A3 family · Accelerator Optimizedx86_64 · Intel/AMD

a3-highgpu-4g

The a3-highgpu-4g machine type has 104 vCPUs and 936 GB of memory. Pricing for this instance starts at $44.24 per hour and $32298.85 monthly in the us-central1 region.

Updated June 18, 2026
vCPUs
104
Memory
936 GB
Network
Up to 200 Gbps
Storage
Local SSD

Save up to $17769/mo on every a3-highgpu-4g

Connect Google Cloud — we'll find the optimal pricing model for your workload.

All pricing options

us-central1 · USD · monthly = 730 hrs
Prices exclude local taxes
TermHourlyMonthlySavings
On-Demand
$44.2450$32298.85 Baseline
1-Year Committed Use Discount (CUD)
$30.6918$22405.04 -31%
3-Year Committed Use Discount (CUD)
$19.4322$14185.50 -56%
Preemptible / Spot
$19.9038$14529.81 -55%
You could save $17769.04/mo on this instance with the right pricing model.
Find my savings

Monthly cost estimator

Estimate your spend based on actual usage.
730 h
1h365h730h (24/7)
1
Estimated monthly
$32298.85
730 hrs × $44.2450/hr × 1

Compute

x86_64 · Intel/AMD
vCPUs i104
Memory i936 GB
Physical processorIntel Sapphire Rapids
Nested VirtualizationNot supported
Sole TenantNot supported
GPU4 GPU
Within A3 family
a3-highgpu-1g
26 vCPU
a3-highgpu-2g
52 vCPU
a3-highgpu-4g
104 vCPU
a3-highgpu-8g
208 vCPU

Lock in this rate across your fleet — typically save 30–40%

Connect your Google Cloud account to apply this pricing logic to every running instance automatically.

Networking

Max egress bandwidth200 Gbps
Tier 1 bandwidthN/A
Enhanced networking iYes (gvnic/virtio)
IPv6 supportYes (Dual-stack VPC)

Storage

Max persistent disks128
Max disk size512 TB
Local SSDsSupported (Scratch Disk)

Compare with another instance

vs.
Speca3-highgpu-4ga3-highgpu-1gΔ
vCPUs10426-75%
Memory936 GB234 GB-75%
Hourly Price$44.2450$11.0612-75%
Monthly Price$32298.85$8074.71-75%
Cost optimization

GCP recommender alerts can cut this bill ~30-40%.

Connect your GCP account. We'll identify every underutilized a3.* instance and show exactly where downsizing, cleaning up orphaned disks, or shifting to CUDs saves money — with zero code changes.