.png)
.png)
 (Custom).png)
.png)
.png)
Synapse Nexus AI Workstation
Synapse Nexus AI Workstation
Built for one thing.
Most local AI infrastructure is about escaping the cloud. Nexus is bigger than that. Running the world's most capable open-source models at the scale serious teams demand.
At its heart is the RTX Pro 5000 Blackwell, with enough memory to run 70B models at production precision. Drives up to 16 concurrent users on 70B-class models - each getting production-quality responses, none of them sharing a queue with the cloud.
This is what it looks like when local infrastructure stops being a compromise.
70B parameters at FP8
The full model, at the quality the original weights were trained at, running on hardware you own. 30 tok/s single user, 107 tok/s across 4 concurrent on Qwen3-32B.
Scale for 16+ concurrent users
When the second person needs access, and the eighth - you don't buy another subscription. Shared infrastructure that scales with how your team works, without a cloud dependency in sight.
Full fine-tune on 30B
Your data never leaves the building. The model you need, trained on your domain, with the headroom to iterate as fast as the team can frame the experiments.
Our support service helps you get up and running, from hardware to your first models.
f6b07ac8-17c6-4bda-a09c-6091f1d3af38.png)
Built for your workload
The RTX Pro 5000 Blackwell is 48GB of GDDR7 ECC memory, 1.34 TB/s bandwidth, fifth-generation Tensor Cores. It was specified to run the models serious teams actually use, at the concurrency serious teams actually need. That decision shapes what Nexus is.
Specialist GPU built for teams, not just benchmarks.
The RTX Pro 5000 Blackwell was chosen to run the models serious teams actually need, at the scale serious teams actually demand.
48GB of GDDR7 ECC memory. 1.34 TB/s bandwidth. Fifth-generation Tensor Cores. Not a gaming card with a Pro badge. A workstation GPU specified from the ground up for sustained inference, fine-tuning, and the kind of concurrency that breaks lesser hardware.
70B models fit natively.
70B FP8 needs ~70GB - tight for a single card at full precision. At Q8 it needs ~35GB and fits with headroom. At Q4 it needs ~40GB and runs comfortably. For most production workloads, 48GB is the right card. The only thing it won't do is 70B FP16, and that needs 140GB, which means two cards regardless of what you buy.
Concurrency that holds under load
At 4 concurrent users, aggregate throughput hits 107 tok/s with per-request latency staying under 1.8 seconds. At 16 concurrent, 248 tok/s aggregate. The card doesn't fall over at team scale - the numbers were measured, not estimated.
ECC memory, not gaming VRAM.
The RTX Pro 5000 Blackwell is GDDR7 ECC — error-correcting memory built for sustained inference workloads, not burst gaming loads. 1.34 TB/s bandwidth. The kind of spec that decides whether a 70B model runs fast or runs at all.
Office thermals. Sustained load.
The RTX Pro 5000 is a workstation card, not a datacentre card. It runs at office-noise levels under sustained inference without liquid cooling or rack infrastructure. It sits where your team sits.
Scale with speed
Your whole team, running simultaneously, getting instant responses.
On an 8 billion parameter model, Chillblast Synapse Node delivers 1,039 tokens per second across 32 concurrent requests. Aggregate throughput scales with every user you add.
What it runs
From fast 8B agents to full 70B inference. What you can actually run on a Synapse Nexus, at what precision, at what speed. Single-user sustained throughput; concurrent figures use vLLM-style batching.
Inference
| Model | Precision | VRAM | Tokens / sec | Concurrent users |
|---|---|---|---|---|
| Llama 3 / Qwen 8B Drafting, agents | FP16 | 16 GB | 220 | 64 |
| Llama 3 32B Production reasoning | Q8 | 34 GB | 80 | 12 |
| Qwen3-32B Benchmarked · coding & reasoning | FP8 | 35 GB | 107 | 16 |
| Llama 3.3 70B | Q4 | 40 GB | 50–70 | 4 |
| Llama 3.3 70B | FP8 | 70 GB | 35 | 2 |
Fine-tuning
| Task | Method | Time |
|---|---|---|
| 8B LoRA fine-tune | Full | 1.2 hours |
| 32B LoRA fine-tune | Full | 6 hours |
| 70B LoRA fine-tune | QLoRA | 20 hours |
| 30B full fine-tune | Native | Comfortable |
Image generation
A model that knows your business
Foundation models know almost nothing about you. For good reason.
Every prompt leaves the building, and every experiment has a cost attached. Synapse Nexus is built for the point where that stops making sense.
Your models run locally, privately, on hardware you own. No dependency on a service that can change its pricing tomorrow.
Data sovereignity
Every inference call, every training run, every fine-tune checkpoint processed and stored on hardware inside your network perimeter.
Your model weights don't touch a third-party GPU. Your training data doesn't transit a cloud region. Your prompts aren't logged by a provider whose retention policy you don't control.
No data processor agreements. No third-party sub-processors. No exposure.
a3974750-21c6-4f13-ae1f-205d8642b1a1.png){kind=icon}
Data agents
Run complex, multi-step data pipelines across your entire organisation: reporting, forecasting, and insight generation serving your whole team concurrently.
2792f579-79bc-4210-8a22-9bd967dc6ded.png){kind=icon}
Employee agents
Trained on everything your organisation knows, serving concurrent users with answers that smaller models can't match.
ef898b20-02e1-4e4c-ae0c-6c9744fa5208.png){kind=icon}
Security agents
Run continuous threat detection and response across your infrastructure, at the quality level where real-time decisions can be trusted.
0b3fa63b-00c5-4c3a-9707-906e2b9ee5cb.png){kind=icon}
Customer agents
Handle high volumes of customer requests concurrently while surfacing trends and patterns across your entire interaction history.
03265121-c10d-49a3-9ec0-27cf9d64368b.png){kind=icon}
Development agents
Fine-tune on your codebase, to generate, review, and document at a quality level that changes how fast your team ships.
f931c92f-2d05-4d24-bdbe-369af6cd1f27.png){kind=icon}
Creative agents
Generate production-quality text and images at scale, serving your whole team concurrently, on a model trained on your brand.
From operating expense to capital asset
At typical team usage, cloud GPU infrastructure and API spend combined runs to £4,500–5,500 a month. That's before you've shipped anything. Synapse Nexus pays for itself in around 9 months.
After that, every model you run, every experiment, every fine-tuning job is yours to run. The infrastructure is paid for. And the difference compounds with every user.
|
Typical Nexus team
A team of 5–15, sharing inference infrastructure, running experiments and fine-tuning jobs, paying API spend on top of cloud GPU rental.
|
|
|---|---|
| Cloud GPU rental Shared H100-equivalent, team hours | £12,000 / year |
| API spend for development work Team usage of Claude / GPT | £12,000 / year |
| Cloud GPU for fine-tuning runs ~200 hours / month | £4,000 / year |
| Total annual spend Nexus displaces | £28,000 / year |
Nexus payback in this scenario: ~9 months.
For teams with heavier API spend, payback can be as short as 7 months. After that, every inference call, every experiment, every fine-tuning job — you own the infrastructure outright.
Yours alone
Most AI work happens on borrowed infrastructure. Every prompt leaves the building, and every experiment has a cost attached.
Nexus is built for the point where that stops making sense.
Your models run locally, privately, on hardware you own. No dependency on a service that can change its pricing tomorrow.
Deliberate throughout
The RTX Pro 5000 Blackwell delivers 1.34 TB/s of memory bandwidth to the GPU. But a bottleneck anywhere else in the system costs you inference performance just as surely - CPU memory bandwidth starving the data pipeline, NVMe throughput limiting model load times, system RAM constraining batch size.
Every component in Nexus was specified against the same workload: sustained, concurrent, 70B-class inference. Nothing in the system is the weak link.
Graphics card RTX Pro 5000 Blackwell
Speed 1.3 TB/s memory bandwidth
Memory 128GB of DDR5-6000 ECC
Processor Threadripper 9995WX
Storage Dual model and OS drives
a1ad44c9-8864-4ab6-911b-13ecc78198ab.png)
Errors caught before they reach your model
Error-correcting memory built for model training that can't afford to fail.
Silent memory errors aren't caught by consumer hardware, and can invalidate a result without warning. You won't know until it's too late.
Error-correcting memory detects and corrects single-bit errors in real time. Every byte that passes through is checked, and every error is fixed. Automatically. In real time.
Space reserved for your models
Your models and your system each get a dedicated drive. Fast storage where it matters, headroom where you need it, and nothing competing for bandwidth.
OS drive
Keeps the system fast. Applications load instantly, pipelines initialise without delay.
Model drive
A dedicated 4TB for your models, datasets, and checkpoints. Room for a working library, with space to spare.
5e31d35d-1850-42b0-8ca5-b00599e798e4.png)
Supporting your configuration
Choosing the right AI infrastructure is a technical decision. Our team understands the workloads, the models, and the tradeoffs, and can help you get from first question to running inference, without the guesswork.
There when you need it
When hardware is part of your production stack, downtime has a number attached to it.
Hand-built and manually tested
Every Nexus is burned in and benchmarked by our engineers before it leaves us. It arrives ready to get to work.
Not sure where to start?
Speak to us before you configure. No obligation - just a conversation about whether the Nexus is the right tool for what you're building.
Configuration support service
Getting a workstation running is one thing. Getting it running the right models, configured for your team's workflows, with inference serving that's ready for production is another.
Our setup service covers the full stack. You tell us what you need to run. We make sure it runs.
Hardware commissioning
Software environment
Model selection
Model deployment
Testing and benchmarking
Find the right machine for your team
Three machines, one architecture. Node for the individual operator who needs serious local inference. Nexus for the team that needs shared infrastructure at scale. Frontier for the team shipping AI into production.
The right one depends on the models you need to run, the concurrency you need to support, and the workloads you can't afford to send to the cloud.
.png)
£8,999.99
From £228.38 per month
.png)
£20,999.99
From £382.28 per month
.png)
Synapse Frontier AI Workstation
£39,999.99
Specifications
Chillblast Synapse Nexus AI Workstation
Processor (CPU)
- AMD Ryzen Threadripper 9980X
- 64 Cores | 128 Threads
- Base Frequency: 3.2 GHz
- Boost Frequency: 5.4 GHz
CPU Cooler
- 360mm AIO Liquid Cooler
- Silverstone XE360-TR5
- 3 x 120mm Fan(s)
Motherboard
- ASUS Pro WS TRX50-SAGE WIFI
- AMD TRX50 Chipset
Network
- Ethernet Up to 10 Gbps (Marvell)
- Up to 2.5 Gbps (Intel)
- 2800 Mbps WiFi
- Bluetooth 5.4
Graphics
- NVIDIA RTX PRO 5000 Blackwell
- 48GB VRAM
Memory (RAM)
- 128GB DDR5 5600MT/s
- 8 x 16GB Modules
Case
- Fractal Define 7 XL Dark TG
- Black with Dark Tint Tempered Glass
- Tempered Glass Side Panel
Physical Dimensions
- Width: 240mm
- Height: 604mm
- Depth: 566mm
Front Ports
- 1 x USB 3.1 Type-C
- 2 x USB 3.0
- 2 x USB 2.0
- 1 x 3.5mm Jack - Headphone
- 1 x 3.5mm Jack - Microphone
Operating System
- Microsoft Windows 11 Home
Power Supply
- Seasonic PRIME TX ATX 3.0
- 1600W
- 80 Plus Titanium
Warranty
- 5 Years Includes:
- 3 Years On Site
- 2 Years Labour
- UK Mainland*
Storage
- 6TB
Rear Ports
- 1 x USB-C 3.2
- 6 x USB 3.1
- 2 x USB 2.0
- 1 x 3.5mm Jack - Line Out
- 2 x RJ-45 Ethernet
Solid State Storage (SSD)
- Samsung 9100 PRO
- PCIe 5.0 x4 NVMe
- Samsung 9100 PRO
- PCIe 5.0 x4 NVMe
Stock Code
- CB-AI-002
Industry-leading 5 year warranty
We’re proud of how much customers love and trust us, and that’s why your PC is protected by an industry-leading 5 year warranty. So just in-case anything goes wrong, we’ve got you covered.
The first 3 years warranty covers parts and labour – we’ll collect your system, repair it and replace any parts as necessary, and return it to you free of charge. The final 2 years covers labour costs on a return to base basis.
Finance options
It's simple to pay monthly with Chillblast. We offer financing on orders over £1,000, powered by Dopple.
Chillblast Limited is a credit broker, not a lender. We are authorised and regulated by the Financial Conduct Authority (FRN 1018196). We will introduce you to Social Money Ltd trading as Dopple (FRN: 675283) who will provide you with access to a panel of lenders who may be able to offer you finance facilities for your purchase.
