AI’s Biggest Winners Aren’t the Biggest Token Spenders

The winners of AI will be the companies that get the best intelligence per dollar spent - not the Tokenmaxxers, but the Token Allocators.

For years I decided how a finite pool of capital was split across a high-frequency trading firm’s desks, forecasting what they’d need and assessing how to deploy it across entities to get the highest returns.

These days, companies are buying AI compute the way those trading desks only wish they could have drawn capital. No limits, no capital charges, no business justification. The race for compute is on with larger token commitments, more agents, more models. The consensus is that whoever spends the most on intelligence wins. They won’t. The winners will be the ones who can tell you what the last million tokens bought. Or even better, what the next million will buy.

Most companies don’t even know how to begin answering those questions.

Having jumped from HFT to AI, I can’t help but wonder: where is the treasury for inference? Where is the optimisation of token spend across workflows, teams, departments, and - at the highest level - across legal entities and jurisdictions? And where is the calculation of the ‘marginal product of tokens’ (MPT) to drive all of the above?

Almost no company realises they’re running a trading book. And they’re already falling behind those who do.

Chamath Palihapitiya recently described his own company's position on the All-In podcast. Its token costs are doubling roughly every 45 days - but productivity is only up by 5%, at best. Each new increment of quality costs a mountain of tokens.

Translate that into the language of a trading floor. The marginal return on incremental capital has collapsed towards zero. You pay exponentially more for almost nothing, yet keep doubling capital usage. A desk running that book gets shut down by lunch.

Striking as above numbers may be, the most surprising part is that they even measure them. Most companies are running the same book and have no appropriate metrics. The CFO who fights over every basis point in the cost of capital has never once been asked what a token returns. Capital gets allocated, yet compute gets spent.

Token allocation should happen on 3 levels. Companies are still playing the tutorial.

Almost every company is doing one of three things.

The most common: sign up for seats, buy the agent, renew the subscription because of a strong demo. No allocation happening at all. One large tech company ran a token leaderboard until it became clear that it was rewarding consumption and not productive output.

More sophisticated companies price each initiative on its own and fund the ones that clear a bar. It looks respectable, and it is a real improvement compared to buying whatever demoed the best. But each initiative clearing a return bar tells you nothing about whether there should be reallocation amongst initiatives - or whether a new workflow, still unbuilt, would grant an even higher return.

With this approach, eventually the sprawl of AI point solutions becomes impossible to ignore, and someone imposes a company-wide cap. It is the first genuinely top-down decision most companies ever make about tokens - and it is made with no demand curve underneath it. As such, the cap ends up throttling the highest-MPT workflow in the building purely because of whose team it sits in.

The first of those approaches allocates nothing at all. The second prices one use case at a time. The third sets the size of the whole pool. Not one of them compares use cases against each other - and that comparison is where the money is.

No CFO would allocate headcount this way - some teams hiring without limit, others frozen by policy, none of them ever compared against each other. Instead, the return on your next hire, or your next dollar of capital, depends entirely on where you put it. So you keep (re)allocating until the return per dollar is the same across every use case. That is MPL/w = MPK/r, with w and r being the price of labour and capital respectively. Cobb and Douglas formalised this nearly a century ago, yet companies behave as though none of it applies to the marginal product of tokens.

Finally, leading companies are starting to behave differently by borrowing ideas from the world’s best capital allocators.

Allocating token spend like a quant

Falling token prices make the allocation problem worse, not better. That is Jevons paradox: when the efficiency of a good increases, you don't spend less on it. Instead, you consume so much more that the bill goes up anyway. This is the reality many companies face today.

AI use cases that were previously too expensive to justify are now candidates for investment. Some of these will create enormous value, others close to none. Prioritising and allocating tokens accordingly will separate the AI winners from the rest of the pack.

Treating tokens like capital, we can think of allocation across 3 levels: firstly, to individual use cases. Secondly, between use cases, workflows, and departments. Thirdly, across the entire company - the overall token pool.

Most companies collapse all 3 into a single AI budget. A trading firm would never.

Token Allocation, Level 1: Individual Use Case

Allocation starts at the use case level. Before a company gives tokens to an agent, workflow, or employee, it needs to know what the next unit of spend is expected to produce. Not what that use case has returned on average, but what the next million tokens are likely to earn.

The same applies to trading firms. Strong-performing strategies don’t necessarily deserve the next dollar. They’re prone to alpha decay, and to capacity constraints as you deploy additional capital against them.

Thus, companies now need to estimate the marginal product of tokens - measuring each use case by its cost per verified outcome: what the company pays in tokens for each success. Almost nobody can produce that number. Companies are finally starting to closely monitor token costs. Few connect those costs to outcomes.

A support triage agent's cost per resolved ticket might sit at $0.40 for the first million tokens, then climb as the agent takes on harder edge cases that require more retries and more back-and-forth human review. When cost per verified outcome starts rising, it is time to assess whether the MPT for this use case is too low.

The return to the company might be 30x on its first million tokens - thirty dollars of verified value for every dollar of tokens - and 9x on its last. Same workflow, same model, same team. A use case that started out as a great investment easily gets pushed beyond its limits.

Assessing the marginal product of tokens is also where the model gets chosen, because not every token is the same. A top-tier model like Claude Opus runs about $25 per million output tokens; a strong open model like Kimi sits near $4, and a fast tier like Grok's around $0.50. Cost per verified outcome prices the outcome rather than the token and simplifies the discussion. If a smaller model resolves a ticket at $0.40 and a top-tier model resolves it at $0.90 with no better verification rate, the cheaper model wins the allocation. Companies can save the ‘frontier tokens’ for the workflows that actually benefit from them.

Notice the two ratios don't match. The top-tier token costs ~six times as much, but the resolved ticket only costs ~twice as much, because the stronger model potentially resolves it in a third of the tokens. Per-token pricing tells you almost nothing on its own. Price per token != price per outcome.

As such, benchmarking MPT per model per use case is a critical input for your Level 1 allocation.

Token Allocation, Level 2: Between Use Cases

Once companies learn to assess the MPT per use case, they need to compare across use cases: where do the next million tokens create the most value? That means comparing support triage against invoice matching, and sales research briefs against an agent nobody has built yet.

The obvious move is to pour tokens into your best performer. But usually you can't.

Let’s work through an example. Say an invoice agent returned 27x on its last million tokens but it already processes every invoice that reaches it. There is no more work for it to do, and the next million tokens would earn nothing. This is the reality of many companies when they start to implement agents. There are binding constraints on the pipeline, and the value of relaxing those constraints might still be 27x (e.g., new regions). But that is the return on relaxing the constraint, not the return on pouring additional tokens into the workflow as-is.

Thus, the next tokens are better spent somewhere that doesn't exist yet. Let’s say a procurement agent not yet built is worth an estimated 15x. We can fund it one of two ways: expand the budget or reallocate from an existing use case. In this example, let’s proceed with reallocation from a low-MPT use case - maybe a sales research agent unnecessarily using the latest model or researching too broadly with 1.3x marginal return.

Every dollar moved from the 1.3x use to the 15x one is effectively worth $10.50. On a $4m token budget, reallocating $200k of existing spend outperforms an additional investment of $2m into the existing, low-MPT use case.

In other words, reallocating 5% beats adding 50% to the total budget!

Companies blindly signing deals with big labs or based on pretty demos fall victim to this. They often miss out on way more than the above example illustrates.

Now tactically, (re)allocation can be a slow process. Assessing switching or implementation costs is not trivial. Going back to the trading world, a desk doesn't move capital the moment a gap appears, especially if it’s a new use case: onboarding a new symbol means connectivity, borrow and risk limits, and standing up a new strategy means research, a backtest and a ramp before it earns anything. Beyond that, as a market maker, you may even need to set up quoting, market, and underlying execution before deploying the more advanced strategies. Tokens behave the same way. A workflow that takes a quarter to build has to clear a higher bar than one you can switch on tomorrow - and the estimate deserves the same haircut treasury applies to any desk forecasting its own capital usage.

At a trading firm, this level of allocation has to be done centrally from a total portfolio view. First for a specific entity or region, then again at a global level. Capital moves dynamically towards the highest-potential strategies.

Something similar will eventually exist inside every Fortune 500 company. A central ‘Inference’ function will own token allocation in the same way that risk committees, treasuries, and global strategy heads run capital. It measures the expected return of each workflow and decides where the next tranche of tokens should go. The “Chief Inference Officer” will therefore look less like an IT executive managing software spend and more like a portfolio manager, with the mandate to maximise the return on the company’s token allocation.

Token Allocation, Level 3: Total Token Pool

The final decision is how many tokens the company should buy at all. A fund deploys capital only when and where the next dollar can still earn an acceptable return. Similar to Level 1, this touches capacity constraints and liquidity bottlenecks, now at a fund level instead of per use case.

Theoretically, this exercise is trivial: grow token spend and budget as long as the initiatives clear your hurdle rate. That is, compare your token returns to those of your capital & labour investments - your steady-state hurdle rate. If nothing left in the pipeline clears your hurdle, you should probably stop increasing token allocation. Revisit once your hurdle or MPT (e.g., through model improvements) become more favourable.

In reality, this is much harder. Even if you do Level 1 and 2 correctly, leveraging a different token budget can take time and be costly. Token resizing is not frictionless. However, compared with the treasurer’s management of capital injections and dividends, token resizing is likely to be a far more dynamic exercise for the Chief Inference Officers of the future.

Efficient allocation creates winners

The next wave of AI winners will be the ones that get the best intelligence per dollar - not the Tokenmaxxers with their fancy leaderboards.

We have come full circle back to Cobb and Douglas. Deploy labour and capital until the return on the next unit of each, per dollar spent, is equal. None of that has changed. There is simply a third term now:

MPL/w = MPK/r = MPT/p

Where p is the price of a token.

We discussed 3 token-allocation levels above. Each solves one part of this equation: the first ensures we can even calculate MPT. The second equalises MPT/p across use cases. The third decides the steady-state equilibrium across factors.

Going through these 3 levels is not only the answer to your company’s rising token costs - it is also a blueprint for how to prioritise every AI initiative.

Most companies can tell you their token spend to the dollar. Almost none can name which of their AI initiatives returns 15x at the margin, and which returns 1.3x.

Want to learn more about assessing business impact of AI use cases, prioritising between them, and how to go about implementation? Subscribe to our newsletter at: www.varickagents.com/newsletter