Aible, Dell, NVIDIA Launch Air-Gapped AI Stack to Eliminate Token Fees

The End to Token Anxiety: Aible, NVIDIA & Dell — Ai4

Ai4

By shifting workloads from foundational cloud models to on-premise hardware and deterministic templates, enterprises reduce operational AI costs tenfold while eliminating data privacy risks.

AI-assisted summary of The End to Token Anxiety: Aible, NVIDIA & Dell — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background This YouTube panel discussion was published on August 19, 2026, by the Ai4 channel. Featuring speakers from Aible, NVIDIA, and Dell Technologies, the session focuses on a complete enterprise generative AI and agentic solution that eliminates unpredictable per-token costs through local, air-gapped infrastructure.

## TL;DR The panel explores how enterprises can escape 'token anxiety' and data security fears by shifting from metered cloud APIs to local, open-weight models running on dedicated hardware. Through real-world case studies like the State of Alaska, Mount Sinai, and the State of Nebraska, the speakers demonstrate how flat-rate pricing, sovereign data governance by design, and targeted model post-training enable rapid agent deployment. The discussion highlights a broader industry transition from monolithic foundation models to networks of specialized micro-models that deliver deterministic, outcome-driven business results.

## Atmosphere & style The session carries an energetic, pragmatic, and highly collaborative tone as industry experts blend high-level strategy with candid personal anecdotes. The dynamic between Sengupta, Lawler, and Bali is conversational yet insightful, demystifying technical barriers with sharp analogies—such as comparing fine-tuned models to surgical scalpels versus broad katanas. Observing the discussion feels like sitting in on an engaging, practical masterclass that demystifies enterprise AI adoption while maintaining a grounded, hands-on focus.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) Aible, NVIDIA, and Dell joined forces to tackle three critical barriers to enterprise AI adoption: strict data security, unpredictable token costs, and model hallucinations. - [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) Relying on metered, cloud-hosted API calls for enterprise-wide deployment creates 'token anxiety' and financial inefficiency compared to local open-weight models. - [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) To eliminate per-token and per-user fees, Aible and Dell introduced a predictable flat-rate 'per server, per year' pricing model. - [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) Sovereign AI and 'governance by design' protect sensitive enterprise data by running workloads on fully air-gapped local hardware like the Dell ProMax GB10. - [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) Mount Sinai safely analyzed sensitive patient data locally on an air-gapped Dell server using Nemotron Guardrails to scrub PHI/PII before private cloud export. - [08:58](https://www.youtube.com/watch?v=WMDPotB7EI8&t=538s) Enterprise agent setups can be simplified using pre-built templates, enabling non-technical staff—such as interns in Nebraska—to build hundreds of custom agents rapidly. - [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) Fine-tuning smaller models via post-training and distillation acts like a surgical 'scalpel,' outperforming massive monolithic foundation models at a fraction of the compute cost. - [16:34](https://www.youtube.com/watch?v=WMDPotB7EI8&t=994s) Aible leverages the NVIDIA Nemotron model family to legally post-train and distill smaller models using outputs from larger models. - [16:34](https://www.youtube.com/watch?v=WMDPotB7EI8&t=994s) True sovereign AI keeps sensitive agent memory and core datasets local while allowing execution across diverse architectures, including BigQuery and edge servers. - [18:09](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1089s) Unconstrained agentic planning loops consume thousands of tokens needlessly; converting agentic workflows into deterministic templates ensures cost control and consistency. - [19:21](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1161s) Generative AI has evolved from standard answer generation into autonomous, outcome-driven agentic workflows that execute complex business processes end-to-end.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the provided transcript segment:

[00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) Solving Enterprise Token Anxiety and Unlocking Scalability

Arijit Sengupta opens the session by introducing Aible as an enterprise AI platform focused on delivering tangible economic impact for business users—such as driving $21 million in sales impact at a Fortune 500 company—by handling large data volumes, running models locally, and leveraging low-cost post-training. Highlighting a real-world example from the State of Alaska, Sengupta explains how organizations attempting to build AI agents are routinely stymied by three primary obstacles: complex data security concerns, unpredictable token costs, and a fundamental lack of trust due to model hallucinations. Aible solves this by allowing enterprises to build, test, and deploy entire agentic workflows on fully air-gapped hardware—such as the NVIDIA DGX Spark or the Dell ProMax GB10—in as little as five minutes without incurring per-token fees.

The conversation sets the stage to examine four critical trends currently shaping the enterprise AI landscape: cost concerns (token anxiety), data sovereignty (data anxiety), autonomous agents or "claws," and post-training. Logan Lawler characterizes the current commercial AI market as "kind of the wild west," admitting that while frontier cloud models like Claude 3.5 Opus have clear utility, relying on metered, cloud-hosted API calls is simply not a scalable financial model when expanding from a single test user to thousands across an enterprise.

Lawler highlights a fundamental industry shift—accented by a joint open-letter initiative signed by NVIDIA CEO Jensen Huang and Dell CEO Michael Dell—toward open-weight models and specialized local hardware. Advanced hardware setups (such as the Dell ProMax GB10, NVIDIA DGX Spark, and Blackwell-generation GPUs with doubled VRAM compared to the Ada generation) enable highly capable open models like Nemotron to run completely on-premise. Lawler argues that if local hardware can deliver performance close to foundational cloud models, paying continuous token fees to cloud providers makes little economic sense.

Sama Bali expands on the psychological toll this dynamic takes on engineering teams, noting that token caps create severe "token anxiety" among developers. When constrained by strict API credit limits, developer behavior becomes strictly transactional—they turn their systems on, complete a single minimal task, and shut them down immediately to conserve balance. Removing token limits through unmetered local AI unlocks the creative freedom necessary for real enterprise innovation. Lawler reinforces this with an analogy, asking why anyone would buy a Ferrari only to drive strictly under the speed limit out of fear of running out of fuel or getting fined. Sengupta emphasizes that running AI workloads locally must encompass the entire end-to-end architecture: the underlying language model, agentic harness, integrations, and structured or unstructured data processing must all reside on the air-gapped device. To offer full financial predictability, Aible partnered with Dell to deliver a flat-rate "per server, per year" pricing model that eliminates per-token, per-user, and per-agent charges.

``` +-------------------------------------------------+ | LOCAL AIR-GAPPED HARDWARE | | (Dell ProMax GB10 / NVIDIA DGX Spark) | | | | +-------------------------------------------+ | | | AIBLE STACK | | | | [Language Models] [Agentic Harness] | | | | [Data Processing] [Integration Tools] | | | +-------------------------------------------+ | +-------------------------------------------------+ | v [FLAT PER-SERVER / YEAR PRICING] (No Per-Token / No Per-User Charges) ```

- **Enterprise Case Study:** The State of Alaska deployed full AI agents on air-gapped devices in under 5 minutes to bypass data security and token cost barriers. - **Hardware & Models Mentioned:** Dell ProMax GB10, NVIDIA DGX Spark, NVIDIA Blackwell-generation GPUs (doubled VRAM over Ada), and open-weight models like Nemotron. - **Business Impact Example:** Aible's localized approach enabled $21 million in sales impact at a Fortune 500 company.

[04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) Sovereign AI and "Governance by Design"

Turning to sovereign AI, Sama Bali underlines that strict data locality and ownership are non-negotiable requirements for highly regulated sectors like healthcare, legal services, and cybersecurity. Keeping data inside a secure, bounded local sandbox eliminates the friction and overhead of sanitizing or proxying sensitive files before sending them off-premises. This localized control simplifies IT governance and accelerates commercialization.

Logan Lawler illustrates the real-world friction caused by corporate data security fears with a personal anecdote from inside Dell. Due to heightened anxiety over data leaving the corporate network, Dell restricted internal access to cloud platforms like Google Drive and blocked the use of external flash drives on laptops. Consequently, Lawler could not even transfer the presentation deck for this panel via a standard USB drive. While acknowledging that blocking flash drives seems extreme, Lawler stresses that it highlights how desperately organizations need granular control over where their data resides and how it is used.

Sengupta reinforces this point by sharing a healthcare case study from Mount Sinai presented at NVIDIA GTC. Mount Sinai wanted to perform analytics on sensitive patient data but faced a classic catch-22: they could not verify whether the dataset was entirely clear of protected health information (PHI) or personally identifiable information (PII) without running an AI analysis first, but they could not send the unanalyzed data to the cloud due to privacy risks. Using a Dell ProMax GB10 completely disconnected from the internet, Mount Sinai ran their full analysis locally, applied Nemotron Guardrails to scrub PHI/PII, and only then pushed the vetted outputs to a private cloud. Sengupta frames this approach as "governance by design"—a structural model where doing the wrong thing is rendered physically impossible by the architecture itself.

Bali adds that data is every company's core "secret sauce." Analyzing raw, unfiltered datasets locally yields deeper context and truth than working with fragmented, pre-scrubbed snippets. Lawler invites audience members to visit the Dell or Aible exhibition booths to touch the physical GB10 hardware during an agent run, noting how you can literally feel the machine warm up under load and cool down afterward while running completely air-gapped.

- **Mount Sinai Healthcare Example:** Used a local, air-gapped Dell ProMax GB10 paired with Nemotron Guardrails to process patient data securely for PHI/PII before transferring results to a private cloud. - **Data Governance Concept:** "Governance by Design" replaces reactive policies by making off-network data leaks physically impossible through air-gapped hardware architecture.

[08:58](https://www.youtube.com/watch?v=WMDPotB7EI8&t=538s) Demystifying Autonomous Agents and "Claws"

The discussion shifts to autonomous agents—frequently referred to in developer communities as "claws" (such as OpenClaw or Hermes). When Lawler polls the audience on who has experimented with tools like OpenClaw or Hermes, only a few hands go up, prompting him to encourage everyone to visit the booth to test them firsthand. Lawler admits that while he loves agentic technology, setting up frameworks like OpenClaw manually requires tedious model configuration, permission mapping, and technical overhead that non-developers struggle to navigate. He praises Aible for making agent deployment virtually "idiot-proof" by offering pre-made templates alongside simple, guided setup flows.

Lawler strongly critiques the enterprise obsession with rigid "use cases" for AI agents, calling it a mistake to force-fit agentic tools into pre-defined boxes. Instead, organizations should treat agents as flexible problem-solving engines tailored to immediate operational needs. To demonstrate, Lawler shares a story from his own team at Dell. After terminating an external marketing agency that managed social media tracking for Dell's workstation ambassador group, Lawler was left without a way to gather key engagement metrics (likes, clicks, and impressions). Rather than spending weeks writing complex API integrations or re-hiring an agency, he built a local agent that automatically monitored ambassador profiles, cross-referenced data files, and scraped engagement metrics without making a single external API call.

Sengupta adds that non-technical users are fully capable of driving this transformation, citing the State of Nebraska, where interns and agency staff—rather than expert data scientists—built 222 custom AI agents in just 90 minutes using Aible.

Sama Bali offers an analogy about baking Christmas Eve cookies with her kids to describe the hidden complexity of autonomous agents. While online recipes promise a quick 15-minute bake, the reality involves spilled flour, tricky measurements, and unexpected cleanup. Similarly, setting up agents appears simple on the surface, but enterprise deployments quickly derail over unaddressed data auditing, security, and permission layers. Platforms like Aible abstract away this friction by building enterprise security directly into the deployment harness.

- **State of Nebraska Deployment:** Non-technical interns and agency staff built 222 functional agents in 90 minutes using Aible. - **Operational Example:** Lawler replaced a fired marketing agency by building a local agent to track social media engagement metrics across ambassador profiles with zero API dependencies.

[14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) Targeted Post-Training: Scalpels vs. Katanas

In the final topic, the speakers tackle model post-training and fine-tuning. Sama Bali cautions against the common enterprise tendency to pull down the largest possible foundation model to solve every problem, explaining that this usually results in severely oversizing the compute solution. By utilizing post-training and distillation, organizations can refine a smaller model on domain-specific data. This approach dramatically reduces hallucinations, guarantees domain grounding, and minimizes compute requirements—freeing up system resources for other pipeline tasks and making model distribution easier.

Sengupta shares a benchmark project Aible executed at NVIDIA GTC, where they post-trained a compact language model for just $5 in compute costs. That $5 post-trained model successfully outperformed a foundational model 100 times its size on domain-specific tasks. Sengupta urges enterprises not to view post-training as a multi-hundred-thousand-dollar endeavor, but rather as an accessible, automated loop where user feedback continually refines the system.

Lawler closes with a striking metaphor comparing model selection to surgical instruments: "If you were going to go and perform open-heart surgery, are you going to use a katana or are you going to use a scalpel?" Large foundational models act like katanas, whereas small, post-trained models function like precision scalpels—running faster, taking up less VRAM, operating locally, and executing specialized tasks with far higher accuracy. He predicts that the future of enterprise AI will not rely on a single massive monolithic model, but on orchestrating hundreds of hyper-specialized, post-trained micro-models working in concert.

- **GTC Post-Training Benchmark:** A small model fine-tuned for $5 in compute costs outperformed a foundational model 100 times larger. - **Architectural Shift:** Enterprise AI is transitioning from massive monolithic models ("katanas") to orchestrated networks of small, domain-specific, post-trained micro-models ("scalpels").

[16:34](https://www.youtube.com/watch?v=WMDPotB7EI8&t=994s) Model Post-Training and Sovereign AI Execution

Arijit Sengupta highlights the inherent data licensing challenges associated with post-training AI models, noting that enterprise developers cannot legally post-train models using outputs directly generated by cloud services like OpenAI or Google Gemini. To resolve this bottleneck, Aible leverages NVIDIA's Nemotron family of models, specifically using the outputs of larger models like Nemotron-4 340B Ultra to post-train Nemotron-4 Super, which in turn post-trains smaller models like Nemotron Nano. This tiered "bootstrapping" approach distills the intelligence of massive foundational architectures down into highly efficient, specialized smaller models.

``` +------------------------------------------------------------------+ | Nemotron-4 340B Ultra | | (Smartest / Highest Capacity Model) | +------------------------------------------------------------------+ | v [Outputs used to Post-Train] +------------------------------------------------------------------+ | Nemotron Super | +------------------------------------------------------------------+ | v [Outputs used to Post-Train] +------------------------------------------------------------------+ | Nemotron Nano | | (Highly Efficient / On-Device Model) | +------------------------------------------------------------------+ ```

Moving to the concept of sovereign AI, Sengupta clarifies that true sovereignty is not confined to a single physical hardware box. Instead, it revolves around flexible architecture and data privacy. Organizations should keep highly sensitive assets, such as agent memory and proprietary data, locally on the user's device or private network. At the same time, processing can occur directly where the data lives. Illustrating this capability, Aible collaborated with Google to execute queries on 5 billion rows of data across 10 million variable combinations entirely inside Google BigQuery—without extracting raw data from the environment—for just $10 in compute costs. Whether deployed strictly on an NVIDIA GB10 workstation, distributed across telecommunications data centers, or placed within hybrid cloud infrastructure, sovereign AI ensures total governance without administrative restrictions.

- NVIDIA Nemotron models are leveraged to bootstrap and post-train smaller, localized models legally and efficiently. - Sovereign AI guarantees that sensitive agent memories and private enterprise data never leave local hardware or controlled environments. - In a joint benchmark on Google BigQuery, Aible processed 5 billion data rows and evaluated 10 million variable combinations for $10 in compute expense. - The Aible execution stack supports flexible deployments spanning standalone NVIDIA GB10 setups, enterprise edge devices, private telco data centers, and public clouds.

[18:09](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1089s) Deterministic Execution vs. Unpredictable Agent Planning

Addressing the practical limits of advanced reasoning models (referred to in the context of Claude or similar agentic architectures), Sengupta cautions against letting fully autonomous agents handle every repetitive enterprise task. Unconstrained planning loops incur extreme token consumption; simple actions or standard greetings like saying "hello" to systems like Hermes can trigger massive backend data loading and step planning, costing anywhere from 20,000 to 30,000 tokens per interaction. Enterprises demand strict repeatability, governance, and predictable cost structures rather than dynamic, uncontrolled reasoning for standard business processes.

To deliver both agility and efficiency, Aible employs a hybrid workflow: teams use flexible agentic models for rapid prototyping to design and validate workflows. Once the ideal pipeline is finalized, the system saves the logic into a deterministic template. Execution at scale then bypasses redundant, costly planning phases, dramatically cutting token usage while enforcing enterprise control.

- Naive or unconstrained agent planning can inadvertently burn 20,000 to 30,000 tokens on trivial operations like preliminary setup or greetings. - Dynamic agentic reasoning is best suited for initial prototyping and workflow discovery. - Finalized enterprise workflows are converted into deterministic templates to eliminate redundant planning overhead and guarantee consistent execution. - Aible provides over 100 pre-built templates out of the box, allowing enterprise users to rapidly build and deploy additional templates tailored to their operational needs.

[19:21](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1161s) Enterprise AI Economics and the Shift to Outcomes

Logan Lawler reflects on his tenure at Dell Technologies, observing that this is the first time in his experience where unit pricing models and tokenomics have actively constrained technology adoption. While autonomous agents are already running background tasks locally on individual workstations without constant human supervision, keeping these intelligent agents and their data safely inside corporate boundaries remains non-negotiable. The joint collaboration between Dell, NVIDIA, and Aible solves this by offering a localized, scalable environment for running, testing, and deploying production-grade AI agents across local workstations and enterprise servers.

Reflecting on the evolution of generative AI since ChatGPT's emergence in 2023, the panelists emphasize a fundamental shift from merely fetching answers to automatically delivering business outcomes. Rather than forcing employees to manually piece together conclusions from raw model outputs, agentic systems take high-level goals, formulate steps independently, and execute end-to-end workflows. This transition relieves professionals from routine administrative tasks, freeing them to focus on high-value initiatives.

- Token costs and unpredictable unit economics have historically stalled the deployment of autonomous enterprise AI solutions. - Dell, NVIDIA, and Aible provide an end-to-end local framework to build, refine, and scale agents securely without exposing sensitive enterprise assets. - Generative AI has transitioned from the "answer generation" paradigm of 2023 into an "outcome-driven" era powered by autonomous enterprise agents.

[21:07](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1267s) Practical Adoption and Final Call to Action

Closing the discussion, the panel urges enterprise technology leaders to cut through theoretical complexity and evaluate agentic tools firsthand. Creating an enterprise agent using Aible's platform is straightforward, with Sengupta observing that "if you think this is hard, you're completely missing the point." Visitors at the event are encouraged to stop by the Dell or Aible booths to build custom agents in real time or test the platform directly on their own terms.

Once a prototype is constructed, enterprise teams should audit their actual token expenditures, evaluate the projected economics at full operational scale, select the appropriate deployment hardware, and continuously iterate on performance.

- Teams can build operational AI agents live at the Dell and Aible event booths or test the software independently. - Organizations should analyze real-world token costs during prototyping before choosing an execution platform for enterprise-wide scaling. - The recommended implementation roadmap centers on starting small, monitoring unit economics, selecting tailored infrastructure, and iteratively refining workflows.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) kind of the wild west — Logan Lawler

> [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) governance by design — Arijit Sengupta

> [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) If you were going to go and perform open-heart surgery, are you going to use a katana or are you going to use a scalpel? — Logan Lawler

> [21:07](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1267s) if you think this is hard, you're completely missing the point — Arijit Sengupta

## Mentioned Aible, NVIDIA, Dell, Dell Technologies, Arijit Sengupta, Sama Bali, Logan Lawler, State of Alaska, NVIDIA DGX Spark, Dell ProMax GB10, Claude 3.5 Opus, Jensen Huang, Michael Dell, Nemotron, Mount Sinai, Nemotron Guardrails, OpenClaw, Hermes, State of Nebraska, NVIDIA GTC, Nemotron-4 340B Ultra, Nemotron-4 Super, Nemotron Nano, Google, Google BigQuery, ChatGPT

## Video description Arijit Sengupta - Founder and CEO, Aible Sama Bali - Sr Manager, AI Solutions Product Marketing, NVIDIA Logan Lawler - Dell Pro Precision AI Strategy, Dell Technologies

Companies are stuck between two choices: use frontier models spending unpredictable amounts on tokens or cobble together open source projects to create your own agentic AI solution. Aible worked closely with NVIDIA and Dell to create a complete genAI and agentic solution that delivers enterprise-strength AI without per-token costs. You can use Aible for standard genAI, agentic analytics, process automation, or even fully autonomous agents - all completely under your control at 10X lower costs.

[Source](https://www.youtube.com/watch?v=WMDPotB7EI8)

中文简报

Aible、Dell与NVIDIA推出物理隔离AI全栈方案以消除Token费用通过将工作负载从云端基础模型转移到本地硬件和确定性模板，企业在消除数据隐私风险的同时，将AI运营成本降低十倍。该小组探讨了企业如何通过从按量计费的云端API转向在专用硬件上运行的本地开放权重模型，从而摆脱“Token焦虑”和数据安全担忧。通过阿拉斯加州、Mount Sinai以及内布拉斯加州等实际案例，演讲者展示了固定费率定价、原生设计的主权数据治理以及针对性的模型后训练如何实现智能体的快速部署。讨论还重点指出了行业正在发生更广泛的转型，即从单一庞大的基础模型转向提供确定性、结果驱动型业务成效的专用微模型网络。 Aible、NVIDIA和Dell联手解决企业采用AI的三大关键障碍：严格的数据安全、不可预测的Token成本以及模型幻觉。 (0:04) 与本地开放权重模型相比，在全企业范围部署中依赖按量计费的云端托管API调用会导致“Token焦虑”和资金效率低下。 (0:04) 为消除按Token和按用户收费，Aible与Dell推出了一种可预测的“每台服务器每年”固定费率定价模式。 (0:04) 主权AI与“原生设计治理”通过在如Dell ProMax GB10等完全物理隔离的本地硬件上运行工作负载，保护敏感的企业数据。 (4:41) Mount Sinai在物理隔离的Dell服务器上安全地本地分析敏感患者数据，并在导出至私有云之前使用Nemotron Guardrails清除PHI/PII。 (4:41)

Reading summary

Aible, Dell Technologies, and NVIDIA argue that enterprise AI adoption is constrained by unpredictable per-token pricing and data exposure fears, advocating for on-premise air-gapped hardware, targeted post-trained models, and deterministic agent templates. Eliminating Token Metering via Local Hardware (0:04) Logan Lawler of Dell and Sama Bali assert that reliance on metered, cloud-hosted API calls introduces severe "token anxiety," forcing engineering teams to operate transactionally and shut down workflows early to conserve API credits. To provide financial predictability, Aible and Dell introduced a flat "per server, per year" pricing structure that removes per-token, per-user, and per-agent charges. This framework deploys the complete enterprise stack—including language models, agent harnesses, integrations, and data processing—onto local, air-gapped systems such as the Dell ProMax GB10 and NVIDIA DGX Spark, utilizing Blackwell-generation GPUs that double the VRAM of prior Ada hardware. Arijit Sengupta of Aible noted that this localized model allowed the State of Alaska to deploy full agentic workflows in under five minutes, while previously generating $21 million in sales impact for a Fortune 500 company without generating ongoing API access bills. Enforcing Governance by Design and Sovereign Data (4:41) Bali and Sengupta argue that strict data locality is essential for highly regulated domains such as healthcare and cybersecurity, where sanitizing data before exporting it to external cloud APIs creates severe operational friction. Sengupta proposed "governance by design," an architectural model where data leakage is rendered physically impossible because the processing hardware is disconnected from external networks. In a benchmark case at Mount Sinai presented at NVIDIA GTC, sensitive patient records were analyzed on an air-gapped Dell ProMax GB10, using Nemotron Guardrails locally to scrub protected health information (PHI) and personally identifiable information (PII) before vetted outputs were transmitted to a private cloud. Sengupta clarified that data sovereignty is not limited to a single workstation box. For larger distributed deployments, Aible partnered with Google to run analytics across 5 billion rows and 10 million variable combinations entirely within Google BigQuery for $10 in compute costs, keeping raw enterprise data within its native environment without extraction. Replacing Monolithic Models with Targeted Post-Training (14:01) Lawler and Bali caution against the enterprise tendency to deploy massive foundational models for specialized tasks, likening it to performing surgery with a katana instead of a scalpel. They contend that compact, domain-specific models require less VRAM, process queries faster, minimize hallucinations, and operate efficiently on local devices. Sengupta highlighted an Aible demonstration at NVIDIA GTC where a compact model fine-tuned for $5 in compute costs outperformed a foundational model 100 times its size on domain-specific benchmarks. To navigate legal restrictions that prevent enterprises from training models on outputs generated by commercial cloud APIs like OpenAI or Google Gemini, Aible utilizes open-weight NVIDIA architectures. By employing a cascading distillation chain—where Nemotron-4 340B Ultra post-trains Nemotron-4 Super, which subsequently trains edge models like Nemotron Nano—enterprises can legally bootstrap high-performing specialized micro-models for on-premise execution. Standardizing Autonomous Agents into Deterministic Templates (18:09) While autonomous agents (often dubbed "claws") offer operational flexibility—such as Lawler automating social media engagement tracking at Dell without third-party APIs, or interns in Nebraska building 222 agents in 90 minutes—unconstrained planning introduces cost and reliability risks. Sengupta pointed out that open-ended agentic loops using reasoning models can inadvertently burn 20,000 to 30,000 tokens on simple greetings or repetitive preliminary planning routines. To balance agility with operational control, Aible utilizes dynamic agentic reasoning during early prototyping to map out workflows, but converts proven pipelines into deterministic, pre-built execution templates for production. This hybrid approach bypasses dynamic, unconstrained reasoning loops during repetitive execution, eliminating redundant token waste while enforcing predictable organizational behavior. The speakers conclude that enterprise AI has evolved from raw answer generation into outcome-driven autonomous workflows. However, achieving scalable adoption requires organizations to audit token economics during early prototyping, transition predictable tasks into deterministic structures, and anchor critical data within sovereign, air-gapped infrastructure.

精炼阅读摘要

Aible、Dell 与 NVIDIA 的代表指出，按 Token 计费的云端 API 模式和数据安全顾虑正在阻碍企业规模化应用 AI；通过在本地物理隔离硬件上运行开源权重模型、结合针对性后训练以及将推理流程固化为确定性模板，企业能够在保证数据主权的同时实现经济可控的自动化成果。 Token 焦虑与固定费率本地算力 (0:04) Logan Lawler 认为，尽管 Claude 3.5 Opus 等前沿云端模型具备明确价值，但按量计费的云端 API 在从单个测试用户推广至企业数千名员工时，在财务上不可持续。Sama Bali 指出，严格的额度上限给开发者带来了“Token 焦虑”，导致工程行为变得高度防范化与事务化，工程师往往被迫在完成单一微小任务后立即关闭系统以节省余额。针对这一痛点，Aible 与 Dell 合作推出了按“每台服务器/每年”计算的固定费率模式，取消了基于 Token、用户数及 Agent 数量的计费限制。在硬件支持上，NVIDIA Blackwell 架构 GPU 的显存容量相较 Ada 架构实现翻倍，配合 Dell ProMax GB10 与 NVIDIA DGX Spark 等设备，使 Nemotron 等开源权重模型得以在完全本地化的物理隔离环境中运行。Arijit Sengupta 指出，Aible 的本地化方案曾帮助一家财富 500 强企业创造了 2100 万美元的销售业务影响，阿拉斯加州政府亦在物理隔离设备上以不到 5 分钟的时间完成了 AI Agent 的构建与部署。物理隔离架构下的数据主权设计 (4:41) 在医疗、法律与网络安全等强监管行业，Sama Bali 强调数据本地存储与归属权不可妥协，将数据保留在受限的本地沙箱内，能够免去在向外部发送敏感文件前所必需的数据脱敏与代理转发开销。Logan Lawler 分享了 Dell 内部的实际管理案例：由于对数据外流的严密风控，公司限制访问 Google Drive 并禁用了笔记本电脑的外接 U 盘，甚至导致他无法使用标准 U 盘拷贝现场演讲幻灯片，这凸显出大型组织对数据驻留位置与使用路径的精细控制需求。在 NVIDIA GTC 大会上展示的西奈山医院（Mount Sinai）案例中，院方希望分析敏感病患数据，却陷入两难：既无法在未运行 AI 分析前确认数据集中是否完全剔除了受保护健康信息（PHI）或个人身份信息（PII），又因合规风险无法将未处理的数据上传至公有云。通过在完全断网的 Dell ProMax GB10 上本地运行分析并结合 Nemotron Guardrails 脱敏工具，院方在本地完成数据清洗后才将合规结果同步至私有云。Arijit Sengupta 将该机制概括为“源头治理”（Governance by Design），即通过物理层面的离线架构从根本上杜绝违规数据外泄的可能。针对性后训练与小模型协作网络 (14:01) 针对企业盲目选用最大型基础模型的现象，Sama Bali 提出警示，指出这往往造成算力资源的严重过度配置；而通过后训练与模型蒸馏，企业可以用领域专有数据提纯较小的模型，在减少幻觉并确保业务契合度的同时大幅降低显存开销。在 Aible 于 NVIDIA GTC 开展的基准测试中，团队仅花费 5 美元算力成本后训练的一个紧凑型语言模型，在特定业务领域内的表现超过了体量大其 100 倍的基础模型。此外，针对企业无法合法使用 OpenAI 或 Google Gemini 等商业云服务输出内容进行模型训练的授权限制，Aible 采用了分层引导（bootstrapping）路径：利用 NVIDIA Nemotron-4 340B Ultra 的输出后训练 Nemotron Super，再进一步向下蒸馏后训练出能够在端侧高效运行的 Nemotron Nano。Logan Lawler 将大模型比作“武士刀”，将专有小模型比作“手术刀”，并预测企业级 AI 的未来形态将不再依赖单一庞大的单体模型，而是由数百个经过专门后训练的微型模型协同工作。自主 Agent 的落地门槛与确定性模板 (8:58) 虽然开发者社区积极探索 OpenClaw 与 Hermes 等被称为“claws”的自主 Agent 框架，但手动配置模型、映射权限与系统集成对非技术人员构成了较高壁垒。Sama Bali 以烘焙曲奇类比，指出 Agent 表面流程看似简单，但落地过程极易因数据审计、安全策略与权限层处理不当而受阻。为展示低代码敏捷性，Arijit Sengupta 引用了内布拉斯加州政府的案例，该机构的实习生与在职人员在无需专业数据科学家介入的情况下，利用 Aible 平台在 90 分钟内构建了 222 个定制 AI Agent；Lawler 亦通过本地 Agent 自动追踪社交媒体互动指标，摆脱了外部营销机构与外部 API 调用。在执行机制上，Sengupta 警告称，若让无约束的 Agent 规划循环处理常规业务，仅日常打招呼或预处理就可能触发冗余的数据加载，单次交互耗费 20,000 至 30,000 个 Token。为此，Aible 采用混合策略：仅在初期原型设计阶段使用动态推理规划，一旦业务逻辑验证完毕便将其固化为确定性模板（内置超 100 种预置模板），在后续规模化执行时直接绕过重复的规划阶段以锁定计算开销并确保执行的一致性。与会者总结指出，企业级 AI 正在经历从 2023 年以“获取解答”为核心的生成式阶段，迈向以“自主交付业务成果”为导向的应用周期。然而，尽管本地化与确定性架构能有效规避 API 计费与外泄风险，企业在实际部署前仍须自主完成从原型阶段的真实 Token 消耗审计，到全流程硬件选型与工作流持续迭代的严格核算。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2026 年 8 月 19 日在 YouTube 平台发布。视频主要记录了 Aible 创始人兼 CEO Arijit Sengupta、NVIDIA AI 解决方案产品营销高级经理 Sama Bali 以及 Dell 战略专家 Logan Lawler 的圆桌对谈，介绍三方如何联手打造兼顾数据主权与极低成本的本地全栈企业级 GenAI 与 Agent 解决方案。 ## TL;DR 三位讲者深入探讨了企业落地大模型时面临的 Token 成本失控、数据合规顾虑及 Agent 构建门槛高等核心痛点。通过结合 Dell 与 NVIDIA 的物理隔离硬件底座、开源权重模型以及 Aible 的全栈自动化工具，企业无需依赖按 Token 计费的云端 API，即可在本地以固定年费实现从原型开发到确定性模版的安全落地。对谈还展示了 5 美元超低成本后训练、蒸馏小模型以及去专业化门槛的实操案例，阐述了 AI 从单纯“获取答案”向端到端“交付业务成果”的范式跃升。 ## Atmosphere & style 整场对谈气氛务实、幽默且极具实战导向，讲者们摒弃了空洞的行业黑话，用烤饼干、开法拉利、手术刀与武士刀等生动比喻直击企业落地的真实困境。Dell 与 NVIDIA 嘉宾坦诚分享内部严苛管控与开发痛点，与 Aible 创始人的技术底气形成默契呼应，互动自然且充满行业前沿的敏锐洞察。现场洋溢着破除技术迷思的务实热情，带给观众一种从被 Token 焦虑裹挟的束缚感走向自主掌控、放手创新的踏实感与掌控力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) 企业级 AI 落地面临四大核心挑战：不可预测的 Token 成本失控、主权 AI 与数据安全合规（如 PHI/PII）、自主智能体构建门槛高，以及高成本模型后训练问题。 - [01:29](https://www.youtube.com/watch?v=WMDPotB7EI8&t=89s) 硬件显存跃升（如 Blackwell 架构）配合本地开源权重模型，使企业能够摆脱按 Token 计费的商业陷阱，转向基于本地服务器固定年费的可预测资产投入模式。 - [01:29](https://www.youtube.com/watch?v=WMDPotB7EI8&t=89s) API 调用的“Token 焦虑”会迫使开发者将创新退化为事务性应付，只有提供本地不限量 Token 资源，才能真正解放团队的试错与迭代潜力。 - [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) 主权 AI 强调数据所有权与本地化，通过物理隔离设备结合 Neotron Guardrails 达成“设计即治理”，让高监管行业无需过度脱敏即可安全分析全量真实数据。 - [08:58](https://www.youtube.com/watch?v=WMDPotB7EI8&t=538s) Aible 颠覆了传统开源框架的高技术门槛，通过预置业务模板与引导式设计，让非技术业务人员能在数分钟内开箱即用搭建自主智能体。 - [10:11](https://www.youtube.com/watch?v=WMDPotB7EI8&t=611s) 智能体开发应摒弃“生搬硬套预设用例”的固有思维，直接围绕业务痛点按需构建；内布拉斯加州的实践验证了一线行政与在校实习生同样具备规模化构建能力。 - [12:33](https://www.youtube.com/watch?v=WMDPotB7EI8&t=753s) 构建企业级 Agent 表面看似简单，实则涉及繁琐的数据审计与权限工程治理，软硬件深度一体化集成的全栈方案能够前置消除落地摩擦。 - [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) 后训练与模型微调践行“手术刀”哲学：针对私有领域进行轻量级小模型蒸馏（如 5 美元完成微调），不仅能显著降低幻觉，更能在特定任务上击败体积大百倍的通用大模型。 - [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) 基于 Nemotron 系列模型构建多层级自举蒸馏（Bootstrap）架构，既解决了严禁使用商业闭源 API 生成训练集的合规约束，又实现了自动化的模型迭代闭环。 - [17:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1025s) 主权 AI 践行“就地计算”（如在 BigQuery 原位完成海量数据分析）原则，在确保核心记忆与敏感业务上下文保留在本地的前提下，提供跨本地与云端的一致性推理体验。 - [17:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1025s) 为了避免动态规划在执行问候或简单任务时空耗数万 Token，Aible 采用双轨机制：在原型期利用动态 CLAW 自由探索，验证后固化为确定性执行模版以锁定成本与合规。 - [19:21](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1161s) 企业将自主智能体完整保留在本地基础设施内部，不仅锁定了核心商业机密不出大楼，更为长效自主运作奠定了坚实的安全底座。 - [19:57](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1197s) AI 演进迎来了从 2023 年“获取静态答案”（Answers）到当前“交付业务成果”（Outcomes）的第二次范式跃迁，由 Agent 端到端闭环解决问题并重塑组织生产力。 - [21:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1265s) 破除对智能体构建的技术迷思，企业应通过现场或亲自动手搭建验证原型，基于真实数据测算 Token 成本，在自主可控平台上实现持续小步迭代与规模化应用。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=WMDPotB7EI8&t=4s) 企业 AI 的落地痛点与四大核心议题 Aible 创始人兼 CEO Arijit Sengupta 开场幽默地表示，Aible 可能是台上大家唯一不太熟悉的公司，但他们一直极度专注于如何让企业业务用户真正从 AI 中获得实际业务价值。Aible 聚焦于处理海量数据、在本地运行大模型、以极低成本进行模型后训练（Post-training），并最终转化为真金白银的经济效益——例如曾帮助一家财富 500 强企业创造了 2100 万美元的销售额增长。 Arijit 提到了阿拉斯加州的实际案例：该州政府尝试构建智能体（Agent）时，面临严峻的数据安全顾虑以及高昂且不可预测的 Token 成本。最终，他们借助 NVIDIA DGX Spark 或 Dell ProMax GB10 等完全物理隔离（Air-gapped）的本地设备，仅用 5 分钟就成功搭建出完整的智能体。这种方式不仅让团队能清晰洞察 Token 消耗模式且无需担心账单失控，还显著降低了模型幻觉与信任问题，顺利实现了规模化落地。Arijit 指出，当前企业级 AI 正在经历四大核心转变： 1. **成本顾虑与 Token 焦虑（Token Anxiety）** 2. **主权 AI（Sovereign AI）与数据安全焦虑** 3. **自主智能体与 Claws（Autonomous Agents / Claws）** 4. **后训练（Post-training）** ```mermaid flowchart TD accTitle: 企业级 AI 落地面临的核心挑战与本地化解决方案 subgraph Pains["企业落地痛点"] P1["Token 成本失控与计费焦虑"] P2["数据外泄顾虑与隐私合规 (PHI/PII)"] P3["智能体构建门槛高与幻觉风险"] end subgraph Solution["Aible + NVIDIA + Dell 本地全栈方案"] S1["物理隔离硬件 (DGX Spark / Dell GB10)"] S2["本地开源权重模型 + Neotron Guardrails"] S3["Aible 自动化开发套件 (5分钟免代码构建)"] S4["固定年费模式 (按服务器/年收费)"] end subgraph Outcomes["业务成果"] O1["零 Token 焦虑，激发无限创新"] O2["数据不出域，实现原生安全治理"] O3["真实业务 ROI (如 2100 万美元销售提升)"] end Pains --> Solution Solution --> Outcomes ``` --- ### [01:29](https://www.youtube.com/watch?v=WMDPotB7EI8&t=89s) Token 焦虑与本地化算力驱动的范式转移在讨论企业面临的成本问题时，Dell 的 Logan Lawler 直言当下的 AI 市场宛如“狂野西部”（Wild West）。他坦言自己也很喜欢使用前沿闭源大模型处理特定任务，但从单用户扩展到上千名企业员工时，按 Token 计费的商业数学逻辑根本无法成立。Jensen Huang 与 Michael Dell 联名签署的关于开源/开放权重模型的倡议，正推动整个行业发生根本性转变。从 Ada 架构升级到 Blackwell 架构后，显存（VRAM）直接翻倍，配合 Dell ProMax GB10 和 NVIDIA DGX Spark 等硬件，使得像 Neotron 等开源权重模型能够在本地高效运行，其能力已非常接近顶尖基础模型。 NVIDIA 的 Sama Bali 进一步从开发者心理学角度解析了“Token 焦虑”对企业创新的扼杀。当开发者受限于 API 调用额度和成本上限时，其工作模式会彻底退化为“事务性应付”——开启系统、急匆匆跑完一两个必需的任务后立刻关闭。而在拥有不限量的本地 Token 资源后，焦虑感随之消散，开发者才能真正放开手脚去探索、迭代和激发核心创意。Logan 形象地打了个比方：“这就好比你买了一辆法拉利，结果却只能按最低限速开，这毫无意义。” Arijit 补充道，许多人误以为只要在本地运行 Agent 外壳就足够了，但只要底层模型仍在调用云端 API，就依然无法摆脱 Token 计费的束缚。Aible 与 NVIDIA、Dell 深度合作，将包括大语言模型、软件框架、分析工具以及结构化与非结构化数据处理系统全部搬到了物理隔离的本地设备上，并推出了按“每台服务器/每年”的固定定价模式，彻底打破了按 Token、按用户或按 Agent 数量计费的不可控成本结构。 - **显存与硬件跃升**：Blackwell 原生架构相比 Ada 世代实现了显存翻倍，使得在本地承载媲美基础大模型的开源权重模型成为可能。 - **计费模式颠覆**：摒弃按 Token/席位抽税的云端 API 模式，转向基于本地服务器年费的固定资产投入模式。 --- ### [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) 主权 AI 与“设计即治理”的企业级数据安全谈及主权 AI（Sovereign AI），Sama Bali 强调数据本地化与数据所有权在医疗健康、网络安全以及法律服务等高度监管行业中是不可逾越的红线。一旦企业能够完全掌控数据，就不再需要为了防止敏感数据外流而设置层层繁琐的代理和脱敏沙箱，从而大幅提升开发与商业化效率，IT 管理部门的工作也得以大幅简化。 Logan Lawler 分享了 Dell 内部极其严苛的数据管控现状作为佐证：由于极度担忧 AI 开发过程中的数据泄露，Dell 内部网络甚至直接禁用了 Google Drive，员工在准备本次大会的演示文稿时甚至无法使用普通 USB 闪存盘导出文件。Logan 感慨道：“数据是企业的命脉，谁都不想拱手送人。”虽然拔掉 U 盘的管控看似极端，但这恰恰印证了企业对数据外泄的巨大焦虑。 ```mermaid flowchart LR accTitle: 西奈山医院基于本地物理隔离设备的数据分析流程 subgraph Airgap["本地物理隔离环境 (Dell ProMax GB10)"] direction TB Raw["原始医疗数据 (含潜在 PHI/PII)"] --> Analysis["Aible 本地化全栈分析"] Analysis --> Guard["Neotron Guardrails 隐私安全检测"] end Guard -->|确认合规脱敏| Cloud["部署推送至机构私有云环境"] ``` Arijit 引用了西奈山医院（Mount Sinai）在 NVIDIA GTC 大会上分享的经典案例：医疗机构往往陷入“不分析就不知道数据里是否含有 PHI（个人健康信息）和 PII（个人身份信息），但直接分析又面临泄密违规”的先有鸡还是先有蛋的死循环。西奈山医院利用完全断网物理隔离的 Dell ProMax GB10 运行 Aible，并配合 NVIDIA Neotron Guardrails 进行合规检测，在彻底确认安全后再将成果推送到私有云中。这种模式将企业的思维方式从“我敢不敢冒险做”转变为“风险已被物理阻断，我可以大胆尝试”，实现了真正的“设计即治理”（Governance by Design）。 - **数据完整性价值**：主权 AI 允许企业直接使用未经阉割、未被过度切碎过滤的全量真实数据，充分挖掘隐藏的业务真相。 - **直观硬件体验**：在展台演示中，运行在 GB10 上的本地完整 AI 栈在闲置时机身冰凉，只有启动自主智能体全力运算时才会升温，全程完全断绝互联网连接。 --- ### [08:58](https://www.youtube.com/watch?v=WMDPotB7EI8&t=538s) 降低门槛：自主智能体与 Claws 的傻瓜化落地在进入自主智能体（Autonomous Agents / Claws）的探讨环节时，Logan 调侃自己虽然每天在展台向观众演示各类 Agent，但他其实很“讨厌”Claw 这个生僻的叫法。当现场询问有多少人实际玩过 Hermes 或 OpenClaw 时，举手者寥寥无几，这反映出当前开源 Agent 框架对普通大众依然存在较高的认知和上手门槛。 Logan 坦言自己并非程序员出身，而是一名硬件与产品营销人员。以 OpenClaw 为代表的传统开源项目通常需要使用者深入理解底层模型参数并进行复杂的配置调试；而 Aible 的突出优势在于提供了开箱即用的预置业务模板，甚至做到了“傻瓜化”（Idiot proof）——非技术人员只需回答几个简短的业务引导问题，系统便能在几分钟内自动生成并调优出满足业务需求的自主智能体，极大地拉近了前沿 Agent 技术与一线业务人员之间的距离。 ### [10:11](https://www.youtube.com/watch?v=WMDPotB7EI8&t=611s) 摆脱固有场景思维：按需定制与普惠化 Agent 构建构建 AI 工具时，若缺乏底层治理与权限管理，不仅会带来难以估量的复杂性，还会引发严重的安全隐患。Logan Lawler 幽默地指出，人们常说 AI 能为人类节省时间，但他感觉自己“现在比以往任何时候都更加努力地工作”，因为在环境搭建和琐碎配置上耗费的时间是无法挽回的，甚至自己的私人邮箱地址在某些不透明系统中泄露了都不得而知。对于业界普遍存在的“寻找典型用例”的执念，Logan 明确表达了反对态度。他认为将业务需求生搬硬套进预设好的用例中是极其受限的；在 Agent 与自主工作流的世界里，核心问题只有一个：“你实际需要解决什么？” 为了说明这一点，Logan 分享了自己的亲身经历：Dell 内部有一个由工作站等设备支持的大使项目（Ambassador Group），此前负责追踪大使社交媒体指标（点赞、点击、曝光量等）的合作代理机构被他解雇后，他急需重新建立整套指标追踪体系。面对突如其来的危机，Logan 本地构建了一个专属 Agent，它无需依赖任何第三方 API 调用，而是直接读取本地参考文件，自主访问各大社交平台抓取公开数据与互动指标，并在本地运行完成分析。 ```mermaid flowchart LR accTitle: 普惠化本地 Agent 自动化工作流 subgraph S1["本地数据与配置"] A["本地参考文件"] end subgraph S2["自主执行 Agent"] B["追踪社交媒体档案"] C["抓取公开互动数据"] end subgraph S3["本地化产出"] D["还原核心分析指标"] end A --> B B --> C C --> D ``` - **内布拉斯加州实践**：内布拉斯加州政府团队借助 Aible 平台，在短短 90 分钟内构建了 222 个 AI Agent。 - **去专业化门槛**：这些 Agent 并非出自资深数据科学家之手，而是由在校实习生和机构一线行政人员搭建，彻底颠覆了“只有高精尖专家才能构建 Agent”的认知壁垒。 --- ### [12:33](https://www.youtube.com/watch?v=WMDPotB7EI8&t=753s) 烤饼干的比喻：消除企业落地 Agent 的复杂性与摩擦谈及 Agent 与自动化工作流（如 NVIDIA NeMo CLAW）的落地难度时，Sama Bali 用一个生动的生活故事作了类比。平安夜他的孩子们想给圣诞老人烤饼干，他本以为这只是个“网上查个食谱、15分钟准备加烘焙”的简单任务，实际操作却手忙脚乱地折腾了一两个小时——面粉洒得满地都是，香草精这类需要极微剂量的调料也难以精确量取。企业部署 Agent 亦是如此：表面上看起来人人皆可快速上手，但真正切入后就会发现前期需要海量的准备工作与严格的数据审计，包括弄清楚数据流向何处、如何做权限隔离以及如何全方位保障资产安全。 Aible 的价值正是站在企业级用户的视角，预先消化了所有工程与安全陷阱，通过极简架构消除这些摩擦。Arijit Sengupta 强调，端到端的紧密合作在此过程中发挥了决定性作用——Aible 团队从第一天起就与 Dell 和 NVIDIA（特别是 NeMo CLAW 团队）开展联合开发冲刺，从底层硬件到中间件全面打通，直接为企业移除了自行试错的技术复杂度。 --- ### [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) 5 美元的后训练革命：小模型蒸馏与“手术刀”哲学在模型微调与后训练（Post-training）方面，Sama Bali 提醒企业切忌盲目追求参数规模最大的模型。加载超大模型往往会导致“过度配置”（Oversizing the solution），带来巨大的算力浪费与运维负担。相反，通过后训练与微调，企业可以先对模型进行蒸馏，再结合私有领域数据深度调优。这种方式不仅能大幅降低幻觉，还能减少推理资源占用，释放系统算力给数据管线的其他模块，让边缘端或本地分发部署更加轻量敏捷。 Arijit Sengupta 以 Aible 在 NVIDIA GTC 大会上的实操项目为例：团队通过针对性的后训练，用一个超轻量小模型在特定任务上击败了比其大 100 倍的庞大模型，而整个后训练的算力成本仅仅花费了 **5 美元**。Logan Lawler 随后给出了一个贴切的比喻：“如果要做开胸心脏手术，你会选择武士刀还是手术刀？”通用基础大模型就像武士刀，而后训练打造的轻量模型则是极其精准的手术刀。针对企业在后训练中面临的“数据合规难、严禁使用商业闭源 API 输出作为训练集”的痛点，Aible 结合 NVIDIA 开源模型生态走出了一条自举蒸馏（Bootstrap）路径： ```mermaid flowchart TD accTitle: 多层级模型自举后训练架构 subgraph L1["旗舰大模型"] A["NVIDIA Nemotron-3 Ultra"] end subgraph L2["中量级模型"] B["Nemotron Super"] end subgraph L3["端侧轻量模型"] C["Nemotron Nano"] end A --"生成高质量蒸馏数据"--> B B --"级联微调更轻量层级"--> C ``` - **合规自举蒸馏**：利用顶层的 Nemotron-3 Ultra 模型输出去后训练下一级的 Super 模型，再逐级蒸馏至端侧运行的 Nano 模型，实现全链路合规且高质量的模型轻量化演进。 - **自动化数据飞轮**：将 NVIDIA 的数据飞轮无缝内置在 Aible 平台内，用户只需提供业务反馈，后续的自动化重训与迭代均由系统后台自动闭环处理。 --- ### [17:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1025s) 主权 AI 与确定性执行：从动态规划到模版化规模落地 Arijit Sengupta 进一步阐述了对“主权 AI”（Sovereign AI）的本质理解：主权 AI 并不等同于将所有东西硬塞进同一个黑盒子里，而是确保最核心、最私密的资产（如 Agent 长期记忆与敏感业务上下文）始终保留在本地设备上绝不外泄。与此同时，数据分析应当“就地执行”（Compute where data lives）。例如 Aible 与 Google BigQuery 深度合作，直接在 BigQuery 内部对 50 亿行数据、1000 万个变量组合完成全量分析，全部算力成本仅耗费 10 美元，数据全程无需搬迁出库。对于推理算力，主权 AI 赋予用户完全的选择权——无论是在本地 GB10 级工作站、电信运营商机房、私有数据中心还是公有云，整套软件栈都能一致运行。 ```mermaid flowchart LR accTitle: 从动态 CLAW 原型到确定性企业级执行 subgraph Phase1["开发阶段：自由探索与规划"] A["动态 Agent / CLAW"] --> B["快速构建业务原型"] end subgraph Phase2["生产阶段：固化与规模化管控"] B --> C["固化为确定性模版"] C --> D["零规划开销 / 严格成本与治理管控"] end ``` 在解决 Agent 规划阶段的 Token 浪费问题上，Arijit 指出，直接向动态 Agent（如 Hermes 等）发送一句简单的打招呼，后台为了执行环境加载与动态规划可能就会空耗 20,000 到 30,000 个 Token。企业在实际生产中绝不希望每次执行都产生未知的动态规划路径和费用，而是需要严格可控的治理标准。Aible 采取的双轨策略是：**在原型设计阶段利用 CLAW 进行高灵活度的快速构建；一旦逻辑验证通过，立即将其固化为确定性执行模版（Deterministic Templates）**。 - **预置与定制模版**：平台开箱自带 100 多套预置模版，并支持企业按需无限拓展沉淀业务模版。 - **杜绝重复消耗**：通过模版化确定性运行，在规模化执行时省去了每一次请求的规划 Token 消耗，兼顾了敏捷构建与严苛的企业级合规控本需求。 --- ### [19:21](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1161s) 终结 Token 焦虑：将 Agent 留在企业内部在总结发言中，Logan Lawler 感慨自己在 Dell 工作多年，这是头一次目睹传统软件架构与商业模式被“Token 经济学（Tokenomics）”所打破与重构。按 Token 计费带来的预算不可预测性与高昂成本，正成为阻碍企业拥抱生成式 AI 的巨大阻力。如今，自主运行的 Agent 已经可以完全驻留在本地电脑后台独立完成复杂任务，无需人员时刻盯梢。Logan 强调，企业最迫切的诉求是确保核心业务逻辑与 Agent 资产“不出大楼”。Dell、NVIDIA 与 Aible 的联合方案正切中了这一命脉，让企业能够在自己的基础设施内掌控一切，实现 Agent 的本地开发、极速迭代以及大规模生产级部署。 ### [19:57](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1197s) 从“获取答案”到“交付结果”：AI Agent 的第二次跃升回溯生成式人工智能的发展历程，2023 年 ChatGPT 横空出世的时刻令人惊叹，当时人们只要输入提示词，无论身处何地、任何时刻都能迅速获取现成答案。然而，技术演进的脚步并未停歇。讲台上讲者们回忆起 2023 年共同探讨 ChatGPT 惊艳表现的场景，指出那时的核心能力还仅仅停留在“提供答案”；而如今，业界迎来了属于智能体（Agent）的全新转折时刻——AI 不再只是给出静态回答，而是能够直接端到端地交付“业务结果”（Outcomes）。这一转变为企业生产力带来了跨越式的升级。在智能体架构下，业务人员只需向 AI 明确交代最终希望达成的业务目标，至于中间具体的执行路径、数据调用与逻辑拆解，完全交由 Agent 自行规划与探索。这种自动化不仅消除了员工陷入繁琐梳理与繁重琐事的负担，真正将人类解放出来，而且正如 Logan 所指出的那样，即便未来大家依然会十分忙碌，但所忙碌的将是“人们真正想做的事情”，从而在更高维度上释放创新活力。随着 NVIDIA、Aible 与 Dell 的深度合作，三方不仅面向开发者，更将这种先进的工程思维赋能给企业端用户，精准满足了大型组织对可控、安全与落地的严苛诉求。 ```mermaid flowchart LR accTitle: 从答案到结果的演进 subgraph S1["2023 生成式 AI 阶段"] A["用户提问"] --> B["获取答案 (Answers)"] end subgraph S2["当前 Agentic AI 阶段"] C["定义最终目标"] --> D["Agent 自主规划与执行"] D --> E["交付业务成果 (Outcomes)"] end ``` - **从 Answers 到 Outcomes**：上一代生成式 AI 解决了海量信息的检索与生成，而 Agentic AI 核心在于自主解决问题并交付实际业务成果。 - **重塑生产力范式**：企业员工无需再受困于“思考具体怎么做”的流程性摩擦，可以将精力集中于真正具备业务价值的核心事项。 - **生态协同合力**：Dell 的硬件底座、NVIDIA 的算力支撑与 Aible 的全流程智能体套件，共同抹平了开发者能力与企业业务需求之间的鸿沟。 --- ### [21:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1265s) 告别“代币焦虑”：动手构建智能体与规模化落地呼吁面对企业用户对构建 Agent 的迟疑与敬畏，现场讨论给出了极为务实的建议：不要把人工智能当成高不可攀的“障眼魔术”，行动是打破技术迷思的唯一方式。讲者直言，这套方案在工程上已经做到“极易上手”（"so blooming easy"），如果觉得构建企业级智能体难如登天，那就“完全抓错了重点”。现场参会者被热情呼吁前往 Dell 或 Aible 的展台，直接让工程师在眼皮底下当场搭建一个 Agent，甚至更推荐参会者亲自上手操作，亲身体验从零到一的极速部署过程。在真正完成上手构建后，企业便能直观审视其背后的代币（Token）消耗与成本模型，进而基于真实的业务数据，科学评估在大规模扩展场景下的实际开销。依托 Aible、Dell 与 NVIDIA 提供的无代币焦虑架构，企业可以在自主可控的基础设施上挑选最适宜的部署平台，随后在实战中不断“小步迭代与持续优化”（"iterate and improve"），以数十倍于以往的成本效益真正将 Agentic AI 引入企业日常生产之中。 - **破除复杂性迷思**：企业智能体的创建门槛已被大幅降低，无论是亲临展台让技术人员实时演示搭建，还是自主上手，都能在短时间内迅速验证可行性。 - **透明的成本评估逻辑**：先动手验证原型，再精准测算 Token 成本，最终选定契合企业自身规模的扩展平台并实施持续迭代。 ## Notable quotes > [01:29](https://www.youtube.com/watch?v=WMDPotB7EI8&t=89s) 这就好比你买了一辆法拉利，结果却只能按最低限速开，这毫无意义。 — Logan Lawler > [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) 数据是企业的命脉，谁都不想拱手送人。 — Logan Lawler > [14:01](https://www.youtube.com/watch?v=WMDPotB7EI8&t=841s) 如果要做开胸心脏手术，你会选择武士刀还是手术刀？ — Logan Lawler > [01:29](https://www.youtube.com/watch?v=WMDPotB7EI8&t=89s) Wild West — Logan Lawler > [08:58](https://www.youtube.com/watch?v=WMDPotB7EI8&t=538s) Idiot proof — Logan Lawler > [04:41](https://www.youtube.com/watch?v=WMDPotB7EI8&t=281s) Governance by Design — Arijit Sengupta > [21:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1265s) so blooming easy — Logan Lawler > [21:05](https://www.youtube.com/watch?v=WMDPotB7EI8&t=1265s) iterate and improve — Logan Lawler ## Mentioned Aible, NVIDIA, Dell, Arijit Sengupta, Sama Bali, Logan Lawler, Jensen Huang, Michael Dell, NVIDIA DGX Spark, Dell ProMax GB10, Neotron Guardrails, NVIDIA NeMo CLAW, Nemotron-3 Ultra, Nemotron Super, Nemotron Nano, Ada, Blackwell, Google BigQuery, Google Drive, Mount Sinai, Hermes, OpenClaw, ChatGPT ## Video description Arijit Sengupta - Founder and CEO, Aible Sama Bali - Sr Manager, AI Solutions Product Marketing, NVIDIA Logan Lawler - Dell Pro Precision AI Strategy, Dell Technologies Companies are stuck between two choices: use frontier models spending unpredictable amounts on tokens or cobble together open source projects to create your own agentic AI solution. Aible worked closely with NVIDIA and Dell to create a complete genAI and agentic solution that delivers enterprise-strength AI without per-token costs. You can use Aible for standard genAI, agentic analytics, process automation, or even fully autonomous agents - all completely under your control at 10X lower costs. [Source](https://www.youtube.com/watch?v=WMDPotB7EI8)

Source quotes

[1:29] Logan Lawler: 这就好比你买了一辆法拉利，结果却只能按最低限速开，这毫无意义。 [4:41] Logan Lawler: 数据是企业的命脉，谁都不想拱手送人。 [14:01] Logan Lawler: 如果要做开胸心脏手术，你会选择武士刀还是手术刀？ [1:29] Logan Lawler: Wild West [8:58] Logan Lawler: Idiot proof [4:41] Arijit Sengupta: Governance by Design [21:05] Logan Lawler: so blooming easy [21:05] Logan Lawler: iterate and improve

Source quotes

[0:04] Logan Lawler: kind of the wild west [4:41] Arijit Sengupta: governance by design [14:01] Logan Lawler: If you were going to go and perform open-heart surgery, are you going to use a katana or are you going to use a scalpel? [21:07] Arijit Sengupta: if you think this is hard, you're completely missing the point

Original transcript

[0:04] Very quickly, Abel is probably the only company you don't know on stage. So, we have been very [0:08] focused on how do you get impact from AI for business users. So, focusing on, you know, [0:14] large volumes of data, running models locally, post- training at very low cost and then executing [0:21] to give economic impact. So, like 21 million in sales impact at a Fortune 500 for example. [0:27] The important part of this that we're going to talk about is that quote from state of Alaska [0:31] where they were trying to build an agent and their problem was they did not know the data security [0:36] concerns. They had token cost concerns and being able to build an agent on an air gap device [0:43] for example the Nvidia DGX Spark or the Dell GB10 Proax GB10 and being able to build that [0:48] whole agent in a matter of five minutes. see the token cost patterns, not worried about the cost, [0:54] reduce the um hallucination uh the trust uh problem and then take it to scale. So that's [0:59] what AEL is very focused on. But today we're going to talk about four key things that are happening [1:04] in the world of enterprise AI. First is cost concerns, token anxiety. Second is sovereignty, [1:12] so data anxiety if you will. Third is clause or autonomous agents. And finally post training. [1:18] And we have two fantastic uh experts to talk on this and I'm going to come into this a little bit [1:23] myself but uh if you look at the cost concerns organizations are running into. Uh Logan I'll [1:29] go to you first. Like what are you facing? What are you seeing in the market? I mean [1:34] I won't say it's the wild west but it kind of is. Everyone knows right? Like I mean [1:39] anything and I'm not against it. I I'll first say that I I love my Opus 5. I love my Fable [1:45] 5 for doing certain things, but at the end of the day, right, like it's not a scalable method [1:51] if you were to scale, for example, with Abel from one user to a thousand, right? I mean, [1:56] that math just doesn't necessarily compute. So I think if if anyone's seen the the letter [2:01] that Jensen did and Michael Dell also signed around kind of openweight open source models, [2:06] there was such a fundamental shift like with the launch of the Dell Promax GB10 or the the DGX [2:12] Spark, but also even in Blackwell native cards and doubling kind of VRAMm kind of genon from the ADA [2:18] to the Blackwell generation where there are models that like Neotron and others that can run locally [2:24] that are very close to foundational model or you know um that level of model and can [2:29] all run locally and save you tokens. So I think the whole point is that if you don't have to use [2:36] the cloud, you don't have to use those models and pay for it. Why why would you? Anything you [2:40] would like to add to that? Yeah, I think from a developer perspective, if you're building tools [2:44] for the enterprise, your developers, if they're um handcuffed by cost, then there's like this token [2:49] anxiety. And what I've seen a lot of developers tell me is when they're um on a credit limit, [2:55] then their work is really transactional. So you go and you turn on your system and you get this one [3:01] or two things that you need to get out and that's it. You turn it off. But when you have unlimited [3:06] tokens and you have local AI working for you and access to uh these resources that aren't metered, [3:12] that anxiety lifts and so you have that freedom to create, to innovate, and to [3:17] try different things. And I think that's where a lot of these really cool ideas really come out. [3:22] When you're handcuffed to transactional um work, it becomes more difficult to be creative. [3:29] Yeah. And you just hit that limit. You're middle of a big project and you hit the limit. I'm like, [3:32] "Oh, I have to wait till tomorrow." Which doesn't make any sense. I want to add one point. The [3:36] analogy I like to use, it's kind of like buying a Ferrari, but then you're just going to drive the [3:40] speed limit. Like why why would you ever do that? Like why would you buy a Ferrari and then because [3:45] you are maxed out, you don't want to spend the cost. you're getting in trouble for spending, [3:49] but like why would you do it, you know, when you can move that work somewhere else? [3:54] And you guys notice that even if you're running a claw or something on a local device, [3:58] but you're using a model on the cloud, you're not getting away from that token. Exactly. Right. So, [4:03] one of the things we did working with Nvidia and Dell is run everything locally. We're not just [4:08] talking about the language model, but the harness, the tools, the dealing with structured data, [4:13] unstructured data, everything on that air gap device, which can then scale to whatever size [4:18] you want. And then we worked with Dell to price it at a per server per year basis. So we are [4:24] not charging by the token, we are not charging by the user, we are not charging by the agent, [4:28] we're charging per server per year. And that allows people to just not worry about, hey, what's [4:34] going to be the unexpected cost? Now let's get to sovereign AI and maybe I'll start with you. Uh [4:41] you know sovereign AI has been talked about a lot and Nvidia has been a leader in that. So what are [4:45] you thinking about? Yeah, I think data locality, data ownership is really critical, especially [4:50] in certain industries like uh healthcare, uh security, um legal services. And so when you have [4:57] control of your data, you really take away the inefficiencies of having to go through the extra [5:01] steps to make sure that your data is protected or you're only sending away the data that is okay [5:07] uh to be sent off premises. When you have that bounded sandbox that you can play in freely, [5:12] that really helps the efficiency and the freedom again for your developers to to work. [5:17] And then from a commercialization perspective, you know, that really streamlines everything as well. [5:21] You don't have to worry about how do you uh make sure that you have the right proxies in place, [5:25] the right sandboxes in place and so it makes the job of your IT administration very easy. Yeah. [5:31] And of course this is the Michael Dale code so maybe we start with D. Yeah, I mean at the end [5:36] of the day, well, I mean, we could talk about the model sovereign AI kind of all day, right? [5:40] But I think it's the downstream impact of when you're like not using something closed, you're [5:48] not using something kind of open source or you're not you're you're sending everything to the cloud. [5:52] I've seen changes. I don't know if anyone from Dell is in this room, but I'll share [5:56] what's happened. We've gotten to the point where we're trying to build things with AI. [6:01] Fun fact, I can't access Google Drive anymore. Do you know why? Because they're afraid of [6:06] stuff getting off the network. Guess what we could not do with this presentation? [6:10] I could not pull this presentation on a a laptop key to take it off network, right? So, [6:16] because of this, I think ancillary things is we are locking down data, making it harder to get to [6:21] purely because of this, because of the anxiety of this, right? And I think with you know [6:28] mean at the end of the day your data is your lifeblood and you do not want to give it away. [6:31] So I mean do I think it's crazy that I can't put a flash drive on my computer? Sure I think that's a [6:35] little crazy but I know why we're doing it and I think it's in relation to this. So at the end of [6:39] the day being able to control the data and where it goes and understanding why it's going there, [6:44] what they're doing with it, how that's affecting your business is pretty critical at the end of [6:48] the day. And on the the point Logan was making uh at GTC we had Mount Si talk about how they did a [6:55] analytics use case on the Dell Pro Max GB10. And what was interesting about it was they were very [7:01] worried about PHIP PII and you know how do you know that your data doesn't have PHIP PII unless [7:07] you analyze it right it's a chicken and egg problem so they were able to put that data on [7:11] a air gap device not connected to the internet at all do their analysis we actually use um [7:17] Neotron um uh guardrails to actually check for is there any PHIP in it [7:23] once they were done with it they were able to push it out to a uh to a private cloud that they were [7:28] um testing it on but that mindset shifts instead of starting from can I do this like am I going to [7:35] put my corporate data at risk am I going to put a stakeholder at risk by doing this [7:40] you can shift to a mindset of let me try I have minimized the risk I'm completely airgapped [7:47] once I have tested and confirmed it works fine now I can scale it wherever I need to right it [7:52] just changes instead of governance is crucially important but this is governance by design, [7:58] right? You're making it hard to do the wrong things by designing it the correct way and then [8:03] you can scale it with at whatever environment you needed to. Hopefully, this resonates with people. [8:09] Yeah. And data is every company's secret sauce. So, that's where you can extract value, [8:13] extract intelligence, extract opportunity. And when you have sovereign AI, your data is [8:19] collectively there in its entirety. So, you're not working with chunks of data that might not have [8:24] that value that you're looking for. So, you're working with unfiltered data and you're getting [8:28] the entirety and the um the truth from that opportunity that's sitting there. [8:32] Yeah. And if you guys want if you haven't been to the Dell booth or the AEL booth yet, go try um [8:38] seeing them do a demo on the GB10 and put your hand on top of it, right? Like you literally see [8:43] it be kind of cold until you start running the agent and it'll get a little hot and then it'll [8:48] take it'll cool back down. But it's really cool that the entire stack, everything you're doing is [8:53] running air gap, not connected to the internet on that device. Completely changes the game. [8:58] Now, Claus, I'm assuming at this point everybody knows what autonomous agents and claws are, but [9:03] uh Logan, I'm going to go to you first given you were doing clause demos in your booth all day. [9:09] I love claws, man. But I hate the word claw, like kind of agent, you know. Uh [9:14] here's the thing is show of hands, and I'm not going to judge. [9:18] How many of you played with like a Hermes or Open Claw before? I hope it's more than [9:23] that because they are in the conference. Like, okay, I'm a little surprising. Okay. [9:29] Well, one quick thing. Go to his booth and try using Hermes and OpenClaw today. So, [9:33] yeah, it's cheeky. Our booth. If you know about soccer, you'll love it. If you don't know about [9:36] soccer, like me being an American, then yeah, you know. Anyways, claws are great. The problem is [9:41] you don't have to be a developer to set them up. However, you need to kind of understand [9:46] how to operate them. And what I love about Able, and I'm not a developer by trade. I'm [9:50] a hardware guy. I've always been like a product marketing sales guy. Um, Able makes it very easy, [9:56] dare I say idiot proof, to set it up. It literally has pre-made templates. And if you don't want to [10:01] do a template, it allows you to set it up from scratch. You answer a couple of questions, you [10:06] kind of refine it, there it is. But in something like OpenClaw, you're configuring your model, [10:11] you're doing all of these extra things that unless you want to spend the time to learn, which I did, [10:16] it it adds just extra complexity, right? Like at the end of the day, [10:20] everyone here, it's funny that people say with AI, we're getting more time back. I actually think I [10:24] work harder than I've ever worked in my entire life to be completely honest. Um, and they do [10:29] deliver some value back to me, but that time and setting it up is time that I'll never get back [10:34] where it allows it to scale very quickly. And then there's not the security, there's not permission [10:39] layers, there's not all the things that we talked about where I mean there might be, you know, [10:43] my personal email address might have leaked out. I have no idea. I was able that won't happen. [10:48] Well, and you were talking about the no use cases point earlier today. Let's talk about that. Right. [10:53] And well, every this is a Logan Waller's personal opinion. This is not Delta. [10:58] Uh, but everyone gets very wrapped up in use cases with like paws and agents [11:04] and I hate that because you're force-fitting what you need it to do like you're someone's [11:09] force feeding it into like here's the things it can do. The world with agents and claws is open. [11:16] It is what do you need? And I'm going to tell you an an example of something I never thought [11:20] it would be able to do. It's kind of a personal story, but make it quick is we have an ambassador [11:25] group at Dell that were uses workstations, GB10, all types of stuff. And the agency I was working [11:30] with, they capture all their social media metrics. So, you know, how many likes, clicks, impressions, [11:34] all this. And that was the value of going through the agency. They were supposed to manage them. [11:37] Unfortunately, I fired them. But that's neither here or there. But I had to recreate the metrics. [11:41] I was freaking out. I was like, how do I recapture all of these metrics? You know what I did? [11:47] basically built an agent that tracks all their profiles, looks at a reference file, [11:51] goes out and clabs posts, likes, all available, all social media data, no API calls, [11:57] no nothing like that, all running locally. Who would have ever thought that was a use case? [12:02] But that's what my use case needed to be. So when you think about agents, don't think about, hey, [12:07] what is someone telling you to do with it? It's what do you need to do with it? And for example [12:11] at the state of Nebraska they built 222 agents in 90 minutes using able right now. What is important [12:19] there was it was interns agency staff it was not expert data scientists building these agents. [12:24] So you have to get past this mindset of I need a lot of expertise to build agents to anybody [12:29] can build it. Actually you guys did a lot of interesting stuff around Nemo claw and claws are [12:33] of course very important to Nvidia. So yeah, and I think the perspective I would like to share is, [12:39] you know, I I remember when my kids asked if we can bake cookies for Santa on Christmas Eve, [12:43] and I was like, "Oh, sure. How hard can that be?" So I pulled up a recipe online and I says, "Oh, [12:47] this can be done in like 15 minutes plus bake time." But when we got into it, it took like [12:53] I don't know, hour, 2 hours. I didn't know like you know flour was going to spill everywhere. [12:57] The vanilla extract was, you know, really hard to measure cuz you only need that little bit of it. [13:01] And so with claws, I think it's kind of the same way because everyone says it's really easy [13:07] and it is to a certain respect, but there are a lot of things that you need to consider and you [13:12] when you get into it, you realize that oh, you needed a lot of preparation, a lot of auditing [13:16] to figure out what data, where is it going, how do I want to secure it, how do I want to protect it? [13:20] And so when you do that, you figure out that there's a lot to it. And the nice thing about [13:25] able is that you've gone through that and put on the hat of an enterprise user and taken all [13:30] those considerations and built around it so that the user can actually really go and see what the [13:35] simplicity of deploying claws and agents is like without having to figure out all the pitfalls that [13:40] might be along the way. Just take the friction out. That's our focus. How and this is where [13:45] working end to end becomes crucial because if we went we we're doing joint sprints with the [13:50] Neimoclot team for example at at um Nvidia. We have worked closely with Dell from day one when we [13:55] started doing this. That changes the equation of how do you take the complexity out. Right. Right. [14:01] Last uh key topic postraining. I guess I'll start with you. What are your thoughts on [14:05] post- training? And you guys invested in thinking machines if I'm not wrong. So um what are you guys [14:11] doing there? I think it's um it's an area that a lot of people should look into because everyone [14:15] thinks that when I use a model I got to pull down this fulls size model or I need to get the biggest [14:19] model I can fit on my system. and you're actually probably oversizing your problem or oversizing the [14:25] solution that is. So when you do post- training fine-tuning, you can take a model and distill it [14:31] and then fine-tune it to your data and now you have the confidence that it is grounded on the [14:35] results that you're looking for. So you reduce hallucination and also you reduce the amount [14:40] of resources that you need to power that model. So then you free up more res resources on your system [14:45] to run other parts of your pipeline and then it's easier to manage. So if you need to distribute it, [14:49] you're distributing less data uh less um u package sizes. So it's overall it's more efficient and [14:56] then the results are going to be very very good in the domain that you post train for. So I've [15:00] seen some of the results that you guys have done and maybe you you can speak more to that. Yeah, [15:04] it was a fun project we did at GDC this year. Um we posted a model and beat a 100 times bigger [15:11] model with a really small model and the post- trainining cost was just $5. And this is the thing [15:17] that people are missing. They think that post training means hundreds of thousands of dollars [15:21] of spend. It was $5 of compute but a 100 times bigger model could be beaten with that. Right? [15:27] Just get into these things. Don't be afraid. These technologies are not hard to use. [15:33] The data flywheel from Nvidia becomes very easy to we just incorporated inside able to make it even [15:39] easier to use. But you just go give feedback and the rest is automated for you. Logan Toddson post [15:45] training. I mean, I think the wave of the future is yes, big foundational models are fantastic, [15:50] but like for example, if you were going to go and perform open heart surgery, are you going to use [15:55] a katana or are you going to use a scalpel? Think of a post-trained model as a very surgical scalpel [16:01] base for that tool that you actually need. And it's quicker, it's smaller, it can run locally, [16:07] it's tuned by you at the end of the day. And I think that's what I like about AEL is that it [16:11] kind of constantly learns. you can post train it for example like Hermes it's not really [16:15] post- training anything like it's just updating kind of the code and stuff like that and being [16:20] able to get that level I think at the end of the day there will be probably in the future a model [16:24] that's post-rained for little very little things and instead of having one f one big model you have [16:30] hundreds of little models that are doing those little tasks and I think that's the future and [16:34] by the way one of the problems with post- training was getting the data for post training because you [16:39] can't post train off of Gemini or OpenAI output It's not licensed for that, but um Nvidia released [16:46] this thing called the Neotron 3 Ultra. You should go try it. It's a pretty big model, right? So, [16:50] we use the output of the Ultra to post train the super, the output of the Super to post train the [16:55] Nano. So, you're kind of basically bootstrapping down smaller and smaller, but you're using the [17:00] smartest guys to train the next level down. Right now, we talked about sovereign AI. One of the [17:06] things I want you to understand is sovereign AI doesn't just mean in one box. Think of it as keep [17:11] your most private information, the agent memory, things of that on the device. You don't want that [17:17] to leave your device because that agent memory has a huge amount of information. But then you run [17:23] and analyze the data where it is. So we have done announcements with Google where we analyze Google [17:28] bitquery data completely inside BigQuery. That data is not coming out there. And we did 5 billion [17:34] rows of data, 10 million variable combinations for $10 of compute cost on Bitquery, right? And then [17:41] when you're running the model, run it locally if it makes sense. Run it on the device or run it on [17:45] a local server or run it on the cloud. Sovereign AI means how do I give you that option? If you [17:53] want to run everything on a GB10, that whole stack runs just on the GB10. If you want to distribute [17:58] it across your telco data center, your private cloud, your public cloud, you can do that too, [18:04] right? So don't think of this as restrictions. These are best practices that can run anywhere. [18:09] And the other part of this is if you look at clause and stuff like that because the claws do [18:13] planning, try doing actually you're going to waste a bunch of tokens. So maybe don't do it. If you [18:18] go and say hello to Hermes, it'll actually often take a bunch of tokens to say hello back to you. [18:23] 30,000. Mine was 20,000 but yeah I'll take your number but the problem is that darn agent is [18:30] doing a lot of data loading a lot of planning before it does anything right so what we found [18:36] was enterprises actually don't want that they want the exact same process executed the exact same way [18:42] they want the governance they want the control so what we do is we use claw for prototyping [18:49] build whatever you want very very quickly once you like it save it of executed at scale in a [18:56] deterministic way. That combination actually works for the enterprise, right? So you're getting the [19:00] governance and cost containment. You're not wasting that planning step every time when [19:04] you do the template. And that's the able side. And we talked about how we get start out with 100 plus [19:10] templates, but you just build as many templates as you want. That's the way to come at this problem. [19:14] These are not either or. These are both, if you will. But let's go into final thoughts. Logan. [19:21] Yeah, I mean, you can read the quote, but I'll say this is I've been at Dell for a long time. [19:27] Um, and it's the first time I've ever seen kind of a thing around pricing and tokconomics really like [19:33] break something, right? Like, and I I just find it so weird, but agents basically, it's so wild that [19:40] I have agents running on my computer right now, right? I'm not watching the screen. They're doing [19:43] everything independently. But the big thing is we don't want them leaving the building. And I [19:47] this is why I love that Dell, Nvidia and Able are collaborating to if you choose to do so, [19:52] keep your agents local, run them, iterate them, take them up production at scale, whether it's [19:57] a server, wherever you want to be, right? So I absolutely love that. And over to you, Daniel. [20:03] Yeah, I think if you think back to the chat GBT moment, it was amazing. We had answers anywhere, [20:09] everywhere, all the time. And I think Logan, you and I, we we were there in 2023. we were there and [20:14] just talking about like wow what you can do with chat GBT and that was just getting answers [20:19] but now we have that moment again and it's not answers but it's outcomes and with agents [20:24] I'm super excited that you can get outcomes whenever you want and so that is an incredible [20:29] um upgrade to just answers you tell the agent what you want to see at the very end and then [20:34] leave it up to the agent to figure it out and do that for you and free you up from having to do [20:38] that figuring out and you can be more productive of course like Logan said you're going to be [20:42] busier but you're going to be busier with things that you actually want to do. And so I think the [20:47] future is bright. There's a lot of opportunity and I'm really excited about um the opportunities [20:52] that Nvidia, Abel and Dell are able to bring not just to developers but to the enterprise because [20:57] Abel has figured out and thought about those things that enterprises need and we bring that [21:01] developer mindset to that enterprise user. And all I would say just go and try it today guys. [21:07] This is so blooming easy like most people think it's a magic trick. So if you go to either the [21:13] Dell booth or the AEL booth, you can make them build an agent in front of you. Or even better, [21:18] just do it yourself. Okay? If you think this is hard, you're completely missing the point. [21:24] Just try it. Once you have tried it, look at the token cost. Decide what the cost [21:29] would be when you scale it. Decide the right platform to scale it on. Let's go. [21:34] And then iterate and improve. Iterate and improve. Thanks everyone. Thank you for [21:38] coming to this session. Hopefully this was useful for everyone. You should. Thank you.