Years building
4 yrs
Trillions of tokens behind it
One intelligence, built around your business.
Labarna orchestrates seven frontier models under one sovereign engine and delivers a complete, working system as the client's own technology in under 30 days — not a wrapper on someone else's model, and not a consultancy. The same engine powers a private, sovereign AI that lives in a person's own pocket.
Every company is being told to adopt AI. What the market ships them is a thin layer over someone else's model, their data on someone else's cloud, a bill that never ends, and nothing they own at the end of it. It looks real in a demo and falls apart in production — because generating the code was never the hard part. Connecting it, migrating onto it, running it safely and keeping it alive is.
Two arms, one engine.
We build the complete system — agents, integrations, the whole platform — and hand it over as the client's own technology in under 30 days. They own the code, the data and the deployment; we run the intelligence connected to it. That is Ghost Architecture.
A private, sovereign AI that lives on the person's own device — their assistant for anything, their data staying theirs, powered by the same engine. Not an account on someone's cloud; their own.
A wrapper is one model behind a prompt. Labarna orchestrates seven frontier models under a sovereign engine built over four years and now in production — dependent on no single lab, and it rides whichever model wins. This is not the AI being pitched everywhere else.
For capital in this region: the intelligence and the data stay independent of any single US lab or cloud — and run on the client's own metal, not rented from someone else's.
The real base the forecast grows from, not a projection.
Roughly $80K/month is recurring — a sticky base that holds independent of new signings. Everything above it is new-client-dependent, which is the single variable the forecast turns on and where the capital goes to work.
Four years and trillions of tokens of building are compressed into one system, now in production. A competitor cannot buy their way level by licensing a better model — the advantage is accumulated, not purchased.
The build never slows. Today compute runs on six dedicated servers plus five in the cloud, with six more cloud NVIDIA servers and two clusters going in to carry the AI load; the roadmap moves all of it onto new owned metal in Dubai and Asia and onto our own model, then rides commoditization down into volume and consumer on economics no one can match.
Cash collected across all lines, USD, from the operating plan. Floor is the case we commit to and defend; ideal is the case we drive for. 2027 collects $3.9M–$5.4M, about 2.5–3.4x 2026, because the 12-month payment route pushes roughly half of each 2027 build's cash into 2028 — which is exactly why 2028 steps to $16M–$30M. Contract value signed runs higher than cash collected each year. The raise-funded line layers on top: $12M closing June 2027 funds a sales team hired ahead of revenue plus the pocket companion, lifting collections to about $6.5M in 2027 and about $36M in 2028.
The plan above is bank-funded and hires out of cashflow — that is the ceiling. Twelve million closing June 2027 removes it: sellers, operators, marketing and compute go in ahead of revenue, and the pocket companion turns on in January 2028. The ideal plan against the raise-funded case:
| 2027 ideal | 2027 raise | 2028 ideal | 2028 raise | |
|---|---|---|---|---|
| New build clients | 32 | 50 | 108 | 120 |
| New build contracts | $9.6M | $15.0M | $32.4M | $36.0M |
| Cash collected | $5.4M | ~$6.5M | $29.8M | ~$36.0M |
| Pocket companion | — | free | — | ~$3.0M |
| Cash at 31 Dec | $2.8M | ~$12M | $18.4M | ~$27M |
| Signed into next year | $5.4M | ~$10M | $13.6M | ~$22M |
The ideal columns are the bank-funded plan with no raise. The raise case hires five sales engineers in June 2027, productive within a month, each closing one build a month — a normal monthly quota with AISCO feeding the funnel — and scales the team through 2028, on top of the founder, the senior closer and Rescue conversions. Builds at the ~$300K average through the June 2028 engine cutover, after which the $25K self-serve tier carries added volume. Year-end raise cash is the organic case plus the $12M, less the metal and the team and marketing put to work. Set rep-by-rep and month-by-month in the workbook.
Two builds a year pay full price; every other build is on the 12-month route, so about half of each 2027 build's cash is collected in 2028. The result is a contracted base that lands before a single 2028 deal is signed.
| 2028 cash already booked in 2027 | Floor | Ideal |
|---|---|---|
| 12-month installments from 2027 builds | $2.87M | $5.45M |
| Sovereign AI care and usage from 2027 clients | $1.17M | $1.98M |
| Total 2028 cash booked in 2027 | $4.04M | $7.43M |
That is the payoff of the monthly route: slower cash in year one, a contracted base in year two — $4.0M to $7.4M of 2028 collections on the books before 2028 selling even starts, plus the recurring subscription base carried in.
The ramp is not hope; it is a machine with several channels feeding it. Every path above the recurring base runs through this.
The pocket companion is distributed by viable marketing on a proven give-it-away-to-seed-the-market playbook: free for the first few months to take ground. Because it runs on the user's own device and the system is already built, the compute cost of that giveaway is negligible. Land the base first, monetize it after.
What everyone now says they want and almost no one can deliver: real AI that is private, on the user's own device, and under their control. Not an account on someone's cloud — their own.
Free on the device through 2027 to take ground; paid from January 2028. The extra $10 connects to the Sovereign Engine for real-time updates and the heavy work the device cannot do alone. On about $150K/month of marketing at roughly $4 a free install, near 450,000 free users land in 2028; at a ~3% paid conversion and a ~$53 blended price, that is roughly $3M recognized in 2028 and about $8.5M of exit run-rate by December — on top of the build and Sovereign AI lines.
Give it away free to take ground, because a free user on their own device costs us almost nothing. Charge the moment they reach for the connected power — the one place real compute is spent is the same place revenue is made. Free to take ground, paid at the point of power.
The system is the headcount. What a bloated org does with hundreds of engineers, the machine does on compute. We don't inflate staff to dress up a story — the spend that scales is tokens, not salaries.
Four years building the system founder-led with contract developers, moving as fast as a founder can. That model was outgrown and retired on purpose.
Building the permanent core in-house — QA, admin, platform operators and a few key hires. Deliberately lean. Outside accounting runs through counsel. The lean cost base is already in the numbers.
Owned GPU collapses marginal cost per build. That is what makes a $25K tier and a zero-marginal-cost pocket app viable at volume — neither can run on rented frontier APIs. Illustrative cost per build:
Today compute runs on six dedicated servers plus five in the cloud, with six more cloud NVIDIA servers and two clusters going in to carry the AI load. At Round 1 close it all moves onto new owned metal in Dubai and Asia, the cloud comes off, and the sovereign companion flips on. End-2027 the stack decouples from frontier APIs onto our own model — sovereignty becomes literal. Round 2 then funds consumer distribution, mass-inference GPU and global reach.
By Round 1 close the owned fleet is eleven NVIDIA DGX frontline AI servers and thirty-four orchestration servers — about $7M of metal, $4M from the raise plus a $3M bank server loan for the additional boxes, roughly 88 top-tier GPUs. We need it because we stop renting intelligence: from end-2027 we run our own model and the Sovereign Engine on owned GPU. That is real load — 140-plus client platforms of agents, the engine's continuous learning and dream cycles, embeddings and ingestion, our own-model fine-tuning, and every paid pocket user who reaches past their device into the engine for the heavy, real-time work. On-device carries the free tier for nothing; the cluster carries what the device can't. Sovereignty is only true if we own enough compute to run it — this is that compute.
Estimates: ~88 GPUs draw about 110 kW, the orchestration tier about 25 kW, so ~135 kW of IT load and roughly 190 kW of facility power with cooling, split across Tier-4 colo in Dubai and Asia. This level of compute carries the own-model decouple and the full 2028 volume and pocket ramp — about eighteen months from Round 1 close — before Round 2 funds the next, much larger cluster for mass consumer scale.
Raising opens January 2027 and Round 1 closes by Q2 — April to June, June at the latest. It holds $60M pre with the $12M raise, so the slice is 16.7% — less dilution, same cash. Round 2 follows ~18 months after close, at $100M into a $300M post. Multiples below are on current revenue (~$1.7M) and on forward revenue.
$72M post ≈ 42x current revenue (~$1.7M), ~14x 2027 forward — early-round pricing on the IP and margin story. First money in holds first selection in any secondary, ahead of later investors.
$300M post is about 10x the ideal 2028 cash (~$30M) and under 8x the raise case — a standard growth multiple, earned if the 2027–28 numbers land. The $100M funds the scale ramp.
Both rounds price rich on today's ~$1.7M of revenue, which is normal for a priced hypergrowth round carrying this IP and margin. The real dependency is new clients: Round 2 at $300M needs 2028 cash in the plan's $16M–$30M range — several times today — so the step-up is earned by acquisition between the rounds, not granted. First money in holds first selection in any secondary ahead of later rounds, and at these terms founders stay above 50% through both, which holds the sovereign posture.
Round 1 is not operating money — the plan funds operations from cashflow. The raise is the accelerant: owned compute, a sales and delivery team ahead of revenue, and the pocket launch.
| Use of funds | Amount |
|---|---|
| Infrastructure — owned metal, Dubai + Asia | $4.0M |
| Sales + marketing — team ahead of revenue, AISCO, pocket launch | $5.0M |
| Delivery + working capital | $2.0M |
| Contingency | $1.0M |
| Total (equity raise) | $12.0M |
| Bank server loan (debt, for metal) | $3.0M |
| Owned fleet (Dubai + Asia) | Qty | Unit | Total |
|---|---|---|---|
| NVIDIA DGX frontline AI servers | 11 | ~$450K | $4.95M |
| Orchestration servers | 34 | ~$25K | $0.85M |
| Data centers — racks, networking, power, redundancy, migration | — | — | $1.20M |
| Total metal | $7.00M | ||
| Funded by equity raise | $4.00M | ||
| Funded by bank server loan | $3.00M |
Roughly 88 top-tier GPUs drawing about 190 kW of facility power across the two sites — enough to run our own model, the Sovereign Engine, and the full 2028 volume and pocket ramp, about eighteen months before the next cluster.
A patented autonomous-payments stack the company owns outright, sold outside the operating plan. None of the numbers above depend on it.
Data room Q2 2027, letters of intent Q3, close Q4 2027; 70% to the company, 30% to the founder. Even the low end lands well above the operating business — optionality the plan is built to stand without.
| 2027 floor | 2027 ideal | 2028 floor | 2028 ideal | |
|---|---|---|---|---|
| New build clients | 19 | 32 | 58 | 108 |
| New build contracts | $5.7M | $9.6M | $17.4M | $32.4M |
| Cash collected | $3.9M | $5.4M | $16.2M | $29.8M |
| Cash at 31 Dec | $2.2M | $2.8M | $10.0M | $18.4M |
| Signed into next year | $2.9M | $5.4M | $7.5M | $13.6M |
From the operating plan and the 2027–2028 growth plan. The raise-funded case and the pocket companion are shown in the sections above.
Planning figures from the operating plan and the 2027–2028 growth plan, plus the raise overlay — modeled, not audited. Current revenue is actual.