OSI-PDM-001 · CONCEPT DEMONSTRATION · OPEN SYSTEMS INC.
REV A · NIGHT OPERATIONS VIEW

Hive / Bees / Pollen

Autonomous machines make more AI training data than the site network can carry. So bees fly it out: a sealed 32 TB cartridge, a 50° optical port on every machine, and a container base that stays put while everything around it changes.

How it works

Several bees, one scheduler

One full cycle runs eight phases: buffer, dispatch, transit, acquire, transfer, return, swap, ingest. Below, two bees run those phases against two machines at once, routed by the Hive scheduler. The rail tracks the phase the highlighted bee is in right now. The hardware moves the data; the scheduler decides.

Colony simulation · 2 bees up · 1 spare
SITE 06:00
M-07 M-12 B3 INGEST HIVE STORE 100G B1 B2 AV MACHINE FLEET · 1 TB/HR EACH HIVE SITE UNIT · APIARY ROOF
SCHED initializing colony…
Buffer
The machine works, making about a terabyte an hour. It all piles up in the onboard Pollen bay, because the site network can't move it.
Architecture

Three rules the design never breaks

Drone vendors will change over the life of the site. These three rules never do.

Rule 01

Data never touches the bee

The Pollen has its own storage, battery, optics, and radio. The drone just carries it, like a truck hauling a locked box. It can't open it, and it never sees what's inside.

Rule 02

Any bee that lifts it, works

No vendor software, no special integration. The mount is bolts and a weight limit. Today it hangs from a drone. Tomorrow it can ride a truck, sit on a pole, or bolt straight onto the machine.

Rule 03

The Hive is permanent

The fiber, the computers, and the ingest shelf go in before the first flight and stay after the last one. The docked cartridges are the storage, and empty slots are the backup buffer if the internet line goes down.

Deployment path

The parts change. The Hive doesn't.

Phase 0
Fiber + container.

The container goes in first: internet, computers, and a secure site network. It's useful on day one, before a single drone flies.

HSU deployedBackhaul live
Phase 1
Proof on mature hardware.

Prove the two hard parts separately: automated flying on off-the-shelf drone hardware, and the optical link on a simple mast rig. Then show them working together.

Ops proofLink physics proofSealed module v1
Phase 2
Open airframes, coordinated autonomy.

Move to open drones running open autonomy software. The blossom, the Pollen, and the Hive don't change at all. Only the drones do.

Autonomy layerUniversal dock
Phase 3
The interface becomes the standard.

Eventually the machine maker builds this into their own equipment. When they do, they build to the published Pollen interface, on the same blossoms, into the same Hive.

Published interface spec
The software story

The scheduler is the product on top

The scheduler decides which bee goes where, and when. It watches every machine's buffer, every bee's battery, and every blossom's health, and plans pickups all day, like a delivery route that never stops updating. Drones come and go over the years. The scheduler and its data stay.

Input

Site state

How full each machine's buffer is, where the machines will pause next, whether each lid works, the weather, each bee's battery, and how many empty cartridges are on the shelf.

Decision

Assignment

The fullest machine gets the readiest bee. Each bee flies at its own altitude so they never cross paths. A broken lid gets skipped instead of found mid-flight. Every decision is written to the log.

Invariant

Adapter per airframe

The scheduler speaks one language. Each drone system plugs in through a small adapter, and switching drone vendors means switching the adapter. The decisions and the data stay put.

Inside the Hive

The building the data lands in

The Hive is a shipping container that goes in before the first flight. One end is the equipment room: the carrier handoff, a 100 Gbps switch, the Pollen ingest shelf, an edge computer, and a UPS, all in one 42U rack. The other end is a working site office with four desks. The ingest bay is a slot in the wall, so cartridge swaps never go through the office.

APIARY PADS ON ROOF BEES DELIVER POLLEN HIVE SITE UNIT INGEST BAY POLLEN SLOTS POWER + DEMARC CARRIER HANDOFF · UPS 42U RACK SHELF · SWITCH · EDGE OPS DESK 4 WORKSTATIONS 100 GBPS EPL · PRIVATE LINE GPU TRAINING CLUSTER HOT DATA · MODEL RUNS S3 OBJECT STORAGE EVERYTHING · FOREVER
The demarc

One clear handoff

The 100 Gbps line is a carrier circuit, and every carrier circuit needs a demarcation point: the exact spot where the carrier's gear ends and ours begins. One panel, one test point. When the line has a problem, you know in minutes whose side it is on. It is also where the circuit gets grounded and surge protected. The panel is sized for two circuits from day one: a second line or a backup path is the cheapest upgrade the Hive will ever get, and the space costs nothing now.

The circuit

A private line, not internet

An EPL is a point-to-point line that behaves like one very long Ethernet cable from the Hive to the datacenter. Fixed bandwidth, no sharing, no public internet anywhere in the path. At hundreds of terabytes a day you need guaranteed throughput, and the security story stays one sentence long: machine, sealed cartridge, private line, training cluster. The data never touches the internet.

The far end

Two destinations

The GPU cluster gets the hot data the models train on. S3 storage gets everything, cheap and forever. A cartridge at the Hive is only wiped and reused after the far end confirms the data is safely written, and safely written ultimately means it is in S3. The Hive only ever holds a few days of data. It is a staging point, not a warehouse.

The drawing reads left to right the same way the data moves: bees drop Pollen in the wall bay, the cartridges plug into the rack and become the storage, the rack hands the traffic to the demarc, and the private line carries it to the two far ends. The amber path is the same color as the beam in the animations on purpose. It is the same data, still moving, on glass now instead of air.

Reference numbers

Why the numbers work

50°
Beacon field of view
15 ft
Hover standoff — a 14 ft capture circle
10→100 Gbps
Wireless fiber · module-selectable tier
32 TB
One Pollen SKU — machine, bee, Hive
45/25
Flight / charge minutes · 64% duty cycle
1 TB/hr
Per-machine generation · up to 16 TB/day
6–7 : 1
Busy machines per bee at 40 Gbps
6 pads
Roof ceiling · ~40 machines per Hive
Offload tiers · module-selectable · any head negotiates to the far end's rate
10 Gbps
~60 GB/min · 5 TB ≈ 83 min
Mast + vehicle sessions · entry blossoms
40 Gbps
~240 GB/min · 5 TB ≈ 21 min
Flight bridge · hover sessions
100 Gbps
~600 GB/min · shift ≈ 13 min
Target standard · fewer trips, fewer bees, fewer pads

The 50° cone gives the drone a 7-foot circle to hover in. Modern drones hold position about ten times tighter than that, so hitting it is easy. Each machine makes a terabyte an hour. A bee visits every five or six hours, picks up about 5 TB, and one bee can cover six or seven busy machines. At the Hive, the docked cartridges are the storage itself, and the 100 Gbps line carries everything to the GPU cluster with room to spare. Nothing is erased until the far end confirms. The 100 Gbps head fits in the same cartridge and empties a full shift in 13 minutes.

Bees make roughly 5 TB pickups, each flying at its own altitude so they never cross paths. At full size, six pads and six bees cover about 40 busy machines making around 640 TB a day, and the 100 Gbps line still carries all of it. Extra slots on a machine cover bad weather. Extra slots on the shelf cover an internet outage. Extra pads cover a broken drone. Every backup on site is the same thing: a spare slot.

Fleet calculator

How many bees?

Set your site below. The calculator works out drones, pads, cartridges, Hives, and the yearly cost against doing it by hand. The number people always guess wrong: drones per machine is a fraction.

Your site

Machines on site20
Data per machine1.0 TB/hr
Working hours per day16 hr
Link speed40 Gbps
10G ≈ 60 GB/min · 40G ≈ 240 GB/min · 100G ≈ 600 GB/min
Drone flight time45 min
Charge time25 min
Round trip to a machine6 min
Machine productive value$500/hr
A 10 minute landing reserve is always kept back.
3 + 1
Bees needed + spare
0.15
Drones per machine
6.7
Machines per bee
3/day
Visits per machine
22 min
Time over each machine per visit
Site data per day320 TB
Pads at the Hive4
Pollen cartridges to stock38
100 Gbps uplink load38%
Hives needed1
Fleet checkBalanced
Hand-carrying drives, per year$0
Mailing drives, per year$0
Running the colony, per year$0
Savings with the colony$0
Each machine makes 16 TB a day. At 40 Gbps a bee picks that up in 67 minutes of hovering, split into 3 visits. One bee can fly about 926 minutes a day, so 3 bees cover all 20 machines, plus one spare on the pad. That is 0.15 drones per machine: the drone is shared infrastructure, like a forklift, not a per-machine accessory.
How the math works. Daily data per machine is rate times working hours. Divide by the link speed to get hover minutes needed per machine per day. Each flight has a fixed budget (flight time minus a 10 minute reserve minus the round trip), which sets visits per day. A bee can fly for flight ÷ (flight + charge) of each day. Total hover minutes plus travel, divided by what one bee can fly, gives the fleet. Faster links cut hover minutes, which cuts bees, pads, and batteries. That is why 100 Gbps is the long-term standard: the fleet shrinks as the silicon matures. The savings rows compare against doing this by hand: a tech swap takes about 35 minutes at $90/hr and stops the machine for about 12 minutes, mailing adds about $60 per package of ten drives plus a float of extra drives in transit, and the colony costs about $60k per Hive per year to operate plus an Open Systems managed service plus $10 per pickup. The real difference is the stoppage: a bee never stops the machine.
Labor savings

What manual handling really costs

The alternative to bees is people: a tech drives to each machine, stops it, swaps the drive, and either carries it back or mails it. Set your labor numbers below. The result is usually a surprise, and the surprise is not the wages. It is the stoppages.

Your labor numbers

Machines on site20
Pickups per machine per day2
Tech time per swap35 min
Tech cost, loaded$90/hr
Machine stopped per swap12 min
Machine productive value$500/hr
Shipping per package (if mailed)$60
Drives per package10
Cost per drive$800
Days a mailed drive is out of service6 days
Open Systems managed service per year$120k
Mailing is priced per package, not per drive: drives are batched, so shipments per year is swaps ÷ drives per package. Mailing also skips the $20k a year site truck — the courier does that leg — but it still stops the machine, because someone still pulls the cartridge. Every drive in transit is a drive you had to buy: the float below is the extra inventory the mail loop forces you to own, amortized over a 3 year drive life, plus 0.2% of shipped drives lost or damaged. Also assumed: the Hive at about $60k a year to operate, an Open Systems managed service to run the colony, and about $10 per colony pickup. Hardware up front is about $250k.
$0
Saved per year with the colony
0 mo
Colony hardware pays for itself in
0 hr
Machine hours stopped per year, manual
Swaps per year0
Tech hours per year0
Packages mailed per year0
Extra drives owned for the mail loop0
HAND-CARRYING COSTS$0
Tech labor for swaps$0
Site truck$0
MAILING COSTS ALONE$0
Postage$0
Drive float + losses$0
LOST PRODUCTIVITY$0
Machine stopped for swaps — charged either way0 hr
Total, hand-carrying$0
Total, mailing$0
Total, colony$0
Why the stoppage line dominates. Wages are a person's time. Stoppage is the machine's time, and an autonomous machine exists to never stop. Every manual swap re-inserts the human the autonomy program spent years removing, and stops production while it happens. A bee hovers overhead while the machine keeps working. Labor savings are real, but the honest headline is simpler: a bee never stops the machine.