Centralization. [1:05] Now, for most of the last decade, we believed AI was limited by models, data, and talent. [1:14] But in 2026, the reality is very different. AI is limited by law chips electrons, and trust. [1:25] Regulation, compute privacy and energy. These are the four horsemen. And I want to be precise. [1:33] These are the threats to AI's progress and not necessarily the harm AI is causing. [1:40] And here, the part that took me the longest to see is that these are not four separate horsemen. [1:48] These are in fact all interlinked with an engine that powers it, which is centralization. [1:56] Privacy failures lead to regulation. Regulation shapes who owns compute, which gets concentrated, [2:07] and concentration of compute leads to gigawatt scale deployments for energy, [2:14] and the energy stress creates a backlash. And this is a loop. And the engine for this loop [2:21] is centralization. We'll come back to that in a second, but first, the first horseman. [2:29] 35 million euros, or 7% of your global income, [2:35] is a maximum penalty for the latest AI Act introduced by EU. [2:42] For comparison, that's bigger than bigger stick than GDPR. Some might argue that [2:48] EU makes more revenue in penalties than in tax revenue. And the timing is almost theatrical, [2:55] right? The Act's main obligations, which is which are high-risk rules and transparencies, [3:02] kind of like all over the place, took effect on August 2nd. While we are here in Vegas, [3:11] at an AI conference, Brussels just activated the most comprehensive AI regulation ever. [3:19] I'm not here to tell you regulation is evil. We need some regulation, but the threat is more [3:25] subtle than that. And while we have one law in Europe, we have several experiments in in America. [3:35] All 50 states have introduced some form of legislation for AI. About 38 states have enacted [3:43] on on on uh on these proposals. About 100 new laws have been created, and the Senate voted 99 to 1 [3:54] against freezing state's AI laws. Of course, we had the administration come out with a very strong [3:59] executive order, but my friends in in the in the legal circles don't think it has any strength. [4:07] What it comes down to is a bunch of patchwork in America. [4:12] Meanwhile, the fights over training data got really expensive. We had the largest [4:17] ever copyright infringement settlement happen between Anthropic. And here's a new front. [4:25] The model weights themselves are the new new new front in this global war on AI. [4:34] Washington is now investigating on the legitimacy of the latest Kimiko 3 weights. [4:41] And as to how they derived these weights. And we have Beijing, which is contemplating [4:47] on export controls for weights. Who knew weights are the new front line? [4:54] And even my home state of California, we had our dear governor veto a new bill that wanted [5:00] to cap how much of compute you can use to uh to train a model. Again, these these caps are [5:09] even though it's vetoed, the the point is everybody wants to regulate, no one agrees [5:14] on what to regulate. And it's a mess. Now, here's the uncomfortable math of regulation. [5:22] Regulation is fixed cost, and like all fixed cost, it it scales sub-linearly. Meaning, a 50-person [5:30] startup has the same cost as a 50,000 people startup. What that means is the biggest players [5:38] with biggest balance sheets will will inevitably win the AI race when it comes to regulation. [5:45] And the hedge is distribution. Open weights keep capability in many hands, right? There's [5:52] no single company to to control the API. A permissionless compute marketplaces, for example, [5:59] are jurisdiction agnostic. And here's a new idea I want to plant now because we're going to come [6:07] back later as part of this presentation called protocol learning. Where the model is trained [6:13] across different nodes on the internet where no single individual or a company actually [6:19] owns the model all the weights, but collectively they can use and train and and use AI. Now, the [6:30] the second paradigm is is the compute. [6:36] As you can see, the it's very well understood the rate at which the the cost is is is gaining is [6:45] really roughly doubling every every 5 months, right? The costs are compounding at about [6:51] 3 and 1/2 times per year. So, GPT-4 cost about $80 million to train. The latest models are running [7:00] this year anywhere between 2 to 500 million dollars to train. And it's expected by next year [7:06] it takes about a billion to 3 billion dollars to train the state-of-the-art models. At that price, [7:13] list of organizations that can play shrinks significantly, right? And it's geographical [7:20] concentration, too. Most of the compute, about 3/4 of the frontier models, are coming from US. [7:27] About 90% of um uh of of frontier models came from from industry, not in not through academia, [7:35] right? So, academia is no longer producing any of the frontier models. It's all industry. [7:41] And and compute becomes stops being an input, it becomes a real motor. [7:49] And the solution? I mean I'll disclose my bias up front. I've been working on a marketplace [7:55] for about for about 6 years. I've been building one. So, don't take my word for it, but judge by [8:00] who's using it. So, we built a open marketplace called Akash that bootstrapped Prime Intellect, [8:06] News Research, uh Brev before they got acquired by Nvidia, Venice AI. They're all dependent on [8:14] open marketplaces to bootstrap themselves. And this is the story and why open marketplaces are [8:19] very important for startup innovation and for small business innovation, right? Uh why does it [8:26] work? Because traditional clouds utilization runs in single digits, whereas uh open marketplaces [8:32] the utilization generally runs anywhere from 60 to 70% because open access means open inventory. And [8:40] planning can be done when you have an open inventory much easier than opaque opaque APIs. [8:46] And here's another stat. I I really love this this number. They're about [8:52] 150 million uh gamers on stream that are gaming at any at a given time. [9:00] And 3/4 of them three out of four folks that are uh on stream uh do have an Nvidia GPU, [9:10] either a 4090 or 5090. Now, these GPUs are providing to be proving to be extremely capable. [9:16] They can run a 70 billion quantized model fairly effectively now on these 4090s and 5090s. [9:24] And that's serious inference machine for about that's free as unused for about 20 hours a day. [9:33] And if a single digit percentage of these cards can come to to the market [9:40] that grows our inference capability exponentially with zero new data center build outs. [9:50] So with Akash home node for example, you boot from USB drive and the gaming rig becomes a [9:56] provider node and node and you earn when it's not used. Now fine, you can say Greg, well, inference [10:04] is one thing, but what about training? Training is computationally heavy, you need co-located [10:09] compute. Well, that was mostly true until about July 14th. A lab, a frontier lab called Pluralis, [10:19] pushed something remarkable to our Arweave. They were able to train 8 billion parameter model [10:26] pre-trained from scratch permissionlessly across 330 nodes across the internet. [10:35] And the minimum requirement to participate in this training run was a 4090 with sub 200 mega [10:41] megabits a second bandwidth. And this training run was was conducted at about 63% efficiency [10:50] uh as you would run on an H100 uh cluster that's interconnected. Sure, 8 billion is not frontier, [10:57] but we we showed what's possible. Now imagine a single digit percentage of the 150 million [11:05] nodes that I talked about coming and participating in a training run over the internet. [11:12] Um over the internet uh in their downtime. And these are in homes, they have power, they have [11:19] energy, and they have cooling and whatnot. So, without net without a new [11:24] data center build out, we can actually train a frontier model if these compute can come online. [11:30] Now the the the the keynote here is the cluster is now the internet. [11:38] And the third horseman, this is my favorite, energy. [11:42] This one you cannot really negotiate with. It's just physics, right? [11:47] These are IEA numbers, International Energy Association. Um the global data center usage in [11:54] 2024 was estimated to be about 215 terawatt hours. That by 2030 is expected to be 945 terawatt hours. [12:07] That's about 3% of entire of global uh energy usage. That's just going for new data centers. [12:16] Uh that's heading to about 1,200 uh terawatt hours by 2035. And AI is named as a biggest driver. [12:24] Zoom into US, which hosts nearly half of it. Uh Berkeley Labs latest midpoint uh that about almost [12:32] 12% of US energy uh production will be used by AI by 2030. These are unsustainable numbers, right? [12:41] These are Those are grid planning uh numbers. You got to upgrade the grid to support this. Now, [12:50] now here's the collision. In 2025, US utilities received about 700 gigawatts [12:58] of data center interconnection requests. That's roughly about 700 nuclear reactors [13:04] worth of new capacity that wants to connect to the grid. For that scale, it's it's more than [13:11] the entire country's peak capacity in 2023. We had a peak capacity about 700 gigawatts in 2023. [13:21] So, the new uh about about sorry, my apologies, about 477 uh at at our peak in 2023. [13:29] The new requests are way above our peak. And that's really um uh the the the scale of the [13:37] the interconnect requests we're looking at, right? Now, on the other other land, the the the [13:41] transmission networks takes anywhere from 4 to 8 years to build out. Let alone the transformer lead [13:47] times. I believe the the number I saw somewhere was about 4 years to get a new transformer. So, [13:56] the the point here is AI is scaling at software speed, and energy is scaling at infrastructure [14:05] speed. These are fundamentally in incompatible scaling uh metrics. [14:13] Uh ERCOT alone, which is the the grid for um for Texas, uh we have a 145 143 GW request [14:22] compared to 85 GW request at at peak. So, it's almost almost almost double, right? In terms of [14:30] how much new capacity is being requested from the grid or or are coming onto the grid. [14:36] So, where [14:40] And and if you look at what that means to the economy, right? PJM, which is the [14:46] the largest um energy producer, it it's in it covers about 65 million Americans, [14:53] experienced about a 75.5 uh percent increase in their cost in the wholesale prices. And that's [15:00] really making consumers unhappy. That's leading to a lot of moratoriums. We have about 12 states that [15:06] are asking for data center moratoriums because of the energy costs. And this is this number did not [15:12] come from an activist. This number came from PJM's independent monitor itself. So, this is a real [15:19] problem, and there's really no way to scale our energy uh landscape. Well, what's the solution? [15:26] Well, if you think about it, we have the biggest energy producers in America. It's called a home. [15:33] A home, a rooftop solar can easily produce about 10 kilowatts that can power a inference cluster. [15:41] And Span Akash Home Node is a attempt to reach reach the home. Span, which is a panel company, [15:48] predicted that they can reach up to 6,000 homes 8,000 homes in about 6 months and actually install [15:56] RTX 4090 RTX 6000 pros, which are extremely good inference machines, and actually [16:02] consume energy the excess energy from the sun to power to power these inference clusters. [16:07] So, we have a solution already to this energy problem. And like I mentioned earlier, we have [16:13] a solution for training over the internet. So, it's not all that bad if you actually [16:18] can think about distributed networks rather than concentrated networks. I'll try to hurry up here. [16:26] Um and the fourth horseman is privacy. Um ChatGPT famous lawsuit between ChatGPT and and and New [16:38] York Times where New York Times sued ChatGPT for for for for for taking their content and [16:45] ChatGPT was court-ordered to disclose about 20,000 conversations they got subpoenaed. [16:51] ChatGPT thinks it's a it's a Well, OpenAI thinks it's a it's a privacy nightmare and it is. [16:58] And they have no way not to give them the data. The point being the it's not that someone is evil [17:05] that they're going to read your chats or whatnot. It's just that the system is designed in a way [17:10] that nothing that you send to the cloud is ever private. Even the temporary and deleted chats [17:17] were disclosed in this subpoena. Um and we look at um we look at some of the interesting [17:25] studies that have been done. There there was a study that was done about [17:28] They took about a million conversations and they they found out about 4% of these conversations [17:34] conversations from from companies included sensitive data from the company. [17:39] And 22% of the uploaded files to these chat GPTs and the cloud services [17:45] included sensitive company data. That's one subpoena away from being disclosed to the public. [17:51] And you know, Clearview and and and the fines were very obvious. Um Clearview got fined about [17:58] you know, tens of bi- millions of euros uh and pounds from from from European regulators. [18:04] And in some industries like like healthcare and and um law, it's nearly impossible to to use [18:13] uh anything remote, anything on the cloud because of the subpoena risk, right? Um [18:20] and now privacy is a deployment gate. Um So, it's it's no longer an option for a lot of the [18:26] companies to put stuff in the cloud. And this the the solution to the problem is, well, the it's not [18:30] the strongest The strongest privacy guarantees uh means uh it should not leave the network. The The [18:36] best private solution Privacy companies and yeah, companies will give you They'll sell you all kinds [18:41] of new techniques to encrypt and unlearn whatnot, but ultimately the best privacy is something that [18:47] never leaves your your network. And fortunately, we have an incredible uh open models that can run [18:54] locally uh where and you have incredible ways to offload your computer your unused compute onto [19:00] networks like Akash. So, privacy can be solved if you can bring compute back home. [19:07] Uh next slide, please. Uh All right, so I want to tie it all back together. If you look at it [19:11] um it in if you look at this uh the system, the privacy of failures trigger enforcement, [19:17] and that's regulation. Regulation reaches for export controls and compliance regimes, [19:22] uh decides who gets compute, concentrated compute demands uh gigawatt campuses that that leads to [19:29] the energy crisis and energy stress produces rate hikes and backlashes and moratoriums and which is, [19:35] you know, leads to more regulation. So, you can see all this is a closed loop and and the [19:41] problem is centralization. And right now we have the tools to decentralize. Next slide, please. [19:46] And, you know, you all saw the pictures about decentralize everything, but instead of a picture, [19:51] I'm just going to give you timelines. Kimik A3 is a frontier model that can be downloaded, [19:56] that can be run anywhere. You don't need Anthropic or OpenAI to run frontier anymore. [20:02] Uh July 14th is when Pluralis released a paper on frontier distributed training over the internet. [20:08] So, training that needs to be collocated is no longer uh a reality. Training can be done over the [20:15] internet. And HomeNode has been in beta uh since February. And this is not some science fiction or [20:22] a science project. This is real. We have a company called Razer uh that ran a viral AI campaign. [20:28] Uh they were able to uh on April 1st, they were able to take your pet's picture and create this [20:33] holographic image. Uh right about 11,000 unique images in about 5 days. Uh they were able to uh [20:42] get the cost up to 1 cent per generation. For comparison, they were paying about 3 to 5 [20:46] 3 to 15 cents on the cloud, and all this was run on consumer GPUs. This is a real [20:52] enterprise company actually running real campaigns and seeing real results. So, none of the thing [20:58] I told you about is a science project. It's actually real right now. Next slide, please. [21:04] So, this is is I want you to photograph. When you go back on Monday, I want you to look at your risk [21:11] metrics across the four four horsemen I talked about. Look at regulation, look at computer, [21:17] look at energy, and look at privacy, and you can understand where your risk is. [21:21] This year, try to diversify. Try to instead of using HRGPT and and open AI for or or Anthropic [21:28] for everything, try to host a model locally, and you yourselves can see how incredible that is. [21:34] And you know, when you have unused GPUs, put it on an Akash home node, so you can actually get some [21:39] you know, return of your investment that can be measured easily. And this decade, [21:43] please back interoperable regulation. We don't want 50 states to have their own law [21:50] own laws and regulations that no company can navigate, and invest in distributed grid-friendly. [21:57] Maybe get a GPU, maybe get some solar, and and and and put it. So, collectively, we can actually [22:04] avoid the AI apocalypse. And with that, I I'm out of my time, so thank you very much.