By 2035, 9 Billion People Will Be Served by 900 Billion Agents

Huawei filed that sentence under a networking trend, as one bullet among ten. Read it as a market size and it is unverifiable. Read it as a ratio and it is checkable, load-bearing, and already behind schedule.

It is a single sentence in Huawei's Intelligent World 2035 report, sitting in Trend 9, the "agentic Internet," between a section about storage and a section about energy.

"By 2035, the global population will reach 9 billion and be served by 900 billion AI agents."

Read that as a market size and it sounds like every other vendor forecast: enormous, unfalsifiable, and safe to ignore. Read it as a ratio and it stops being a forecast at all. It becomes a staffing plan.

Huawei is a networking company. Trend 9 is about networks, and in the next paragraph it explains what a network needs to do about it: connect every person on Earth, "as well as a staggering 900 billion agents," through something Huawei names the agentic Internet, in order to absorb "the 100-fold increase in data traffic in the next 10 years."

That is the part worth stopping on. A vendor with a decade of optical and wireless roadmaps just told the market that its growth case is agents, and that the volume is not 10% or 30% above human traffic. It is 100 times the current shape of the internet.

The same week, Reuters reported the forecast's other half: that agents will account for more than 90% of global AI token traffic by 2035, that the report "makes no call to slow development," and that Huawei lists security and privacy protection for autonomous agents as one of ten technologies required to support it.

One company's forecast is not evidence of anything. So here is the test we ran instead: take Huawei's ratio, put it beside the only independent measurements that already exist, and see whether the world is on schedule. It is not on schedule. It is ahead.

In February 2026, agent traffic on OpenRouter passed human traffic. It has since grown roughly 14x. That means the crossover everyone treats as a 2035 question already has a date, and the date is behind us.

This is where most coverage will go wrong. A forecast of "90% of traffic by 2035" will be filed as a prediction, compared against other predictions, and forgotten by Thursday, because nobody can be held to a number ten years out. But the variable that decides the headline is not on a ten-year clock. It is on a six-month clock, and it has already moved.

There is a second reason the number is smaller than it looks. Agent traffic and human traffic are not the same kind of unit, and traffic share is not population share. Goldman Sachs modeled the structure in a 41-page note on May 5: an average LLM chatbot session is roughly 1,000 tokens, and the 2025 average query runs about 1,715 tokens. An agent does something categorically different. It runs a loop.

Parse intent, plan, call tools, read results, validate, retry, update state, repeat. Goldman's simulated email-monitor agent runs 24 hourly scans, a 50-loop triage pass, a 12-loop reply pass, and a scheduling sub-loop. That is a single agent doing nothing you would recognize as work, and it is burning a hundred times a chat session before it has made a decision.

So when Huawei says 90% of token traffic, it is not saying 90% of activity. It is saying that the parts of the economy that talk to machines will be outnumbered, unattended, and continuous. Half your traffic being agents and half being people is entirely compatible with most people still using AI the way they do now.

Goldman built the forecast from the bottom up, simulating travel booking, email monitoring, coding, data entry and call centers in pseudo-code, counting tokens per task and tokens per day, then layering adoption curves on top. Knowledge-worker adoption lands near 12% by 2030 and 37% by 2040. The output of that exercise is the number the infrastructure industry is actually building against.

Divide 120 quadrillion tokens a month by the number of seconds in a month and the industry's 2030 target comes out at roughly 46 billion tokens per second, sustained, forever, across everything. That is not a spike. That is a utility.

Two things follow from that table, and they point in opposite directions.

The first is that the demand case for the build-out is no longer speculative. It is the arithmetic that lets $725 billion in annual hyperscaler capex, up roughly 77% year over year, look like a plan instead of a bet. Whoever signs off on a data center needs exactly this number, and now they have it, with a bank's name attached.

The second is that the same numbers demolish the pricing model. Goldman notes compute cost per token from leading silicon is falling 60% to 70% per year. Compounded across the nine years between now and Huawei's 2035, even the conservative end of that band is a collapse of roughly 1 in 3,800. At 65% a year, it is 1 in 12,688. And the money being spent assumes tokens go from cheap to nearly free while volume goes from large to infinite.

This is the sentence that reframes the whole story: 90% agent traffic is not a prediction of the future. It is a precondition of the present. Diverging prices and rising volumes only both work if the customer is software. The forecast and the capex are the same claim. Huawei is not forecasting agent dominance so much as describing, in a networking company's vocabulary, the only world in which anyone's current spending closes.

Now do what Huawei did and take the number literally, but follow it one step further than the report does. If a hundred agents per human actually arrive, every one of them needs the things a thing needs before it can act on your behalf. An identity. A permission set. Somewhere to run continuously. A record of what it did. A way to be stopped.

Huawei's own report concedes this is unsolved. Security and privacy protection for autonomous agents is one of the ten technologies the report says are required. Item one of ten, from the company selling the network. That is not a talking point. That is a requirements list with an empty box at the top.

Put a number on the box. Take 900 billion agents, assume each makes a modest 100 tool calls a day, and assume each call writes a single kilobyte of log.

And the alternative to storing it is worse. An agent economy that keeps no log is not a rogue-agent story, it is an unattributable one. When an agent moves money, rewrites a record, or sends something it should not have, the only three questions that matter are who it acted for, what it was permitted to do, and what it did. Without a per-call record, none of them has an answer.

That is why the safety debate keeps failing to land. It is argued in the register of catastrophe, and catastrophe is easy to wave away. Storage and retention are boring, measurable, and impossible to wave away. The honest version of the autonomy problem is a capacity problem: who keeps the record, who deletes it, who may read it, and who is accountable when it is gone.

The bill nobody scheduled a hearing for

On September 15 the House moved the Ratepayer Protection Act under suspension of the rules. It concerns electricity for large data centers. That is the visible half of the cost of running a hundred agents per human, and it is now a live legislative question with sponsors and a floor vote.

The invisible half does not have a sponsor. When an agent runs on the machine in front of you instead of in a facility, the compute bill arrives as depreciation. We have the receipts on one desk, running a handful of agents for three weeks:

• 37 TB written to disk in 21 days, about 65 GB a day, from logging and state alone
• RAM growth peaking near 911 GB per hour during a leak window
• A CPU held at 98°C under sustained agent load

Nothing here is exotic. It is what the architecture does. The upstream trackers have the same story in public, filed as bug reports rather than as cost: memory issues 11315, 67433, 79815, 84960 and 86984 on anthropics/claude-code; disk-write issues 30612, 32496, 35401 and 35482 on openai/codex.

Scale that by Huawei's number and the arithmetic is not ambiguous. Ninety-two petabytes a day cannot be written to a laptop. It can only be written to a facility, by software that lives there. The forecast is not just a demand signal for networks and power. It is a demand signal for somewhere for the agent to live.

Goldman's model separates agents into two species, and the gap between them is the most useful number in the note.

The 24x and 55x growth figures in the chart in section 02 are produced almost entirely by the bottom row. On-demand agents do not get you to 120 quadrillion tokens a month. Persistent ones do.

So the honest question raised by Huawei's 900 billion is not which model wins, or how good agents get. It is whether the thing running your agent is something you switch on, or something that stays on. A model is a component. An environment is the product. The distinction sounds like positioning until you try to run something continuously and watch it compete with your own laptop for memory, disk and thermals.

Which is the whole reason the ratio matters. A hundred agents per human is not a fleet anyone operates by hand, opening windows. It is closer to a population, and populations need somewhere to live: state, addressability, a record, and a way to reach the person they work for. That is the layer this decade will actually be decided on, and it is a layer the current debate is barely discussing.

1. Is it a ratio or a total? Totals are unfalsifiable and chosen for press coverage. Ratio it against something already known, as we did with 900 billion over 9 billion, and it becomes checkable. Huawei's number only bites as 100:1.
2. Does the crossover already have a date? Anything framed as a 2035 milestone should be checked against 2026 measurement. In this case the answer was that the milestone had already passed and the forecast was six months behind reality.
3. Whose cost structure requires it to be true? Huawei sells the network. The hyperscalers carry the capex. Every participant in this forecast is spending against it. That does not make it false. It tells you which parts will be repeated loudly and which parts will be omitted.
4. What is the vendor's own requirement list, and who has filled it? Huawei's list is public and its first item is unsolved. The most reliable disclosure in any vendor report is the section describing what does not exist yet.

Three tests, not three opinions

One. Does the always-on number hold up? Goldman's >100,000 tokens per day per user is the multiplier behind the whole curve. Watch for the note itself, and for anyone reproducing the figure without citing the note. If the number softens, the 24x softens with it.

Two. Does 100:1 survive contact with identity? A hundred agents per human requires a hundred sets of credentials, permissions and revocations. Watch for the first serious standard, incident disclosure, or regulator that treats an agent as a principal with its own audit trail. An empty box at the top of Huawei's list is an opportunity, and someone is going to take it.

Three. Who publishes the first number that would prove them wrong? Every participant here is a seller of compute, networks, models or environments, including us. The industry's forecasts will improve the day one participant publishes a falsifier with the same font size as the headline. Watch for that, and expect to wait.

Grades. A primary document, filing, or first-hand measurement. B reported by a major outlet or a detailed single-source account. C secondary aggregation we could not trace to a primary document, used only where it is not carrying weight.

Derivations are ours and labelled. The 100:1 ratio, the 46 billion tokens per second, the 1 in 12,688 price collapse, the 92 petabytes per day and the 2035 extrapolation are all our arithmetic on the inputs in the table above. Assumptions are stated where they appear. None of them is a vendor projection and none should be quoted as one.

What would prove this wrong. If per-token cost decline flattens to single digits per year, the pricing argument in section 02 collapses. If always-on agent token load is an order of magnitude below Goldman's figure, the 24x and 55x curves collapse with it and Huawei's 90% becomes a networking sales pitch rather than a structural claim. If agent traffic growth on the major gateways decelerates through 2027, section 01 inverts. Write them down now so the scorekeeping is honest later.