Distributed AI inferencing, measured in production.

Cross-border ADAS model serving at 38 ms p95

Replaced a single-region GPU fleet with Aether's sovereign edge tier to serve perception and prediction models to fleets in Germany and eastern China under GDPR + PIPL constraints.

p95 latency

38 ms

from 112 ms

Inference cost

-47%

per 1M tokens-equiv

Data residency

100%

in-region

Workload

Vision transformer (1.8B) + trajectory model, 42k QPS peak

Deployment

6 Sovereign Edge PoPs · 2 Core IDC (Frankfurt, Shanghai)

"Aether let us keep German driving data in Frankfurt and Chinese data in Shanghai while running the same model version everywhere."

— VP, Autonomous Driving Platform

24-region LLM serving without egress lock-in

Migrated a 70B-parameter coding copilot from a hyperscaler to Aether's hotel edge network to reach 24 metros with in-country inference and no egress fees.

TTFT

-62%

median

Serving cost

-38%

per token

Regions live

24

in 9 weeks

Workload

70B MoE, streaming completions, 180M req/day

Deployment

3 Core IDC + 41 Hotel Edge sites, KV-cache tiering

"The hotel edge tier is what unlocked LATAM and MEA for us. Nobody else had capacity where our users actually were."

— Head of Infrastructure

Diffusion-based relighting shipped into artist workstations

Wired Aether inferencing into Nuke and Houdini so artists in Vancouver, London and Seoul call the same relighting and matte-generation models with sub-100ms first-token time.

Iteration time

-71%

per shot

Artist throughput

+2.4x

shots/day

GPU utilization

84%

24h avg

Workload

SDXL-class diffusion + control nets, burst 8k jobs/hr

Deployment

Core IDC (Ashburn, Frankfurt) + 12 Sovereign Edge

Real-time fraud scoring at 6 ms in-region

Deployed a graph + sequence ensemble across Singapore, Jakarta and Manila with strict MAS residency, replacing a homegrown Kubernetes fleet.

Scoring latency

6 ms

p99

False positives

-31%

vs prior stack

Uptime

99.995%

12-month

Workload

Ensemble scoring, 380k TPS sustained

Deployment

Sovereign Edge (SG, ID, PH) · dual-active

HIPAA-compliant radiology inference at the bedside

Hotel-edge nodes co-located with 34 hospital campuses run CT/MR triage models with PHI never leaving the covered entity boundary.

Report turnaround

-58%

median

PHI egress

0

bytes

Sites onboarded

34

in 6 months

Workload

3D UNet + report generation LLM (13B)

Deployment

34 Hotel Edge + 2 Core IDC (Ashburn, Dallas)

"The BAA and the sovereign-edge topology together were the reason our compliance team approved this in weeks, not quarters."

— CISO

Sovereign Arabic-English translation for public services

Delivered a fully in-country LLM inference stack under UAE PDPL for citizen-facing services, with air-gapped model updates.

Response latency

42 ms

p95

Sovereignty

100%

in-country

Concurrent users

1.8M

peak

Workload

Bilingual 34B model, 12M requests/day

Deployment

2 Core IDC (Dubai, Riyadh) + Sovereign Edge