Scaling AI Development with DGX Cloud: ServiceNow and SLB Production Deployments

nvidia · Tech · 2024 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: enterprise and scientific foundation-model infrastructure

ServiceNow and SLB needed to train, evaluate, and productize domain foundation models quickly without turning large GPU-cluster operations into the bottleneck.

Constraints:

• ServiceNow wanted frontier-like reasoning performance from smaller models so enterprise agent workloads would not depend on expensive frontier models for every task.
• SLB's energy workflows require multimodal scientific outputs rather than generic text-to-text responses.
• Both organizations needed high GPU utilization, reliable cluster operations, optimized networking and storage, and support for fast product iteration.
• SLB's customers have proprietary data and trust requirements that may require customer-specific fine-tuning or on-prem deployment.
• The model infrastructure had to fit into existing tool portfolios, physics simulators, RAG pipelines, and agentic orchestration layers.

Architecture

NVIDIA DGX Cloud on AWS supplied a managed full-stack training platform combining GPU clusters, optimized networking, Lustre storage, Run:ai scheduling, EKS integration, elastic cloud capacity, and enterprise support. ServiceNow used it to train and productionize small Apriel models for enterprise agent scaffolding, while SLB used it to accelerate seismic and petrophysical foundation models that integrate with scientific applications, customer data, and existing physics/modeling tools.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["DGX Cloud on AWS"] --> B["Run:ai scheduler"]
  A --> C["Lustre storage"]
  A --> D["EKS orchestration"]
  A --> E["ServiceNow foundation-model lab"]
  E --> F["ServiceNow agentic scaffolding"]
  F --> G["Customer policies and tools"]
  A --> H["SLB data and AI platform"]
  H --> I["Seismic foundation model"]
  H --> J["Petrophysical foundation model"]
  H --> K["RAG ingestion pipelines"]
  H --> L["Physics-based simulators"]
  I --> M["Geoscientist workflows"]
  J --> M
  L --> M

Components

Data flow

ServiceNow trains compact foundation and reasoning models on DGX Cloud, evaluates them with high-utilization Run:ai scheduling, then puts them behind ServiceNow's agentic scaffolding, policies, tools, and guardrails. SLB trains seismic and petrophysical domain models on DGX Cloud, combines them with public, proprietary, customer, and document/RAG data sources, and uses them inside scientific workflows alongside physics-based simulators and existing applications.

Why this architecture

The shared architecture externalized hard cluster operations while preserving flexibility for different model and deployment needs. ServiceNow needed small efficient models surrounded by agentic scaffolding; SLB needed domain-specific multimodal models integrated into scientific workflows. DGX Cloud provided the managed GPU stack, scheduler, storage, networking, elasticity, and support so each company could spend engineering effort on models and products rather than cluster maintenance.

Models used: ServiceNow Apriel models · 5B parameter model · 15B parameter reasoning model · seismic foundation model · petrophysical foundation model · domain foundation models

Scale

• ServiceNow targets 5B and 15B parameter Apriel models.
  “There was a 5-billion-parameter model that we trained, and then we trained a 15-billion-parameter reasoning model.”transcript, stated
• ServiceNow's 15B model is compared against much larger 235B and 600B models.
  “we are at 15-billion-parameter scale. DeepSeek R1 is about 600 billion. And then Quinn is about, you know, you see it's about like 235 billion.”transcript, stated
• ServiceNow reported close to 100,000 open-source downloads for the model.
  “the open-source community also seems to love it because we got like close to 100,000 downloads.”transcript, stated
• ServiceNow reports almost 100% GPU utilization across training, data synthesis, and evaluation.
  “We are able to get almost like 100% utilization either with training models or synthesizing data or evaluating.”transcript, stated
• ServiceNow reports close to zero downtime over roughly the last year.
  “we've had like over the last year or so, close to like zero downtime, like maybe a few hours on one day”transcript, stated
• SLB operates in more than 120 countries.
  “we operate in over 120 countries, as well as helping people with their new energy needs.”transcript, stated
• SLB shipped two major product releases over two years.
  “we've just done two major project product releases, Generative AI and Agentic Assistant in the last two years.”transcript, stated
• SLB's seismic model progressed from 2D to 2.5D to 3D.
  “we moved from a 2D model to a two-and-a-half-D to a 3D model for the Seismic Foundation model.”transcript, stated

Alternatives considered (roads not taken)

Operate large in-house clusters

Rejected because: Cluster maintenance was a full-time job and made hardware refresh harder.

Use frontier models for every ServiceNow agentic workload

Rejected because: The cost would make common employee workflows impractical at enterprise scale.

Use only RAG for SLB specialty scientific domains

Rejected because: SLB separates corporate-memory retrieval from scientific-output models, using fine-tuned domain foundation models for seismic and petrophysical outputs.

Train one generic SLB model for every customer

Rejected because: Major customers often need fine-tuning on their own data to match their geological settings and build trust.

Patterns

latency-optimization (infrastructure)

DGX Cloud bundled optimized GPU, networking, storage, scheduling, and support so teams could reduce infrastructure delay and accelerate model-product delivery.

cost-aware-routing (model-ops)

ServiceNow designs small models for most workloads while reserving frontier models for the most complex cases.

fine-tuning-adaptation (model-ops)

SLB trains and fine-tunes seismic and petrophysical foundation models for scientific outputs instead of relying on generic text models.

tool-use-orchestration (agent-architecture)

ServiceNow treats the language model as one part of an agentic stack that includes tools, policies, triggers, and guardrails.

safety-policy-enforcement (governance)

ServiceNow supports customer-defined policies and guardrails around agentic workflows.

human-in-the-loop-gate (governance)

SLB frames the models as assistants for geoscientists whose trust and domain expertise remain central.

Tradeoffs

Training infrastructure

Chose DGX Cloud on AWS over continuing to operate only self-managed hardware clusters.

Rationale: Managed cloud infrastructure reduced cluster-operation burden and gave access to newer hardware and optimized stack components.

ServiceNow model strategy

Chose small specialized models plus agent scaffolding over frontier models for every workflow.

Rationale: Small models reduce cost while the surrounding platform handles policies, tools, documents, and orchestration.

SLB knowledge strategy

Chose RAG for documents and fine-tuned domain models for scientific outputs over one retrieval pattern for every data type.

Rationale: Corporate memory and specialty scientific prediction require different treatment.

SLB deployment model

Chose hybrid on-prem, SaaS, commodity cloud, and DGX Cloud paths over one universal cloud deployment.

Rationale: Customer sensitivity, product maturity, and global SaaS delivery have different infrastructure needs.

Gotchas (what bit them)

Running large GPU clusters can become the product bottleneck.

Root cause: Maintaining cluster hardware, job management, orchestration, networking, and uptime requires specialized operational effort.

Resolution: ServiceNow moved to DGX Cloud on AWS with integrated storage, scheduler, networking, and support.

Frontier models are too expensive for broad enterprise-agent usage.

Root cause: Per-employee and per-use-case frontier-model calls accumulate cost quickly.

Resolution: ServiceNow trained small models intended to match frontier-level performance for most platform use cases.

Scientific foundation models must earn expert trust.

Root cause: SLB's outputs affect uncertain geoscience models and must fit workflows used by Ph.D.-level scientists and engineers.

Resolution: SLB trains on public, proprietary, and customer data and positions the models alongside existing physics and ML tools.

Customer data sensitivity complicates global model deployment.

Root cause: Some SLB customers cannot move proprietary data into shared cloud services.

Resolution: SLB expects to mix on-prem, SaaS, commodity cloud, and DGX Cloud patterns as products mature.

Old corporate documents are difficult RAG inputs.

Root cause: Historical reports span decades, changing language, and OCR-heavy archival formats.

Resolution: SLB uses OCR, semantic ingestion, automatic vectorization, and out-of-the-box RAG for corporate memory.

Evaluation

The session reports operational and model-progress evidence rather than a single benchmark harness: GPU utilization, downtime, model-size/performance comparisons, open-source downloads, product-release velocity, throughput gains, and movement from 2D to 3D seismic modeling.

Metrics: Almost 100% utilization · Close to zero downtime over roughly a year · 5B and 15B ServiceNow models · 15B model compared to 235B and 600B models · Close to 100,000 downloads · Two major SLB product releases in two years · 2D to 2.5D to 3D seismic-model progression

Outcomes

• ServiceNow reports close to 100% cluster utilization with Run:ai.
• ServiceNow reports close to zero downtime over roughly a year.
• ServiceNow released 5B and 15B Apriel models and reports close to 100,000 downloads.
• SLB accelerated two major product releases over two years.
• SLB moved its seismic foundation model journey from 2D to 2.5D to 3D.
• Both companies describe DGX Cloud as allowing teams to focus on models and products rather than infrastructure.

ServiceNow reports 5B and 15B models, almost 100% GPU utilization, close to zero downtime, and close to 100,000 downloads. SLB operates in more than 120 countries, shipped two major AI product releases in two years, and advanced seismic modeling from 2D to 2.5D to 3D.