PULSE.SH - 5-PAGE MULTI-PAGE ARCHITECTURE
RELEASE 4.2 // SUB-MILLISECOND GPU CONTAINERS

Deploy Serverless GPUs & Real Time Telemetry in 80ms.

Zero cold-start Firecracker microVMs, NVLink H100 inference pools, and eBPF packet observability in one unified developer cloud.

Launch Live Sandbox $ npx pulse-cli@latest deploy
pulse-cloud://prod-eu-central-fra01
ACTIVE CLUSTERS
fra-h100-010.4ms
iad-us-east0.7ms
hnd-tokyo-040.6ms
sin-ap-south0.9ms
THROUGHPUT
142,890 req/s
LIVE eBPF TRACE STREAM 99.999% 200 OK
[14:22:01.004] POST /v1/embeddings (fra-01)200 OK • 1.8ms
[14:22:01.019] POST /v1/chat/completions (iad-02)200 OK • 6.4ms
[14:22:01.042] GRPC /kv.StateSync/Replicate (hnd-04)SYNC • 0.4ms
[14:22:01.088] POST /v1/rerank-batch (sin-01)200 OK • 2.1ms
[14:22:01.112] SNAPSHOT /vm/checkpoint-nvme (fra-01)DONE • 14.0ms
NODE HEALTH
VRAM Allocation68.4 / 80 GB
NVLink Bandwidth890 GB/s
Cold-Boot P9978 ms
$ pulse scale --min 0 --max 128
✓ Autoscale policy active (80ms wake)

Scale-to-Zero GPU Pools

Never pay for idle silicon. Memory snapshots hibernate to NVMe fabric and restore on the first incoming TCP packet.

eBPF Kernel Sensor: ENABLED
Zero-overhead HTTP/2 & gRPC tracing

Kernel-Level Observability

Inspect every syscall, SQL query, and tensor allocation in real time without adding SDK bloat to your containers.

Anycast BGP • 320 Tbps Edge Capacity
DDoS L7 Mitigation: ACTIVE

Global Anycast Mesh

Requests route automatically to the nearest healthy bare-metal pop across North America, Europe, and Asia-Pacific.

INFRASTRUCTURE SAVINGS CALCULATOR

Cut Cloud Compute Bills by up to 64%.

Drag the slider to compare monthly H100 & Ephemeral vCPU spend against legacy hyperscalers.

PULSE.SH Bare-Metal$2,850 / mo
AWS EC2 P5 Instance$7,350 / mo
Google Cloud A3 VM$6,900 / mo
HOBBY & PROTOTYPE
$0 / mo

Includes $25 monthly serverless compute credit and 3 global regions.

DEDICATED RACK
Custom

Private InfiniBand spine, SOC2 Type IIBAA compliance, and 24/7 slack engineers.

ZERO-BOILERPLATE SDK

Spin Up 8x H100 Workers in 7 Lines of Code.

Native TypeScript, Python, Rust, and Go SDKs with automatic weight caching on local NVMe RAID arrays.

14.2 GB/s
Safetensors Cold Load
0 ms
Idle Standby Billing
// deploy-worker.ts
import { PulseCluster } from '@pulse-cloud/sdk';
const pool = await PulseCluster.spawn({
  gpu: 'H100-SXM5-80GB',
  regions: ['fra-01', 'iad-02', 'hnd-04'],
  autoscale: { min: 0, max: 64, wakeMs: 80 }
});
BARE-METAL POP TOPOLOGY

Global Low-Latency Datacenters

3.2 TBPS INFINIBAND SPINE • 100% HYDRO & NUCLEAR POWERED
FRA-01 // FRANKFURT● 0.4ms

512x H100 SXM5 • DE-CIX Peering

IAD-02 // ASHBURN● 0.7ms

1,024x H100 SXM5 • US-East Fiber

HND-04 // TOKYO● 0.6ms

384x H100 SXM5 • APAC Backbone

SIN-01 // SINGAPORE● 0.9ms

256x H100 SXM5 • Equinix SG3

LHR-03 // LONDON● 0.5ms

384x L40S + H100 • LINX Direct

SJC-02 // SILICON VALLEY● 0.4ms

768x H100 SXM5 • US-West Core

HARDWARE ISOLATION

Defense-in-Depth for Sovereign AI Workloads

ENCLAVE 01

NVIDIA Confidential Compute

Hardware-enforced VRAM memory encryption across PCIe and NVLink buses.

ENCLAVE 02

SOC2 Type II & HIPAA

Audited annually with automated BAA execution for clinical & financial teams.

ENCLAVE 03

Private WireGuard Mesh

Zero-config encrypted L3 peering between your AWS/GCP VPC and Pulse nodes.

ENCLAVE 04

Ephemeral Disk Scrub

Cryptographic key destruction within 50ms of container scale-down.

Infrastructure & Billing FAQ

How does Pulse.sh achieve 80ms GPU cold starts?

We snapshot initialized CUDA contexts and model weights directly to local PCIe Gen5 NVMe arrays and restore them via custom Firecracker DMA mappings.

Do you charge for ingress/egress bandwidth?

No. Unlike hyperscalers, all inbound and outbound data transfer across our 42 Anycast edge locations is 100% free and unmetered.

PAGE 02 // GLOBAL BARE-METAL TOPOLOGY (infrastructure.html)

3,200 Gbps InfiniBand H100 & B200 Superclusters

Zero virtualization tax. Direct PCIe Gen5 NVMe weight caching and RDMA over Converged Ethernet across 42 Anycast edge PoPs.

US-EAST-VA // ASHBURN

1,024x NVIDIA H100 SXM5

80GB HBM3 per GPU • 3.2 Tbps Quantum-2 InfiniBand interconnect • PUE 1.09 liquid-cooled.

EU-CENTRAL // FRANKFURT

512x NVIDIA GH200 Grace Hopper

EU GDPR Sovereign enclave • 480GB LPDDR5X + 96GB HBM3e unified coherent memory.

AP-NORTHEAST // TOKYO

768x NVIDIA L40S Inference

Optimized for sub-12ms FP8 LLM token streaming and real time voice synthesis.

pulse-cloud://prod-eu-central-fra01
ACTIVE CLUSTERS
fra-h100-010.4ms
iad-us-east0.7ms
hnd-tokyo-040.6ms
sin-ap-south0.9ms
THROUGHPUT
142,890 req/s
LIVE eBPF TRACE STREAM 99.999% 200 OK
[14:22:01.004] POST /v1/embeddings (fra-01)200 OK • 1.8ms
[14:22:01.019] POST /v1/chat/completions (iad-02)200 OK • 6.4ms
[14:22:01.042] GRPC /kv.StateSync/Replicate (hnd-04)SYNC • 0.4ms
[14:22:01.088] POST /v1/rerank-batch (sin-01)200 OK • 2.1ms
[14:22:01.112] SNAPSHOT /vm/checkpoint-nvme (fra-01)DONE • 14.0ms
NODE HEALTH
VRAM Allocation68.4 / 80 GB
NVLink Bandwidth890 GB/s
Cold-Boot P9978 ms
$ pulse scale --min 0 --max 128
✓ Autoscale policy active (80ms wake)

Scale-to-Zero GPU Pools

Never pay for idle silicon. Memory snapshots hibernate to NVMe fabric and restore on the first incoming TCP packet.

eBPF Kernel Sensor: ENABLED
Zero-overhead HTTP/2 & gRPC tracing

Kernel-Level Observability

Inspect every syscall, SQL query, and tensor allocation in real time without adding SDK bloat to your containers.

Anycast BGP • 320 Tbps Edge Capacity
DDoS L7 Mitigation: ACTIVE

Global Anycast Mesh

Requests route automatically to the nearest healthy bare-metal pop across North America, Europe, and Asia-Pacific.

INFRASTRUCTURE SAVINGS CALCULATOR

Cut Cloud Compute Bills by up to 64%.

Drag the slider to compare monthly H100 & Ephemeral vCPU spend against legacy hyperscalers.

PULSE.SH Bare-Metal$2,850 / mo
AWS EC2 P5 Instance$7,350 / mo
Google Cloud A3 VM$6,900 / mo
HOBBY & PROTOTYPE
$0 / mo

Includes $25 monthly serverless compute credit and 3 global regions.

DEDICATED RACK
Custom

Private InfiniBand spine, SOC2 Type IIBAA compliance, and 24/7 slack engineers.

PAGE 03 // DEVELOPER SDK & CLI REFERENCE (developers.html)

Deploy Custom Containers & vLLM Endpoints in 14 Seconds

First-class TypeScript, Python, and Rust SDKs with OpenTelemetry trace auto-instrumentation and zero-downtime blue/green rollouts.

// 1. Install Pulse CLI & Authenticate
$ curl -fsSL https://pulse.sh/install.sh | sh
$ pulse auth login --sso

// 2. Deploy autoscaling H100 inference worker
$ pulse deploy --gpu h100:2 --region iad \
    --image vllm/vllm-openai:latest \
    --model meta-llama/Llama-3.3-70B-Instruct

✓ Snapshot cached on NVMe (4.2s)
✓ Endpoint live: https://llama70b.pulse.run (78ms cold start)
BUILT-IN OBSERVABILITY

eBPF Kernel & CUDA Warp Profiler

Inspect GPU SM occupancy, KV-cache fragmentation, and tail P99 token latency without adding a single line of wrapper code.

● P99 TTFT: 29.4ms
● KV-Cache Utilization: 84.2%
INFRASTRUCTURE SAVINGS CALCULATOR

Cut Cloud Compute Bills by up to 64%.

Drag the slider to compare monthly H100 & Ephemeral vCPU spend against legacy hyperscalers.

PULSE.SH Bare-Metal$2,850 / mo
AWS EC2 P5 Instance$7,350 / mo
Google Cloud A3 VM$6,900 / mo
HOBBY & PROTOTYPE
$0 / mo

Includes $25 monthly serverless compute credit and 3 global regions.

DEDICATED RACK
Custom

Private InfiniBand spine, SOC2 Type IIBAA compliance, and 24/7 slack engineers.

ZERO-BOILERPLATE SDK

Spin Up 8x H100 Workers in 7 Lines of Code.

Native TypeScript, Python, Rust, and Go SDKs with automatic weight caching on local NVMe RAID arrays.

14.2 GB/s
Safetensors Cold Load
0 ms
Idle Standby Billing
// deploy-worker.ts
import { PulseCluster } from '@pulse-cloud/sdk';
const pool = await PulseCluster.spawn({
  gpu: 'H100-SXM5-80GB',
  regions: ['fra-01', 'iad-02', 'hnd-04'],
  autoscale: { min: 0, max: 64, wakeMs: 80 }
});
BARE-METAL POP TOPOLOGY

Global Low-Latency Datacenters

3.2 TBPS INFINIBAND SPINE • 100% HYDRO & NUCLEAR POWERED
FRA-01 // FRANKFURT● 0.4ms

512x H100 SXM5 • DE-CIX Peering

IAD-02 // ASHBURN● 0.7ms

1,024x H100 SXM5 • US-East Fiber

HND-04 // TOKYO● 0.6ms

384x H100 SXM5 • APAC Backbone

SIN-01 // SINGAPORE● 0.9ms

256x H100 SXM5 • Equinix SG3

LHR-03 // LONDON● 0.5ms

384x L40S + H100 • LINX Direct

SJC-02 // SILICON VALLEY● 0.4ms

768x H100 SXM5 • US-West Core

PAGE 04 // PER-SECOND COMPUTE BILLING (pricing.html)

Pay Only for Active Milliseconds. $0 Egress Fees.

ACCELERATOR SKUVRAMINTERCONNECTPULSE.SH / HRHYPERSCALER AVG
NVIDIA B200 Blackwell SXM192 GB HBM3e1.8 TB/s NVLink 5$3.40 / hr$8.90 / hr
NVIDIA H100 SXM580 GB HBM3900 GB/s NVLink 4$1.90 / hr$4.90 / hr
NVIDIA L40S Ada48 GB GDDR6PCIe Gen4 x16$0.88 / hr$2.25 / hr
AMD Instinct MI300X192 GB HBM3896 GB/s Infinity$2.15 / hr$5.40 / hr
ZERO-BOILERPLATE SDK

Spin Up 8x H100 Workers in 7 Lines of Code.

Native TypeScript, Python, Rust, and Go SDKs with automatic weight caching on local NVMe RAID arrays.

14.2 GB/s
Safetensors Cold Load
0 ms
Idle Standby Billing
// deploy-worker.ts
import { PulseCluster } from '@pulse-cloud/sdk';
const pool = await PulseCluster.spawn({
  gpu: 'H100-SXM5-80GB',
  regions: ['fra-01', 'iad-02', 'hnd-04'],
  autoscale: { min: 0, max: 64, wakeMs: 80 }
});
BARE-METAL POP TOPOLOGY

Global Low-Latency Datacenters

3.2 TBPS INFINIBAND SPINE • 100% HYDRO & NUCLEAR POWERED
FRA-01 // FRANKFURT● 0.4ms

512x H100 SXM5 • DE-CIX Peering

IAD-02 // ASHBURN● 0.7ms

1,024x H100 SXM5 • US-East Fiber

HND-04 // TOKYO● 0.6ms

384x H100 SXM5 • APAC Backbone

SIN-01 // SINGAPORE● 0.9ms

256x H100 SXM5 • Equinix SG3

LHR-03 // LONDON● 0.5ms

384x L40S + H100 • LINX Direct

SJC-02 // SILICON VALLEY● 0.4ms

768x H100 SXM5 • US-West Core

HARDWARE ISOLATION

Defense-in-Depth for Sovereign AI Workloads

ENCLAVE 01

NVIDIA Confidential Compute

Hardware-enforced VRAM memory encryption across PCIe and NVLink buses.

ENCLAVE 02

SOC2 Type II & HIPAA

Audited annually with automated BAA execution for clinical & financial teams.

ENCLAVE 03

Private WireGuard Mesh

Zero-config encrypted L3 peering between your AWS/GCP VPC and Pulse nodes.

ENCLAVE 04

Ephemeral Disk Scrub

Cryptographic key destruction within 50ms of container scale-down.

Infrastructure & Billing FAQ

How does Pulse.sh achieve 80ms GPU cold starts?

We snapshot initialized CUDA contexts and model weights directly to local PCIe Gen5 NVMe arrays and restore them via custom Firecracker DMA mappings.

Do you charge for ingress/egress bandwidth?

No. Unlike hyperscalers, all inbound and outbound data transfer across our 42 Anycast edge locations is 100% free and unmetered.

PAGE 05 // SOVEREIGN SECURITY & RESERVED CLUSTERS (security.html)

Reserve Dedicated H100/B200 Pods with Custom BAA & SLA

Speak directly with a Principal Infrastructure Architect to configure private WireGuard L3 peering, HIPAA/SOC2 Type II compliance packs, and committed volume discounts.

✓ SOC2 Type II & ISO 27001 Certified
✓ Hardware TPM 2.0 & Confidential VRAM Encryption
✓ 15-Minute Dedicated Slack Engineer Response
HARDWARE ISOLATION

Defense-in-Depth for Sovereign AI Workloads

ENCLAVE 01

NVIDIA Confidential Compute

Hardware-enforced VRAM memory encryption across PCIe and NVLink buses.

ENCLAVE 02

SOC2 Type II & HIPAA

Audited annually with automated BAA execution for clinical & financial teams.

ENCLAVE 03

Private WireGuard Mesh

Zero-config encrypted L3 peering between your AWS/GCP VPC and Pulse nodes.

ENCLAVE 04

Ephemeral Disk Scrub

Cryptographic key destruction within 50ms of container scale-down.

Infrastructure & Billing FAQ

How does Pulse.sh achieve 80ms GPU cold starts?

We snapshot initialized CUDA contexts and model weights directly to local PCIe Gen5 NVMe arrays and restore them via custom Firecracker DMA mappings.

Do you charge for ingress/egress bandwidth?

No. Unlike hyperscalers, all inbound and outbound data transfer across our 42 Anycast edge locations is 100% free and unmetered.

pulse-cloud://prod-eu-central-fra01
ACTIVE CLUSTERS
fra-h100-010.4ms
iad-us-east0.7ms
hnd-tokyo-040.6ms
sin-ap-south0.9ms
THROUGHPUT
142,890 req/s
LIVE eBPF TRACE STREAM 99.999% 200 OK
[14:22:01.004] POST /v1/embeddings (fra-01)200 OK • 1.8ms
[14:22:01.019] POST /v1/chat/completions (iad-02)200 OK • 6.4ms
[14:22:01.042] GRPC /kv.StateSync/Replicate (hnd-04)SYNC • 0.4ms
[14:22:01.088] POST /v1/rerank-batch (sin-01)200 OK • 2.1ms
[14:22:01.112] SNAPSHOT /vm/checkpoint-nvme (fra-01)DONE • 14.0ms
NODE HEALTH
VRAM Allocation68.4 / 80 GB
NVLink Bandwidth890 GB/s
Cold-Boot P9978 ms
$ pulse scale --min 0 --max 128
✓ Autoscale policy active (80ms wake)

Scale-to-Zero GPU Pools

Never pay for idle silicon. Memory snapshots hibernate to NVMe fabric and restore on the first incoming TCP packet.

eBPF Kernel Sensor: ENABLED
Zero-overhead HTTP/2 & gRPC tracing

Kernel-Level Observability

Inspect every syscall, SQL query, and tensor allocation in real time without adding SDK bloat to your containers.

Anycast BGP • 320 Tbps Edge Capacity
DDoS L7 Mitigation: ACTIVE

Global Anycast Mesh

Requests route automatically to the nearest healthy bare-metal pop across North America, Europe, and Asia-Pacific.