Workloads

Built for the way modern AI actually runs

Match each workload to the right GPU tier — from budget experimentation to enterprise-scale training — on infrastructure engineered to put its own heat to work.

Abstract crystalline data structure representing large language model inference.

LLM Inference

Run text-generation, chatbot, RAG, and API inference workloads with GPU options matched to model size, latency, and budget.

Recommended GPUs

  • Intel Arc Pro B70 — budget / high-VRAM value for quantized or experimental workloads (slower alternative).
  • RTX 5090 — fast consumer-grade inference, limited / secondary option.
  • RTX PRO 6000 Blackwell — professional inference and larger models.
  • H200 / B200-class — enterprise-scale large-model inference.
Best for
  • Chatbots & APIs
  • RAG pipelines
  • Low-latency serving
Join the waitlist
Source code on screen representing model fine-tuning and training.

Fine-Tuning & LoRA Training

Fine-tune open-source models, LoRAs, embedding models, and domain-specific AI systems.

Recommended GPUs

  • RTX PRO 6000 Blackwell — serious fine-tuning and professional CUDA workflows.
  • H200 / B200-class — large training jobs.
  • Intel Arc Pro B70 — budget experimentation where software compatibility is acceptable.
Best for
  • LoRA adapters
  • Domain tuning
  • Embeddings
Join the waitlist
Glowing circuit board representing scalable AI agent backends.

AI Agents & Automation

Host persistent AI agents, coding agents, research agents, and tool-using workflows with scalable GPU backends.

Recommended GPUs

  • RTX PRO 6000 Blackwell — reliable production agent workloads.
  • RTX 5090 — fast but limited secondary capacity.
  • H200 / B200-class — high-throughput enterprise agents.
Best for
  • Coding agents
  • Tool use
  • Long-running tasks
Join the waitlist
Vibrant abstract gradient representing generative image and video models.

Image, Video & Generative Media

Run diffusion models, video generation, rendering-adjacent AI, upscaling, and creative AI pipelines.

Recommended GPUs

  • RTX 5090 — fast creative workloads when available.
  • RTX PRO 6000 Blackwell — professional generation, stability, and large VRAM.
  • H200 / B200-class — large-scale batch generation.
Best for
  • Diffusion
  • Video gen
  • Upscaling
Join the waitlist
Laboratory pipette and sample tubes representing scientific research compute.

Scientific & Research Compute

Support biology, chemistry, climate, simulation, data analysis, and university research workloads that need high-memory compute.

Recommended GPUs

  • RTX PRO 6000 Blackwell — data science, HPC prototyping, genomics, and simulation.
  • H200 — high-memory HPC and generative AI research.
  • B200 / larger Blackwell-class — future large-scale research clusters.
Best for
  • Simulation
  • Genomics
  • Climate models
Join the waitlist
Data-center server racks with cabling representing batch compute jobs.

Batch Jobs & Compute-Heavy Tasks

Run scheduled jobs, rendering-style tasks, data processing, experiments, and queued compute workloads.

Recommended GPUs

  • Intel Arc Pro B70 — low-cost queued jobs where speed is less critical.
  • RTX PRO 6000 Blackwell — high-performance professional workloads.
  • H200 / B200-class — enterprise batch and cluster jobs.
Best for
  • Scheduled runs
  • Data processing
  • Experiment queues
Join the waitlist
The difference

Compute Power With a Second Purpose

Traditional GPU servers turn electricity into heat, then spend more energy removing that heat. AM ThermaLink is designed to capture that thermal output and reuse it for useful heating applications such as pools, domestic hot water, and buildings.

A swimming pool warmed by recovered server heat.

Pool Heating

  • AI workload runs on the GPU
  • GPU produces heat as a byproduct
  • Heat is captured through liquid cooling
  • Transferred to pool water via a sealed exchanger
Modern buildings representing space heating from recovered server heat.

Building Heat

  • AI workload runs on the GPU
  • GPU produces heat as a byproduct
  • Heat is captured through liquid cooling
  • Delivered to space-heating loops on-site
A clean bathroom representing domestic hot-water systems fed by recovered heat.

Hot Water Systems

  • AI workload runs on the GPU
  • GPU produces heat as a byproduct
  • Heat is captured through liquid cooling
  • Preheats domestic hot water where a fit exists

A lower-waste, heat-reuse approach: reduced wasted thermal energy and a design intended to reduce the environmental footprint of compute. Actual heat recovery depends on site conditions, workload, and demand.

How it works

From Workload to Useful Heat

A straightforward path from choosing a job to putting recovered heat to work.

1

Choose a workload

Inference, fine-tuning, agents, generative media, research, or batch jobs.

2

Select the right GPU tier

Match model size, latency, and budget to the recommended GPU.

3

Deploy or join the waitlist

Launch available capacity now, or reserve future high-memory tiers.

4

Compute runs on AM ThermaLink

Your job executes on liquid-cooled, heat-aware infrastructure.

5

Captured heat is reused

Recovered server heat is delivered to useful systems where possible.

6

Two useful outputs

You get compute; the building receives useful thermal energy.

Need GPU compute that does more than sit in a data center?

Join Compute Waitlist Pre-order GPU Capacity Ask About Hosting a System

Early capacity will prioritize partners with clear workloads, consistent usage, and interest in sustainable compute infrastructure.

FAQ

Common questions

Is Intel Arc Pro B70 good for large AI models?
It can be a strong budget option because it offers 32 GB of VRAM at a lower cost, especially for quantized models, testing, and memory-heavy experiments. However, it is slower than higher-end NVIDIA cards and may require more compatibility work because many AI workflows are optimized for CUDA.
Which GPU should I choose for serious AI workloads?
RTX PRO 6000 Blackwell is the preferred near-term professional option because of its 96 GB VRAM, NVIDIA software ecosystem, and suitability for inference, fine-tuning, data science, and compute-heavy work.
Why is RTX 5090 listed as secondary?
RTX 5090 is very fast, but it is a consumer GPU with 32 GB VRAM and should not be positioned as the main enterprise compute product. It is useful for smaller fast jobs, creative AI, and limited capacity.
When will H200 or B200 capacity be available?
H200 and B200-class capacity is shown as future / waitlist capacity. Pricing and access will be set with early partners.
How is AM ThermaLink different from a normal GPU cloud?
AM ThermaLink is designed around heat reuse. Instead of treating GPU heat only as waste, the system is built to capture thermal energy and reuse it for pools, buildings, and hot-water systems where possible.