Adaptation Curve

Adaptation Curve · operator’s view from the curve

The Bandwidth Ladder — and where orbital breaks it

Four figures tracing the communication hierarchy. Every figure traces to AI Compute Bandwidth Ladder.xlsx. Current = B200 / NVLink 5 / HBM3e / 800G (mid-2026); Forward = Rubin / NVLink 6 / HBM4 / NVL576 (2026–27). Not investment advice.

FIG 1The bandwidth ladder

Per GPU or per link, log scale. Each rung down is ~an order of magnitude slower than the one above. The orbital rung sits below terrestrial cross-rack — the whole thesis in one row.

1 GB/s10 GB/s100 GB/s1 TB/s10 TB/sOn-die20 TB/sIntra-package /chiplet (die-to-die)10 TB/sOn-package memory(HBM)7.7 TB/sScale-up — in-rack(NVLink)1.8 TB/sScale-out — acrossracks100 GB/sOrbital scale-out —sat-to-sat laser(ISL)12.5 GB/s↓ ~18× cliff↓ ~8× cliffTop-to-bottom: ~1,600× — and the shape is invariant across generations
Current (B200-gen)Forward (Rubin-gen) reachThe two cliffs that matter
~1,600× drop from on-die SRAM to a satellite laser link. Notice the ratios barely move from Current to Forward — faster silicon moves the whole staircase up; it doesn't flatten the cliffs.
FIG 2The scissors: model size vs. the rack

Frontier total parameters (left, log) against the NVLink scale-up domain (right) — the 8 → 72 → 576 staircase. Models got bigger because the coherent fast domain got bigger.

20202021202220232024202520262027100B300B1T2TTotal params872576GPUs / scale-up domainGPT-3PaLMGPT-4*DeepSeek-V3Kimi K2NVL8NVL72NVL576* Undisclosed param counts shown as estimates (params_disclosed=false in DB)
The 8→72 step (GB200 NVL72, 2025) is a 9× expansion of the coherent domain — and it lands the same year ultra-sparse trillion-parameter MoE models became routine. Cause and effect, not coincidence.
FIG 3The orbital cliff — terrestrial vs. orbital

Two ladders, identical down to the scale-up rung. Terrestrial bottoms out at ~100 GB/s cross-rack; orbital drops a further rung to ~3–25 GB/s sat-to-sat laser. On Earth a model spilling the rack drops once; in orbit it drops twice.

TerrestrialOn-die · 20 TB/sDie-to-die · 10 TB/sHBM · 7.7 TB/sScale-up · 1.8 TB/sCross-rack · 100 GB/s← bottom rungOrbitalOn-die · 20 TB/sDie-to-die · 10 TB/sHBM · 7.7 TB/sScale-up · 1.8 TB/sSat-to-sat laser · 12.5 GB/s← bottom rungIdentical to the scale-up rung — then orbital drops a further ~8×
The inversion: on Earth the constraint has migrated down toward power. In orbit power is abundant (continuous solar) — and networking becomes the wall. Orbital compute makes the bottom rung the whole story.
FIG 4Why orbital training is gated — the causal chain

It isn't one bottleneck; it's a chain. Keeping it honest: physical limits on the satellite cap the fast domain, which forces the model across the slow link.

Launch · mass · power· radiator limitsSmall per-satellitescale-up domain (onerack)Frontier model exceedsone domainInter-satellite laserlink (ISL) bindsPower is the easy part in orbit. The networking tier that's nearly free on Earth becomes the binding constraint.
The steelman: a tight ~1 km formation with DWDM optics (Google's Project Suncatcher, ~1.6 Tbps bench) rebuilds the DC fabric in orbit — turning the bandwidth problem into a formation-flying problem. The constraint changes disciplines; it doesn't vanish.
Sources: Dwarkesh × Dylan Patel (communication ladder, HBM-vs-DDR shoreline math, 20× Hopper→Blackwell); NVIDIA GB200 NVL72 / Vera Rubin NVL144/576 disclosures; Starcloud & SpaceX laser-ISL announcements (May 2026); Google Project Suncatcher (arXiv 2604.07760). All figures dated in AI Compute Bandwidth Ladder.xlsx — verify at publication.