What is ServiceNow data-center strategy through 2027?

ServiceNow's 2026-27 data-center strategy is three-pronged: (1) hyperscaler primary on AWS + Azure + GCP for region breadth and elastic GPU capacity, (2) sovereign cloud builds in EU/UK/Saudi/India/Australia to clear regulator-driven RFP gates, and (3) named GPU partnerships (NVIDIA Blackwell + AWS Trainium2) to host Now LLM and AI Agent Studio inference at a unit cost that doesn't blow up gross margin.

Every region decision is a cost-vs-compliance tradeoff: each new sovereign region adds 50-150bps of GM drag for ~12-18 months until utilization catches up, which is why CFO Gina Mastantuono keeps repeating "infrastructure leverage" on earnings calls. The owned-DC footprint is being held flat (Toronto, Amsterdam, San Jose, Equinix sites) while net-new capacity goes to hyperscaler regions and a small number of sovereign builds.

The Workflow Data Fabric and RaptorDB sit on this hybrid stack, with data residency enforced per-tenant via region pinning. By FY27, expect ~70% of net-new capacity on hyperscalers, ~25% on sovereign clouds, ~5% owned.

The Footprint Today

AWS regions (primary hyperscaler):

• Us-east-1, us-east-2, us-west-2 (commercial US)
• AWS GovCloud (US-East, US-West) for FedRAMP High + DoD IL5
• Eu-west-1 (Ireland), eu-central-1 (Frankfurt), eu-west-2 (London)
• Ap-southeast-2 (Sydney), ap-northeast-1 (Tokyo), ap-south-1 (Mumbai)
• Trainium2 capacity in us-east-2 for Now LLM inference

Azure regions:

• East US, West US 2 for commercial multi-cloud customers
• Azure Government (US Gov Virginia, US Gov Texas) for FedRAMP High parity
• UK South for UK GovCloud workloads
• Germany West Central for EU sovereign-adjacent

GCP regions:

• Us-central1, us-east4 for commercial spillover + GenAI co-location
• Europe-west4 (Netherlands) for EU multi-cloud
• Smaller footprint than AWS/Azure, used selectively for AI workloads

Sovereign cloud regions:

• Germany sovereign cloud (BSI C5 + future EUCS)
• France SecNumCloud-aligned (via partner)
• UK GovCloud High (live, MOD-eligible)
• Saudi Arabia (PIF / Vision 2030 alignment, Riyadh)
• India (MeitY-empaneled, Mumbai + Hyderabad)
• Australia IRAP PROTECTED (Canberra + Sydney)

Owned data centers (held flat):

• San Jose, CA (HQ-adjacent, legacy)
• Toronto, ON (Canadian residency)
• Amsterdam, NL (legacy EU)
• Equinix colos in 8+ metros for low-latency edge

What Drives Expansion 2026-28

• Sovereign cloud regulations — EUCS (EU Cloud Services scheme) finalization, UK NCSC guidance, Saudi NCA cloud framework, India DPDP Act enforcement, Australia Hosting Certification Framework refresh
• FedRAMP High demand — DoD IL5 and IL6 RFPs increasingly require ServiceNow as ITSM standard; capacity in AWS GovCloud + Azure Gov is the gating factor on bookings
• GPU capacity for Now LLM inference — Now Assist + AI Agent Studio inference is doubling QoQ; NVIDIA Blackwell H200/B200 contracts and AWS Trainium2 commits go out 18-24mo
• Named-customer requirements — Deutsche Bank, BMW, Saudi Aramco, BHP, Tata Consultancy Services all driving region-specific buildouts as deal conditions
• Workflow Data Fabric residency — federated query across customer-owned data lakes requires per-region presence to keep query plans inside sovereign boundaries
• AI Control Tower telemetry — agent observability data is high-volume and sticky; needs regional capacity to avoid egress costs

The Sovereign Cloud Strategy

• Germany sovereign cloud — BSI C5 Type 2 attested today; positioning for EUCS "High" tier when finalized; Deutsche Telekom + SAP partnerships in play
• France SecNumCloud — partner-led delivery (via OVHcloud/Outscale), targets defense + intelligence accounts
• UK GovCloud High — MOD-cleared, used by HMRC, NHS Digital expansion, Ministry of Justice
• Saudi Arabia — Vision 2030 mandate, PIF investment alignment, Aramco + STC + Saudi Vision Realization Office deployments
• India MeitY DPDP — empaneled as cloud service provider; TCS, Infosys, HDFC, Reliance Jio as anchor tenants
• Australia IRAP PROTECTED — Defence + Home Affairs + ATO; Canberra-region buildout active
• Named-customer wins per region justify the GM drag during ramp

The GPU + AI Inference Strategy

• NVIDIA partnership — multi-year H200/B200 commit announced March 2024, expanded 2025 to include Blackwell Ultra; co-developed Now LLM with NVIDIA NIM microservices
• AWS Trainium2 adoption — second-source GPU strategy to reduce NVIDIA dependency; Now LLM 4o-class models being ported to Trainium2 for inference cost reduction
• GPU capacity planning — 18-24mo forward contracts; co-located in AWS us-east-2, Azure East US, and a small NVIDIA DGX Cloud footprint
• Now LLM hosting — exclusively on ServiceNow-controlled GPU pools (no third-party LLM hosting for customer data); inference latency target sub-300ms p95
• AI Agent Studio inference cost optimization — dynamic routing between Now LLM (cheap) and partner LLMs (Anthropic, OpenAI via private endpoint) based on task complexity

The Cost Discipline

• Each new sovereign region adds 50-150bps of GM drag for the first 12-18 months until utilization climbs above 60%
• Multi-cloud cost premium — Azure + GCP run roughly 8-15% more expensive than AWS-primary at equivalent capacity, justified only by customer concentration or sovereign requirement
• Sovereign cloud premium pricing — list price uplift of 20-35% on sovereign SKUs absorbs most of the infrastructure cost; net GM drag is the timing gap
• CFO Gina Mastantuono commentary — repeated framing on "infrastructure leverage as we scale Now Assist"; expect 50-100bps of GM expansion in FY27 from GPU utilization improvements
• Trainium2 substitution math — every 10% of Now LLM inference moved from H200 to Trainium2 saves an estimated 30-40% per-token cost

What's NOT In The Strategy

• No aggressive China expansion — geopolitical risk + data residency requirements make China-mainland a hard pass; Hong Kong via partner only
• No LATAM ex-Brazil sovereign cloud — Mexico, Argentina, Chile served from AWS us-east-1 + Sao Paulo; no dedicated sovereign builds planned through 2027
• No own-DC build-out beyond named existing sites — capex discipline; the Toronto/Amsterdam/San Jose footprint is held flat, all growth is hyperscaler
• No Oracle Cloud expansion — OCI not in the multi-cloud roadmap despite Oracle's GenAI push; ServiceNow staying on AWS/Azure/GCP triumvirate
• No edge-compute / on-prem appliance push — unlike Salesforce Data Cloud's Hyperforce-on-prem flirtation, ServiceNow is pure cloud; edge is partner territory

Region x Cloud Provider Status Matrix

Driver to Region to Outcome Flow

FAQ

What is the three-pronged data-center strategy through 2027? ServiceNow runs hyperscaler-primary on AWS, Azure, and GCP for region breadth and elastic GPU capacity; builds sovereign clouds in the EU, UK, Saudi Arabia, India, and Australia to clear regulator RFP gates; and uses named GPU partnerships (NVIDIA Blackwell plus AWS Trainium2) to host Now LLM and AI Agent Studio inference at controlled unit cost.

By FY27, the mix targets roughly 70% of net-new capacity on hyperscalers, 25% on sovereign clouds, and 5% owned.

How much gross-margin drag does each new sovereign region add? Each new sovereign region adds 50-150bps of gross-margin drag for the first 12-18 months, until utilization climbs above 60%. Sovereign SKU list-price uplift of 20-35% absorbs most of the infrastructure cost, so the net drag is really the timing gap during ramp.

This is why CFO Gina Mastantuono keeps repeating "infrastructure leverage" on earnings calls.

Which named customers are driving region-specific buildouts? Deutsche Bank, BMW, Saudi Aramco, BHP, and Tata Consultancy Services are all cited as driving region-specific buildouts as deal conditions. In India, TCS, Infosys, HDFC, and Reliance Jio anchor the MeitY-empaneled tenancy, while Saudi deployments tie to Aramco, STC, and the Saudi Vision Realization Office.

Named-customer wins per region justify the GM drag during ramp.

What is the GPU and AI inference strategy? ServiceNow signed a multi-year NVIDIA H200/B200 commit (announced March 2024, expanded in 2025 to Blackwell Ultra) and co-developed Now LLM with NVIDIA NIM microservices. It adopted AWS Trainium2 as a second source to reduce NVIDIA dependency, targeting sub-300ms p95 inference latency.

AI Agent Studio dynamically routes between cheap Now LLM and partner LLMs (Anthropic, OpenAI via private endpoint) based on task complexity.

How much does shifting inference to Trainium2 save? Every 10% of Now LLM inference moved from H200 to Trainium2 saves an estimated 30-40% per-token cost. GPU capacity is planned on 18-24 month forward contracts, co-located in AWS us-east-2, Azure East US, and a small NVIDIA DGX Cloud footprint.

Now LLM is hosted exclusively on ServiceNow-controlled GPU pools with no third-party LLM hosting of customer data.

Bottom Line

ServiceNow's 2026-27 infrastructure strategy is disciplined multi-cloud hyperscaler-primary, with surgical sovereign-cloud builds where regulator gates justify the GM drag, and aggressive GPU dual-sourcing (NVIDIA + Trainium2) to keep Now Assist inference economics workable. The owned-DC footprint is frozen, capex is going to hyperscaler commits, and the named CFO commentary on "infrastructure leverage" telegraphs FY27 GM expansion as GPU utilization climbs.

The differentiator vs. Salesforce/Workday isn't the footprint itself — it's the willingness to take 50-150bps of short-term GM pain for sovereign builds that unlock multi-year named-customer ARR. (see also: q1613, q1626, q1627)