Meta Knowledge: Energy Systems

Energy density has two dimensions: energy per kilogram (by mass) and energy per liter (by volume). From firewood to coal, to oil and gas, to nuclear fuel, you climb a steadily rising ladder of density. Each rung sharply cuts the weight and volume needed to carry the same energy, unlocking new forms — coal put the steam engine on rails; only the high density of oil made flight possible.

Uranium's energy density is roughly two million times that of coal — a golf-ball-sized lump could, in theory, supply one person's entire lifetime of energy. Batteries are the mirror image: to fly an airliner across an ocean on batteries, the batteries would outweigh the plane itself. For the same energy need, a car can absorb the weight penalty; a plane cannot. That single physical constraint is why the energy transition is wildly easier in some industries than others.

In biology, fat is an animal's high-density fuel (~37 MJ/kg) while glycogen is low-density quick-access — which is exactly why long-migrating birds store fat, not sugar. In information theory it maps to "information density": for a fixed bandwidth, the compression ratio sets how much you can send. In computing it becomes "compute density" — how many operations per watt, per square millimeter — deciding what fits in a phone versus what must live in a data center.

Model deployment has its own density ladder: at a given level of intelligence, parameter count and memory footprint decide whether it runs in the cloud, at the edge, or on-device. Distilling a capability into a smaller model is, in essence, raising its "intelligence density" — pushing what once required a data center down into a phone. Every round of quantization, distillation, or pruning is a step up this ladder, unlocking a new deployment form.

EROI = energy obtained ÷ energy invested to obtain it. Early Middle Eastern gushers had EROIs around 100:1 — invest one barrel, recover a hundred. As we drill deeper into deep water, shale, and tar sands, the ratio slides to 10:1 or worse. What a civilization can actually spend is its "net energy" — total minus what the energy sector consumes on itself. The lower the EROI, the more labor and capital are locked into "spending energy to get energy," leaving thinner margins for education, medicine, art. Studies estimate modern society needs EROI above roughly 7–10:1.

An energy source can be "net positive" and still make a society poorer. A biofuel at 1.3:1 does yield net energy, but to power an entire society on it, nearly all labor would have to go into energy production — like an agrarian society where 95% farm and the remaining 5% can't sustain any civilizational upper layer. Each historical jump in energy surplus (wood → coal → oil) tracked a jump in the share of people freed from the land, and with it the rise of cities, division of labor, and a knowledge class.

In economics this is the energy version of "surplus product" — the ceiling on social complexity is set by net surplus. In biology it maps to net foraging return: if a lion's energy spent per hunt approaches the energy captured, it starves — which in turn shapes a predator's territory size and rhythm. In investing it is simply the real rate of return after costs — however high the gross yield, if costs are higher, scaling up only loses more.

Every tool and workflow has its own EROI: the time it saves minus the time spent building, maintaining, and debugging it. Many automations look "net positive" yet maintenance eats most of the gain, leaving a surplus too thin to justify rollout. An AI workflow with an EROI of 1.2 means most of your energy goes into feeding the workflow itself rather than creating value with it — it produces, and it also drags.

Grid frequency (50/60 Hz) is the real-time gauge of supply and demand: when consumption exceeds generation, frequency drops; when it's the reverse, it rises. Stray too far and equipment trips protectively, cascading into blackouts. Traditional grids use dispatchable coal and hydro to "follow load" — burn a bit more when demand rises. But wind and solar are non-dispatchable: they run on weather, not on demand's commands. The higher the renewable share, the harder it is to match consumption by adjusting generation.

California produced the famous "duck curve": at midday, solar floods the system and wholesale prices go negative — generators pay people to consume; then as the sun sets and the evening peak arrives, the grid must ramp up enormous backup in just a few hours. The counterintuitive result: the more renewables you install, the more flexible fossil plants you need on standby, and total emissions often fall by less than capacity growth suggests. Germany's curtailed wind, negative prices, and reliance on neighbors' grids are all faces of the same balancing problem.

This is the "real-time consistency" problem in distributed systems — a system without buffers (inventory) must stay strongly synchronized, and any delay propagates straight into failure. In biology it maps to homeostatic control of blood sugar and temperature: the body too is a "grid" that must balance energy production and use in real time, with insulin as its "frequency regulator." In supply chains it is zero-inventory just-in-time — most efficient, least robust to shocks.

A high-concurrency service is itself a grid: requests (load) fluctuate in real time, processing capacity (generation) must follow, and without enough buffer the "frequency" collapses — latency spikes, then avalanche. Autoscaling is "dispatching generators"; rate limiting, graceful degradation, and queues are the buffer bolted onto an inventory-less grid. The intermittency of renewables mirrors the burstiness of traffic: you can't control when it arrives, only absorb the swings on the system side.

Storage must satisfy capacity, power, duration, cost, and lifespan all at once — and these conflict. Lithium batteries suit the "few hours" scale, but storing a whole winter's electricity is made hopeless by their cost and self-discharge; pumped hydro is limited by terrain; hydrogen loses more than half its energy on the round trip. No single technology spans the full range from seconds to seasons — a dual constraint of physics and economics.

The hidden advantage of fossil fuels is precisely that they are storage: a barrel of oil is tens of millions of years of sunlight compressed into a tank that keeps for decades with almost no loss. We never needed a separate "storage" industry because the fuel was the inventory. Switching to wind and solar swaps a system that "comes with a warehouse" for one with none — generation is solved, and only then does the warehouse problem surface. This is why "renewables are cheaper than fossils" headlines so often mislead: the cost of generating power and the cost of round-the-clock reliable supply are two different things.

In computer architecture this is the "memory hierarchy" — registers, RAM, disk — where no single medium is fast, large, and cheap at once, so you approximate with tiers. In biology it maps to layered energy storage: ATP for the instant, glycogen for hours, fat for the long haul — the body uses different media for different time scales. In finance it is liquidity management: allocating across checking, term deposits, and long-term assets is, at bottom, stocking the right inventory for each time scale of spending.

Cache design is a storage gamble: in-memory cache is fast but costly and volatile, while local disk and object storage are slow but cheap and durable — different time scales and costs. Since no single tier wins, you tier your caches. Inference KV cache, vector stores, and hot/cold data tiering are all the same problem: store an "expensive-to-compute-once" result for reuse across time, forever trading off how long, how fast, and how dear to store it.