A Carnot Bank
05 May 2026Take a seventy-lakh housing loan at 9% over twenty years. You repay roughly one and a half crore — your seventy lakh, plus eighty more in interest. Stretch it to thirty years and you cross two. The bank gives the depositor whose money you’re using two percent simple interest. The spread is the institution. The spread is also where the conservation law breaks.
Carnot’s engine is a thought experiment. No real engine reaches its efficiency. It still organizes every conversation about heat and work because it tells you which losses are structural and which are choices. I want to draw the same picture for banking — not as a model anyone will build, but as a diagnostic. And the diagnostic, applied honestly, says something harder than I expected when I started drawing.
The first law before the second
Every real engine obeys the first law before it obeys the second. Put one joule in, you get at most one joule out. Carnot, Otto, Diesel, Stirling, the idealizations and the messy real cousins — none of them let you manufacture energy inside the cycle. This isn’t an efficiency claim. It’s the floor under the entire edifice. Break it and you’re not doing thermodynamics anymore. You’re doing magic.
A bank that breaks even is dead. A bank that takes one rupee and pays out one rupee has no reason to exist, no shareholders to pay, no bonus pool to distribute, no buildings to lease, no Circle Offices to staff. So banks don’t break even. They take one rupee and pay out one rupee plus interest, and the interest came from the borrower, whose interest came from their margin, whose margin came from their customers, and you trace the chain and at some point you arrive at a journal entry where the spread was written into existence. The ledger is where conservation gets suspended.
This isn’t fraud in the criminal sense. It’s the institution functioning exactly as designed. Specific journals, specific ledgers, specific incantations by the correctly credentialed wizards — chartered accountants, statutory auditors, RBI inspectors — and the suspension is blessed. Basel III is the catechism for the blessing. Tier capital, risk-weighted assets, capital adequacy ratios — the whole apparatus exists to formalize the conditions under which the conservation law may be ritually broken.
Every other engine in the universe is held to the first law. Banks aren’t, because banks aren’t engines. They’re rituals that look like engines. The Carnot frame doesn’t fail when applied to finance. Finance fails the Carnot frame. That’s the frame doing what frames are for — telling you what kind of object you’re actually looking at.
What an actual cycle would look like
If you wanted to build something that did obey the first law — a bank-shaped object that took in money and paid out money without conjuring spread from incantation — the architecture is constrained.
Two account types on the deposit side: savings, and term (TD or RD). Two on the loan side: overdraft, and term loan. In the cycle:
- Savings: no interest. It’s a wallet, not an investment.
- Term loans: no interest. The bank funded the asset; you’re repaying.
- Overdrafts: interest, but only on the availed limit. You drew working capital, you serve the cost.
- Term deposits: interest paid to the depositor.
OD interest funds RD/FD interest. Energy in equals energy out, minus operating losses (staff, premises, the genuine cost of running the engine). The book is smaller than what we run today — no corporate lending, no merchant-banker games, retail, agri, and MSE only. Corporates want to pay each other in interest, fine, that’s not a utility function. Send them to Mumbai.
Note what this isn’t. It isn’t more efficient than current banking. It’s less profitable, by design, because profit in the current system is the measure of how successfully the conservation law was broken. A first-law-obedient bank can’t grow faster than its operating surplus. That’s a feature.
The reservoirs are the wrong temperature
A heat engine’s efficiency depends on the temperature gap between reservoirs. The bigger the gap, the more work you can extract.
For a bank, the analogous gradient is trust between depositor and borrower. The branch knows the borrower. The branch verifies the business — not by reading a freshly-minted CA’s PowerPoint where every line item grows 20% YoY, but by walking into the godown, looking at the inventory, watching the customer across three monsoons.
Modern banking has flattened the gradient. Officers rotate every three years explicitly to prevent the relationships that would make verification cheap. The official reason is anti-collusion. The actual function is anti-continuity. Without continuity every loan is a stranger lending to a stranger, and the only thing that closes the gap is paperwork — the working substance bleeding heat.
You can’t run a heat engine between two reservoirs at the same temperature. Anonymized banking is isothermal. That’s why the staff has to bleed to make the books close — Canara Bank, one of India’s largest public sector banks, runs roughly ten thousand branches with sixty-three thousand officers, six per branch, a ratio that holds across the sector. The only reason any of it appears profitable is that the staff is being burned as fuel and the conservation law is being suspended at the ledger. Take away either and the engine stops.
Some inefficiency is the work
Carnot says some irreversibility is structural — no real engine reaches the bound. The sharper claim, when you extend the engine to knowledge work, is that a minimum irreversibility is what makes the engine a knowledge engine at all. Push below that floor and you haven’t built a more efficient bank. You’ve built a different machine that happens to share the name.
Toyota Production System works because a bolt is a bolt. Variance is bounded, failure modes enumerate, the feedback loop is short. Just-in-time collapses inventory because inventory is pure waste when the next step is deterministic. Knowledge work isn’t deterministic. The slack — half-formed ideas, redundant reading, the third draft thrown away, the colleague bounced off who didn’t end up mattering, the file sat with overnight — is the production process. You can’t JIT it because you don’t know what you’re producing until you’ve produced it.
This is the deeper cut against Circle Offices. CO is lean-agile logic misapplied: every layer justifies its throughput, every officer hits metrics, every report feeds the next report. Fine on a factory floor. In a bank where the actual product is judgment about a borrower’s business across cycles, it strips out exactly the slack that makes judgment possible. The branch manager who has time to walk the godown twice, chat with the customer’s accountant, sit with the file overnight — that’s not inefficiency. That’s the production process. Eliminate it as waste and NPAs rise, which the system reads as evidence that more reports are needed, which strips more slack, which raises NPAs further. The factory metaphor eats itself.
Knuth said premature optimization is the root of all evil. Same lemma, different domain. You can’t optimize what you don’t yet understand, and most knowledge work is the not-yet-understanding phase.
The named irreversibilities
So there are real losses, and there are losses the institution has chosen to incur while pretending they’re physics.
The 90-day NPA rule. A loan goes bad on day ninety. Not because anything happens on day ninety. Because the RBI prudential norm says so. Day eighty-nine the borrower’s truck is broken; day ninety-one the borrower is a defaulter. The branch manager knows the truck. The rule doesn’t care. Continuous risk gets binarized at an arbitrary cutoff because the system can’t trust the manager to make the call.
Centralized restructuring. A genuine extension request goes to HO. Two months later HO says yes or no, by which time the business has either died or recovered without help. The local actor with information can’t act; the remote actor without information must.
Circle Offices. A fourth layer of hell sandwiched between HO, RO, and branch. The CO produces no loans, services no customers, visits no godowns. It exists to translate branch reports into RO reports and RO reports into HO reports. A region of the engine where work is done but no heat is moved.
CA-conjured projections. The MSE applicant brings a balance sheet certified by someone who was a chartered accountant nine months ago and who has projected 20% revenue growth into perpetuity. The bank prices risk off this fiction. Measurement noise injected into the cycle.
The rotation regime. Every other loss assumes the gradient exists; rotation removes the gradient itself.
What the diagnostic says
Carnot’s bank will never be built. Not because the engineering is impossible but because the institution it would replace doesn’t want to be replaced, and the institution it would replace is the one writing the rules.
That’s fine. Carnot wasn’t a target either. The Carnot cycle was a diagnostic, and so is this. It tells us three things, in escalating order of accusation.
First: staff exhaustion isn’t a labor problem, it’s a thermodynamic one. The reservoirs have been flattened, so the working substance is being burned to produce the work the gradient should have produced for free.
Second: the institution has misclassified its own losses. The 90-day rule, centralized restructuring, Circle Offices, the rotation regime — these are choices dressed up as physics. The slack they strip out isn’t waste. It’s the production process. Lean-agile logic eats the engine it was meant to optimize.
Third, and this is the one that matters: the system is profitable because it suspends the first law of thermodynamics in its ledgers and calls the suspension prudent regulation. Every other engine in the universe takes one joule in and gives at most one joule out. Banks take one rupee in and give back one rupee plus interest, and the interest is conjured by correctly credentialed wizards reciting Basel III over the right journal entries. The conservation law is for everyone else. Banking is for the wizards.
The gap between the Carnot bound and the reality isn’t the institution’s inefficiency. It’s the institution’s claim to be exempt from the laws that govern every other engine humans have ever built. A Carnot bank is what you’d construct if you took that exemption away.
I don’t know whether anyone will. But I know what the frame says, and the frame doesn’t lie.
Co-drafted with Claude (Anthropic). The polemic is mine; the prodding was his.