The language model
GPU memory decides which model fits, how much context it can hold, and how many people can use it at once.
AI servers built for the work they run.
NVIDIA and AMD in our lab · Every system bespoke
We design and build bespoke workstations and servers around the models and agentic workflows you will actually run, on NVIDIA or AMD, then install them in your office or data center. We are not a partner of any chip, hardware, or cloud vendor, so you stay in control of every layer.
We work for you, not a vendor.
Palm Hill is not a partner of any chip, hardware, or cloud vendor. That is about control: you decide what runs, where it runs, and when it changes, and every layer stays replaceable, from the chip to the cloud.
| Layer | How you stay in control |
|---|---|
| Chips | NVIDIA or AMD, chosen by running your workload on both in our lab. The next machine can use either. |
| Parts | Industry-standard components through standard supply channels, so no single supplier decides how the machine is repaired or replaced. |
| Serving software | Open-source vLLM, llama.cpp, and Ollama run on either maker’s GPUs and answer on OpenAI-compatible endpoints, so your applications don’t change when the hardware does. |
| Models | Open-weight models you hold yourself, so no provider can retire or change them under you, and commercial models where they test better. WolfPackAI puts both behind one gateway. |
| Where it runs | Your office, a colocation facility, Azure, AWS, Google Cloud, or a specialist GPU cloud, under your own accounts. Services ship as containers, so they can move. |
An agent is several models, not one.
An agentic workflow plans with a language model, looks things up, and checks its work before acting. Each piece puts different demands on the machine, so we start from the workflow and choose parts after.
Retrieval
Embedding models and a search index for private RAG. The index wants fast local storage and plenty of system memory.
Speech and vision
Whisper-class transcription, segmentation, and pose estimation share the GPU with the language model, or get a card of their own.
Checks before acting
Small System One models and classic classifiers answer in milliseconds, so they can screen a step before the agent takes it.
No catalog. Every machine is built for one workflow.
We start from the models, the people who will use them, and the room the machine will live in. The GPU maker and count, memory, storage, networking, power, and cooling all follow from that.
At a desk
A workstation for one person or a small team, in an ordinary office.
In a server room
Rack servers for a data center, server closet, or colocation facility, with the power and cooling to match.
At the edge
Small systems for a branch office, or a site that has to stay air-gapped.
Which maker fits depends on the models. NVIDIA has the broadest software support, and AMD now runs the major serving tools, but not every model performs the same on both. We measure your workload on each rather than assume.
Built to be fixed, by us or anyone else.
A private AI server should outlast the vendor that sold it. Ours are built from industry-standard components bought through standard supply channels, not a sealed appliance.
- Standard x86 motherboards, PCIe GPUs, power supplies, drives, and network cards
- Replacement parts available through the same channels any IT shop uses
- A written record of the parts, firmware settings, and software versions in each machine
- Serviceable by us, your IT team, or your MSP
Tested before you build
We run candidate models on our own NVIDIA and AMD machines with samples of your work, so the system is sized to measured speed and memory use, not a spec sheet.
Checked where it will live
A multi-GPU machine can need its own circuit, and server fans are loud. We check power, cooling, noise, and the network in the room before anything ships.
From workflow to working machine.
-
Test
Run candidate models against samples of your work on our NVIDIA and AMD machines, and measure speed and memory use.
-
Specify
Choose the GPUs, memory, storage, networking, power, and cooling for this workflow and the room it will run in.
-
Build and install
Build it, test it under load, and install it on-site, connected to your network or fully air-gapped.
-
Keep it running
Monitor GPU memory, temperature, and power, and handle maintenance and parts. Or hand it over to your IT team or MSP.
What we watch once it is live is on the approach page. What runs on it: WolfPackAI, NoPilot, private RAG, and the open-weight models we deploy.
Down to the circuit board.
Edgett Hilimire designs the systems and builds the machines himself. His background runs from software to the electronics underneath it.
- Custom circuit boards and embedded firmware for microcontrollers
- Sensor devices for equipment monitoring
- Component-level repair and modification of computer hardware
- Custom server and workstation builds, including power and cooling
- An internal GPU cluster for language model inference and training at a healthcare technology company
- Data centers, networks, and phone systems for a claims processing operation
Why it matters
The person choosing the model is the person specifying the machine. When a card, a fan, or a power supply misbehaves, the same person can trace it down to the component.
Questions about AI hardware
Do we need new hardware to run private AI?
Not always. Smaller open-weight models and embedding models run on CPUs and existing workstation GPUs. We test on samples of your work first and tell you if what you already have is enough.
Do you partner with NVIDIA or AMD?
No. Palm Hill is not a partner of any chip, hardware, or cloud vendor. We recommend whatever tests best for your workflow, and you keep control: you can change the hardware, the model, or where it runs without asking any vendor.
How do you decide what to build?
Every system is built for one workflow. We run candidate models on our own NVIDIA and AMD machines with samples of your work, measure speed and memory use, and specify the machine from what we measured.
Where can it run?
In your office, a server closet, your data center, or a colocation facility, and we check power, cooling, noise, and the network before a machine arrives. The same software can run in Azure, AWS, Google Cloud, or a specialist GPU cloud, so moving later doesn't mean starting over.
Can it run without an internet connection?
Yes. We build air-gapped systems where models, software, and updates are carried in and nothing calls out. Chat411’s local path uses a local model, media can be encrypted, and neither PSTN nor a cloud model provider is required. Its cloud realtime path requires an outbound connection and streams caller audio to a provider, xAI by default.
Who maintains it after installation?
We can monitor GPU memory, temperature, and power and handle maintenance and part replacement. Or we hand it over to your IT team or MSP. Because it is built from standard components, they can service it too.
Tell us the workflow. We will build the machine.
Bring the models you want to run, how many people will use them, and where the machine will live. We will tell you what we would test on each maker’s hardware.
NVIDIA is a trademark of NVIDIA Corporation. AMD is a trademark of Advanced Micro Devices, Inc. Palm Hill AI is not affiliated with NVIDIA or AMD.