# Open-Weight Self-Host Sizing Kit

Version 1.1, November 14, 2026. From jwatte.com. Companion to the post "Open-Weight AI vs ChatGPT and Claude: Models, Costs, Hardware" at https://jwatte.com/blog/open-weight-ai-vs-chatgpt-claude-self-host-hardware/

Changed in 1.1: added the sources for Xiaomi's MiMo V2.6 Pro, which the post now covers.

What this is: worksheets and checklists for deciding whether to keep using a hosted AI service, use an open-weight model from a host, or own a machine that runs one. Fill in the blanks marked [ ]. Every price and specification below was read on November 14, 2026 (UTC) and will change; recheck before you rely on it. This kit is not legal, financial, procurement, electrical or engineering advice. Electrical work on panels, circuits and 240 volt outlets is for a licensed electrician under your local code.

We have no relationship with any vendor named here and were not paid. Prices are list prices on the read date, not quotes.


## 1. Decision checklist: stay hosted, hosted open weights, or own

Answer each line. The first path whose lines are all true is your starting point.

Path A. Stay on ChatGPT or Claude
- [ ] The best available answer quality matters more than cost for your main tasks. Done when you can name the tasks where a wrong answer is expensive.
- [ ] Sending this data to OpenAI or Anthropic is allowed by your contracts and policies. Done when the person responsible for data has said yes in writing.
- [ ] Your monthly AI bill is affordable. Done when you have last month's actual total.

Path B. Use an open-weight model from a host
- [ ] Some tasks are routine and high volume (sorting, extracting, first drafts). Done when you have listed them with a monthly count.
- [ ] A hosted open model passes your acceptance test (section 4) on those tasks. Done when the scoring sheet shows it passing the tasks that matter.
- [ ] The host's terms allow your data and use. Done when you have read the host's data and retention terms. Remember any host sees your prompts.

Path C. Own a machine
- [ ] At least one of these is true: data must not leave your building; the system must run with no internet; volume is large and steady. Done when you can write which one, and why.
- [ ] You have a room that can supply the power and cooling (section 7). Done when an electrician has confirmed the circuit.
- [ ] Someone will run, update and back up the machine, and someone can cover when they leave. Done when two names are written down.
- [ ] Owning beats hosting at your volume (section 3). Done when the worksheet shows your tokens per month above the break-even line.


## 2. Memory sizing worksheet

Formula (weights only):

  Weight memory in GB = parameters in billions x bits per weight / 8

Common bits per weight: BF16 = 16, FP8 = 8, NVFP4 or MXFP4 = 4 (plus small scale factors), 2-bit = 2.

Then add working memory for conversations. It grows with context length and with the number of people served at once. Wherever the model card states a hardware requirement, use the card, not this formula.

Fill in:
- Model: [ ]
- Total parameters (billions): [ ]
- Precision the card ships: [ ]
- Weight memory = [ ] x [ ] / 8 = [ ] GB
- Card's stated hardware (copy it): [ ]
- Machine you plan to use and its total GPU or unified memory: [ ] GB
- Room left for conversations = machine memory minus weight memory = [ ] GB

Filled example 1, a desk machine. Nemotron 3.5 Super VL, 120B parameters, NVFP4.
- 120 x 4 / 8 = 60 GB of weights.
- The card lists one DGX Spark (128 GB unified memory) or one RTX PRO 6000 (96 GB).
- NVIDIA's DGX Spark recipe serves one conversation at a time.

Filled example 2, a data-center model. Nemotron 3 Ultra 550B A55B, BF16.
- 550 x 16 / 8 = 1,100 GB of weights.
- The card's minimum is eight B200-class GPUs (8 x 180 = 1,440 GB) and says this fits weights plus working memory "with headroom".
- In NVFP4 it is 550 x 4 / 8 = 275 GB before scale factors; the card's minimum is four B200 or eight H100.

Filled example 3, a model that only fits the biggest box. Kimi K3.1 Pro, 2.8T parameters, MXFP4 weights per its card.
- 2,800 x 4 / 8 = 1,400 GB before the higher-precision parts.
- An 8x B200 box has 1,440 GB: no room left. An 8x B300 box has 8 x 288 = 2,304 GB.
- Hugging Face lists the published repository at about 1,561 GB, more than the 8x B200 box holds.
- This is our arithmetic; the card text we read did not state hardware.


## 3. Cost and break-even worksheet

Formulas:

  Owned cost per month = purchase price / months of use + electricity per month + other monthly costs
  Electricity per month = watts x hours per day x 30 / 1,000 x price per kWh
  Blended price per million tokens = (3 x input price + output price) / 4   (assumes 3 input tokens per output token; change the ratio to match your logs)
  Tokens per month needed to break even = owned cost per month / blended price per million tokens

Assumptions to fill in:
- Purchase price: $[ ]
- Months of use (36 for 3 years, 60 for 5 years): [ ]
- Machine watts (maker's rating): [ ] W. If only GPU ratings are known, sum them and label it a floor.
- Hours per day: [ ]
- Electricity price per kWh: $[ ]. US averages for August 2026 from the EIA: 18.16 cents residential, 14.47 cents commercial.
- Room, cooling or colocation per month: $[ ]
- Staff time per month: $[ ]
- Spares and support per month: $[ ]
- Software licenses per month (for example NVIDIA AI Enterprise if you use NIM in production, listed at $4,500 per GPU per year): $[ ]
- Closed model you would otherwise use, input and output price per million: $[ ] and $[ ]
- Host price for the same open model, input and output per million: $[ ] and $[ ]
- Measured tokens per second on your machine (only if you measured it): [ ]

Filled example (electricity and purchase only; the other lines are left for you):
- Machine: Qubrid 8x B200 appliance, sale price $564,000 on November 14, 2026.
- Power: Qubrid does not state draw. NVIDIA's own 8x B200 system, the DGX B200, is rated 14.3 kW maximum. Used as a stand-in: it is NVIDIA's figure for its own system and is not the seller's specification. Ask the seller for the real draw.
- Electricity: 14,300 x 24 x 30 / 1,000 = 10,296 kWh x $0.1447 = $1,489.83 per month.
- 3 years: 564,000 / 36 = $15,666.67 + $1,489.83 = $17,156.50 per month.
- 5 years: 564,000 / 60 = $9,400.00 + $1,489.83 = $10,889.83 per month.
- Against GPT-6.1 Terra ($1 input, $5 output; blended (3 x 1 + 5) / 4 = $2.00): 17,156.50 / 2.00 = 8,578 million tokens per month over 3 years; 5,445 million over 5 years.
- Against DeepSeek V4.1 Flash 1017 hosted on Fireworks ($0.30, $1.20; blended $0.525): 32,679 million per month over 3 years; 20,743 million over 5 years.
- Against Claude Haiku 5.5 ($0.10, $0.50 for prompts up to 100,000 tokens; blended $0.20): 85,782 million per month over 3 years.
- We had no measured throughput for this model on this machine, so we do not convert tokens to users. If you measure one, users is a separate calculation you should test, not assume.

Renting instead (730 hours is an always-on month):
- Monthly rent = hourly price x 730 x number of GPUs.
- Example: one RTX Pro 6000 at $2.69 an hour (RunPod) = $1,963.70. Eight B200 at $4.89 an hour (DeepInfra) = $28,557.60.


## 4. Fair-comparison method and acceptance test

How to pair an open model with a closed one:
- [ ] Same independent source and same date for both scores. Done when both numbers come from one leaderboard read on one day.
- [ ] Same reasoning setting, or the highest each offers, stated. Done when the setting is written next to each score.
- [ ] Similar context length, inputs (text, image, audio) and tool calling. Done when you have listed any difference.
- [ ] Vendor-reported scores labeled as the vendor's claim. Done when no vendor number sits in a column with independent numbers.
- [ ] Where no independent result exists, write "no independent result found" and the date.
- [ ] Write what you give up. Done when the pairing has a one-line "what you give up" note.

Acceptance test template (20 real tasks):
1. Pick 20 tasks from last month's real work. Mix easy and hard. Remove names and private details if the test leaves your building.
2. Write what a correct answer must contain for each.
3. Run each task on your current hosted model and on the candidate open model, with the same instructions.
4. Hide which answer came from which model. Have a person score both.
5. Decide in advance which tasks must pass.

Scoring sheet (copy one line per task):

  Task # | Task in one line | Must pass? (Y/N) | Answer A score 0 to 2 | Answer B score 0 to 2 | Notes

Score 0 = wrong or unsafe, 1 = usable with edits, 2 = correct as is.
Done when every must-pass task scores 2 on the candidate, or you have decided in writing to accept a lower score.


## 5. On-premises setup checklist and offline checklist

Setup (who: your IT person unless noted):
- [ ] Choose serving software. Ollama or LM Studio for one machine and a few people; vLLM or SGLang for many users; NVIDIA NIM where a supported container is wanted (production use needs an NVIDIA AI Enterprise license per NVIDIA's NIM FAQ); llama.cpp for CPU or mixed hardware. Done when one model answers a test question.
- [ ] Copy the serving command from the model card or the project's own documentation for your exact hardware. Done when the command and its source page are saved with the date.
- [ ] Expose an OpenAI-compatible endpoint and put a gateway in front for keys, limits and logs. Done when an existing tool works after changing only its base address.
- [ ] Set up a chat screen with one account per person. If you use Open WebUI, read its license: branding may not be removed once more than 50 people use it in 30 days, unless you have permission or an enterprise license. Done when each person can sign in.
- [ ] Decide what to log and for how long. Done when the setting matches a written policy.
- [ ] Run the acceptance test (section 4). Who: owner. Done when it passes.
- [ ] Time answers with your expected number of people working at once. Done when you have a written result.

These commands were copied from the projects' documentation on November 14, 2026, and were not run by us:
- Ollama: ollama run nemotron3
- LM Studio CLI: lms server start
- llama.cpp server: llama-server -hf <user>/<model>:<tag>

Offline (air-gapped):
- [ ] Download weights, containers and installers on a connected computer. Done when each file has a recorded checksum.
- [ ] Move them on encrypted removable storage. Done when the checksums match on the offline machine.
- [ ] Confirm each license allows offline use. Done when the license text is saved beside the weights.
- [ ] For NIM: NVIDIA's air-gap guide prepares a cache while connected, then runs with no outbound network and no API keys. Ask NVIDIA how production licensing is checked offline. Done when you have NVIDIA's answer in writing.
- [ ] Turn off phone-home features where the software has them. Settings from each project's documentation: vLLM VLLM_NO_USAGE_STATS=1 or DO_NOT_TRACK=1; Hugging Face HF_HUB_OFFLINE=1 and HF_HUB_DISABLE_TELEMETRY=1; Ollama OLLAMA_NO_CLOUD=1; Open WebUI OFFLINE_MODE (download its embedding model first). Done when each setting is applied and logged.
- [ ] Set up a local update mirror, a local time source, local logs and local backups. Done when a restore from backup has worked once.
- [ ] Write the procedure for bringing in a new model version, including who approves it. Done when it is signed.
- [ ] Prove it. Unplug the network cable, restart, run the acceptance test. Done when every must-pass task passes unplugged.


## 6. License review checklist (questions, not answers)

For each model you plan to use:
- [ ] What is the license name and version, and where is the full text? Saved a copy with the date?
- [ ] Does it allow commercial use? Under what conditions?
- [ ] Does it treat offering the model as a service to others differently from internal use? Are there revenue or user thresholds?
- [ ] Must you display a name or notice in your product?
- [ ] Is there a usage policy attached? What does it forbid?
- [ ] Does it restrict use by company size, by sector or by country?
- [ ] Does it end your rights if you bring a patent or copyright claim?
- [ ] Is the model derived from another model whose license also applies?
- [ ] Are the training datasets under separate licenses that matter to you?
- [ ] Is access gated behind a click-through you must accept?
- [ ] Does anything in it depend on being online?
- [ ] Where is the maker based, and do your contracts or policies limit software from that country? Ask counsel.
- [ ] Has the license changed since the version you downloaded? Who checks, and how often?


## 7. Hardware site worksheet (power, cooling, noise, weight, room)

Formulas:

  Watts = volts x amps
  Continuous limit = circuit watts x 0.8   (breakers are sized at 125% of a continuous load, one running 3 hours or more)
  Heat in Btu per hour = watts x 3.412   (1 kWh = 3,412 Btu, EIA)
  Tons of cooling = Btu per hour / 12,000
  Battery backup energy in watt-hours = watts x minutes needed / 60   (before inverter losses)

Reference values:
- 120 V 15 A circuit = 1,800 W, continuous 1,440 W.
- 120 V 20 A circuit = 2,400 W, continuous 1,920 W.
- 240 V 20 A circuit = 4,800 W, continuous 3,840 W.
- 240 V 30 A circuit = 7,200 W, continuous 5,760 W.
- A resting adult gives off about 105 W (356 Btu per hour). A portable heater is 750 to 1,500 W. A clothes dryer is 1,800 to 5,000 W.
- A workstation maker's own advice (Bizon X5500 G2 page): one 20 A, 120 V outlet on a dedicated circuit for its one-GPU build; two such outlets, each on its own dedicated circuit, for its four-GPU build; and consult your electrician.

Outlet and circuit check:
- [ ] Machine's rated watts from its spec sheet: [ ] W. Done when the source page is saved.
- [ ] Circuit voltage and breaker amps: [ ] V, [ ] A. Done when an electrician has identified the circuit.
- [ ] Everything else on that circuit: [ ] W. Done when listed.
- [ ] Machine watts plus other loads is below the continuous limit. Done when an electrician agrees.

Filled example, a home machine (DGX Spark):
- Rated 240 W power supply (NVIDIA).
- Heat: 240 x 3.412 = 819 Btu per hour, about two resting adults.
- Electricity: 240 x 24 x 30 / 1,000 = 172.8 kWh x $0.1816 = $31.38 a month around the clock; 8 hours a day = $10.46.
- Fits a 120 V 15 A circuit (1,440 W continuous) with room for other loads. NVIDIA declares 29 dB(A) at the operator position at full load.
- Battery for 10 minutes: 240 x 10 / 60 = 40 Wh before losses.

Filled example, an appliance (8x B200 class):
- Qubrid does not state draw, heat, weight, airflow or noise. Every figure in this example is NVIDIA's for its own DGX B200, used as a stand-in. None of it is the seller's specification; ask the seller for the real figures.
- NVIDIA's DGX B200 is 14.3 kW maximum, with six power supplies at 200 to 240 V, 16 A each, C19/C20 locking cords.
- Heat: NVIDIA states 48,794 Btu per hour. Our arithmetic: 14,300 x 3.412 = 48,792. That is 48,794 / 12,000 = about 4 tons of cooling.
- Circuits: 14,300 / 5,760 = about 2.5 continuous 240 V 30 A circuits of capacity. Your electrician or data center designs the actual feeds.
- Electricity at the business rate: 14,300 x 24 x 30 / 1,000 = 10,296 kWh x $0.1447 = $1,489.83 a month.
- Physical: NVIDIA lists the DGX B200 at 142.4 kg, 10 rack units, 1,550 cubic feet per minute of airflow, 10 to 35 C operating temperature.
- Battery for 10 minutes: 14,300 x 10 / 60 = 2,383 Wh before losses.

Cooling and room:
- [ ] Heat load in Btu per hour: [ ]. Done when compared with the room's cooling capacity.
- [ ] Cooling capacity available: [ ] Btu per hour. Done when the HVAC person has confirmed it.
- [ ] Noise: maker's stated level [ ] dB, and where people sit. NIOSH advises hearing protection at 85 dBA and above. Done when the machine is not in a room where people work, if its level is unknown or high.
- [ ] Weight and floor: [ ] kg, rack units [ ]. Done when the rack and floor rating are confirmed.
- [ ] Room locks and access list. Done when keys or badges are assigned.
- [ ] Surge protection and grounding. Done when an electrician has checked both.

Questions for an electrician:
- What is on this circuit now, and what is its breaker size?
- Can this circuit carry [ ] W continuously, or do we need a new one?
- Do we need a 240 V circuit, and what outlet type matches the machine's cord?
- Is the panel big enough for the new circuit?
- Is the ground correct, and where should surge protection go?
- Does local code require a permit or inspection?

Questions for a colocation provider:
- How many kW can one cabinet draw, and at what voltage and amperage?
- What is the price per cabinet or per kW per month, and what is included?
- Can you cool [ ] Btu per hour in one cabinet? Is the facility air-cooled or liquid-cooled?
- What are the weight limits per cabinet?
- What remote hands, access hours and security do you provide?
- What happens to power during maintenance? Are A and B feeds separate?
- What is the contract term and the exit process?


## Sources (all read November 14, 2026, UTC)

- NVIDIA Nemotron page, https://developer.nvidia.com/topics/ai/nemotron (03:34:47)
- Nemotron model cards on Hugging Face, for example https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 and https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 (03:35:29 to 03:35:30)
- Kimi K3.1 Pro card, https://huggingface.co/moonshotai/Kimi-K3.1-Pro (03:37:44)
- Qubrid appliance pages, https://www.qubrid.com/ai-appliances/qubrid-llm-gen-ai-appliance-server-8-x-nvidia-b200-gpu-air-cooled and its B300 and RTX PRO 6000 siblings (03:34:48 to 03:34:49)
- NVIDIA DGX B200 user guide, https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html (03:44:01)
- NVIDIA DGX Spark, https://www.nvidia.com/en-us/products/workstations/dgx-spark/ (03:43:39)
- Bizon X5500 G2, https://bizon-tech.com/bizon-x5500.html (03:45:42)
- Kimi K3.1 Pro repository size, https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro (04:40)
- OpenAI API pricing, https://developers.openai.com/api/docs/pricing (03:38:35)
- Anthropic API pricing, https://platform.claude.com/docs/en/about-claude/pricing (03:38:38)
- Fireworks pricing, https://fireworks.ai/pricing (03:40:07)
- DeepInfra pricing, https://deepinfra.com/pricing (03:40:08)
- RunPod pricing, https://www.runpod.io/pricing (03:40:27)
- NVIDIA AI Enterprise pricing, https://docs.nvidia.com/ai-enterprise/planning-resource/licensing-guide/latest/pricing.html (03:41:11)
- NVIDIA NIM FAQ, https://docs.api.nvidia.com/nim/docs/product (03:41:47)
- NVIDIA NIM air-gap deployment, https://docs.nvidia.com/nim/large-language-models/latest/deployment/air-gap-deployment.html (03:41:28)
- Open WebUI license, https://github.com/open-webui/open-webui/blob/main/LICENSE (03:46:50)
- Ollama, https://ollama.com/library/nemotron3 and https://docs.ollama.com/faq (03:47:04)
- vLLM usage stats, https://docs.vllm.ai/en/latest/usage/usage_stats.html (03:47:05)
- Hugging Face environment variables, https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables (03:47:06)
- llama.cpp server, https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md (03:47:07)
- LM Studio CLI, https://lmstudio.ai/docs/cli (03:47:07)
- Open WebUI environment settings, https://docs.openwebui.com/reference/env-configuration/ (03:47:28)
- EIA Electric Power Monthly Table 5.6.A, https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a (03:41:48)
- EIA British thermal units, https://www.eia.gov/energyexplained/units-and-calculators/british-thermal-units.php (03:42:23)
- IAEI on NEC 210.20(A), https://iaeimagazine.org/2016/may2016/100-vs-80-choosing-the-right-ocpd-solution/ (03:42:21)
- IAEI appliance table, https://iaeimagazine.org/electrical-fundamentals/energy-use-of-home-appliances/ (03:43:19)
- Cornell thermal notes, https://www.ergo.human.cornell.edu/studentdownloads/DEA3500notes/Thermal/thcondnotes.html (03:43:18)
- ACCA on tons of cooling, https://hvac-blog.acca.org/?p=16063 (03:43:20)
- NIDCD, How loud is too loud, https://www.nidcd.nih.gov/health/how-loud-too-loud (03:42:36)

All of the above were read again in an independent review on November 14, 2026, 04:31 to 04:41 UTC, and the prices were rechecked in a desktop browser at 04:22 to 04:26 UTC.

Added in version 1.1 (read November 14, 2026, UTC):

- MiMo V2.6 Pro card, https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL (18:11:30)
- MiMo V2.6 Pro repository size, https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.6-Pro-RL?blobs=true (18:11:30)

Prices and specifications change. Recheck every number on the seller's or maker's page before you buy or sign.