Your GPUs work the night shift.
SwingShift turns your firm's idle GPU workstations into a private, on-prem AI cluster every night and weekend. Each machine is back in Windows, exactly as its engineer left it, before the workday starts. No cloud spend. No data leaving the building.
One quiet cycle, every night.
Each workstation alternates between two operating systems that never touch each other's data. Your team submits jobs during the day and collects results in the morning.
Windows steps aside
Once the PC has been idle for 30 minutes with no CAD, render or simulation running, the agent hibernates Windows. Open apps and files are preserved exactly as left.
Night mode boots
A hardened Linux image boots from its own encrypted drive. It unlocks only when that machine's TPM and your on-prem key server both agree.
Batch AI runs locally
The on-prem controller dispatches queued jobs to each box by GPU tier. Models run with no internet route and no access to the Windows disk.
Handed back, as left
The box wipes job scratch space, resets the GPU and reboots. Windows resumes from hibernation, and the engineer's desktop is exactly where they left it.
Every failure path ends in Windows. Night mode sets "next boot = Windows" the moment it starts, so a crash, power cut or hung job still lands the machine back in the day OS. A hardware watchdog, a controller-forced power cycle and a power-button early return back that up.
Built for work that can wait until morning.
SwingShift is an overnight batch cluster. Submit through an OpenAI-compatible Batch API, so the tools and scripts you already use work unchanged.
Summarize & extract
Specs, RFPs, contracts and reports turned into summaries, tables and structured data overnight.
Searchable archives
Index drawings, project archives and standards with embeddings so your team can search them with AI.
First drafts
Proposal sections, meeting notes and code-review comments, ready for a human pass in the morning.
Translation
Whole document sets translated overnight, without sending client material to a third party.
Weekend fine-tuning
LoRA fine-tuning and model evaluation on your own data, using the long weekend window on your biggest cards.
Not for daytime chat. If your team needs interactive AI during the day, we quote a single always-on GPU server alongside your night cluster.
Designed to pass your IT review.
Every deal goes through IT and security, so the design starts there. SwingShift keeps three things apart: your employees' Windows data, the night workloads, and the outside world.
- Employee data is never touchedNight mode never mounts the Windows disk, and the NTFS driver isn't installed. The day agent reads only idle time and process names.
- Workload data stays on siteNight nodes have no internet route. The controller runs on your premises, and nothing is sent back to SwingShift.
- You hold every keyDrives are LUKS2-encrypted and bound to each machine's TPM and your on-prem Tang server. Recovery keys are escrowed with your IT team, never with us.
- A stolen box reveals nothingOff your network, or with Secure Boot tampered with, the night drive stays locked.
- Every job is traceableAn append-only audit log records who submitted what, which machine ran it, and when, without storing prompt text.
- No standing vendor accessSupport happens only through remote sessions you approve. Uninstall and drive wipe are available at any time.
Bring the workstations you already own.
NVIDIA pro cards, Ampere or newer, with at least 16 GB of GPU memory. Every box gets a dedicated NVMe drive for night mode, so Linux and models stay physically separate from Windows.
| Tier | GPU memory per box | Example cards | What it runs well |
|---|---|---|---|
| Entry | 16 GB | RTX A4000 | 7–8B models, embeddings, classification, simple extraction |
| Standard | 20–24 GB | RTX 4000 Ada, RTX A4500, RTX A5000 | Up to ~14B at 8-bit or ~24B at 4-bit; solid summarization |
| Prosweet spot | 40–48 GB | 2× 20 GB cards, RTX A6000, RTX 6000 Ada | ~30B-class models, or two independent workers on dual-card boxes |
| Max | 96 GB | Blackwell-class 96 GB | 70B-class at 8-bit, longer context, weekend fine-tuning |
Also required: UEFI with Secure Boot, TPM 2.0, wired Ethernet, Wake-on-LAN or Intel vPro, and Windows 10/11 Pro or Enterprise.
What are your idle GPUs worth?
A workstation used 45–55 hours a week sits idle for about 110 of the week's 168 hours. Here is what that time costs with SwingShift, compared with renting the same GPU hours in the cloud.
- GPU-hours reclaimed / month
- Electricity / month
- SwingShift subscription / month
- Your total / year
- Same hours in the cloud / year
Prove it on your floor, then subscribe.
Start with a paid pilot on up to 20 machines. If the mornings are clean and the work gets done, keep going on a simple per-workstation subscription.
Paid pilot
- Up to 20 workstations
- Hardware, BitLocker and hibernate check per model
- Encrypted night drive, agent, image and controller installed
- One workload running end to end
- Weekly report: clean mornings, jobs, GPU-hours, savings
Subscription
- Entry $50 · Standard $75
- Pro $100–125 · Max $150
- Image and agent updates
- Signed model bundles
- Support
Controller appliance
- Pre-built on-prem controller
- Key server, job queue and dashboard
- Or run it free on your own VM
Daytime chat server
- One always-on GPU server
- Interactive AI during working hours
- Same on-prem, no-cloud design
*Example: a Pro-tier box at $100/month plus ~$25 of electricity (0.35 kW at $0.15/kWh) over ~475 idle hours comes to about $0.26 per GPU-hour.
Firms with GPUs that sleep, and data that can't leave.
Architecture & AEC
Many GPU workstations, client NDAs and large document archives. Typical first workload: a searchable project archive and RFP responses.
Engineering consultancies
Civil, MEP and structural firms with heavy spec and code documents and small IT teams. Typical first workload: spec review and summarization.
VFX, animation & games
Big GPU fleets and strict content confidentiality. Typical first workload: asset tagging, script and document processing.
University labs
Idle lab machines and research data rules. Typical first workload: fine-tuning and evaluation runs over the weekend.
The questions IT asks first.
What happens if someone has a render or simulation running overnight?
The machine doesn't switch. The agent checks a configurable deny list of processes (CAD, simulation, render, backup), confirms no keyboard or mouse input for 30 minutes and GPU use under 5%, and re-checks every 5 minutes through the night. Employees can also click "Keep my PC on tonight".
What if someone comes in early?
A press of the power button triggers an immediate hand-back. The box is back in Windows, with their session resumed, in under three minutes.
Will this trigger BitLocker recovery prompts?
It can on some hardware if BitLocker is bound to boot-order measurements. Before any deployment we validate each workstation model through repeated cycles and recommend a PCR 7-only BitLocker profile. Recovery keys must be escrowed before install. "No recovery prompt" is a must-pass test.
Does Linux share the Windows drive?
No. Night mode lives on a dedicated, encrypted NVMe drive. The only change to Windows is hibernate and one firmware boot entry. Windows users see an unreadable partition in Disk Management, and it is disclosed in the employee notice.
Can one large model span several workstations?
No. Office Ethernet is too slow, so each box works independently. A box with two cards can pool them for one larger model or run one worker per card. 70B-class models are scheduled on 96 GB boxes.
How do we remove it?
Your IT team can uninstall at any time. Wiping a box means destroying its encryption keys and removing it from the controller. You hold the break-glass and recovery keys throughout.
Ten workstations. Sixty days. Every morning on time.
The pilot succeeds when 99% of mornings have every box back in Windows on time, no Windows user data is touched, and your workload finishes overnight at the volume we agreed. It starts with a 30-minute call with your IT team.