Internal one-stop shop — SKUs, specs, pricing, networking, regions, data centers, inference products, and roadmap.
| GPU Model | Config | VRAM | vCPU | RAM | Boot Disk | Scratch Disk | Size Slug | Virtualization | Mode | Status |
|---|---|---|---|---|---|---|---|---|---|---|
| RTX 4000 Ada | 1× | 20 GB | 8 | 32 GB | 500 GiB NVMe | — | gpu-rtx4000ada-20gb | Passthrough | On-Demand | GA |
| L40S | 1× | 48 GB | — | — | 500 GiB NVMe | — | gpu-l40s-48gb | Passthrough | On-Demand | GA |
| RTX 6000 Ada | 1× | 48 GB | — | — | 500 GiB NVMe | — | — | Passthrough | On-Demand | GA |
| H100 SXM5 | 1× | 80 GB | — | — | 720 GiB NVMe | 5 TiB NVMe | gpu-h100x1-160gb | Passthrough + FabricMgr svc VM | On-Demand | GA |
| H100 SXM5 | 8× | 640 GB | 160 | 1,920 GB | 2 TiB NVMe | 40 TiB NVMe | gpu-h100x8-640gb gpu-h100x8-640gb-contracted | Passthrough + 4× NVSwitch + FabricMgr | On-Demand Contract | GA |
| H200 SXM5 | 1× | 141 GB | 24 | 240 GB | 720 GiB NVMe | 5 TiB NVMe | gpu-h200x1-141gb | Passthrough + FabricMgr svc VM | On-Demand | GA |
| H200 SXM5 | 8× | 1,128 GB (1.1 TB) | 192 | 1,920 GB | 2 TiB NVMe | 40 TiB NVMe | gpu-h200x8-1128gb gpu-h200x8-1128gb-contracted | Passthrough + 4× NVSwitch + FabricMgr | Sold Out Contract | GA |
| B300 (Blackwell) | 1× | 288 GB HBM3e | TBD | TBD | — | — | gpu-b300x1-… | Passthrough + DOCA + CX-8 NICs | Contract | GA (single ~3/31/26) |
| B300 (Blackwell) | 8× | 2,304 GB (2.25 TB) | TBD | TBD | — | — | gpu-b300x8-2304gb-contracted | Passthrough + 2× NVSwitch + DOCA + CX-8 | Contract | Multi-node ~mid-May 2026 |
| GPU Model | Config | VRAM | vCPU | RAM | Boot Disk | Scratch Disk | Size Slug | Virtualization | Mode | Status |
|---|---|---|---|---|---|---|---|---|---|---|
| MI300X | 1× | 192 GB HBM3 | 20 | 240 GB | — | — | — | SR-IOV (VF) — intentional | On-Demand | GA |
| MI300X | 8× | 1,536 GB | 160 | 1,280 GB | 2 TiB NVMe | 40 TiB NVMe | — | SR-IOV (VF) + Infinity Fabric | On-Demand Contract | GA |
| MI325X | 8× | 2,048 GB (2 TB) | 160 | 1,280 GB | 2 TiB NVMe | 40 TiB NVMe | — | SR-IOV (VF) + Infinity Fabric | Contract Only | GA |
| MI350X | 8× | 2,304 GB (2.25 TB) | TBD | TBD | 2 TiB NVMe | 40 TiB NVMe | gpu-mi350x8-2304gb | SR-IOV (VF) + Infinity Fabric (CDNA 4) | Contract Only | GA (RIC1+ATL1, Feb/Mar 2026) |
| MI355X ⚡ | 8× | ~2,304 GB (est.) | — | — | — | — | gpu-mi355x8-2304gb (est.) | SR-IOV (VF) — expected same as MI350X; liquid-cooled | Coming Soon | Q3 2026 est. |
| GPU | Config | Per GPU/hr (12-mo contract) | Node/hr (12-mo) | Monthly/node (est.) | Savings vs On-Demand | Notes |
|---|---|---|---|---|---|---|
| H100 SXM5 | 8× | $1.99/GPU/hr | $15.92/hr | ~$11,462/mo | ~33% off | vs $23.92/hr on-demand |
| MI300X | 8× | $1.49/GPU/hr | $11.92/hr | ~$8,582/mo | ~25% off | vs $15.92/hr on-demand |
| MI325X | 8× | $1.69/GPU/hr | $13.52/hr | ~$9,734/mo | — | Contract-only SKU; no OD baseline |
| H200 SXM5 | 8× | Contact Sales | Contact Sales | — | — | Contract required; confirm with sales |
| B300 | 8× | Contact Sales | Contact Sales | — | — | Market range from $5.65/GPU/hr (12-mo); DO pricing TBD |
| MI350X | 8× | Contact Sales | Contact Sales | — | — | Contract-only; confirm with sales |
| Product | Pricing Model | Rate | Notes |
|---|---|---|---|
| Serverless Inference | Per-token (per model) | See Model Catalog — varies by model | No GPU provisioning cost; billed per token only when running inference |
| Dedicated Inference | Per GPU-hour | Same rates as GPU Droplet hourly for selected GPU | B300, H100, H200, MI300X, MI325X, etc. — mirrors GPU Droplet pricing |
| Inference Hub | No extra charge | $0.00 | Platform access is free; pay only for Serverless or Dedicated usage |
| BYOM Model Storage | Flat monthly | $5.00/mo | Model weights stored in service-managed Spaces location |
| GPU SKU | Intra-node Interconnect | Intra-node BW | Switch/Fabric | Inter-node | GPU Fabric BW | Host Ethernet | Public | Private | NIC Model | NIC–GPU Pairing |
|---|---|---|---|---|---|---|---|---|---|---|
| H100 SXM5 (8×) | NVLink | 900 GBps | 4× NVSwitches | RoCEv2 | 3.2 Tbps (8×400G) | 4×100 Gbps | 10 Gbps | 25 Gbps | CX-7 (1/GPU) | 1:1 GPU↔NIC rail |
| H200 SXM5 (8×) | NVLink | 900 GBps | 4× NVSwitches | RoCEv2 | 3.2 Tbps (8×400G) | 4×100 Gbps | 10 Gbps | 25 Gbps | CX-7 (1/GPU) | 1:1 GPU↔NIC rail |
| B300 (8×) | NVLink (NVSwitch) | TBD | 2× NVSwitches | RoCEv2 | est. 3.2 Tbps | TBD | 10 Gbps | 25 Gbps | CX-8 onboard (2 VF NICs/GPU) | 1:1 — DOCA drivers required |
| MI300X / MI325X (8×) | Infinity Fabric (on-die) | 896 GB/s (bidirectional) | N/A — on-die | RoCEv2 | 3.2 Tbps (8×400G) | — | 10 Gbps | 25 Gbps | SR-IOV VF (1/GPU) | 1:1 GPU↔NIC rail (VF mode) |
| MI350X (8×) | Infinity Fabric (CDNA 4) | TBD | N/A — on-die | RoCEv2 | est. 3.2 Tbps | — | 10 Gbps | 25 Gbps | SR-IOV VF (1/GPU) | 1:1 GPU↔NIC rail (VF mode) |
| MI355X (8×) | Infinity Fabric | TBD | N/A — on-die | RoCEv2 | TBD | — | — | — | Same as MI350X (expected) | — |
| RTX 4000 / L40S / RTX 6000 (1×) | N/A (single GPU) | N/A | None | N/A | N/A | — | 10 Gbps | 25 Gbps | — | N/A |
hipIpcOpenMemHandle (cross-process GPU memory on same node) may fail in VF mode. AMD ticket open. No customer-facing workaround confirmed yet.gpu-h100x8-base image (poorly named but supports all hardware including B300). The gpu-h100x1-base image lacks DOCA drivers and will not work with B300.| GPU SKU | Config | Available Regions | Mode | Max Cluster Size | SLA | Spin-up | Notes |
|---|---|---|---|---|---|---|---|
| RTX 4000 Ada | 1× | TOR1 | On-Demand | 1 | 99.5%/mo | <1 min | TOR1 only per official docs |
| L40S / RTX 6000 Ada | 1× | TOR1 + others TBD | On-Demand | 1 | 99.5%/mo | <1 min | Confirm full region list |
| H100 SXM5 | 1× / 8× | NYC2NYC3TOR1ATL1 | Sold Out / Contract | 512 GPUs / 64 nodes | 99.5%/mo | <1 min | On-demand capacity fluctuates; spot via grafana capacity monitor |
| H200 SXM5 | 1× / 8× | NYC2ATL1 | Sold Out / Contract | 512 GPUs / 64 nodes | 99.5%/mo | <1 min | Very limited on-demand; mostly contracted |
| B300 | 1× / 8× | RIC1 | Contract Sold Out | TBD | 99.5%/mo | TBD | Single-node GA ~3/31/26; multi-node ~mid-May 2026; RIC1 only |
| MI300X | 1× / 8× | NYC1TOR1ATL1 | On-Demand / Contract | 512 GPUs / 64 nodes | 99.5%/mo | <1 min | Dedicated Inference regions: NYC1, TOR1, ATL |
| MI325X | 8× | ATL1SFO2SFO3NYC1 | Contract Only | 512 GPUs / 64 nodes | 99.5%/mo | — | Active firmware upgrade program underway |
| MI350X | 8× | RIC1ATL1 | Contract Only | TBD | 99.5%/mo | — | ATL1 (Feb 2026) + RIC1 (Mar 12, 2026). On-demand option TBD. |
| MI355X ⚡ | 8× | MEM1 | Coming Soon | TBD | — | — | MEM1 cluster referenced Apr 2026. Liquid-cooled racks. Q3 2026 est. |
| Region | Location | Launched | GPU SKUs | Purpose / Design | Key Infrastructure | Notes |
|---|---|---|---|---|---|---|
| ATL1 | Atlanta-Douglasville, GA | June 2025 | H200 MI325X, MI300X, MI350X | Largest DO DC at launch; AI/ML optimized; AMD Developer Cloud partnership | VAST storage; multi-room colo; full DO stack (DOKS, DBs, App Platform, LBaaS) | 80% of CPTO teams involved in buildout. Inference available (NYC1, TOR1, ATL). |
| RIC1 | Richmond, VA | Mar 12, 2026 | B300 MI350X | Purpose-built next-gen GPU; high-density GPU pod design | High-density B300 + MI350X pods; VAST storage; full network stack; DOKS-ready; Jammy kernel for B300 | B300 private preview completed in 6 weeks (vs 8-week "speed of light" target). Single-node GA ~3/31/26. |
| NYC2 | New York, NY | Legacy | H100, H200 | Legacy DC; core DO stack | H100/H200 nodes (largely sold out) | Sparse GPU capacity; H200 test nodes used by virt team |
| TOR1 | Toronto, Canada | Legacy | H100, RTX 4000 Ada MI300X | Multi-GPU region; RTX 4000 Ada exclusive to TOR1 | Standard DO stack; Canadian data residency | RTX 4000 Ada only available in TOR1 as of Mar 2026 |
| SFO2/3 | San Francisco, CA | Legacy | MI325X | AMD GPU region; active firmware upgrade efforts | MI325X fleet nodes | MI325X firmware upgrades underway; SFO3 validated Jan 2026 |
| NYC1 | New York, NY | Legacy | MI300X | Legacy DC; MI300X + Dedicated Inference launch region | Standard DO stack | One of 3 Dedicated Inference private preview regions (NYC1, TOR1, ATL) |
| MEM1 ⚡ | Memphis, TN (est.) | Q3 2026 est. | MI355X | Dedicated MI355X cluster; liquid-cooled racks | TBD | ⚡ Referenced in Apr 2026 Moonshot AI POC thread. Confirm DC name/location with engineering. |
| Product | Status | GPU Infra | Pricing | API / Access | Supported GPUs | Key Features | Docs |
|---|---|---|---|---|---|---|---|
| Serverless Inference | GA | Latest NVIDIA GPUs (DO-managed; runs on DI platform) | Per-token per model | inference.do-ai.run /v1/chat/completions | NVIDIA (Blackwell + others, managed by DO) | No GPU provisioning; OpenAI-compatible API; Day-0 model launches; unified billing; prompt caching; streaming; security-hardened defaults | Docs |
| Dedicated Inference | GA (Apr 30, 2026) | Private, reserved GPU instances | Per GPU-hour (same as GPU Droplet rates) | Control Panel + API | B300, H100, H200, MI300X, MI325X, etc. | No noisy-neighbor; auto-scaling; custom concurrency & sequence lengths; scale-to-zero; BYOM; LoRA adapters; speculative decoding; private & isolated; 1-click from Serverless/Playground | Docs |
| Inference Hub | Public Preview (Mar 2026) | Same as above | $0 for Hub; pay per usage | DO Control Panel UI | Serverless + Dedicated | Model Catalog (search/filter); Playground; built-in code snippets; single pane of glass for both modes; model evaluation; intelligent routing (roadmap) | Docs |
| GPU Droplets | GA | Raw VM with dedicated GPU(s) | Per GPU-hour | DO API, CLI, Control Panel, Terraform | All SKUs (H100, H200, B300, MI300X, MI325X, MI350X, RTX series, L40S) | Full root access; DOKS integration; VAST/NFS shared storage; multi-node up to 512 GPUs; 99.5% SLA; pre-installed CUDA/ROCm/PyTorch/TensorFlow | Docs |
| Model | Provider | API Model ID | Launch Date | Day-0? | Use Case |
|---|---|---|---|---|---|
| GLM-5 | Zhipu AI | glm-5 | Mar 19, 2026 | Day-0 (GTC) | Deep reasoning, long-context, agentic |
| Kimi-K2.5 | Moonshot AI | kimi-k2.5 | Mar 19, 2026 | Day-0 (GTC) | Multi-step reasoning, multimodal |
| MiniMax-M2.5 | MiniMax | minimax-m2.5 | Mar 19, 2026 | Day-0 (GTC) | High-volume production, coding, agents |
| Nemotron 3 Super | NVIDIA | nemotron-3-super | Mar 19, 2026 | Day-0 (GTC) | Fast multilingual reasoning, agentic (120B hybrid) |
| Arcee Trinity Large Thinking | Arcee AI | trinity-large-thinking | Apr 2, 2026 | No | Agentic, long-horizon, multi-turn tool calls |
| Opus 4.7 | Anthropic | opus-4.7 | Apr 17, 2026 | Day-0 | Advanced reasoning, agentic, coding |
| GPT Image 2.0 | OpenAI | openai-gpt-image-2 | Apr 23, 2026 | Day-0 | Image generation — /v1/images/generations |
| GPT-5.5 | OpenAI | openai-gpt-5.5 | Apr 28, 2026 | Day-0 | Autonomous multi-step agent tasks, coding |
| GPU / Feature | Vendor | Config | VRAM | Region | Target Date | Status | Notes / Source |
|---|---|---|---|---|---|---|---|
| B300 — Multi-node GA | NVIDIA | 8× multi-node | 2,304 GB HBM3e | RIC1 | ~Mid-May 2026 | In Progress | Single-node GA ~3/31. Multi-node ~mid-May. Contract only + sold out. — #gpu-program Jan 2026 |
| MI355X — GA | AMD | 8× | ~2,304 GB (est.) | MEM1 | Q3 2026 (est.) | Coming Soon | Liquid-cooled racks. "Next quarter, DO will deploy MI355X GPUs." — Feb 2026 press release |
| B300 — On-Demand | NVIDIA | 1× / 8× | 288 GB / 2,304 GB | RIC1 | TBD | TBD | Currently contract-only + sold out. On-demand when capacity expands. |
| MI350X — On-Demand | AMD | 8× | 2,304 GB | RIC1 ATL1 | TBD | TBD | Currently contract-only. On-demand under discussion. — #gpu-program Jan 2026 |
| GPU Contract Customer Experience | — | — | — | All | H1 2026 | In Progress | 2026 priority: improve contract GPU customer experience as GPU footprint expands rapidly. Led by Jenni Griesmann. — #gpu-program Jan 2026 |
| MXFP4/NVFP4 Precision Support | NVIDIA / AMD | — | — | — | TBD | TBD | Requires B300 or MI350X. Targets FP8 quality gap for memory-intensive inference. — DO Blog Apr 2026 |
| AMS3 (Amsterdam) GPU | TBD | — | — | AMS3 | 2026+ | TBD | DO product page mentions AMS3 as a current/upcoming GPU region. First Europe GPU location. GPU type TBD. |