Buy Low, Use High

I have spent a month arguing that intelligence is becoming a line on the household bill and that the line will be called usage credits. This piece takes the step that makes finance people wince. Treat usage credits as an asset. Put a portion of what you would otherwise invest into them. Buy them when they are cheap. Use them when they are dear.

The first objection is the right one, so I will make it myself. Read the terms.

• ExpiryOne year from purchase at OpenAI; one calendar year from the confirmation notice at Anthropic. Unused balance is forfeited.
• TransferForbidden at both. OpenAI says any attempt to sell or trade credits may result in revocation of the credits and termination of the account.
• Cash valueNone. Not legal tender, not currency, not a stored-value account, not property. Non-refundable except where the law insists.
• PriceFloating. Credits are dollars. Each token is priced at the rate on the day it is used, “as may be updated from time to time.”

Everything that makes a financial asset an asset has been deliberately removed. You cannot hold it, you cannot sell it, and you have not locked in a price. Lawyers at both companies went to real trouble to make a credit not be property.

So why call it one? Because of the single property they left in: convertibility. A credit converts, on demand, into the scarcest input in the economy, and the thing it converts into can be kept, sold, and compounded even though the credit itself cannot. Farmers have always understood this kind of asset. Seed corn: no bank lends against it, no exchange prices it well, it rots if you hold it, and its entire return is realized in a field. Nobody argues that seed corn is not an asset. It is the asset that makes the farm.

Liquidity normally means how fast a thing turns back into money. By that test credits are the least liquid thing you can buy. The exit is closed by contract.

There is a second test, which finance rarely runs because nothing used to pass it: how fast does money turn into labor? A dollar of credits becomes an hour of skilled work in the time it takes to press return. No search, no hire, no contract, no notice period, no waiting for morning. No asset in history has converted money into labor at that speed.

So credits are liquid in exactly one direction: forward, into work; never back, into cash. Call it one-way liquid. It is not a defect of the instrument. It is the instrument telling you which return you are permitted to earn. You cannot earn the trader’s return on a thing that cannot be traded. You can only earn the maker’s.

Here I have to correct my own instinct, because “buy now while they are cheap” hides a wrong premise if you read it the obvious way. Tokens are not getting more expensive. For a fixed level of capability they are getting cheaper faster than any input in the history of computing. Andreessen Horowitz measured the cost of a constant-quality model falling about tenfold a year, from $60 per million tokens in late 2021 to six cents by late 2024. And since credits are dollars rather than tokens, prepaying would not protect you from a price rise anyway. If that were the argument, it would be wrong twice.

Three other things are rising, and they are the whole case.

The frontier

Cheap tokens are last year’s intelligence. The price of the best available model has not fallen; it has held in a band for five years. The same a16z piece notes in passing that OpenAI’s o1 cost the same per output token as GPT-3 did at launch, $60 per million. Anthropic’s price list today runs Sonnet 5 at $2 in and $10 out per million tokens, Opus 5 at $5 and $25, and a tier above Opus, Fable, at $10 and $50. Sonnet’s price was announced as introductory and scheduled to rise 50% on 1 September; Anthropic cancelled the increase in August and made the low price permanent. Read that as competition holding the floor down. The ceiling did not move. The same page notes that Anthropic’s newer tokenizer produces about 30% more tokens for the same text, so the same sentence costs more to say to a better model. If you want the frontier, and for anything that matters you will, you are not a beneficiary of LLMflation. You are paying roughly what the frontier has always cost.

The queue

Demand exceeds supply, and the binding constraint keeps moving: from GPUs to memory to electricity. Apollo’s June note reports H100 rental rates that fell from about $3.00 an hour in early 2023 to $1.85 in late 2025 and then rose back to roughly $2.40 by March 2026, spot DRAM prices up about eightfold since early 2025, and the two big turbine makers nearly sold out through 2029. Its conclusion is the sentence this essay rests on: if demand keeps outpacing supply, the immediate answers are materially higher prices or usage rationing for the leading-edge models, and early signs of both are already emerging. Anthropic sold priority capacity against committed spend; its documentation now says new commitments are no longer available for purchase, which I read as a company that would rather not promise capacity it may not have. On 5 October CME Group lists futures on H100 and B200 rental rates. Nobody lists a futures contract on a thing whose price is calmly and predictably falling. Futures exist for things whose price hurts.

Your appetite

Menlo Ventures found enterprise spending on model APIs went from $3.5 billion to $8.4 billion in six months, more than doubling while unit prices fell. Apollo puts the mechanism plainly: an agent that reasons, tests, and uses tools can consume a hundred to a thousand times the tokens of a chat reply. The bill rises because the amount of intelligence worth buying grows faster than its unit price falls, and it will keep doing so as long as each new capability is worth more than it costs.

So the corrected sentence is this. Tokens get cheaper. Intelligence gets dearer, because the intelligence you will want is at the frontier, behind a queue, in a quantity that keeps growing. “Cheap now” was never about the token price. It is about the conversion ratio: what a dollar of intelligence can still be turned into before everyone else turns theirs.

Suppose the terms were different. Suppose you could buy credits at a discount and sell them when the queue got long, or trade the compute futures that list next month. What would you earn? The spread. A commodity spread is a few percent, competed toward zero by people with more capital and faster machines than you, and it is bounded above by the market’s opinion of a token.

Now look at the other return. This essay cost a few dollars of credits and will be read for years. The morning script cost less than a coffee and has run every day since. A codebase, a book, a business process that works while you sleep: none of these carry a price tag related to what the tokens cost. The return on a token is not what the next buyer will pay for it. It is the value of the thing the token became, and that value is untethered from the input.

Arbitrage earns the spread. Transmutation earns the thing.

Arbitrage prices the token. Transmutation prices what the token becomes. One is bounded by the market for compute; the other is bounded by what you can imagine making. The labs, by forbidding resale, have told you which game you are in. Take the hint.

If the asset cannot be stored, the portfolio question changes from how much to hold into how fast to convert. Five rules I follow.

1. Make it a capital line, not a subscription

   Decide a share of what you would otherwise invest each month and buy credits with it. Not an expense you minimize, an allocation you deploy. It is the only position in the portfolio whose return you control directly.

2. Keep a rolling reserve

   Never more than twelve months of spend on hand, never less than next quarter’s plan. Expiry is a governor, not a threat. It forces the conversion the return depends on.

3. Buy low means buy on the cut

   Price cuts and cancelled increases are a calendar, not a forecast. When the price of the model you actually use drops, top up the reserve. When a lab tries a 50% increase and blinks, that is a buying week.

4. Use high means the marginal task

   Spend frontier tokens on the task with the highest conversion ratio: the thing that becomes an asset, runs while you sleep, or replaces a week. Do not burn Fable on what Haiku does. That is the one arbitrage that is allowed, and it happens inside your own account.

5. Reinvest what the tokens made

   Time, cash, a product, a process. A portion goes back into the reserve. The compounding is not in the price of the credit. It is in the conversion.

Already true

The terms at both labs make credits expiring, non-transferable, and priced at use. The frontier tier costs roughly what it cost five years ago while the floor collapses. The world’s largest derivatives exchange is listing compute futures. Enterprise bills doubled in six months while unit prices fell. My own balance drains to zero every year, on purpose.

What has to happen

Conversion has to stay within reach of people who do not write code, which means tools like Cowork have to keep turning a dollar into finished work without an engineer in the loop. The shortage has to persist long enough that the queue and the frontier premium matter, which the supply chain suggests it will through at least 2027. And expiry has to stay at a year or lengthen; a lab that shortens it is telling you the reserve should shrink.

Where I am probably wrong

If supply catches up faster than the analysts think and the frontier premium collapses into the floor, credits become a prepaid expense with a fuse and the correct allocation is zero: pay as you go. And if the labs open transfer, a secondary market will appear overnight and someone will make money on the spread. They will still make less than the person who used them.