Arithmetic
Does owning the hardware actually cost you less?
Enter what you run today. This compares three years of per-token billing against three years of owning and operating the equivalent capacity, including power, maintenance, and someone to administer it. It will often tell you to keep renting.
What you run today
What you would build instead
Workstation-class node for a single team or an edge site. The throughput figure is an estimate with no audited benchmark behind it — treat it as a placeholder until measured on your workload.
Site and operating assumptions
Keep renting. Owning costs more at this volume.
Across 36 months, owning runs $232,964 above what you are paying now, before floor space, cooling capital, and the wait for switchgear. It does not turn in your favour at any volume this configuration can serve, so the configuration is the thing to revisit.
Your volume uses 3.7% of one node. Capacity is 13,676M tokens per month; the mark is 70%.
- Rented, per month
- $200
- Owned, per month after build
- $4,811
- Build cost, one time
- $66,960 · 1 node
- Effective cost per 1M tokens, rented
- $0.40
- Effective cost per 1M tokens, owned
- $13.34
- Rented, 36 months
- $7,200
- Owned, 36 months
- $240,164
- Difference
- costs $232,964
No break-even exists within this configuration’s capacity. It would need to change before the comparison means anything.
Bring these numbers to the call. Arguing with them is the fastest way to find out whether the case is real.
Throughput for the eight-GPU nodes is MLPerf Inference v5.0, server scenario, Llama 2 70B. Other model classes are scaled from that anchor and are estimates. Hardware pricing is the 2026 integrated-system range from OEM channel reporting. Token rates are published list prices as of September 2026. The workstation-class node is an estimate with no audited benchmark behind it.
Not included: networking beyond the node, storage, floor space, cooling capital, utility interconnection, and the lead time on switchgear. On current evidence those decide the project more often than the token arithmetic does. Every figure here is an input to a conversation, not a quotation.
Rate card last reviewed September 2026. Token prices move; if this page has not been reviewed in a while, treat every figure as indicative and ask.
Reference build
What an entry deployment actually looks like
This is a reference configuration, not a product. Every engagement starts by sizing against your workload, your power, and the space you actually have. The numbers below are what that sizing produces, and they are the numbers your facilities team will ask for first.
- Usable accelerator memory
- 384GB
- Peak electrical draw
- 9.4kW
- Thermal load
- 32,000BTU/hr
- Floor footprint
- 0.6 × 1.2m
Reference entry deployment. Sized per engagement; this is what the drawing set looks like.
Fit
When to call us
The data cannot leave the building.
Regulated workloads, contractual restrictions, or IP you are not willing to send through someone else's API. You need the model next to the data, with a defensible boundary around both.
We design the boundary first, then the compute that fits inside it, and document it in a form your auditor will accept.
The inference bill became a capital question.
Usage is steady, predictable, and large enough that renting GPUs now looks like paying rent on a machine you could own.
The assessment models owned against rented over three years, including power, cooling, and the operations cost most comparisons leave out. Sometimes the answer is keep renting. We will tell you that.
The hardware exists. Nobody runs it.
A rack was bought, a project started, and the person who understood it has left. It is drawing power and producing nothing.
We take over administration remotely, document what is there, and either bring it back into service or tell you what it is worth as scrap.
Budget is approved. The site can't take it.
The capital was signed off, the hardware is quoted, and then facilities asked about switchgear, cooling, and the utility interconnect. Now the project is stuck on questions nobody on the AI side can answer.
The assessment answers the power, thermal, and space questions against your actual building before a single part is ordered, and tells you what the site can support today versus what it needs first.
Price
The assessment
One engagement, priced in advance, credited in full against the build if you proceed. Deployment and managed operations are quoted from the assessment, not from a rate card, because we will not know what they cost until we have seen your site.
A recommendation not to buy is a valid outcome. If the arithmetic says rent, or the site cannot take the load, the report says so and the engagement ends there — you keep the report, the sizing, and the power and thermal findings either way. That is what the fee buys: a decision you can defend to a board, not a proposal.
Feasibility assessment
$15,000fixed fee · 2 weeks
Workload profiling, sizing, bill of materials, a power, thermal and space check against your actual site, the security boundary, and a fixed price for the build.
- Sizing model against your actual workload
- Bill of materials with current lead times
- Power, cooling, and space requirements
- Security boundary and management plan
- Fixed price for the build
- Three-year owned-versus-rented comparison
- A written recommendation — including not to proceed, where that is the answer
Sequence
How an engagement runs
- 01
Scoping call
Thirty minutes, no charge. If this is not worth doing, that is where we say so.
- 02
Feasibility assessment
Two weeks, fixed fee. Workload profiling, sizing, bill of materials, power and thermal check, security boundary, and a fixed build price.
- 03
Design and quote
You receive the full design and a fixed price for the build. The assessment fee is credited against it if you proceed.
- 04
Build and burn-in
Procurement, assembly, and staged load testing before anything ships to your site.
- 05
Handover
Installation, network and security configuration, runbook, monitoring, and your team trained to operate it.
- 06
Managed operations, if you want it
Remote administration on a monthly basis. If you do not want it, we leave and the system is still yours to run.
Scope
What we don't do
- Resell hyperscaler capacity, or take vendor margin we have not disclosed to you
- Act as general contractor for facility construction
- Build, train, or fine-tune your models
- Sell strategy decks
- Take the work when the honest answer is that you should rent it
Accountability
Who does the work
Preston ZenPrincipal
Security and infrastructure engineer, and the original creator of 1337sheets.com. Builds and operates server infrastructure across the United States, Europe, and Japan, sources and integrates hardware internationally, and works alongside the local electricians and technicians who will still be on site long after handover. Leads Kaizen Labs from a Japan and Ukraine hybrid office.
- OSCE³
- Offensive Security Certified Expert 3 defcon.org (opens in a new tab)
- DEF CON 34
- Four-hour workshop — “Battle-Tested Broadcasts: RF Insights From Ukraine”, on RF detection and direction finding defcon.org (opens in a new tab)
- DEF CON 33
- Four-hour workshop — “Disrupting Digital Dataflows”, on reverse engineering real-world electronic warfare techniques from Ukraine at the hardware level defcon.org (opens in a new tab)
- Ukraine, since 2022
- Volunteer work on drone systems, RF detection, EW countermeasures, and humanitarian logistics with Dronarnia, BeeTA, American Made Freedom, and Shield of Freedom defcon.org (opens in a new tab)
Specialist work — red teaming, physical security assessment, structured cabling — is subcontracted to named specialists, named to you before the engagement starts. You have one contract, one point of contact, and one party accountable for the outcome.
Start
Book a scoping call
Thirty minutes, no charge, on Monday, Wednesday or Thursday. Bring what you are trying to run and where it would live; you will leave the call knowing whether this is worth paying to assess.
A scoping conversation, not the assessment. We establish what you are trying to run, where it would live, and whether a private deployment is the right answer. If it is not, we tell you on the call. The assessment is a separate, fixed-fee engagement, sold on that call.
No charge · 30 minutes · Monday, Wednesday, Thursday
Please do not share credentials, keys, customer data, or site security details before or during the call. We will agree a secure exchange method if the project proceeds.
Or send it in writing
If you would rather not book yet, tell us what you are trying to run and where. If it is not a fit, we will say so in the reply rather than book a call to say it.