shekhar mudarapu · shakesnotbot · 2026

Your LLM stack is an operating system with no scheduler tokens fell 90% and spend doubled anyway

Cheaper tokens didn't shrink the bill — they grew it, because cheap makes you use far more (Jevons).

The fix isn't a better model; it's the one OS component everyone skipped: a scheduler that decides what a request costs before it runs. The essay.

01 · the paradox

Why cheaper tokens cost more

Economists have a name for it: Jevons Paradox — make a resource cheaper and you use so much more that total spend rises.

Coal did it in the 1800s; tokens do it now.

Two things drive the inversion: an agent burns 5–25× the tokens of a chat turn, and the frontier models agents need still cost ~100× commodity ones.

spend ↑ price/token ↓

"Falling prices are what's driving spend up, not down." Gartner already expects 40% of agentic projects to be cancelled by 2027 on cost overruns.

02 · the missing part

Karpathy's mapping drops the one thing that makes it an OS

The popular picture maps the LLM to a computer. It's a good picture — and it's missing the scheduler, the part that decides who gets the machine when everything wants it at once.

normal OS LLM stack today apps want work scheduler admits · orders · limits CPU · RAM · disk agents want tokens missing layer forecast · admit · route model · context · tools

An operating system is not just a pile of resources. It is the layer that decides what runs, where it runs, and when it must stop.

weights= CPU
context window= RAM
retrieval= file system
tools / MCP= system calls
agents= long-running apps
scheduler= missing
the resource it should schedule

"An OS isn't defined by having a CPU and memory — a calculator has those. It's defined by what it does when many things want the machine at once."

The resource here has a name: the token.

03 · what a token really is

Three things at once — and you can't stockpile it

Every token you emit is simultaneously three resources. That's why a monthly cost dashboard can't govern it — the decision has to happen at the instant of generation.

a dollaron a 100× price ladder; output costs 3–7× input
a milliseconddecode is sequential — tokens are your p99 latency
a safety surfaceeach one can leak PII or carry an injection

And unlike compute or storage, a token is non-storable — consumed the instant it's made, no reserve, no arbitrage. You can't buy tokens cheap and save them; you can only decide, in the moment, whether to spend one and on which model.

04 · why now

Agents made cost quadratic

Chatbots were safe to ignore: one turn, one call, bounded cost.

Agents loop — reason, call a tool, read the result, reason again — and each turn re-sends the whole transcript.

Cost is 1 + 2 + 3 + … + N.

Double the steps, quadruple the bill, with no upper bound.

cost = Σ steps = O(N²)
read on air

"An agent without a budget gate and a pre-run forecast is a runaway process with a corporate credit card. In classic OS terms, it's a fork bomb — and the thing that stops a fork bomb isn't a better invoice, it's the scheduler."

05 · measure vs. decide

Tokenomics measures the bill. Token architecture decides it.

Dashboards and chargebacks tell you what you spent after you spent it — always too late to change it. Token architecture puts four controls on the hot path, before emission:

budgets, not billsallocate per tenant; hard-fail over budget, don't discover it in finance
tiering by defaultroute by capability — hard cases to frontier, the rest to cheap
cache is architecturethe cheapest token is one you never recompute
governance on the railthe budget gate is also where PII/policy checks live

The payoff is concrete: published benchmarks show 80/20 routing (80% to a cheaper tier) cuts blended output cost from ~$25 to ~$9 per million tokens — a ~64% reduction — at near-parity quality, because the hard 20% still gets the frontier model.

06 · where are you

The five levels of token awareness

A ladder for how far an org has moved from reading invoices to scheduling tokens. Almost everyone is on the bottom two rungs.

1 · invoice reading
you find out what you spent from the monthly bill.
2 · dashboard monitoring
real-time cost-per-token dashboards — still downstream.
3 · ad-hoc optimization
routing and caching, project by project, inconsistent.
4 · data-plane budgets
token budgets enforced as a first-class control in serving. Very few are here.
5 · token-scheduled OS
forecast → admit/deny → route → meter → govern, all on the hot path. Cost becomes a control surface the platform team owns.
the takeaway

The gap between the best model and the next closes every quarter.

"Models are commodities. Systems are the moat."

The edge is no longer which model you call — it's how efficiently you schedule tokens across all of them.

zoom out · why it matters now

Why the scheduler question got urgent this month

Agents went mainstream in 2026, inference passed training as the cost center, and GPUs became the bottleneck — this very week two labs capped each other and ran out of capacity (see Kimi K3). When compute is the scarce resource, the org that schedules tokens well doesn't just save money; it's the only one that can afford to run agents at scale.