← Back to Daily Builds

Daily build note · October 27, 2026

Server Closet

Pick a model size, quantization, users and uptime to see the GPU memory a private AI deployment needs, the cheapest hardware tier that fits (workstation GPU to 8x datacenter server), rough capex and running cost, and the breakeven month against renting cloud GPUs or paying per token.

AI Tools Interactive demo live Static HTML/CSS/JS

What shipped

Server Closet is a sizing calculator for "AI in a box": a firm that won't send documents to a shared cloud API picks a model class (7B to 120B), a quantization (FP16, 8-bit or 4-bit), a context length, users, request volume and uptime, and gets back the GPU memory it needs (weights plus KV cache plus runtime overhead), the cheapest hardware tier that fits and keeps up (one workstation GPU, a 48GB workstation card, a Mac Studio-class box, or a 2x, 4x or 8x 80GB datacenter server), rough capex, monthly power and support cost, and a cumulative-cost chart against the cloud with the breakeven month marked. You can compare against dedicated cloud GPUs (a private tenancy, the default), a per-token open-model API, or a frontier API. Verdict chips include NEEDS A SERVER, WORKSTATION DEPLOYMENT, AIR-GAP READY, CLOUD NOT ALLOWED, PAYS BACK < 12 MO, CLOUD IS CHEAPER, KV CACHE HEAVY, OVERSIZED and PRIVACY, NOT PRICE. A tier ladder shows why each option fits or doesn't. There's a markdown report stamped in Central time, and state saves under server-closet-v2.

How to try

  • The demo opens on Law firm: 40 lawyers on a 70B model at 8-bit with 24x7 redundancy needs about 97 GB of GPU memory, which means two 2x 80GB servers at about $177,500. Against renting the same GPUs privately (about $9,400 a month), it pays back around month 24.
  • Load Clinic to see a 13B model on workstation GPUs pay back in under a year, and Defense contractor to see an air-gapped deployment where the cloud comparison is only hypothetical (CLOUD NOT ALLOWED).
  • Small dev shop fits on two 48GB workstation cards. School district has 150 people using it at once, so the KV cache outgrows the model weights and renting comes out cheaper.
  • Switch the comparison to Per-token open-model API and the case for buying hardware on cost alone goes away. That's why the PRIVACY, NOT PRICE chip is there.

Why this (vs neighbors)

This comes from the Startup Opportunities list ("Dell for private AI": AI in a box for firms that won't use the cloud). The first question any buyer asks is "what box, and what does it cost?", and the honest answer depends on memory math most firms have never done. Earlier builds like Agent Visitors and Agent Ledger look at what agents do. This one answers what hardware it takes to run them privately.

Scope trim

No live price feeds, no vendor quotes, no benchmark database, no fine-tuning or RAG storage sizing, and no multi-model serving. MoE models are treated as dense 120B for memory.

Limitations

  • Every price and throughput figure is a rough illustrative estimate, not a quote. Real hardware prices, GPU-hour rates and power costs change month to month.
  • Throughput uses a simple memory-bandwidth model at 25 output tokens per second per user. Real serving stacks (batching, speculative decoding, paged KV) can do much better or worse.
  • Compliance (HIPAA, CMMC, ITAR) depends on the whole control environment, not just where the model runs.

Suggested next steps

  1. Add editable hardware prices and a vendor-quote import.
  2. Add RAG storage, embedding model and backup sizing to the bill of materials.
  3. Turn the report into a one-page proposal a reseller could hand to a managing partner.