Integrated · the network
Transmitting a bit costs more than computing one. So the network is built around the inference, not the other way around. Every tile is a node and a relay. The mesh and the bus are the same graph.
Orbital mesh · 24 nodes · inter-satellite packets · one gateway downlink
Live three.js animation
I · The forcing function
The physics is the same: compute, comms, and storage are three names for moving information against a thermodynamic gradient. In vacuum, the joule-per-bit and joule-per-op are not separately optimizable.
Free-space loss, antenna gain, channel capacity, and the Shannon limit set the joule price of every delivered bit. The link budget is not a knob — it is a wall. The architecture that closes around it is the one that ships.
On a small spacecraft, transmitting a kilobit can cost more joules than a million compute operations. The single biggest energy lever on an orbital data center is: never transmit what you can avoid.
Run the inference at the tile that captured the data. Send the classification, the coordinate, the flag — not the megapixel image, not the raw radar return. The downlink carries the answer, not the question.
When every tile is both compute and relay, the inter-satellite mesh is the system bus. There is no separate "comms layer" carrying packets between a separate "compute layer." Both run on the same physical graph.
II · The link-budget price tag
A modern LEO crosslink at 1 Gbps with reasonable margins lands somewhere around 10⁻¹⁰ – 10⁻⁹ J/bit at the transmitter. The same compute substrate, performing a useful integer addition, runs around 10⁻¹⁶ – 10⁻¹⁴ J/op.
That is a four-to-six order-of-magnitude difference. The arithmetic decides where the architecture lives: the network is built around the inference, not around the bandwidth pipe.
III · Worked examples
Pain
A 10 m/pixel SAR satellite generates ~3 TB/day of raw data. A direct-downlink architecture pays for transmitting all of it, with most bits never producing useful signal.
Integrated answer
Ship-detection, deforestation flagging, and anomaly clustering run at the tile. The downlink carries the answer — roughly 10 MB/day of structured output instead of 3 TB. Three to five orders of magnitude less radio energy per useful bit.
comms energy collapse
Pain
A large model partitioned across a constellation needs every tile to share intermediate activations with neighbors. Off-mesh, each cross-tile message is a multi-hop downlink-uplink relay through a ground station — orders of magnitude more J/bit than the compute itself.
Integrated answer
The mesh routes between tiles directly. The activation tensor moves over the inter-satellite optical link, not the radio downlink. Compute and the inter-process bus are on the same envelope.
no compute/bus split
Pain
Disseminating a model update or firmware version across a constellation, classical store-and-forward, costs N × downlink-uplink-cycles to reach all N tiles.
Integrated answer
The update lands once on the gateway tile. The mesh propagates it as gossip across the farm — every tile is a relay. Network-scale broadcast cost approaches zero per recipient.
gossip in O(log N)
IV · Three tiers, one envelope
A classical orbital architecture has three subsystems: a payload that records, a payload that processes, and a payload that downlinks. Each one is a separately-engineered box with its own thermal, power, and mass budget. The boundaries between them are Cartesian, not physical.
Under the integrated envelope, the three become one: storage is unified persistent memory at the tile, compute is in-memory primitive dispatch at the tile, and comms is the optical mesh routing from the tile's edge. Three subsystems collapse into one die, one radiator, one address space, one fault domain.
The savings are not marginal. Removing the inter-subsystem interfaces removes their thermal interfaces, their mechanical mass, their power-conversion losses, and the engineering hours that maintain them.
Continue