The Physical Stack

Infrastructure megatrends

The Physical Stack

Five bets. One story. The atoms being poured to power an AI-and-electrification century.

Software scaled on someone else's servers. The next century scales on steel, silicon, concrete, copper, and orbit. These are not separate industries — they are one physical stack.

Camera path
Scroll to descend

Start in orbit. Descend through campus power. Enter the reactor core. Assemble the cell. Land on the street.

01 · Orbitscroll = depth

LEO Constellations

Starlink-type low-Earth-orbit mesh

7,000+

active Starlink sats in LEO — and climbing

The network is no longer a tower on a hill. It is a moving shell of routers 550 km above you, replenished every few years.

Depth strataL1/4
  1. Shell geometryL1

    Thousands of co-orbiting nodes as a statistical surface

    Altitude bands stack like nested free-fall highways

    Coverage is density × inclination, not a single bird overhead

  2. Bus & powerL2

    Flat-panel mass, solar wings, thruster budget

  3. Aperture physicsL3

    Phased-array elements forming steerable beams

  4. Carrier → symbolL4

    RF/optical energy becoming packet symbols

Mechanism beats (4)
  1. 1
    Shells, not birds

    Thousands of satellites occupy altitude bands (shells). Coverage is a statistical property of density + inclination, not a single bird overhead.

  2. 2
    Phased-array steers beams

    Ground terminals electronically steer without moving parts. Beams hand off as satellites race across the sky at ~27,000 km/h.

  3. 3
    Laser inter-sat links

    Optical crosslinks turn the constellation into a spaceborne backbone — traffic can hop sat-to-sat without touching a ground gateway.

  4. 4
    Launch cadence is the product

    Reusability collapsed cost-to-orbit. The competitive moat is replenishment rate: how fast you can fill and replace a shell.

Scale

A full mega-constellation is a city of computers in free-fall

vs. a few hundred GEO sats that covered the 20th century

Stakes

Spectrum, debris, and gateways

Whoever owns dense LEO capacity owns a second internet layer for AI edge, defense, and regions fiber never reached.

Deep dive this layer
spatial stageOrbit → micro
loading depth
SHELL 01 · ~340 km
180sats
Depth · Constellation

Coverage is a property of shell density, not a single bird overhead.

Thousands of free-fall routers form altitude bands around Earth

10 Mm
10³–10⁴ km shell diameter
scroll 0%
macroorders of magnitudemicro
  1. Constellation
  2. Orbital shell
  3. Satellite bus
  4. Phased array
  5. RF / optical link
  6. Bit stream
02 · Campusscroll = depth

Hyperscale Data Centers

AI training & inference campuses

100+ MW

per campus — some AI builds push toward gigawatt scale

A modern AI campus is not a building with servers. It is a power plant that happens to compute — electricity in, tokens out, heat rejected at industrial scale.

Depth strataL1/4
  1. Campus as plantL1

    Substations, halls, cooling yards as one thermal machine

    Megawatts enter at the fence; heat leaves through chillers

    Interconnection queues now gate AI capacity more than GPUs

  2. Aisle thermodynamicsL2

    Hot/cold corridors, liquid loops, rack as a furnace

  3. Package & fabricL3

    HBM stacks bonded to accelerator silicon

  4. Gate → tokenL4

    Nanometer switches flipping into model activations

Mechanism beats (4)
  1. 1
    Racks as thermal machines

    GPU racks draw 40–120+ kW each. The design problem is power delivery + heat extraction before it is FLOPS.

  2. 2
    Liquid cooling takes over

    Air hits a wall. Direct-to-chip liquid and rear-door heat exchangers move heat into facility water loops, then to chillers or free cooling.

  3. 3
    PUE is the scoreboard

    Power Usage Effectiveness tracks overhead. Best campuses approach ~1.1; every tenth of a point is megawatts of non-compute waste.

  4. 4
    Interconnect is the new chassis

    Training clusters live or die on NVLink / InfiniBand / Ethernet fabrics. The network is as capital-intensive as the GPUs.

Scale

One hyperscale campus can draw like a small city

≈ tens of thousands of homes continuous load

Stakes

Grid interconnection queues

Whoever secures power + land + fiber first locks in a decade of training and inference advantage.

Deep dive this layer
spatial stageCampus → micro
loading depth
Depth · Campus

A hyperscale site is a power plant that happens to compute.

Halls, substations, and cooling yards as one thermal machine

5.0e+2 m
10²–10³ m footprint
scroll 0%
macroorders of magnitudemicro
  1. Campus
  2. Server hall
  3. GPU rack
  4. Accelerator
  5. Die / transistor
  6. Bit / token
03 · Corescroll = depth

Nuclear / SMR Reactors

Firm carbon-free baseload for the AI grid

24/7

firm power — the attribute intermittent renewables cannot fake alone

AI and electrification need always-on megawatts. Small modular reactors aim to industrialize nuclear the way factories industrialized everything else: repeatable units, not one-off cathedrals.

Depth strataL1/4
  1. Civil boundaryL1

    Pad, containment, defense-in-depth shells

    SMR shrinks the pad; safety language stays multi-barrier

    Licensing and site are as hard as reactor physics

  2. Vessel & latticeL2

    Pressure boundary around fuel geometry

  3. Pellet chemistryL3

    Ceramic UO₂ energy density in centimeters

  4. Neutron economyL4

    Neutrons moderate, absorb, or induce the next fission

Mechanism beats (4)
  1. 1
    Fission → heat → steam → electrons

    Neutrons split heavy nuclei; heat raises steam (or drives advanced coolants); turbines spin generators. The physics has not changed — the packaging has.

  2. 2
    Containment layers

    Fuel cladding, reactor vessel, containment building. Defense-in-depth is the design language: multiple independent barriers to release.

  3. 3
    SMR = factory modules

    50–300 MWe units built in factories, shipped, and stacked. The bet is learning curves and parallel construction vs. decade-long mega-projects.

  4. 4
    Load-follow & co-location

    New designs target co-location with data centers and industrial heat loads — nuclear as a private baseload peer, not only a utility asset.

Scale

One SMR footprint can be a fraction of a gigawatt plant

same firm output, modularized site plan

Stakes

Licensing, fuel, and first-of-a-kind cost

If SMRs clear the first commercial hump, they become the only scalable firm zero-carbon option that matches AI load shapes.

Deep dive this layer
spatial stageCore → micro
loading depth
PAD · SMR vs GW FOOTPRINT
Depth · Plant / SMR pad

SMRs shrink the pad; physics of fission does not shrink.

Containment sits on a civil pad sized for safety and heat rejection

3.0e+2 m
10²–10³ m civil works
scroll 0%
macroorders of magnitudemicro
  1. Plant / SMR pad
  2. Containment
  3. Pressure vessel
  4. Fuel assembly
  5. Fuel pellet
  6. Fission event
  7. Neutron economy
04 · Cellscroll = depth

Battery Gigafactories

Cells → modules → packs → grid storage

GWh

annual cell output per plant — the unit that replaced 'factory'

Electrification is a materials-and-manufacturing problem wearing a product costume. The gigafactory is the industrial form factor that made lithium-ion a commodity curve.

Depth strataL1/4
  1. Factory throughputL1

    Coating lines, formation, GWh/year as the unit

    Dry rooms and calendars are the real bottlenecks

    Electrification rate-limits on process yield, not slogans

  2. Pack hierarchyL2

    Cell → module → pack structure + BMS + cooling

  3. Electrode stackL3

    Anode | separator | cathode at micron coatings

  4. Ion & latticeL4

    Li⁺ hops through electrolyte into crystal host sites

Mechanism beats (4)
  1. 1
    Anode · separator · cathode

    A cell is a controlled electrochemical sandwich. Lithium ions shuttle between electrodes through a separator soaked in electrolyte.

  2. 2
    Jelly-roll / stack → cell

    Electrodes are wound or stacked, sealed, filled, formed. Formation cycling is slow capital — time in chamber, not just line speed.

  3. 3
    Cell → module → pack

    Mechanical structure, thermal management, and BMS turn chemistry into a product that survives vibration, crash, and abuse.

  4. 4
    Learning curves compound

    Every doubling of cumulative production historically crushed $/kWh. Process yield and cathode chemistry dominate the next leg.

Scale

A gigafactory is a city block of dry rooms and precision coating

output measured in GWh/year, not units/hour alone

Stakes

Critical minerals & geographic concentration

Whoever owns cell capacity + mineral offtake owns the rate limit on EVs, grid storage, and mobile robotics.

Deep dive this layer
spatial stageCell → micro
loading depth
Depth · Gigafactory

Output is measured in GWh/year — the industrial unit of electrification.

Coating, winding, formation as continuous process cities

5.0e+2 m
10²–10³ m dry rooms & lines
scroll 0%
macroorders of magnitudemicro
  1. Gigafactory
  2. Pack
  3. Module
  4. Cell
  5. Electrode stack
  6. Li⁺ shuttle
  7. Crystal host
05 · Streetscroll = depth

Autonomous Vehicles

Sensor stacks, maps, and edge cases

360°

multi-modal sensing — the car as a mobile robot

Autonomy is not 'better cruise control.' It is a robotics stack — perception, prediction, planning, control — deployed on public roads where the long tail is the product.

Depth strataL1/4
  1. Operational domainL1

    Street scene with agents, weather, rules

    The robot shares space with non-cooperative humans

    The long tail of edge cases is the actual product

  2. Sensor geometryL2

    Lidar, cameras, radar as fused viewpoints

  3. Sample → trackL3

    Rays, pixels, points becoming object tracks

  4. Intent → controlL4

    Latent futures collapsing into actuator setpoints

Mechanism beats (4)
  1. 1
    Sensor suite

    Cameras for semantics, radar for velocity through weather, lidar for dense geometry. Fusion is the hard part — not any single modality.

  2. 2
    What the car sees

    Point clouds and segmentation frames become object tracks. The model predicts where agents go next, not just what they are now.

  3. 3
    Planner under constraints

    A decision stack chooses trajectories under traffic rules, comfort, and risk bounds. Edge cases are where policy meets physics.

  4. 4
    Fleet learning loop

    Miles driven feed rare-event mining. The moat is data + simulation + validation infrastructure as much as the on-car model.

Scale

Robotaxi fleets turn cities into continuous validation labs

billions of simulated miles + millions of real ones

Stakes

Liability, regulation, and the long tail

Autonomy is where the physical stack meets human streets — the last mile of the AI century is still asphalt.

Deep dive this layer
spatial stageStreet → micro
loading depth
Depth · Street scene

Autonomy is robotics on public roads — the long tail is the product.

Vehicle is one agent among humans, signals, weather

80 m
10¹–10² m operational design domain
scroll 0%
macroorders of magnitudemicro
  1. Street scene
  2. Vehicle body
  3. Sensor suite
  4. Ray / return
  5. Pixel / point
  6. Feature / intent
  7. Control bit

Connective tissue

One stack. Five surfaces.

Orbit feeds bandwidth. Campuses convert power into intelligence. Reactors and cells keep the electrons firm and portable. Vehicles move the century on public streets. The product is the whole chain — not any single module.

Built by Transient Labs · Agentic Systems and Product Studio

The Physical Stack · stack.transientlabs.ai