Google processed more than 3.2 quadrillion tokens across its surfaces in May 2026, up from 9.7 trillion in May 2024.1 That is roughly 330x in twenty-four months. Microsoft disclosed more than 100 trillion tokens in a single quarter while OpenRouter's weekly volume went from 5 trillion to 25 trillion in six months.2 Dell modeled inference reaching one quadrillion tokens by 2028, revised the estimate to 57 quadrillion, and still suspected it was too low.3
Every one of those numbers is presented as evidence of demand.
None of them is evidence of work.
The precedent nobody wants to name
In 1999, the metric was clicks. Then page views, then eyeballs, then registered users. Each one measured activity. Each one was real. The servers really did serve the pages, and the users really did click. What the market did was treat activity as though it were monetizable demand, and then build valuation models on the extrapolation.
The activity survived. The valuations did not.
A token is a unit of text that passed through a model. It is not a unit of value delivered, a task completed, or an hour of labor displaced.
Tokens have the same structural weakness. A token is a unit of consumption, and the AI industry has spent two years reporting consumption as if it were output.
What a swarm actually spends
A conventional interaction is short:
prompt → inference → answer
A multi-agent system is not:
prompt → planner → N agents → tool outputs → agent critiques
→ revisions → arbitration → synthesis → verification
Twenty agents at 50K tokens of context and output across several iterations turns a task that a well-scaffolded single agent handles in 100K–200K tokens into millions. The spend breaks into four buckets:
- duplicated context
- redundant reasoning
- coordination traffic
- discarded branches
Every agent re-ingests the same repository, the same requirements, the same tool results. Then agents generate intermediate text whose only consumer is another agent.
You are paying inference prices for the communication fabric itself.
The multiplier is already published
Anthropic said it out loud. Their multi-agent research system consumes roughly 15x the tokens of a chat interaction; single agents about 4x.4 They published this as an economic constraint, not a selling point, and they were explicit that tasks requiring shared context and dense interdependencies are a poor fit for the architecture.
Coding is exactly such a task. Cognition published "Don't Build Multi-Agents" the same week.5
That 15x is the engineered case: an orchestrator-worker design, dedicated context windows, a separate citation pass, production evaluation. It is a floor for competent multi-agent systems, not a ceiling for the ones being deployed. A subagent that recursively spawns subagents, or a tool that returns oversized results, multiplies from there.
Measuring the waste
The right baseline is not zero. It is minimum competence — the tokens a well-designed single-agent-plus-tools workflow needs to reach the same verified outcome.
Waste, against a competent baseline:
\begin{aligned} W &= 1 - \frac{1}{A} \end{aligned}A task that a good agent solves in 400K tokens and a swarm solves in 1.5M carries W ≈ 73%. The inverse is more useful for financial modeling — token amplification:
\begin{aligned} A &= \frac{1}{1-W} \end{aligned}75% waste is 4x amplification. 90% is 10x. 95% is 20x.
My working estimates, offered as hypotheses to be tested rather than measured industry statistics:
| Architecture | Estimated avoidable token overhead |
|---|---|
| Good single agent + tools | 10–30% |
| Agent + verifier/reviewer | 20–45% |
| Small orchestrated system (3–5 agents) | 35–60% |
| Large multi-agent workflow | 50–80% |
| Open-ended swarm / recursive agents | 70–95%+ |
I will not claim that 70% of AI tokens are wasted. Nobody has the industry-wide telemetry to defend that number. But "current agentic architectures exhibit 2–5x token amplification relative to optimized workflows" is a defensible hypothesis, and it is testable.
Coordination, not capability, is the binding constraint
The Berkeley-led MAST work analyzed more than 1,600 execution traces across seven popular multi-agent frameworks and identified fourteen distinct failure modes across specification, inter-agent misalignment, and verification.6 The finding that matters for this argument: in a number of cases, the same model in a single-agent configuration outperformed the multi-agent version. Follow-on analysis of that taxonomy puts roughly 42% of failures in specification, 37% in coordination, and 21% in verification,7 and reports error amplification as high as 17x in uncoordinated systems against roughly 4.4x where a centralized validation bottleneck exists.8
These are not model limitations. Better base models do not fix them. They are distributed-systems engineering failures, and every one of them is billed by the token.
The incentive runs the wrong direction
Semiconductor providers are delivering per-token inference cost reductions on the order of 60–70% per year.9 GPT-3.5-equivalent quality fell from $20 to $0.07 per million tokens in about eighteen months.10
When inference approaches free, throwing more agents at a problem is cheaper than engineering a better one. The path of least resistance becomes:
cheap tokens → more agents → more retries → more tokens
instead of:
better architecture → fewer inference calls
And here is the part that should interest anyone underwriting this buildout: nobody in the chain is paid to remove the waste. The customer pays for the retries. The provider books them as revenue. The analyst books the revenue as demand. The capex model books the demand as a curve.
Coordination overhead is indistinguishable from adoption at every layer of the stack above the API call.
The honest counterargument
Anthropic's own data cuts the other way in one important respect: token usage explained roughly 80% of performance variance in their research evaluations. Spending tokens is not incidental to quality in breadth-first search — parallel exploration across independent context windows is the mechanism.
Where a task genuinely decomposes into independent directions, the multiplier buys something real.
That defense holds for research, due diligence, literature review, competitive intelligence. It does not hold for tightly coupled work where every agent needs the same state. And tightly coupled work — coding above all — is where the largest agentic token volumes are currently being generated.
What it does to the forecast
Goldman Sachs projects token consumption growing roughly 24x between 2026 and 2030 to about 120 quadrillion per month, with agentic workloads the majority of it.11
That forecast contains two multipliers braided together. One is genuine adoption. The other is architectural amplification. The first is durable. The second is an engineering defect that shrinks as orchestration matures — better context selection, persistent state, deterministic tools, retrieval, caching, prompt caching, smaller specialist models, structured inter-agent protocols, and materially fewer LLMs talking to LLMs.
If a meaningful fraction of agentic token demand is amplification, then part of the curve being extrapolated is an efficiency deficit, not a market. The industry could produce far more AI-generated economic output in 2030 while consuming fewer tokens per unit of it. Measuring AI adoption by tokens consumed would then look like measuring the PC revolution by CPU cycles burned.
A badly optimized program is not evidence of greater economic activity.
The experiment that settles it
Take a fixed set of SWE-bench-style tasks. Run each through four architectures: single agent with tools, agent plus critic, a five-agent orchestrated system, and an open-ended swarm. Normalize by verified completion — tests passing, diffs accepted. Report tokens per successful task and dollars per successful task.
That is a week of work for anyone with a lab and a budget. The result would replace every estimate in this essay with a measurement.
The fact that no vendor has published it is itself a data point.
The metric
Token Efficiency = Verified Useful Work ÷ Inference Tokens Consumed
For coding agents, the numerator is accepted changes, tests passed, bugs fixed, engineering hours actually displaced. In dollar terms: verified economic output per inference dollar.
None of this invalidates the buildout. Compute demand is real and the adoption is real. What it threatens is a narrower thing: valuations priced on today's per-token revenue multiplied by tomorrow's projected token volume, when both terms are moving, and one of them contains a defect that engineering is actively removing.
Watch for the first CFO who asks what the accepted-diff-per-dollar number is. That question will get answered eventually, by someone, in a footnote. The interesting part will not be whether the data centers were justified.
It will be which vendor's quarterly token growth turns out to have been a line item called coordination overhead.