# The Deskside Advantage: Secure, Scalable AI (Presented by Dell Technologies)

## Abstract

Key takeaways Be among the first to see what's next for Dell Deskside Agentic AI. Dell Technologies and NVIDIA will preview new innovations designed to help organizations deploy, secure, and scale agentic AI workloads with greater confidence and operational efficiency. Tokenomics matters. Autonomous reasoning loops can multiply token consumption exponentially, making AI cost forecasting and control just as critical as model performance. Local execution changes the equation. Running production agentic AI models at the deskside improves latency, strengthens data control and privacy, and delivers cost predictability—while reducing reliance on public-cloud APIs. Real-world validation. Discover how Baker Hughes leveraged NVIDIA PhysicsNeMo and the Dell Pro Precision with GB300 to train a physics-informed AI surrogate for compressor families—achieving 99.5% median accuracy in predicting 3D flow fields while cutting inference times from hours to fractions of a second, transforming batch engineering into interactive, real-time design exploration.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:05 – 00:00:08]** Hi, everyone, and thanks for joining us today.

**[00:00:08 – 00:00:12]** We're going to talk about how Dell AI Factory and NVIDIA

**[00:00:12 – 00:00:16]** is transforming the desk site into a real engine of AI innovation.

**[00:00:16 – 00:00:19]** We'll cover some new announcements, walk through how they work,

**[00:00:19 – 00:00:21]** and most importantly, we're gonna share a customer story

**[00:00:21 – 00:00:25]** that brings all of it to life.

**[00:00:26 – 00:00:28]** My name is Bruno Maddaloni.

**[00:00:28 – 00:00:30]** I'm a senior solutions architect and AI advisor

**[00:00:30 – 00:00:34]** at Dell Technologies, and I design and deliver end-to-end AI

**[00:00:34 – 00:00:38]** solutions to our customers based on Dell infrastructure and, obviously,

**[00:00:38 – 00:00:41]** NVIDIA Accelerated Compute and the whole software ecosystem.

**[00:00:41 – 00:00:44]** And today I'm joined by Richard from Baker Hughes.

**[00:00:44 – 00:00:45]** Richard, would you like to introduce yourself?

**[00:00:45 – 00:00:49]** Yes, I'm Richard Borge, and I lead an AI and Accelerated

**[00:00:49 – 00:00:51]** Computing team at Baker Hughes.

**[00:00:51 – 00:00:52]** Awesome, thank you.

**[00:00:52 – 00:00:55]** And together we walk you through the solution architecture,

**[00:00:55 – 00:00:59]** what we're presenting today, highlight some key benchmarks,

**[00:00:59 – 00:01:03]** and most importantly, demonstrate how you can make your desk

**[00:01:03 – 00:01:07]** site an engine for AI innovation.

**[00:01:09 – 00:01:12]** So, what's behind that?

**[00:01:12 – 00:01:17]** Dell AI Factory and NVIDIA is putting serious AI capabilities

**[00:01:17 – 00:01:20]** onto the desk side of employees and businesses, basically.

**[00:01:20 – 00:01:24]** And we're helping organizations to move past basic AI systems

**[00:01:24 – 00:01:30]** more to autonomous multi-step workflows, everything running

**[00:01:30 – 00:01:34]** locally on your premise and under your enterprise control.

**[00:01:34 – 00:01:37]** So companies like Baker Hughes, by the way, are already using Dell

**[00:01:37 – 00:01:39]** workstations to do exactly that.

**[00:01:39 – 00:01:44]** So they're running AI simulation models to go faster, iterate

**[00:01:45 – 00:01:48]** through the work faster, and evaluate more designs.

**[00:01:48 – 00:01:52]** And the advantages of local AI executions are clear.

**[00:01:52 – 00:01:56]** You have more control over your costs, over your data,

**[00:01:56 – 00:02:00]** and very important, you have no dependencies on cloud providers,

**[00:02:00 – 00:02:02]** SaaS providers, or anything else.

**[00:02:02 – 00:02:04]** Everything is in your hand.

**[00:02:07 – 00:02:11]** And there's a pattern we see in nearly every enterprise right now.

**[00:02:11 – 00:02:15]** There's like a software developer or a small team doing an AI

**[00:02:15 – 00:02:18]** proof of concept, and everything works fine, everybody's happy,

**[00:02:18 – 00:02:21]** it's super cool, and leadership suddenly says, great, let's

**[00:02:21 – 00:02:23]** scale that to production.

**[00:02:23 – 00:02:25]** And that's usually where such projects stall.

**[00:02:25 – 00:02:28]** Traditionally, a lot of the times, scaling AI projects

**[00:02:28 – 00:02:32]** meant booking some cloud or SaaS resources, throwing everything

**[00:02:32 – 00:02:34]** in there, and hoping for the best.

**[00:02:34 – 00:02:37]** But this introduces a whole new set of problems; higher

**[00:02:37 – 00:02:41]** costs, latency, data leaving your premises, and sometimes

**[00:02:41 – 00:02:44]** even the IT team having to manage an environment they

**[00:02:44 – 00:02:46]** didn't even plan for.

**[00:02:47 – 00:02:51]** So this is exactly what Dell DeskSide Agentic AI is solving.

**[00:02:52 – 00:02:56]** We've built five ready-to-deploy agentic AI solutions that

**[00:02:56 – 00:03:01]** you can deploy on your DeskSide, basically, ranging from autonomous

**[00:03:01 – 00:03:05]** agents, software assistance, video search, and many more.

**[00:03:05 – 00:03:08]** And the products this is running on are tiered into

**[00:03:08 – 00:03:10]** three different tiers.

**[00:03:10 – 00:03:13]** The first one is the small tier, enabling one employee,

**[00:03:13 – 00:03:18]** one software developer, speaking the Dell Pro Max, for example,

**[00:03:18 – 00:03:20]** or the Dell Pro Precision.

**[00:03:20 – 00:03:23]** If you want to move that further to enable a whole team, or

**[00:03:23 – 00:03:25]** even a whole department, you can have a look at the Dell

**[00:03:25 – 00:03:28]** Precision 9 T4, for example.

**[00:03:28 – 00:03:33]** Or if we take that one step even further and want to enable

**[00:03:33 – 00:03:37]** a whole enterprise rollout, you can have a look at the GB300.

**[00:03:37 – 00:03:41]** That's huge AI capabilities in a box that fits under your

**[00:03:41 – 00:03:44]** desk, if you have a high desk.

**[00:03:45 – 00:03:52]** The GB300 can handle up to 96 concurrent users,

**[00:03:52 – 00:03:54]** concurrent meaning hitting the box at the same time.

**[00:03:54 – 00:03:56]** And this scales very linearly.

**[00:03:56 – 00:03:59]** You can scale from one box up to over a thousand boxes,

**[00:03:59 – 00:04:04]** deploying the next one with the next blueprint, basically.

**[00:04:05 – 00:04:10]** So if your team has AI working in the lab already and is trying

**[00:04:10 – 00:04:13]** to bridge the gap from this lab to the production environment, this

**[00:04:13 – 00:04:16]** is basically the bridge for that.

**[00:04:18 – 00:04:20]** Sorry about that.

**[00:04:20 – 00:04:25]** And Baker Hughes, by the way, is one of the largest energy

**[00:04:25 – 00:04:29]** technology companies, and they're facing engineering challenges

**[00:04:29 – 00:04:31]** every engineering company knows.

**[00:04:31 – 00:04:32]** They have simulation models they have to run.

**[00:04:32 – 00:04:34]** This takes some time.

**[00:04:34 – 00:04:36]** They want to iterate faster and obviously improve

**[00:04:36 – 00:04:37]** the time to market.

**[00:04:37 – 00:04:41]** And they did exactly that with Dell DeskSight Agendic AI.

**[00:04:41 – 00:04:43]** Richard, I would hand it over to you.

**[00:04:44 – 00:04:46]** Thank you very much, Bruno.

**[00:04:46 – 00:04:50]** So Computational Fluid Dynamics, or CFD, is the gold standard

**[00:04:50 – 00:04:54]** for simulating fluids in the engineering design process.

**[00:04:54 – 00:04:57]** Over the next 10 minutes, I'll walk you through the challenge we

**[00:04:57 – 00:05:02]** solved, the model we built, and how it achieved near-CFD accuracy in

**[00:05:02 – 00:05:07]** a tenth of a second on designs it had never seen before, and why it

**[00:05:07 – 00:05:09]** took three companies to get there.

**[00:05:09 – 00:05:13]** But first, a quick word on who Baker Hughes is, because the

**[00:05:13 – 00:05:17]** challenge we solved comes straight from the machines we build.

**[00:05:18 – 00:05:21]** Baker Hughes is an energy and technology company.

**[00:05:21 – 00:05:24]** We design and deploy some of the most advanced technologies to serve

**[00:05:24 – 00:05:28]** energy and industrial customers looking for more reliable, more

**[00:05:29 – 00:05:32]** efficient, and cleaner solutions.

**[00:05:32 – 00:05:36]** And our brand promise is rewriting the energy equation.

**[00:05:36 – 00:05:39]** This is the link between the energy sources the world relies

**[00:05:39 – 00:05:43]** on and the industrial outcomes they make possible.

**[00:05:43 – 00:05:47]** And today, we're living in what we refer to as the demand

**[00:05:47 – 00:05:50]** decade, where the world needs historic amounts of additional

**[00:05:50 – 00:05:55]** energy to power economies, industries, and the AI data

**[00:05:55 – 00:05:57]** centers many of you are building.

**[00:05:57 – 00:06:02]** And it needs to be secure, affordable, and sustainable.

**[00:06:02 – 00:06:05]** Delivering that energy requires machines, and one machine shows up

**[00:06:05 – 00:06:08]** almost everywhere, the compressor.

**[00:06:09 – 00:06:14]** These machines move and pressurize gas in carbon capture systems,

**[00:06:14 – 00:06:19]** on-site power gen, and pipelines, among many other applications.

**[00:06:19 – 00:06:23]** So let me show you why designing one's so hard.

**[00:06:24 – 00:06:27]** Baker Hughes has designed and deployed 5,000

**[00:06:27 – 00:06:30]** compressors since the 1960s.

**[00:06:30 – 00:06:33]** On the far right is a centrifugal compressor.

**[00:06:33 – 00:06:39]** The way it works is low pressure gas comes in, a rotating impeller

**[00:06:39 – 00:06:43]** adds energy to that gas, and then a stationary diffuser

**[00:06:43 – 00:06:46]** turns that energy into pressure.

**[00:06:46 – 00:06:48]** We then stack these stages in series to

**[00:06:48 – 00:06:50]** keep raising the pressure.

**[00:06:50 – 00:06:55]** And designing it follows a loop that many engineers will recognize.

**[00:06:55 – 00:06:59]** First, you have the geometry in a CAD design, you mesh that geometry,

**[00:07:00 – 00:07:04]** you run CFD, and then the engineer evaluates the results, goes through

**[00:07:04 – 00:07:06]** a lengthy optimization process.

**[00:07:06 – 00:07:08]** And here's the challenge.

**[00:07:08 – 00:07:13]** Just one geometry variant and a single operating point,

**[00:07:13 – 00:07:15]** it can take 12 hours.

**[00:07:15 – 00:07:19]** And for a new stage design, this may be thousands of variants.

**[00:07:19 – 00:07:23]** So the simulation sets the pace, not the engineering.

**[00:07:23 – 00:07:27]** And highly skilled engineer teams are only able to evaluate

**[00:07:27 – 00:07:31]** a small slice of the design space, and they spend their time

**[00:07:31 – 00:07:34]** setting up, waiting on solver runs.

**[00:07:35 – 00:07:38]** AI surrogates have been the answer to this for years.

**[00:07:38 – 00:07:41]** The real question was, could you design one that generalized

**[00:07:41 – 00:07:47]** across stage families, was near safety accurate, and didn't require

**[00:07:47 – 00:07:50]** an entire design study to train on?

**[00:07:50 – 00:07:54]** So let's see how we built it.

**[00:07:54 – 00:08:00]** We built a physics-informed graph neural network built

**[00:08:00 – 00:08:05]** on MeshGraphNet and trained on our CFD simulations.

**[00:08:05 – 00:08:10]** It predicts the full 3D flow fields directly from the simulation mesh.

**[00:08:10 – 00:08:12]** And it works in a three-step process,

**[00:08:12 – 00:08:15]** encode, process, and decode.

**[00:08:16 – 00:08:20]** For encoding, each node and edge turns into a compact numerical

**[00:08:20 – 00:08:25]** representation that includes the design and the operating point.

**[00:08:25 – 00:08:28]** This information is then passed to neighboring nodes until

**[00:08:28 – 00:08:30]** it spans the whole geometry.

**[00:08:30 – 00:08:34]** Then lastly, this information is decoded into temperature,

**[00:08:34 – 00:08:39]** pressure, and velocity, along with system-level key performance

**[00:08:39 – 00:08:44]** indicators, or KPIs, like efficiency, pressure ratio,

**[00:08:44 – 00:08:50]** that engineering teams must know to make a design decision.

**[00:08:50 – 00:08:54]** In the bottom right is what makes it generalizable and trustworthy.

**[00:08:54 – 00:08:58]** We train on two goals at once, match the CFD results and

**[00:08:58 – 00:09:02]** satisfy the steady RANS equations.

**[00:09:04 – 00:09:06]** And this ensures that our predictions remain

**[00:09:06 – 00:09:10]** physically consistent, especially on geometries

**[00:09:10 – 00:09:13]** outside of the training set.

**[00:09:13 – 00:09:17]** We also used NVIDIA's physics AI framework, which allowed

**[00:09:17 – 00:09:22]** us to put the governing equations directly into the model.

**[00:09:22 – 00:09:26]** And because we included both the design and operating points

**[00:09:26 – 00:09:31]** in this, one model now applies to an entire stage family,

**[00:09:31 – 00:09:34]** not one model per stage.

**[00:09:35 – 00:09:38]** And right here, we have memory.

**[00:09:38 – 00:09:43]** For the memory, one of these high-fidelity CFD simulations

**[00:09:43 – 00:09:45]** is about 8 million points.

**[00:09:45 – 00:09:49]** And it simply didn't fit on an A100 hardware that we were

**[00:09:49 – 00:09:51]** using at the time.

**[00:09:51 – 00:09:53]** So we engineered around it.

**[00:09:53 – 00:09:56]** We used mixed precision, gradient checkpointing,

**[00:09:56 – 00:09:58]** and adaptive mesh sampling.

**[00:09:58 – 00:10:01]** And it worked; we shrunk it down, but it also took months

**[00:10:01 – 00:10:06]** of engineering effort to do it, all because of a hardware limitation.

**[00:10:06 – 00:10:11]** So let's talk a little bit more about how well it worked.

**[00:10:12 – 00:10:17]** These models, they need to be most effective; they need to

**[00:10:17 – 00:10:22]** address generalization, performance prediction, and data efficiency.

**[00:10:23 – 00:10:26]** However, in this domain, many of these models can only address

**[00:10:26 – 00:10:29]** one, maybe two of these.

**[00:10:29 – 00:10:34]** For generalization, it's normally one model per stage design.

**[00:10:34 – 00:10:38]** Our model is valid across a 15-stage family.

**[00:10:38 – 00:10:44]** And we held three stages completely blind, made up of 62 simulations.

**[00:10:44 – 00:10:48]** For performance prediction, the models can generally predict

**[00:10:48 – 00:10:52]** the 3D flow fields, but they fall short on the system level

**[00:10:52 – 00:10:56]** KPIs, and an engineer can't make a decision based on that.

**[00:10:57 – 00:11:01]** Our model, using that blind set I just referred to, it achieved

**[00:11:01 – 00:11:05]** or exceeded 99.5% median accuracy

**[00:11:07 – 00:11:12]** On both the 3D flow fields and the system-level KPIs.

**[00:11:12 – 00:11:18]** For data efficiency, every single training point is a 12-hour run.

**[00:11:18 – 00:11:22]** So a model that needs hundreds or thousands of data points to train

**[00:11:22 – 00:11:25]** on eats away at its own savings.

**[00:11:25 – 00:11:29]** This model was trained on only 52 simulations.

**[00:11:29 – 00:11:31]** And here's the part that really matters in practice.

**[00:11:32 – 00:11:33]** They weren't hand-picked.

**[00:11:33 – 00:11:37]** They were chosen at random across 12 stage designs and

**[00:11:37 – 00:11:39]** their associated operating points.

**[00:11:39 – 00:11:43]** Generally, an engineer who already knows where the machine

**[00:11:43 – 00:11:48]** performs well and where it doesn't selects this training set.

**[00:11:49 – 00:11:52]** So this model is both data efficient, and it doesn't

**[00:11:52 – 00:11:55]** need the answer before it starts.

**[00:11:56 – 00:11:59]** Now, let's go back to that design loop where we had the

**[00:11:59 – 00:12:01]** 12-hour solve time.

**[00:12:01 – 00:12:03]** We need to compress that.

**[00:12:03 – 00:12:06]** Using this model that we trained, the inference time is now

**[00:12:06 – 00:12:08]** a tenth of a second.

**[00:12:08 – 00:12:12]** That is roughly 363,000 times faster than a 16 CPU CFD run.

**[00:12:18 – 00:12:21]** Additionally, to put that in perspective, if you needed

**[00:12:21 – 00:12:26]** to run a 100,000 case Monte Carlo design study, this would go from

**[00:12:26 – 00:12:31]** 2,000 CPU years to an afternoon.

**[00:12:31 – 00:12:35]** That is engineering design at conversational speed.

**[00:12:35 – 00:12:38]** You ask the question, and you get the answer while you're

**[00:12:38 – 00:12:41]** still thinking about it.

**[00:12:41 – 00:12:46]** And CFD still gets the final say in the designs that are

**[00:12:46 – 00:12:50]** chosen, but now engineers can afford to ask far more

**[00:12:50 – 00:12:53]** questions before they get there.

**[00:12:54 – 00:12:58]** And none of this came from just one company.

**[00:12:58 – 00:13:02]** It takes multiple companies to achieve this kind of result.

**[00:13:03 – 00:13:07]** Baker Hughes brought the challenge, the physics and AI model,

**[00:13:07 – 00:13:12]** and decades of compressor design and validation data, along with the

**[00:13:12 – 00:13:14]** standard for what correct means.

**[00:13:14 – 00:13:18]** Otherwise, you have a fast model with nothing to be right about.

**[00:13:19 – 00:13:22]** NVIDIA brought the physics and AI framework, which allowed

**[00:13:22 – 00:13:26]** us to put the governing equations directly into the model.

**[00:13:26 – 00:13:31]** Otherwise, you have a curve fit that breaks on new geometries.

**[00:13:31 – 00:13:34]** And, do you remember all of those months we spent shrinking

**[00:13:34 – 00:13:38]** that graph down to fit in an A100?

**[00:13:38 – 00:13:42]** Well, Dell, they brought the ProPrecision with GB300 and

**[00:13:43 – 00:13:47]** Coherent Memory, which now fits the full resolution graph,

**[00:13:47 – 00:13:52]** or about 4.5 million points versus 0.6 million points on an A100.

**[00:13:55 – 00:13:59]** And while this research is still in the early phase,

**[00:13:59 – 00:14:01]** the potential is broad.

**[00:14:01 – 00:14:05]** Engineering that moves at conversational speed can help

**[00:14:05 – 00:14:10]** compressor technology to move faster, to meet the evolving

**[00:14:10 – 00:14:13]** demands of the energy and industrial sectors.

**[00:14:13 – 00:14:17]** So when the world asks for more energy, faster engineering

**[00:14:17 – 00:14:18]** helps answer that call.

**[00:14:18 – 00:14:20]** Thank you, and back to Bruno.

**[00:14:20 – 00:14:23]** Thank you very much, Richard, for that very cool view

**[00:14:23 – 00:14:26]** behind the curtains.

**[00:14:28 – 00:14:31]** We just discussed briefly how and why DeskSight is emerging

**[00:14:31 – 00:14:35]** as such a serious AI platform now, and especially how Baker

**[00:14:35 – 00:14:38]** Hughes used that to their advantage and transformed their business

**[00:14:38 – 00:14:40]** and processes in that regard.

**[00:14:40 – 00:14:43]** If you want to learn more about all of that with Dell,

**[00:14:43 – 00:14:47]** Visit dell.com slash desk site agentic AI, feel free to reach

**[00:14:47 – 00:14:49]** out to the Dell representatives.

**[00:14:49 – 00:14:53]** We also offer our complimentary AI accelerator workshop.

**[00:14:53 – 00:14:56]** This is usually where we sit down with customers, we identify

**[00:14:56 – 00:14:59]** your high-value use cases, and we actually build a roadmap

**[00:14:59 – 00:15:01]** to achieve those.

**[00:15:01 – 00:15:04]** And last but not least, if you're looking for a flexible

**[00:15:04 – 00:15:07]** consumption model, we can offer that as well with APEX services.

**[00:15:07 – 00:15:10]** For example, we can offer PC as a service, meaning You're

**[00:15:10 – 00:15:14]** transforming all that capex cost that may be upfront too much

**[00:15:14 – 00:15:18]** into operational costs you can draw out over a longer period of time.

**[00:15:18 – 00:15:22]** Thank you so much for joining us today, and we're looking forward

**[00:15:22 – 00:15:24]** to develop UAI journeys together.

**[00:15:24 – 00:15:26]** Thanks, everybody. Thank you.

