There is a number going around that AI infrastructure will need 24 times more tokens by 2030. It gets cited as the bullish case, the aggressive forecast, the one that makes data-center people sweat. On a recent 20VC segment, someone called it way too low.
I think they are right, and not for the reason most people assume.
The 24x figure is not too low because demand is hot. It is too low because it is the wrong shape. It models AI usage the way we use it today: one human, one prompt, one answer, repeated more often by more people. That is a line going up and to the right. But the workflow that draws it is already on its way out.
The Forecast Is Modeling a Workflow That Is Disappearing
We have watched forecasts miss this badly before. In 2024, by one industry estimate, real generative-AI usage reached roughly 667 trillion tokens, against earlier projections somewhere near 20 trillion [7]. That is not a rounding error. It is more than an order of magnitude, and it happened because the forecasts extrapolated a curve the technology was quietly redrawing underneath them.
The redraw this time has a name: parallel agents. Cheaper inference plus parallelism means a system can spend far more tokens without making you wait any longer [3]. That second consideration, waiting time, helps explain why parallel work changes consumption. For two years the real governor on token consumption was not the model and not the price. It was you, sitting there reading the last answer before you typed the next prompt.
That governor is being unbolted right now. Native parallel-agent support is the kind of thing we did not have a couple of months ago — genuinely new, not a roadmap promise. And the forecasts that miss it will keep projecting a one-prompt-at-a-time world that exists mostly because human attention runs on a single thread.
One to Ten Agents Is Not 10x
Going from one agent to ten does not multiply your token consumption by ten. It multiplies it closer to 250. That is the line I most wanted to pull on, because it is either the whole insight or a great piece of podcast theater.
It is a podcast assertion, not a measured result, so let me show the work and tell you which number I actually trust. The credible anchor comes from Anthropic, which instrumented its own multi-agent research system: a single agent uses about 4x the tokens of a chat, and an orchestrated multi-agent system about 15x [1]. That is already multiplicative. Adding agents does not add tokens — it multiplies them.
So where does the rest of the climb to 250 come from? Mechanically, from a stack. A parallel workflow is not ten tidy copies of one call. It is a decomposition call that splits the task, N worker calls running at once, an aggregation call that stitches them together, and the retries when branches fail [4]. Then the part nobody budgets for: in frameworks like Google’s ADK, parallel sub-agents run as independent branches with no automatic sharing of state, so each one re-pays its own context and instructions [5]. Ten agents are not splitting one context window between them. They each rent their own.
The layers do not add — they stack. Every branch re-pays its context, then you rank across all of them, then you retry a fraction, and each step scales with N rather than sitting beside it. That is why the cost compounds instead of summing, and why a number well above 10x is the expected result, not an exaggeration.
And the spending is not waste. On hard, decomposable research tasks, Anthropic found that token usage alone explained about 80% of the variance in multi-agent performance. Spending more tokens in parallel is the mechanism for getting a better answer. So I read 250x not as a measured fact but as the measured 15x plus stacked ranking passes, retries, and per-branch context. The exact figure is arguable. The shape is not — and anything multiplicative makes a linear forecast wrong at the root, not at the margin.
Anthropic’s own system is the clean version of this. Hard research that overflowed a single chat got handed to a lead agent that broke it apart, fanned the pieces out to parallel subagents, and synthesized what came back. The bill ran to roughly 15x a chat’s tokens — and bought a good answer, because each subagent returned an independent perspective the lead could weigh and distill. The multiplier paid off because something on top of it was doing the choosing.
The Ceiling Was Never the Model
We prompt Claude or ChatGPT one turn at a time. Read the reply, think, prompt again. And somewhere along the way we started assuming that rhythm is how the tools work. It is not. It is how we work. The single-prompt habit is a sixteen-core processor running everything on one core: the unused capacity stays behind the pace at which a person can read and respond.
Pull the human pacing out and a second multiplier is already compounding underneath. The length of tasks that autonomous agents can finish has been doubling roughly every seven months [8]. Longer tasks mean more steps, more tool calls, more tokens per finished job — entirely independent of how many agents you run. The curve is not one exponential. It is exponentials riding on exponentials.
There is a fair objection sitting right here, and it is worth more than a wave-off. Maybe we prompt one step at a time not because we are slow but because we want to check each step before we commit to the next. In regulated or high-stakes work, that checkpoint is not friction — it is the entire point. Run ten branches with no human between the steps and you have removed the safeguard along with the bottleneck.
True. But notice what kind of ceiling parallelism actually lifts. It removes the pacing ceiling, not the trust one. You can still verify — you just do it once, over ten finished candidates, instead of step by step over a single line of reasoning. The checkpoint does not vanish. It moves to the end, where the stakes are clearest.
From Generation to Curation
That move — verify the finished set, not each step — is the whole shift compressed into a sentence. Instead of generating one answer, you generate ten competing expressions of a solution in parallel, have the model rank them, and keep or recombine the best [11]. The operative question stops being “can AI do this?” and becomes “which of these ten is the right one?” You are no longer only the producer. You are the curator.
Which sounds flattering right up until you remember that curation means reading ten drafts to find the one paragraph worth keeping. The scarce skill is no longer producing the answer. It is the rubric that picks the winner — and that is an engineering problem, not a vibe. I think of it as the Multiplier Stack: the same five layers read two ways, as a model for forecasting a parallel workflow’s true demand, and as a design checklist for making that spend buy capability instead of chaos.
- Count the fan-out (N). The visible multiplier, the “ten agents.” It sets the floor, not the ceiling. Pilot at three, five, ten; find where more diversity stops adding anything. N is a cost decision, not a free parameter.
- Add the orchestration overhead. The router call, the per-branch context re-payment, the coordination traffic. It scales with N, and it is where linear forecasts break.
- Add the evaluation pass. The model-as-judge that ranks and recombines. This is the step that turns the multiplier from waste into value — and it is itself a token cost, because judging ten candidates means reading all ten. Write the rubric before you generate.
- Add the retry and recursion tail. Failed branches retry; candidates spawn sub-agents of their own; ties trigger another round. This tail is what drags the measured 15x up toward the seed’s 250x. Set a stopping rule, or the tree keeps growing on its own.
- Forecast multiplicatively, then clip by cost. Total demand is a product of those layers across goals and users, then capped by cost-per-token — which is exactly what makes a multiplicative curve look deceptively linear when budgets are tight.
It does not fit everywhere, and pretending otherwise is how you get burned. In regulated work each branch has to be auditable, so end-of-pipeline curation can’t stand in for verifying the middle. If you cannot actually measure “best candidate,” generate-then-curate just buys you ten ways to be wrong. And on a small budget the multiplier is simply unaffordable — the math only works when inference is cheap relative to what the task is worth.
The Objections That Keep This Honest
The strongest pushback is not about the number. It is about supply. Power, chips, data, and a latency wall are real ceilings, and frontier training runs in 2030 are estimated at one to five gigawatts [12]. Clip a multiplicative demand curve with a hard supply limit and it looks linear in practice — so maybe 24x already prices in the agentic future. This is the objection that could actually sink the thesis, so it deserves a real answer rather than a dodge.
The answer is to separate two things people keep fusing. Supply sets the level. Demand sets the shape. Everything I am arguing is about shape: it explains why linear forecasts undershoot, not that consumption runs to infinity. And falling cost-per-token is precisely what loosens the supply clip over time, so the multiplier arrives gradually rather than all at once — the same cost dynamic I wrote about in The Coming Cost Shock, seen from the demand side.
Then there is the failure-mode objection: the higher token bill may not buy reliable work. Run ten agents naively and you do not get ten times the work. In reported Silicon Valley deployments, practitioners said the headline problem was the cost of inference and called the coordination chaotic — interdependencies they could not fully reason about [9]. Tokens climbed; reliable value did not. That is the 250x bill with none of the capability.
It is also the exact failure this thesis predicts: the multiplier only converts to value with a disciplined layer doing the choosing. Nine of ten candidates being mediocre is not a hole in the argument. It is the reason the human role moves to selection in the first place. Which is also why taste — the ability to tell the good one from the nine near-misses stops being a soft skill and starts being the job.
What To Do Before the Curve Catches You
If you forecast compute, stop modeling token demand as linear per-user growth. Build it as a product of fan-out, overhead, evaluation, and retries — and pressure-test it against the surprises, the 667-trillion-token years that linear models never see coming.
If you operate, write the evaluation rubric before you fan out, and instrument the whole stack: router tokens, per-branch context, evaluation, retries, each on its own line, with a per-goal budget and a stopping rule. That discipline is the difference between buying capability and buying noise.
If you invest, watch cost-per-token. It is the governor that decides how far the multiplier runs before firms cap concurrency [10]. When generation is cheap and parallel, evaluative judgment — taste, applied to ten candidates instead of one — becomes the scarce input everyone has to buy.
The people who win the next year will not be the ones running the most agents. They will be the ones who wrote the best rubric for firing nine of them.
References
- [1] Anthropic Engineering (2025). How we built our multi-agent research system. https://www.anthropic.com/engineering/built-multi-agent-research-system
- [3] Andrew Ng (2025). Agents Running in Parallel Get There Faster. DeepLearning.AI (The Batch). https://www.deeplearning.ai/the-batch/agents-running-in-parallel-get-there-faster
- [4] Amazon Web Services (2024). Workflow for Parallelization — Agentic AI Patterns. AWS Prescriptive Guidance. https://docs.aws.amazon.com/prescriptive-guidance/latest/agentic-ai-patterns/workflow-for-parallelization.html
- [5] Google (2025). Parallel agents — Agent Development Kit (ADK). https://google.github.io/adk-docs/agents/workflow-agents/parallel-agents/
- [7] Jim McGregor, Tirias Research (2025). The Unstoppable Growth of Generative AI (AI Outlook Part 1). Forbes. https://www.forbes.com/sites/tiriasresearch/2025/06/20/the-unstoppable-growth-of-generative-ai-ai-outlook-part-1/
- [8] METR (2025). Measuring AI Ability to Complete Long Tasks. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
- [9] CNBC (2026). Silicon Valley’s AI agent hiccups: Wasted tokens and ‘chaotic’ systems. https://www.cnbc.com/2026/04/19/siiicon-valley-ai-agent-openclaw-problems.html
- [10] NVIDIA (2025). Cost per token is the metric that defines inference TCO. https://www.nvidia.com/en-us/data-center/resources/cost-per-token-inference/
- [11] Together AI (2024). Parallel Workflow. https://docs.together.ai/docs/parallel-workflows
- [12] Epoch AI (2024). Can AI Scaling Continue Through 2030? https://epoch.ai/blog/can-ai-scaling-continue-through-2030