FOR FEDERAL CONTRACTORS
A sealed AI machine inside your facility.
Staff put documents into it that they could never put into ChatGPT. CUI never leaves the building. Installed in a day, owned outright.
THE PROBLEM
01
PROHIBITED.
Controlled data is barred from public AI tools. CUI in ChatGPT is a reportable DFARS 252.204-7012 incident, with a 72-hour clock.
02
SHADOW USE.
Written policy does not stop it. Staff use whatever finishes the job fastest. The fix is a sanctioned tool on your own network.
03
ENFORCEMENT.
DoD suspended CMMC's Phase 2 deadline in July, citing costs; the safeguarding rules never paused. In June a Navy logistics contractor paid $507,144 over false compliance claims.
TWO SCENARIOS
Tuesday, 4:40 PM.
Your proposal is due at five. A contracts manager is staring at a forty page RFP with a distribution statement on every page, and the summarizer that would save the hour is one paste away. They know the policy. The deadline does not care about the policy.
With a Node on the rack, the sanctioned tool is the fast tool. The RFP goes into the machine on your own network, the summary and the compliance matrix come back, and the proposal files at 4:58. Nothing crossed the perimeter. No incident, no 72-hour clock, no story.
Thursday, 9:15 AM.
The delivery schedule lives in four spreadsheets that do not agree, and reconciling them eats one morning every week. Someone finally asks the machine for a tool that does it. Twenty minutes later a small working program sits on the shared drive: it reads all four files, flags every disagreement, and prints the morning report. Nobody wrote a requisition. Nobody hired a contractor. The person who asked is not a programmer.
The fastest tool in the building should be the one your people are allowed to use. Tuesday was one percent of what it does.
THE NODE
A sealed machine on your own network, configured and maintained by Splay inside your facility. Staff use it to draft proposals and compliance matrices, write working scripts and small tools, and search the documents you cannot upload: proposals, drawings, contract data.
It needs no cloud account and no internet connection, and nobody pays per seat. We maintain the machine without ever holding your files. The whole setup is written down, step by step; any competent IT shop could run it without us.
NOT A CHATBOT
The free tools taught everyone that AI answers questions. A Node runs the same kind of AI behind the big-name tools, on your own floor, and it produces finished work.
WORKING SOFTWARE
Scripts, small tools, whole programs. Staff describe the tool in plain English and run what comes back.
WEB PAGES
An intranet page, a form, a working dashboard, from a paragraph of description.
THE DOCUMENTS THAT WIN WORK
Proposal sections, compliance matrices, SOPs, contract-clause reviews, in your own templates.
DRAWINGS AND PHOTOS
With a vision cartridge loaded, staff show it scanned prints and pictures of parts.
Every Node does all of this; the bigger machines do it with stronger hands. The pilot exists so you can watch it happen on your own problems.
THE MACHINES
Five machines and a bespoke build, priced flat and in public. Each one is sealed: no path out to the internet, serving your staff over your own network, with the strongest open model that fits its memory. The drawings are illustrations; the hardware is listed on each card.
The ladder is one design at five sizes, and the small machines grow into the big ones: a second Spark and one cable turn a NODE 01 into a NODE 02, and two more turn it into a NODE 03. Nothing is replaced and nothing is migrated; the machine you bought becomes part of the machine you upgrade to, and what you paid counts in full against the price. Past the Sparks, the bigger machines share the same network and the same work, and upgrades into them are credited by quote. NODE 03 itself runs two ways: linked into one machine carrying GLM-5.2 for a small team that wants the strongest model, or split into four Sparks serving up to sixty staff on the smaller models.
NODE 03 and NODE 04 differ on speed, not intelligence. The Sparks are a highway with many lanes and a low speed limit: they carry a whole office, unhurried. The Blackwell machines are a race track with one lane and no limit: a Spark moves memory at 273 gigabytes a second, and each of NODE 04's four cards moves it at 1,792. NODE 04 runs GLM-5.2 too when you want the strongest model; its card names MiniMax-M3 because that pairing feels instant. If your staff sit waiting on answers, the wait is what NODE 04 removes.
63
INTELLIGENCE / FRONTIER = 100
- MODEL
- MINIMAX-M2.7
- HARDWARE
- DGX SPARK, 128GB
- SERVES
- 1-15 STAFF
$14,500 INSTALLED
YEAR 1 CARE INCLUDED, THEN $600/MO
67
INTELLIGENCE / FRONTIER = 100
- MODEL
- DEEPSEEK V4 FLASH
- HARDWARE
- 2X SPARK, LINKED, 256GB
- SERVES
- 16-30 STAFF, TWO MACHINES
$32,500 INSTALLED
YEAR 1 CARE INCLUDED, THEN $900/MO
85
INTELLIGENCE / FRONTIER = 100
- MODEL
- GLM-5.2
- HARDWARE
- 4X SPARK, LINKED, 512GB
- SERVES
- 1-10 STAFF ON GLM-5.2, OR 60 SPLIT
$59,500 INSTALLED
YEAR 1 CARE INCLUDED, THEN $1,200/MO
73
INTELLIGENCE / FRONTIER = 100
- MODEL
- MINIMAX-M3
- HARDWARE
- 4X RTX PRO 6000, 384GB
- SERVES
- 15-75 STAFF, FAST
FROM $129,500 INSTALLED
YEAR 1 CARE INCLUDED, THEN $1,500/MO
85
INTELLIGENCE / FRONTIER = 100
- MODEL
- GLM-5.2, BEST OPEN MODEL
- HARDWARE
- DGX STATION GB300, 748GB
- SERVES
- 15-75 STAFF
FROM $195,000 INSTALLED
YEAR 1 CARE INCLUDED, THEN $2,500/MO
The ladder covers most offices. When it does not cover yours, we build to order: memory, speed, and models sized to your work, quoted in writing after a call.
BY INQUIRY
JACK@SPLAY.LLC, SUBJECT: BESPOKEIntelligence: the strongest closed model = 100, on the Artificial Analysis index, JULY 2026. Raw scores 38, 40, 51, 44, 51 against the frontier's 60. On trackingai.org's offline Mensa-style IQ test, frontier models measure 129 to 136, and GLM-5.2, the NODE 03 and 09 model, measures 123; the other flagships have no published score. The model inside swaps like a cartridge. Offices that require American-made models run gpt-oss, Llama, or Nemotron; no model can send anything out either way. NODE 01 sits under the $15,000 federal micro-purchase threshold; purchase cards clear it. Spark machines upgrade by joining, and what you paid credits in full from NODE 01 through 03; upgrades into 04 and 09 credit by quote. NODE 04 and 09 are built to order: 50% deposit, quotes hold 30 days. Care is flat per machine, adjusting up to 5% annually. Prices as of July 22, 2026, subject to change.
Pick a model. See where it lands.
Published open models against the machines. The verdict is arithmetic over each model’s published size and each machine’s installed memory, nothing else. The flag is the developer’s home country.
NODE 01 / 128GB
NO FIT
NODE 02 / 256GB
NO FIT
NODE 03 / 512GB
Q4
NODE 04 / 384GB
Q3
NODE 09 / 748GB
Q4
The estimate covers the model itself plus 5% overhead; working memory is extra. Sizes are the published figures.
THE MATH
| SPLAY NODE | GOV-CLOUD AI | PUBLIC AI TOOLS | |
|---|---|---|---|
| PERMITTED FOR CUI | YES | YES, AFTER MIGRATION | NO |
| YOUR DATA LIVES | IN YOUR BUILDING | MICROSOFT GOV CLOUD | ON VENDOR SERVERS |
| SETUP | FROM $14,500 | $50,000+ MIGRATION | NONE |
| RUNNING COST | $0 YR 1, THEN $600/MO | ~2X LICENSING/SEAT | $25-30/SEAT/MO |
| COST TO ADD A SEAT | $0 | ANOTHER LICENSE | ANOTHER SEAT |
Public tools remain fine for public work. The Node covers the work they are barred from. Gov-cloud figures are typical GCC High migration plus licensing uplift and vary by tenant. Hardware alone does not certify you.
Your headcount. The math.
The gov-cloud band is GCC High class licensing plus migration, and varies by tenant. The Node figure is list price plus 24 months of care at the year-one rate, for the cheapest machine that clears your headcount and your intelligence floor.
DAY ─── OF THE CMMC REWRITE
DoD suspended the Phase 2 certification deadline on July 13, citing the cost burden on small contractors. DFARS 252.204-7012 remains in force.
THE RATE CARD
All of it. If a figure is not on this page, it is not a price.
The clean room
A CUI enclave, designed and built inside your facility: its own walled-off network, a bay for your Node, screens that can view but never copy, and a documentation pack for your assessor.
$25,000 TO 10 SEATS
$45,000 TO 25 SEATS
Fixed price; the machine is priced separately. You own it, and your IT runs it with our manual.
Services
* Credited in full against any machine. Courses are operator training: getting employee-grade work out of the machine, on your own examples. Custom models are tuned and tested on your hardware, and your files stay with you. Extra on-site work is $2,400 a day.
On retainer
Model drop: each quarter's best open model, installed and briefed on-site. Included with NODE 04/09 care. AI officer: a standing operator who builds the prompts, workflows, and small tools your staff ask for.
HOW IT WORKS
The call.
20 minutes. You talk, we scope. No slides.
The survey.
A readiness survey at your facility. $2,400, credited in full against any machine.
The pilot.
A pilot machine on your bench for 30 days. $2,000, credited in full. Then you decide.
A fleet of four pilot machines exists. Every machine is a copy of one documented setup, with a step-by-step manual; any competent IT shop could run it without us.
QUESTIONS, ANSWERED
Is this CMMC certification?
No, and no machine is. Hardware alone does not certify you. The Node gives your staff a permitted place to work with controlled data, and the clean room engagement includes a documentation pack for your assessor.
The strongest open models are Chinese. Can we run American models?
Yes. The model inside swaps like a cartridge. Offices that require American-made models run gpt-oss, Llama, or Nemotron; NVIDIA's Nemotron 3 Ultra scores 38 on the same public index the ladder uses, level with NODE 01's flagship. The machine is sealed either way: no model of any origin can send anything out.
Can it actually write software?
Yes. Staff describe a tool in plain English and the machine writes the code; small tools and scripts come back in minutes and run on your own network. It is not a replacement for your software vendor on big systems: it is closer to a very fast junior developer. The pilot exists so you can test it on your own problems.
Who runs it after the install?
Your IT does, with our manual. Every machine is a copy of one documented setup. If you would rather we watch it, managed care starts at $600 a month and year one is included with every machine.
What happens if Splay disappears?
You own the machine outright, and the whole setup is written down, step by step. Any competent IT shop could run it without us.
Can we try it before we buy it?
A pilot machine sits on your bench for 30 days for $2,000, credited in full against any machine. So is the $2,400 readiness survey.
What does it cost to add a seat?
$0. The machine is the cost. NODE 01 serves 1 to 15 staff; when you outgrow a machine you buy the next machine, not the next license.
How many people can use it at once?
Everyone gets their own chat over your network; nobody shares a screen. What limits a machine is how many answers it is writing at the same instant. NODE 01 is bench-measured at ten simultaneous conversations, and an office of fifteen rarely has five going at once. The bigger machines carry more, and we size each person's working room at install so the machine stays quick. The serves range on each card is that sizing, done conservatively.
What happens when we outgrow it?
You grow the machine. The Spark machines join by one network cable: a second Spark turns a NODE 01 into a NODE 02, four make a NODE 03, and what you paid is credited in full against the upgrade. Upgrades into the bigger machines are credited by quote. No machine on the ladder is a dead end.
Do you see our data?
No. The machine has no path out to the internet, and your files stay in your building. Custom models are tuned and tested on your hardware, and support visits happen on-site, at the rate on the card.
Why are your prices public?
Because a buyer who checks must find every figure where we said it would be. The intelligence numbers come from a public index with the raw scores printed in the footnote, and every price carries its edition date: July 22, 2026.
TALK TO US
The people who answer this address are the ones who install the machines. We are based in Virginia and install on site across Northern Virginia and the Washington DC area. 20 minutes, no slides.