OptimalNets Unveils Software to Double Multi-Tenant GPU Speeds

ON.AI - Faster Jobs. Same GPUs. More Revenue with OptimalNets — Ai4

Ai4

Standard networks degrade multi-tenant AI cluster performance by treating packets uniformly, but autonomous job-aware routing can double workload completion speeds and expand billable capacity on existing silicon.

AI-assisted summary of ON.AI - Faster Jobs. Same GPUs. More Revenue with OptimalNets — original source. Check the original for full context.

Source published: 2026-08-21

Detailed notes

## Background Published on YouTube by Ai4 on August 21, 2026, this presentation features OptimalNets CEO Lior Friedman discussing how network-level job awareness can accelerate AI workloads by 2x to 5x on multi-tenant clusters. The session highlights how infrastructure providers can turn committed GPUs and power into increased billable capacity without adding new hardware or increasing power consumption.

## TL;DR Standard multi-tenant AI clusters suffer from 'job-blind' network switching, which routes packets based on generic averages and forces disparate AI workloads to stall together. OptimalNets introduces ON.AI, an autonomous data path control plane running on top of existing network switches to dynamically optimize fabric parameters based on live workload telemetry. This software-driven approach cuts Job Completion Time (JCT) by 2x to 3x, boosts tokens-per-watt efficiency by 30% to 50%, and unlocks substantial billable capacity for GPU cloud operators. By adapting continuously in software, ON.AI future-proofs multi-year hardware investment cycles against rapidly evolving AI model architectures.

## Atmosphere & style The presentation delivers a crisp, pragmatic, and executive-level technical pitch tailored for infrastructure architects and data center operators. Friedman presents with authoritative clarity, systematically dismantling conventional networking assumptions before laying out a structured, data-backed solution. The tone feels grounded, business-savvy, and engineering-driven, moving logically from structural pain points to clear ROI and deployment pathways.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) Traditional network fabrics lack workload context, breaking cohesive AI jobs into disjointed packets and causing all concurrent jobs to crawl at a mediocre pace. - [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) Multi-tenant GPU clouds cannot use static manual network tuning, forcing them to configure for a generic 'average job' that leaves infrastructure optimized for none. - [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) ON.AI functions as an autonomous, software-defined control plane on existing network hardware, requiring zero additional capex, GPUs, or power draw. - [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) Implementing job-aware network control achieves a 2x to 3x reduction in Job Completion Time (JCT) across AI workloads. - [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) Cluster efficiency increases by 30% to 50% more tokens generated per dollar per watt. - [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) The ON.AI engine operates via a four-stage continuous feedback loop: telemetry collection, traffic analysis, parameter optimization, and cyclic network injection. - [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) A three-phase deployment model—Recommendation, Semi-Autonomous, and Fully Autonomous—allows operators to validate optimizations before granting automated control. - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) Data center providers can recapture up to 50% more billable capacity from their committed GPU investments and long-term power agreements. - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) ON.AI supports open networking architectures to maximize price-to-performance while avoiding proprietary vendor lock-in. - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) Dynamic software adaptation future-proofs physical AI infrastructure, bridging the gap between frequent AI model releases and 3-to-5-year hardware replacement cycles.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) The Challenge of Blind Networks in Multi-Tenant AI Clusters

Traditional networking architectures fundamentally fail to understand the workloads they transport. In standard enterprise and cloud environments, hundreds of employees across diverse departments—from marketing to R&D—as well as autonomous AI agents are constantly submitting disparate AI tasks to shared clusters. Conventional AI infrastructure takes these cohesive workloads, breaks them down into billions of disjointed packets, and routes them across thousands of GPUs. Because standard network hardware treats all packets uniformly, every workload advances at roughly the same mediocre pace, forcing virtually every user to wait longer than necessary for their results.

OptimalNets addresses this structural inefficiency with **ON.AI**, an autonomous data path control plane designed specifically for AI clusters. Rather than allowing all tasks to crawl forward simultaneously and finish in a delayed cluster at the very end, introducing job awareness enables the network to prioritize and complete individual jobs as quickly as possible. This paradigm shift returns actionable results to users earlier, recaptures millions of dollars in idle GPU capacity, and significantly boosts business productivity.

- AI infrastructure natively understands packets rather than complete computational jobs. - Multiplexing unrelated workloads across identical, uncalibrated paths causes broad job latency. - Introducing job awareness ensures workloads complete sequentially and rapidly rather than stalling together.

```mermaid flowchart TD accTitle: Standard Blind Switching vs. Job-Aware ON.AI Data Path accDescription: Comparison showing how traditional networks mix disjointed packets while ON.AI uses a continuous control loop to maintain job awareness.

subgraph Standard["Standard Cluster Networking (Job-Blind)"] direction TB A1["Users & Agents Submit Jobs"] --> B1["Schedulers Distribute to GPUs"] B1 --> C1["Jobs Sliced into Raw Packets"] C1 --> D1["Packets Mixed in Network Fabric (Tuned for the 'Average Job')"] D1 --> E1["Uniform Slow Progress & High Latency"] end

subgraph Optimal["OptimalNets ON.AI Control Plane"] direction TB A2["GPU Inter-Connect Telemetry"] --> B2["ON.AI Model Engine (Understands Jobs & Network Dynamics)"] B2 --> C2["Dynamic Parameter Optimization"] C2 --> D2["Network Applies Job-Aware Policies"] D2 --> E2["2x–3x Faster Job Completion Time"] end ```

---

[01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) Why Shared GPU Clouds Suffer from the "Average Job" Penalty

In a typical AI data center, workloads originating from disparate teams and agentic pipelines are mapped onto GPU hardware via orchestrators and job schedulers. However, when those GPUs communicate with one another to execute distributed training, fine-tuning, or inference, the underlying interconnect splits the jobs into packets and commingles them across the fabric. In doing so, the network completely loses all context regarding which packet belongs to which job, its operational priority, or its latency sensitivity.

``` +-----------------------------------------------------------------------------+ | Divergence in AI Infrastructure Optimization | +-----------------------------------------------------------------------------+ | Dedicated LLM Provider Clouds | Multi-Tenant GPU Clouds & Neoclouds | +----------------------------------------+-------------------------------------+ | • Single, static, known model workload | • Unpredictable user submissions | | • Hardware manually pre-tuned | • Heterogeneous prompts and agents | | • High utilization for one specific job| • Tuned for the generic "average" | | • Not built for multi-tenant diversity | • Result: "Optimized for none" | +----------------------------------------+-------------------------------------+ ```

While elite foundational LLM providers can manually hard-tune their proprietary networks because they run known, static model architectures, that approach breaks down in multi-tenant environments. Hyperscalers, enterprise on-premise data centers, and emerging neoclouds must serve a fluctuating stream of unknown user prompts, diverse model sizes, and agentic workflows. Because these providers cannot anticipate what workloads will arrive, their networks are statically configured for a generic "average job." As a result, the network is truly "not optimized to none," leading to degraded performance, poor utilization, and prolonged user wait times when submitting prompts.

---

[04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) ON.AI: A Model-Driven Autonomous Data Path Control Plane

ON.AI acts as a software-defined, autonomous data path control plane running on top of existing physical hardware. By activating ON.AI, operators achieve a 2x to 3x reduction in Job Completion Time (JCT) while expanding system output by 30% to 50% more tokens per dollar per watt. Crucially, these efficiency gains require no capital expenditures on new GPUs, additional memory, or specialized switching hardware, while concurrently reducing overall operational power consumption.

```mermaid flowchart LR accTitle: ON.AI Autonomous Control Loop accDescription: Cyclic telemetry analysis and parameter injection loop deployed on top of existing switches.

subgraph Hardware["Existing Infrastructure Fabric"] NET["Physical Switches & GPU Network"] end

subgraph ControlPlane["ON.AI Virtual Control Layer"] T1["1. Continuous Telemetry Collection"] T2["2. ON.AI Engine Traffic Analysis"] T3["3. Job-Aware Parameter Calculation"] T4["4. Cyclic Network Injection"] end

NET --> T1 T1 --> T2 T2 --> T3 T3 --> T4 T4 --> NET ```

The underlying intelligence of ON.AI is a proprietary AI model built and trained from scratch to comprehend three critical dimensions: the characteristics of AI workloads, the mechanics of high-speed network fabrics, and the dynamic interplay between them. The platform operates via an ongoing, four-stage closed loop: 1. **Continuous Telemetry Collection:** Ingests live telemetry directly from the running network fabric. 2. **Engine Traffic Analysis:** Analyzes dynamic traffic requirements against workload profiles using the ON.AI engine. 3. **Parameter Optimization:** Computes tailored network parameters aligned with active job needs. 4. **Cyclic Network Injection:** Continuously pushes dynamic configuration updates back to the network, imbuing standard switches with real-time job awareness.

To accommodate risk-averse IT and infrastructure teams that hesitate to alter configurations on expensive GPU clusters, ON.AI supports a phased, non-disruptive deployment pathway: - **Phase 1 (Recommendation Mode):** The system generates passive optimization recommendations, showing administrators exactly what settings to alter without touching production state. - **Phase 2 (Semi-Autonomous Mode):** Operators review and approve suggested control updates with automated push mechanisms. - **Phase 3 (Fully Autonomous Mode):** The control plane closes the loop independently, executing continuous real-time optimizations to extract peak cluster performance.

---

[07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) Core Value Propositions and Future-Proofing GPU Infrastructure

OptimalNets delivers four fundamental business and technical benefits to data center operators, cloud service providers, and end users:

```mermaid flowchart TD accTitle: OptimalNets Core Value Pillars accDescription: Four key pillars of the ON.AI value proposition for GPU infrastructure operators.

VP1["Up to 50% Higher Billable Capacity Denser utilization of committed GPU capex and locked-in power contracts"] VP2["2x Faster Job Completion Halves turnaround latency for interactive prompts and agentic workflows"] VP3["Open Networking Support Avoids proprietary lock-in for superior cost-to-performance economics"] VP4["Continuous Future-Proofing Dynamic adaptation to daily model releases on 3–5 year hardware cycles"]

VP1 --- VP2 VP2 --- VP3 VP3 --- VP4 ```

- **Up to 50% Increase in Billable Capacity:** Infrastructure providers make massive upfront capital investments in GPUs and commit to long-term power purchase agreements. ON.AI allows operators to pack workloads significantly more densely onto their committed power and silicon envelope, translating previously stranded resources into billable revenue. - **Faster End-User Turnaround:** End customers and automated agents experience prompt completion in roughly half the time, directly improving user satisfaction and application responsiveness. - **Support for Open Networking Standards:** The platform is built to support open network fabrics, maximizing price-performance and freeing infrastructure teams from vendor lock-in. - **Resilience Against Model Churn (Future-Proofing):** Modern AI models evolve on a weekly or daily basis, each introducing novel communication patterns and tensor-parallel requirements. Because physical data center hardware undergoes upgrade cycles only once every 3 to 5 years, ON.AI bridges the gap by continuously learning new model behaviors in software, ensuring existing hardware remains optimized for next-generation architectures.

The session concludes with an open invitation for cluster operators and GPU cloud architects to engage directly with the OptimalNets engineering team via LinkedIn, email, or at their conference location (Booth K522).

## Notable quotes > Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. — Lior Friedman

> 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. — Lior Friedman

> [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) Traditional networking architectures fundamentally fail to understand the workloads they transport. — Lior Friedman

> [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) Because standard network hardware treats all packets uniformly, every workload advances at roughly the same mediocre pace, forcing virtually every user to wait longer than necessary for their results. — Lior Friedman

> [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) Introducing job awareness ensures workloads complete sequentially and rapidly rather than stalling together. — Lior Friedman

> [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) the network completely loses all context regarding which packet belongs to which job, its operational priority, or its latency sensitivity. — Lior Friedman

> [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) Because these providers cannot anticipate what workloads will arrive, their networks are statically configured for a generic "average job." As a result, the network is truly "not optimized to none" — Lior Friedman

> [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) ON.AI acts as a software-defined, autonomous data path control plane running on top of existing physical hardware. — Lior Friedman

> [04:22](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=262s) 2x to 3x reduction in Job Completion Time (JCT) while expanding system output by 30% to 50% more tokens per dollar per watt. — Lior Friedman

> [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) Because physical data center hardware undergoes upgrade cycles only once every 3 to 5 years, ON.AI bridges the gap by continuously learning new model behaviors in software, ensuring existing hardware remains optimized for next-generation architectures. — Lior Friedman

## Mentioned Lior Friedman, OptimalNets, ON.AI, Ai4, LinkedIn

## Video description Lior Friedman - CEO, OptimalNets

Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. On every multi-tenant, multi-job cluster, unrelated workloads share the same network. Standard infrastructure applies one generic setting to all of them, so every job underutilizes the GPUs and the committed power behind it, and runs slower than it can. I'll show benchmark data across inference, fine-tuning, and training workloads: 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. Come by if you run GPU infrastructure, or if you're a customer who wants faster jobs and should be asking your provider for it.

[Source](https://www.youtube.com/watch?v=GNi9Hf9nmUo)

中文简报

OptimalNets推出新软件使多租户GPU速度翻倍标准网络因对数据包进行无差别处理而降低了多租户AI集群的性能，但自主任务感知路由可在现有芯片上使工作负载完成速度翻倍并扩大可计费容量。标准多租户AI集群饱受“无法感知任务”（job-blind）的网络交换之苦，这类交换根据通用平均值路由数据包，迫使不同的AI工作负载一同停滞。OptimalNets推出了ON.AI，这是一种运行在现有网络交换机之上的自主数据路径控制平面，可根据实时工作负载遥测数据动态优化网络结构参数。这种软件驱动的方法将任务完成时间（JCT）缩短至原来的1/2到1/3，将每瓦Token效率提升30%至50%，并为GPU云运营商释放可观的可计费容量。通过在软件层面持续自适应，ON.AI能够保障长达数年的硬件投资周期，使其从容应对快速演进的AI模型架构。传统多租户 GPU 集群将连贯的 AI 任务拆分为离散数据包并在网络中混洗，静态配置的“平均任务”策略导致所有并发任务普遍面临延迟与低利用率。 (0:04) OptimalNets 推出软件定义自主控制平面 ON.AI，依托专有模型在现有网络硬件上运行四阶段持续反馈闭环，无需追加 GPU、新硬件或额外功耗。 (4:22) 通过引入作业感知网络控制，ON.AI 可将任务完成时间（JCT）缩短 2 至 3 倍，同时让每美元每瓦特生成的 Token 数量提升 30% 至 50%。 (4:22) 系统提供推荐、半自主到全自主的三阶段平滑部署方案，帮助数据中心运营商从既有 GPU 投资与电力合同中释放高达 50% 的可计费容量。 (7:19)

Reading summary

OptimalNets argues that multi-tenant AI clusters suffer substantial performance and utilization losses because conventional network fabrics route packets blindly for a generic 'average job.' To eliminate this bottleneck, the company offers ON.AI, an autonomous software-defined control plane that introduces real-time, job-aware traffic optimization to existing GPU infrastructure. Packet-Level Switching and the Average Job Penalty (1:40) In standard multi-tenant enterprise data centers and neoclouds, hundreds of users and autonomous agents submit disparate workloads across shared clusters. Conventional networking architectures divide these jobs into disjointed packets and route them across GPU fabrics without context regarding job boundaries, priority, or latency sensitivity. While dedicated foundational model providers can manually hard-tune fabrics for static, proprietary workloads, multi-tenant operators must serve unpredictable, fluctuating model submissions. Because operators cannot anticipate incoming traffic, fabrics are statically configured for a generic 'average job.' OptimalNets asserts that this makes networks optimized for no workload in particular. As packets from unrelated tasks intermingle uniformly, all workloads advance at a slow pace and finish in delayed clusters, leaving expensive GPU capacity idle while driving up user turnaround times. Closed-Loop Autonomous Control on Existing Switches (4:22) OptimalNets addresses this structural gap through ON.AI, a software-defined data path control plane deployed directly over existing physical switches without requiring capital expenditures for new GPUs, memory, or specialized network hardware. The platform is driven by a proprietary model trained to interpret AI workload profiles, high-speed fabric mechanics, and their dynamic interactions. It executes an ongoing four-stage control loop: gathering live network telemetry, evaluating traffic against active job profiles, calculating optimized network parameters, and injecting dynamic configuration updates back into standard switches. To accommodate risk-averse IT and infrastructure teams hesitant to alter production GPU environments, ON.AI supports a phased rollout. Operators can begin in a passive recommendation mode that suggests configuration changes without modifying cluster state, advance to a semi-autonomous mode requiring human approval for automated pushes, and ultimately transition to fully autonomous real-time closed-loop optimization. Utilization Gains, Open Fabrics, and Model Churn (7:19) OptimalNets claims that job-aware network control cuts Job Completion Time (JCT) by 2x to 3x and boosts efficiency by 30% to 50% more tokens per dollar per watt while lowering power usage. For data center operators operating under locked-in power purchase agreements and heavy GPU capital expenditures, the company projects up to a 50% increase in billable capacity by monetizing previously stranded compute. The architecture is designed to support open networking standards, which OptimalNets states provides superior price-performance economics while avoiding proprietary vendor lock-in. Additionally, the software acts as a buffer against rapid model churn: while physical data center hardware is replaced on 3- to 5-year cycles, AI models and tensor-parallel communication patterns evolve weekly or daily. OptimalNets maintains that continuous software-level learning allows legacy clusters to adapt dynamically to next-generation model architectures. OptimalNets positions ON.AI as a non-disruptive, software-driven solution to reclaim stranded compute and accelerate multi-tenant AI workloads without hardware refreshes. However, the presentation represents vendor claims, and the source notes provide no independent benchmarking data or technical specifics regarding the computational overhead of running the ON.AI model engine alongside active production clusters.

精炼阅读摘要

OptimalNets提出，当前多租户AI集群的网络架构缺乏对作业层面的感知，导致集群资源利用率低下；该公司推出的ON.AI软件定义控制平面通过动态闭环优化，宣称可在不更换底层硬件的前提下缩短作业完成时间并释放闲置算力。多租户网络无感知与“平均作业”瓶颈 (0:04) 传统企业与云端AI基础设施在承载多租户计算任务时，调度器会将来自不同业务部门和自主智能体的异构任务切分成数以十亿计的离散数据包。然而，标准底层网络交换硬件仅在数据包层级执行均等转发，完全丢失了数据包所属的完整计算作业、执行优先级以及延迟敏感度等上下文信息。这种对作业内容无感知的调度方式，迫使所有任务以同样迟缓的节奏同步推进，导致所有提交任务的用户均面临不必要的等待延迟。与能够针对已知静态模型结构进行底层硬件硬调优的专有基础大模型提供商不同，多租户超大规模云、企业私有数据中心以及新兴GPU云必须应对类型不可预测的用户输入和复杂的工作流。由于无法预先获知具体的负载模式，运营商通常只能采用静态参数配置网络，以求迎合通用的“平均作业”。OptimalNets指出，这种折中方案的实际后果是网络对任何特定作业都未实现优化，进而导致整体吞吐能力下降、算力资源闲置以及用户等待时间大幅延长。模型驱动的四阶段自主控制与渐进部署 (4:22) 为消除网络盲区，OptimalNets开发了名为ON.AI的软件定义自主数据路径控制平面。该系统依托专门训练的自研AI模型运行，旨在实时解析AI计算特征、高速网络交换机理以及二者之间的动态映射关系。该平台通过四个阶段构建闭环运行逻辑：持续采集运行中网络架构的实时遥测指标；由分析引擎比对流量动态与工作负载画像；根据活跃作业需求计算定制化网络参数；周期性地将优化策略下发注入物理网络，使标准交换机具备实时的作业感知能力。考虑到基础设施与IT运维团队对调整高价值GPU生产集群配置的风险顾虑，ON.AI设计了渐进式交付路径。系统初始运行时处于“推荐模式”，仅被动输出参数优化建议而不变更生产状态；随后可过渡至由人工复核并确认执行的“半自主模式”；在建立充分置信度后，最终升级为闭环运行的“完全自主模式”，在无需人工干预的情况下进行持续的实时动态调优。硬件免更换增效、开放架构与模型演进适应 (7:19) 根据OptimalNets公布的数据，通过在现有物理硬件上激活ON.AI控制平面，运营商可实现作业完成时间缩短2至3倍，并将每美元每瓦特的Token产出提升30%至50%，且无需追加采购GPU、内存或专用网络硬件。对于面临昂贵资本支出和长期电力采购协议的算力中心，该方案据称可将可计费算力容量提升最高达50%，并将终端交互式提示词与智能体任务的响应延迟降低约一半。此外，该平台支持开放网络架构标准，以避免专有硬件厂商的技术绑定并提升算力性价比。面对前沿AI模型架构以日或周为周期频繁迭代、而物理数据中心硬件更新周期普遍长达3至5年的结构性矛盾，OptimalNets主张通过纯软件控制平面的自适应学习能力，动态适配各类新型张量通信模式，使既有基础设施能够持续兼容新一代模型架构。 OptimalNets展示了通过软件定义控制平面消除多租户网络盲区、提升既有GPU集群产出的技术路径；不过，其宣称的2至3倍作业加速与50%算力扩容等关键指标主要基于其方案设想与内部测试，在高度复杂、多供应商混合的大规模生产环境下的实际调优开销与稳定性仍有待产业实践验证。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2026 年 8 月 21 日发布在 YouTube 平台。视频内容为 OptimalNets 首席执行官 Lior Friedman 的演讲，重点介绍了其专为 AI 集群打造的自主数据路径控制平面产品 ON.AI，阐述其如何在不增加硬件、软件或能耗成本的前提下，使多租户网络具备任务感知能力，并将计算任务完成速度提升 2 至 5 倍。 ## TL;DR 传统多租户 GPU 集群往往将网络参数设置为妥协折中的“平均值”，导致底层网络无法理解具体计算任务的语义，造成严重的算力与电力浪费。OptimalNets 推出的纯软件控制平面 ON.AI，通过专用 AI 模型持续采集物理网络遥测数据，并动态下发针对具体工作负载的最优网络配置。该方案在完全复用既有硬件的前提下，将任务完成时间缩短至原来的 2 到 5 倍，每美元/每瓦特产出 Token 数提升 30% 至 50%。同时，系统提供从建议模式到全自主模式的平滑部署路径，不仅使可计费算力容量提升高达 50%，还赋予长生命周期的基础设施持续适配未来前沿模型的能力。 ## Atmosphere & style 整个演讲节奏明快、直击要害，展现出典型的硅谷 B2B 深度技术硬核路演风格。Lior Friedman 的表达自信笃定且极具商业说服力，开篇通过短片敏锐切入当前 AI 算力中心面临的普遍延迟痛点，迅速建立共鸣，随后层层剖析从调度层到物理数据包的架构断层。现场氛围兼具技术严谨性与务实商业导向，以清晰的数据对比和闭环逻辑，为身处算力军备竞赛中的基础设施运维者提供了极具吸引力的破局思路。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) 传统 AI 集群网络缺乏任务感知（Job Awareness）能力，仅在数据包级别无差别分发流量，导致不同任务相互拖慢并延长等待时间。 - [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) OptimalNets 提出“具备任务感知的网络”理念，让网络理解任务语义与优先级，促使各计算任务按需尽快交付，避免最后时刻集中拥堵。 - [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 上层调度器（Orchestrator）在计算节点建立的任务边界，在数据切片进入底层物理网络后彻底丢失，导致网络层与计算层严重脱节。 - [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 多租户 GPU 云与企业私有集群因无法预判异构任务，被迫采用折中的“平均化网络参数”，实质上造成了全集群算力与带宽的严重浪费。 - [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) ON.AI 在零硬件采购、零新增能耗的既有集群上，可将任务完成速度提升 2 至 5 倍，每美元/每瓦特产出 Token 数提升 30% 至 50%。 - [05:20](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=320s) ON.AI 基于专有 AI 模型打造，通过“遥测采集、智能分析、动态参数生成、网络执行”实现全流程闭环自适应控制。 - [05:20](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=320s) 系统支持“仅提供建议、半自动、全自主”三级阶梯式上线模式，打消 IT 运维团队对频繁改动底层昂贵 GPU 网络的顾虑。 - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) ON.AI 能帮助数据中心释放高达 50% 的可计费算力容量（Billable Capacity），显著提高既有 GPU 采购与电力配额的投资回报率。 - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) 全面支持开放网络（Open Networks）生态架构，帮助企业解绑昂贵的专有私有网络生态并降低 TCO。 - [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) 基于 AI 持续学习通信行为的特性使旧硬件具备面向未来的自适应能力（Future-proof），完美弥合 3 至 5 年硬件生命周期与天级迭代 AI 模型之间的断层。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) AI 集群的网络盲区与任务感知（Job Awareness）缺失 OptimalNets 首席执行官 Lior Friedman 介绍了其专为 AI 集群打造的自主数据路径控制平面产品——ON.AI。演讲伊始，他通过一段演示短片揭示了当前 AI 算力中心面临的普遍困境：无论是市场部门还是研发团队，成百上千的员工以及新型 AI Agent 正在源源不断地向集群提交各种各样的 AI 任务，所有人都在急切等待计算结果。然而，现有的底层 AI 基础设施从根本上并不理解这些“任务”（Jobs）的语义和业务逻辑，仅仅是将它们暴力拆解为数十亿个网络数据包，漫无目的地穿梭分散在成千上万颗 GPU 之间。当所有任务在网络中以近乎相同的节奏、无差别地混杂推进时，导致的结果就是几乎每一个使用者都要经历漫长的排队等待。传统的网络层只懂得转发数据包，并不知晓哪个任务优先、哪个任务需要更紧急的带宽调度。OptimalNets 的核心理念正是为网络注入智能化，构建“具备任务感知的网络”（Job-aware Network），使得各任务能够按需尽快完成并即时返回结果，而不是所有任务被动地在最后时刻一同堵塞交付，从而为企业节省数以百万计的 GPU 采购成本，并释放巨大的商业生产力。 - 现有网络底层缺乏“任务感知”能力，仅在数据包（Packet）级别进行无差别分发。 - 多任务混合导致集群中所有工作负载同步变慢，延长了用户与 AI Agent 的等待延迟。 - 通过让网络理解计算任务，能够在不增加硬件成本的前提下大幅缩短任务完成周期。 --- ### [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 传统多租户 GPU 集群的痛点与“平均化”配置瓶颈 Lior 深入剖析了标准 AI 数据中心的高层运作逻辑：来自各类用户、团队与 Agent 的任务经由上层的编排调度器（Orchestrator / Scheduler）分配至具体的 GPU 算力节点；随后，GPU 之间需要进行密集的跨节点通信以协同完成模型计算。一旦进入连接这些 GPU 的物理网络，整块任务就会被切片并混杂成海量的数据包，调度器在计算层建立的任务边界在网络层彻底丢失。 ```mermaid flowchart TD accTitle: AI 集群网络数据流与 ON.AI 闭环控制 subgraph ComputeLayer["计算与调度层"] A["用户与 AI Agent 任务"] --> B["调度器 / Orchestrator"] B --> C["GPU 集群计算节点"] end subgraph NetworkLayer["物理网络传输层"] C --> D["数据包混杂传输 (切片打散)"] D --> E["网络参数动态下发"] end subgraph ControlPlane["ON.AI 控制平面"] D -. "持续遥测采集 Telemetry".-> F["ON.AI 专用 AI 模型分析"] F -. "生成最优配置".-> E end ``` 这种架构在不同场景下面临截然相反的处境。少数头部超大规模 LLM 服务商（Hyperscalers）可以针对自己已知、固定的训练或推理负载，对整个底层基础设施进行极致的静态手工调优。但在当下的绝大多数场景中——无论是 GPU 云提供商、Neo-clouds，还是各大企业的私有 GPU 数据中心，基础设施都必须承载无数未知的异构任务。运维团队无法预判下一个 Prompt 来自谁、调用什么模型，因此只能将网络参数设置为妥协折中的“平均值”。这种“针对平均任务调优”的做法在实际运行中等同于“对任何任务都没有优化”，直接导致网络吞吐受阻、算力浪费，用户在敲击回车发送提示词后只能面对漫长的延迟。 - 头部大模型厂商可针对固定作业定制调优，但多租户 GPU 云和企业私有集群无法预知工作负载。 - 采用通用/平均配置导致 GPU 算力无法被充分打满，网络成为最大的性能瓶颈。 - ON.AI 作为自主数据路径控制平面，可将任务完成时间缩短至原来的 2 到 3 倍（部分场景达 2 到 5 倍）。 - 在完全相同的既有硬件（相同 GPU、内存与网络）且不增加能耗的前提下，实现每美元、每瓦特产出 Token 数（Tokens per dollar per watt）提升 30% 至 50%。 --- ### [05:20](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=320s) ON.AI 的核心技术机制与闭环部署流程为了打破网络层与应用层的隔阂，OptimalNets 从零开始训练并构建了一个专有的新型 AI 模型。该模型能够深度理解各类计算任务的特征、物理网络的状态，以及任务与网络之间的实时交互规律。在实际运行中，系统通过四个连续阶段实现闭环自适应控制：持续实时采集网络底层的细粒度遥测数据（Telemetry），交由 ON.AI 智能分析引擎研判流量需求，随后动态生成并向网络下发最优调优参数，从而将传统的哑管道网络改造为具有实时任务感知能力的智能网络。 ```mermaid flowchart LR accTitle: ON.AI 四阶段闭环自适应流程 S1["1. 持续遥测采集 (Network Telemetry)"] --> S2["2. 智能流量分析 (ON.AI Engine)"] S2 --> S3["3. 动态参数生成 (Optimization)"] S3 --> S4["4. 下发网络执行 (Job-aware Control)"] S4 --> S1 ``` 针对 IT 与网络运维团队普遍抗拒在高昂、脆弱的 GPU 硬件上频繁变更配置的心理，ON.AI 被设计为一个纯软件的虚拟控制层，部署极为平滑安全。系统支持阶梯式上线策略，确保控制权始终掌握在企业团队手中： 1. **建议模式（Recommendation Only）**：系统仅输出优化建议与推荐参数，供 IT 人员评估审查； 2. **半自动模式（Semi-Automatic）**：关键策略人工确认，常规调整自动执行； 3. **全自主模式（Complete Autonomous）**：实现毫秒级/秒级的全闭环自动调优，最大化释放集群性能。 --- ### [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) 四大核心价值与面向未来的网络自适应能力 Lior 将 ON.AI 带来的核心商业与技术价值提炼为四大维度： 1. **提升高达 50% 的可计费算力容量（Billable Capacity）**：数据中心在 GPU 采购与电力容量配额上投入了巨额资本承诺，ON.AI 能大幅提高算力与电力的使用密度，使相同规模的机房能够承载并结算更多业务。 2. **终端用户任务交付速度翻倍**：显著缩短端到端任务耗时，用户或自主 Agent 提交 Prompt 后获得响应的时间缩短近半。 3. **全面拥抱与支持开放网络（Open Networks）**：解绑昂贵的专有私有网络生态，为企业提供更高性价比（ROI）的架构选型自由。 4. **面向未来新模型的持续自适应性（Future-proof Infrastructure）**：数据中心硬件的迭代周期通常长达 3 至 5 年，而前沿 AI 算法模型几乎以“天”为单位快速演进，每次模型变动都会对网络通信模式提出全新要求。由于 ON.AI 本身基于 AI 模型驱动，它能够持续学习新模型的通信行为及其与网络的交互，使旧硬件具备无缝适配未来前沿模型的能力。演讲最后，Lior 呼吁希望为其 GPU 网络注入智能化的数据中心管理者、云提供商及企业客户展开深入合作，并通过现场展位 K522、领英或邮件进行交流洽谈。 ## Notable quotes > Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. — Lior Friedman > 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. — Lior Friedman > [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) 现有的底层 AI 基础设施从根本上并不理解这些“任务”（Jobs）的语义和业务逻辑，仅仅是将它们暴力拆解为数十亿个网络数据包，漫无目的地穿梭分散在成千上万颗 GPU 之间。 — Lior Friedman > [00:04](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=4s) 构建“具备任务感知的网络”（Job-aware Network），使得各任务能够按需尽快完成并即时返回结果，而不是所有任务被动地在最后时刻一同堵塞交付，从而为企业节省数以百万计的 GPU 采购成本，并释放巨大的商业生产力。 — Lior Friedman > [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 调度器在计算层建立的任务边界在网络层彻底丢失。 — Lior Friedman > [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 这种“针对平均任务调优”的做法在实际运行中等同于“对任何任务都没有优化”，直接导致网络吞吐受阻、算力浪费，用户在敲击回车发送提示词后只能面对漫长的延迟。 — Lior Friedman > [01:40](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=100s) 在完全相同的既有硬件（相同 GPU、内存与网络）且不增加能耗的前提下，实现每美元、每瓦特产出 Token 数（Tokens per dollar per watt）提升 30% 至 50%。 — Lior Friedman > [05:20](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=320s) 系统通过四个连续阶段实现闭环自适应控制：持续实时采集网络底层的细粒度遥测数据（Telemetry），交由 ON.AI 智能分析引擎研判流量需求，随后动态生成并向网络下发最优调优参数，从而将传统的哑管道网络改造为具有实时任务感知能力的智能网络。 — Lior Friedman > [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) 数据中心硬件的迭代周期通常长达 3 至 5 年，而前沿 AI 算法模型几乎以“天”为单位快速演进，每次模型变动都会对网络通信模式提出全新要求。 — Lior Friedman > [07:19](https://www.youtube.com/watch?v=GNi9Hf9nmUo&t=439s) 由于 ON.AI 本身基于 AI 模型驱动，它能够持续学习新模型的通信行为及其与网络的交互，使旧硬件具备无缝适配未来前沿模型的能力。 — Lior Friedman ## Mentioned Lior Friedman, OptimalNets, ON.AI, Ai4, GPU, Orchestrator, AI Agent, Neo-clouds, Hyperscalers ## Video description Lior Friedman - CEO, OptimalNets Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. On every multi-tenant, multi-job cluster, unrelated workloads share the same network. Standard infrastructure applies one generic setting to all of them, so every job underutilizes the GPUs and the committed power behind it, and runs slower than it can. I'll show benchmark data across inference, fine-tuning, and training workloads: 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. Come by if you run GPU infrastructure, or if you're a customer who wants faster jobs and should be asking your provider for it. [Source](https://www.youtube.com/watch?v=GNi9Hf9nmUo)

Source quotes

Lior Friedman: Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. Lior Friedman: 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. [0:04] Lior Friedman: 现有的底层 AI 基础设施从根本上并不理解这些“任务”（Jobs）的语义和业务逻辑，仅仅是将它们暴力拆解为数十亿个网络数据包，漫无目的地穿梭分散在成千上万颗 GPU 之间。 [0:04] Lior Friedman: 构建“具备任务感知的网络”（Job-aware Network），使得各任务能够按需尽快完成并即时返回结果，而不是所有任务被动地在最后时刻一同堵塞交付，从而为企业节省数以百万计的 GPU 采购成本，并释放巨大的商业生产力。 [1:40] Lior Friedman: 调度器在计算层建立的任务边界在网络层彻底丢失。 [1:40] Lior Friedman: 这种“针对平均任务调优”的做法在实际运行中等同于“对任何任务都没有优化”，直接导致网络吞吐受阻、算力浪费，用户在敲击回车发送提示词后只能面对漫长的延迟。 [1:40] Lior Friedman: 在完全相同的既有硬件（相同 GPU、内存与网络）且不增加能耗的前提下，实现每美元、每瓦特产出 Token 数（Tokens per dollar per watt）提升 30% 至 50%。 [5:20] Lior Friedman: 系统通过四个连续阶段实现闭环自适应控制：持续实时采集网络底层的细粒度遥测数据（Telemetry），交由 ON.AI 智能分析引擎研判流量需求，随后动态生成并向网络下发最优调优参数，从而将传统的哑管道网络改造为具有实时任务感知能力的智能网络。 [7:19] Lior Friedman: 数据中心硬件的迭代周期通常长达 3 至 5 年，而前沿 AI 算法模型几乎以“天”为单位快速演进，每次模型变动都会对网络通信模式提出全新要求。 [7:19] Lior Friedman: 由于 ON.AI 本身基于 AI 模型驱动，它能够持续学习新模型的通信行为及其与网络的交互，使旧硬件具备无缝适配未来前沿模型的能力。

Source quotes

Lior Friedman: Faster job completion turns your committed GPUs and power into more billable capacity — and happier customers. Lior Friedman: 2–5x faster job completion once the network becomes job-aware. That speed costs nothing extra — no new hardware, no new software stack, no new power draw. [0:04] Lior Friedman: Traditional networking architectures fundamentally fail to understand the workloads they transport. [0:04] Lior Friedman: Because standard network hardware treats all packets uniformly, every workload advances at roughly the same mediocre pace, forcing virtually every user to wait longer than necessary for their results. [0:04] Lior Friedman: Introducing job awareness ensures workloads complete sequentially and rapidly rather than stalling together. [1:40] Lior Friedman: the network completely loses all context regarding which packet belongs to which job, its operational priority, or its latency sensitivity. [1:40] Lior Friedman: Because these providers cannot anticipate what workloads will arrive, their networks are statically configured for a generic "average job." As a result, the network is truly "not optimized to none" [4:22] Lior Friedman: ON.AI acts as a software-defined, autonomous data path control plane running on top of existing physical hardware. [4:22] Lior Friedman: 2x to 3x reduction in Job Completion Time (JCT) while expanding system output by 30% to 50% more tokens per dollar per watt. [7:19] Lior Friedman: Because physical data center hardware undergoes upgrade cycles only once every 3 to 5 years, ON.AI bridges the gap by continuously learning new model behaviors in software, ensuring existing hardware remains optimized for next-generation architectures.

Original transcript

[0:04] So, hi. I'm Leo from uh Optimal Nets. And we built an autonomous data pass control plan for AI [0:14] clusters. The product name called ON AI. And now I'll just show you a short video clip describe the [0:22] problem, and then we'll go through the solution. Okay? Hundreds of employees, different AI jobs, [0:32] from marketing to R&D. And now, it's not just employees, AI agents are submitting jobs, too. [0:40] And they are all waiting for the results. But the AI infrastructure does not understand those jobs. [0:47] It breaks them into billions of packets and sends them across thousands of GPUs. When every [0:53] job advances at roughly the same pace, nearly every employee wait longer to get the results. [1:01] A smarter approach is to create a job awareness network. Instead of returning all the results [1:07] at the end, each job returns its result as soon as possible. Save millions on GPUs. [1:14] Gain millions in business productivity. Optimal Nets, add intelligence to networks. [1:23] So, as we saw, network understand do not understand jobs, they understand packets. [1:29] And uh we understand jobs, and we bring this awareness into the networks. [1:37] And uh [1:40] now let's see how AI cluster, standard AI cluster, works very high level. So, there are the themes, [1:50] users, agents, all are sending jobs into this AI cluster, into this data center. [1:58] And all of those jobs are allocated into GPUs that we all know. These are done by [2:08] orchestrator schedulers. But eventually, when those jobs arrive to those GPU, the GPUs need to [2:16] communicate with each other in order to process those jobs. And then when it goes there to the [2:21] network that connects all of those GPUs, those jobs are sliced and split into separate packets. [2:32] And those packets are mixed inside the network. And therefore, it loses job awareness. [2:49] When an infrastructure, an AI infrastructure, knows the job [2:59] in advance, it can be tuned to achieve high performance. [3:05] This is what the biggest LLM providers operate like. Okay, they tune their infrastructure [3:13] to be optimized to the job that they are going to run on it. And this is okay. [3:20] However, when we are talking here in this conference about GPU clouds, hyperscalers, [3:30] organization data centers for GPUs, neo clouds, all of those [3:37] infrastructures need to serve multiple unknown jobs. Okay? They don't know [3:47] what each of their users will send, what are the prompts, who are the agents, and so on. [3:54] Therefore, they are not optimized to none of them. They are tuned for the average [4:01] job. Meaning they are not optimized to none. And therefore, their the performance are poor. And [4:13] how we filled it, we know that when we press enter after clicking our prompts, it take a lot of time. [4:22] So, when AI when AI is an autonomous data path control plan, [4:29] when AI is active on those infrastructures, it deliver two up to three times [4:36] faster job completion time. Meaning you your customers will get the results faster. [4:44] And if you have your own data center organization, then you will get 30 50% more tokens per dollar [4:55] per watt, which is huge numbers. And we are doing it on the same infrastructure [5:02] that you already have. Same GPUs, same GPUs, same memory, same network, and less power. [5:12] So, this is what on AI brings to the table. Now, how we do it? [5:20] We built and design and train a new model. New AI model, obviously, from scratch, [5:28] that understand jobs, understand networks, and understand the interaction between them. [5:35] And then, we implemented four steps. The first step [5:40] is continuously learns network behavior. We continuously collect telemetry from the network. [5:50] Then we analyze it with Onai AI engine. We analyze the traffic requirements [5:57] and we send back parameters into the network. And now this network understand jobs. [6:10] So deployment, Onai deployment is easy, it's safe and it's under your control. [6:17] Onai is a virtual layer sit on top of existing infrastructure. [6:23] It gets telemetry from the network send its result back to the network [6:29] in a cyclic time, in a cyclic order, all the time. It's not a one one show. And your IT teams [6:38] that most of the time don't like to make changes to this expensive hardware. They have all the [6:46] need in order to control this process. So at the beginning it will be like recommendation only. [6:55] The system Onai will recommend what are the best parameters, what need to be changed in the network [7:00] to give their customer better performance. Later it can be semi-automatic. And at the end [7:09] it will be complete autonomous. And then obviously the results the performance will be the best. [7:19] So Optimal's value proposition. [7:24] We deliver up to 50% increase in billable capacity. What does it mean billable capacity? [7:32] It means the data center already invest a lot of money in GPUs. Most of the time they also bought [7:42] um, capacity of power with commitments. And now with only AI they can use it much more dense way. [7:57] Okay? Number two, can deliver faster job completion to the customer, to the end user. [8:06] Because when you they click the prompt or the agent, they get they will they get the [8:15] the results much faster, half of the time. And number three, we support open networks. [8:24] So open networks mean in this case better value for money. [8:33] And number four, [8:37] we're talking about adaptive, adaptation to future models. [8:43] Every new models need have new requirements from the infrastructure, but the infrastructure [8:49] are changing maybe once in three years, in five years, but new models arrive almost daily. [8:58] So because only AI is based on an AI model, it learns new models, it learns the interactions, [9:06] it learns your network continuously. So actually you get future proof [9:16] infrastructure. [9:21] So if you like to be and to add intelligent to your network, please talk to us. And ask me, [9:31] LinkedIn, email, or visit our booth K522. Thank you for your time. Any question? Welcome.