OpenIE · The Stack · The Opportunity

The efficient path was always there.

Every workload sits in a cell. Most workloads are running in a cell that costs orders of magnitude more energy than the physics floor of the same answer. The gap between current cost and the floor is the opportunity — visible, measurable, choosable.

Where the joules go now Back to the overview

I · Where the joules go now

The gap, by orders of magnitude.

General-purpose silicon runs at roughly 1010 above the Landauer floor. That is the figure that explains data-center thermal envelopes, GPU power draw, the cooling bill.

Generative inference per output bit runs at 1018 to 1020. The extra eight to ten orders are model-bloat: forward-passing every parameter to produce one token, then doing it again for the next, then again for the next.

The Landauer floor itself is a thermodynamic minimum, not an engineering target. Hitting it exactly requires reversible logic and zero impedance. Approaching it within a few orders is realistic with disciplined hardware.

FIG. O1 · THE GAP Three substrates, twenty orders of magnitude apart. 10⁰ 10² 10⁴ 10⁶ 10⁸ 10¹⁰ 10¹² 10¹⁴ 10¹⁶ 10¹⁸ 10²⁰ Landauer floor k_B T · ln 2 ≈ 2.87 × 10⁻²¹ J/bit General-purpose silicon ~10¹⁰ × Landauer Generative inference, per output bit ~10¹⁸–10²⁰ × Landauer the gap = the opportunity openie · stack · opportunity
Fig O1Distance from the floor on a log scale. Most workloads can move down.

II · Route before compute

The most efficient computation is the one that doesn’t run.

Every query has a cell on the cost surface. Closed-form answers belong in Z1 wells — arithmetic, parsing, lookup, deterministic control. Pattern-bound answers belong in Z2 basins — retrieval, classification, small-model inference. Open-ended synthesis belongs in Z3 plateau — generative work without named grounding.

Routing decides which cell handles each query. The decision itself is a lookup — itself a Z1 well. The router is the table, not the model.

Most everyday queries have closed-form answers. Routing them to the well-shaped cell saves orders of magnitude on energy per query. The savings are not from compressing the model. They are from not running it on a workload that doesn’t need it.

FIG. O2 · THE ROUTING Three lanes. Most traffic belongs in the well. INPUT QUERIES Z₁ · WELL closed-form arithmetic · parsing · lookup deterministic control ~10⁰–10¹⁰ ABOVE FLOOR lowest-cost cell Z₂ · BASIN pattern-bound retrieval · classification small-model inference ~10¹²–10¹⁵ ABOVE FLOOR Z₃ · PLATEAU open-ended synthesis generative no grounding 10¹⁸+ ROUTING ITSELF IS A Z₁ WELL — THE ROUTER IS THE TABLE openie · stack · opportunity
Fig O2Most queries have closed-form answers. Route them there and the bill collapses.

III · The ladder of savings

Three stacked levers.

Lever 1 — Route to the right cell. For workloads that didn’t need the plateau, the savings are 108 to 1010 — the entire gap between Z3 and Z1 on the cost surface.

Lever 2 — Cache the lookup. Repeat work pays the substrate cost once and the retrieval cost on every subsequent call. Savings: 102 to 104 on workloads with significant repetition. The memory cascade names the eight retrieval tiers; tier L0 pays effectively nothing.

Lever 3 — Match silicon to the cell. General-purpose silicon is 1010 above Landauer. Cell-aligned silicon — designed for the specific primitive composition the cell uses — can recover several orders of that gap. Reversible logic, in-memory compute, and adiabatic switching all live in this lever.

Stacked, the three levers approach the physics floor. The ceiling is not the Landauer bound. The ceiling is how disciplined the routing and the hardware choices are.

FIG. O3 · THE LADDER Three levers, stacked. START ~10¹⁸ above floor LEVER 1 route ÷ 10⁸–10¹⁰ AFTER ROUTE ~10¹⁰ above floor LEVER 2 cache ÷ 10²–10⁴ AFTER CACHE ~10⁶–10⁸ LEVER 3 · CELL-ALIGNED SILICON · ÷ 10²–10⁴ ~10²–10⁴ above floor stacked total: 10¹²–10¹⁸× openie · stack · opportunity
Fig O3Three levers multiply. The right workload, routed and cached on cell-aligned hardware, approaches the floor.

IV · What it looks like in practice

Already visible on the map.

The pieces of the efficient path are not theoretical. They show up across the family of stacks, each one labeled and measurable.

Lever 1 · routing

Same task, four cells

Extract invoice numbers from 500 PDFs. The deterministic-output cells flatten at $0.02 and $37.50 forever. The stochastic-output cells climb linearly per batch. Same task on the same substrate; the bill is what tells the cells apart.

agent.openie.dev/cost →

Lever 2 · caching

The memory cascade

Eight retrieval tiers from L0 register through L7 generative reconstruction. A query enters at L0; only what each tier cannot answer falls to the next. The lowest-cost answer is the one already on the shelf.

memory.openie.dev/instrument →

Lever 3 · silicon

The cost function

E(x) = Σ θ(p) · μ(p, H). Change the primitive composition or change the hardware impedance, and per-input energy moves. The function names both levers in one expression.

synthesis.openie.dev →

The floor

The permitted-act set

258 primitives in 33 families. Every workload composes out of this set. Knowing which primitives a cell uses is how the joule cost gets calculated, not estimated.

compute.openie.dev →

The eight family stacks together carry the working instruments. Read across them and the opportunity stops being abstract.

Closing

The map is here. The cells are labeled.

The efficient path is a choice. The eight stacks make the cells visible. The cost function makes the savings calculable. Routing makes the savings available on every query the system sees.

None of the three levers requires new physics. Each one is well-studied in its own discipline. What is available now is the assembly — reading them together as a single design surface.

A teaching artifact. Read it where it’s useful.

Back to the overview