armada.plavno.io · inference infrastructure

Production AI inference on GPUs you can actually afford.

Armada runs your models on a focused lineup of RTX 6000, RTX 5090, RTX 4090 and H20 GPUs — with fast cold starts, autoscaling, and per-minute billing. The cost-performance of consumer and inference-optimized silicon, packaged for production.

Start deploying Talk to an engineer
142ms median TTFT
2,140tok/s throughput on RTX 5090
1.4s typical cold start
Trusted by teams shipping AI in production
NORTHWIND HELIX LABS CORTEX/AI OBELISK PLAVNO STRATUM
01 Why Armada

Built for inference, priced for the real world.

Most platforms still charge datacenter-GPU rates for workloads that run perfectly well on consumer and inference-optimized silicon. Armada doesn’t.

Fast model runtimes

Optimized server runtimes deliver low time-to-first-token and high sustained throughput across LLMs, diffusion and audio models.

  • 142 ms median TTFT
  • 2× throughput vs naive serving

Effortless autoscaling

Scale from zero to thousands of replicas. Cold starts measured in seconds, warm starts in milliseconds.

  • 1.4 s typical cold start
  • 0→N autoscaling, scale-to-zero

Pay for what you use

Per-minute billing — not per-hour, not per-day. Scale-to-zero means you pay nothing for idle endpoints.

  • per-minute granularity
  • $0 idle cost at zero replicas
our differentiator

Affordable GPU lineup

RTX 6000, RTX 5090, RTX 4090, and H20 — selected because they deliver the best cost-per-token for real inference workloads. Skip the H100 tax.

  • up to 72% cheaper than H100/B200 instances
  • 4 GPUs, picked for inference economics
02 GPU lineup

Four GPUs. Picked for cost-per-token.

Datacenter-class cards are built for training. We selected a lineup tuned for the economics of serving: high throughput per dollar, the right amount of VRAM for the models you actually deploy, and per-minute billing across the board.

RTX 4090 $0.60/hr*

The workhorse

Best for 7B–13B LLMs, ComfyUI / SDXL, embeddings, Whisper.

VRAM
24 GB GDDR6X
Memory BW
1.0 TB/s
Tensor FP8
660 TFLOPS
Inference perf
1.0× baseline
most popular
RTX 5090 $0.90/hr*

The sweet spot

Best for 30B–70B LLMs, Flux / diffusion at scale, real-time audio.

VRAM
32 GB GDDR7
Memory BW
1.8 TB/s
Tensor FP4
1.8 PFLOPS
Inference perf
1.9× vs 4090
RTX 6000 $1.20/hr*

The big-VRAM card

Best for 70B+ LLMs, long context, multi-LoRA serving, video gen.

VRAM
48 GB GDDR6 ECC
Memory BW
960 GB/s
Tensor FP8
728 TFLOPS
Inference perf
1.4× vs 4090
H20 $1.50/hr*

The inference rig

Best for 70B+ inference at scale, batched serving, KV-cache heavy workloads.

VRAM
96 GB HBM3
Memory BW
4.0 TB/s
Tensor FP8
296 TFLOPS
Inference perf
2.4× vs 4090

*Illustrative rates. Final pricing depends on region and commitment — see full pricing.

03 Capabilities

Every modality, one platform.

LLMs, diffusion, audio, voice agents, embeddings, custom checkpoints — deployed the same way and served through the same OpenAI-compatible interface where possible.

LLM inference

Qwen, DeepSeek, Llama, gpt-oss, Mistral, Gemma. vLLM and TensorRT-LLM runtimes, OpenAI-compatible endpoints, structured outputs and tool use.

vLLMOpenAI APItool use

Image & video

SDXL, Flux, Stable Video Diffusion, ComfyUI workflows. Bring your graph; deploy it as an autoscaling endpoint with cached weights.

FluxSDXLComfyUI

Transcription & diarization

Whisper-large-v3, NeMo and Parakeet runtimes with speaker diarization, word-level timestamps and streaming.

Whisperdiarizationstreaming

Real-time voice

Low-TTFB pipelines for voice agents: STT → LLM → TTS with WebSocket streaming. Sub-300 ms turn-taking at production scale.

WebSocket<300 msbarge-in

Embeddings

BGE, E5, Nomic, OpenAI-compatible embedding endpoints. Batched inference, optional reranking, p99 under 50 ms.

BGErerankersp99 <50ms

Custom & fine-tuned

Bring a checkpoint, a Docker image, or a fine-tune. We’ll build it, cache the weights, and give you an autoscaling endpoint.

any ckptDockerLoRA
04 How it works

From model to endpoint, in three steps.

Pick a model. Pick a GPU. Get an autoscaling API endpoint — with logs, metrics and observability wired in from day zero.

  1. 01

    Pick a model

    Start from a tuned preset (Qwen, DeepSeek, Llama, Flux, Whisper) or bring your own checkpoint, Docker image, or ComfyUI graph.

  2. 02

    Choose a GPU

    RTX 4090, 5090, RTX 6000 or H20 — we’ll surface the right one for your model size and target latency, with the hourly cost up-front.

  3. 03

    Get an endpoint

    Autoscaling, scale-to-zero, OpenAI-compatible. Logs, metrics, traces and per-request timings in the dashboard.

terminal 
# Install
$ pip install armada

# Authenticate
$ armada login

# Deploy from a HF repo
$ armada deploy deepseek-ai/DeepSeek-V3 \
    --gpu h20 \
    --replicas 2 \
    --scale-to-zero

✓ Endpoint live at https://api.armada.plavno.io/v1/deepseek-v3
✓ Streaming logs to ~/armada/deepseek-v3.log
✓ Metrics: dashboard.armada.plavno.io/deepseek-v3
Structured logsper-request, JSON, streamed
Per-token metricsTTFT, TPOT, throughput, queue
OpenTelemetrytraces exported to your stack
Cost telemetry$/1k tokens, per endpoint, live
05 Pricing

Only pay for the compute you use, down to the minute.

No commitments to start. Scale-to-zero means idle endpoints cost nothing. When you’re ready for committed-use discounts, we have those too.

GPU VRAM Best for Hourly* Per-minute
RTX 409024 GB7B–13B LLMs, SDXL, embeddings$0.60$0.0100
RTX 509032 GB30B–70B LLMs, Flux, voice$0.90$0.0150
RTX 600048 GB70B+ LLMs, long context$1.20$0.0200
H2096 GB70B+ at scale, batched serving$1.50$0.0250

*Illustrative rates. Includes per-minute billing, autoscaling, logs and metrics. Egress is free up to 1 TB / month.

06 By the numbers

Teams ship faster on Armada.

2× throughput vs naive serving
−50% time-to-first-token
$0 idle cost · scale-to-zero
14k+ models deployed in production
“We replaced our H100 fleet with Armada’s 5090s and cut inference cost by 68%, with better TTFT on our 70B model. The migration took an afternoon.”
Maya Okafor Head of ML Platform · Northwind
“Per-minute billing is the headline, but the real win is the runtime — sub-200 ms TTFT on a 32B model with sane defaults. We stopped writing serving code.”
Jens Holm Founding engineer · Helix Labs
“Cold starts under two seconds with scale-to-zero. Our voice agent backend used to cost $14k/month; on Armada it’s under $4k and faster.”
Priya Raman Staff engineer · Cortex/AI
Start today

Deploy your first model in under five minutes.

Free trial credits. No credit card. Per-minute billing when you scale.

Start deploying Talk to sales
$ pip install armada && armada deploy your-model --gpu rtx-5090