AI Air Cargo Slips to 85,000t a Month From 100,000 Peak as 2027 AI Spend Hits $3.6trn

The Analysis

I price air capacity for a living, so the number that stuck with me after Miami was not the $3.6 trillion. It was 85,000.

Rotate put monthly cloud-computing air volumes at just under 100,000 tonnes at the April peak, falling to about 85,000 tonnes by August. Ryan Keyrouse, in the same conversation at TIACA's Air Cargo Forum on 28 October, put worldwide AI spend at $2.7 trillion in 2026 and $3.6 trillion in 2027 while tonnage flattened. Cloud and AI supplied 33% of air cargo demand growth between January and July. His closing line was the one I wrote down twice: price 2027 capacity off that, not off 2025.

Two true statements pointing in opposite directions is the whole story. Money is accelerating. Kilos are not. Everything an importer has to decide about 2027 sits in that gap, and the gap is where both the risk and the negotiating room are.

Take the three numbers apart before anything else. Monthly volume peaked at just under 100,000 tonnes in April and slid to roughly 85,000 tonnes by August, a 15% fall in four months. The vertical's monthly growth rate had been climbing since January 2025, topped out near 10% in mid-2025, and then came off that peak. Against that, global AI spend runs from $2.7 trillion this year to $3.6 trillion next, a 33% step up in twelve months. And in January to July, cloud and AI made up 33% of the growth in air cargo demand. A peak, a deceleration, and still a third of the growth. Most coverage printed those as a contradiction. They are not a contradiction. They are a capex curve and an opex curve being read as if they were the same line.

AI air cargo is a delivery business for capital equipment, and capital equipment arrives in waves. The first wave is racks, accelerators, switches, optics, cooling manifolds, power shelves, test gear. Heavy, high value, and genuinely urgent, because a data hall earns nothing until the last rack is powered. Once a hall is commissioned, the next dollars go into power purchase agreements, land, substations, water, software, and staff, and none of that boards an aircraft. So the tonnage line tracks a commissioning schedule rather than a demand curve, and a commissioning schedule front-loads. April was not the month the world stopped wanting compute. It was the month the first big tranche of halls got finished.

Run the arithmetic on the old growth path and the size of the mistake becomes obvious. Ten percent a month compounds to 1.1 to the twelfth power, which is 3.14 times in a year. If the mid-2025 rate had simply held, the vertical would be printing around 314,000 tonnes a month right now. It is printing 85,000. Even a gentle 3% a month would put it 43% above today's level. So a 2027 plan built on the sentence AI keeps growing like 2025 does not have a margin of error, it has a factor of three inside it. That is not a forecast risk. That is a different business.

Who eats the difference depends on where you sit. A carrier that converted freighter capacity or wrote charters against AI electronics is now holding aircraft against a flatter curve, and the first symptom is not a rate collapse, it is a quiet reallocation of flying to other commodities. A forwarder with block space on those lanes has the same problem one step removed, and usually with a take-or-pay clause attached. The general cargo shipper, the one who got pushed out of the market in 2025 and paid peak rates for space, is now being asked to sign 2027 at levels set by a curve that has already turned. Those three groups are being told the same story and should be making opposite decisions about it.

The 33% figure deserves its own paragraph, because it is the one most likely to be quoted out of context for the next six months. It describes January to July. If the tonnage is flat in the second half, AI's contribution to growth goes to roughly zero, which means the rest of the market has to carry the whole number on its own. Put it as arithmetic: if total demand grows 4% and AI adds no incremental tonnes, then non-AI cargo has to grow faster than 4% to deliver that 4%. Nothing in the data says it will. A third of your growth coming from one vertical is fine while the vertical climbs and lethal the quarter it stops, and the quote everyone is repeating was measured before the stop.

Here is the part the coverage did not say. The gap between the dollars and the kilos is not an anomaly to be explained away, it is the intended shape of the buildout. Annualise the August number and you get about 1.02 million tonnes a year of AI-related air cargo. Divide $2.7 trillion by that and you land near $2,650 of AI spend per kilo that actually flies. The number is crude, and I would not put it in a board pack without a footnote, but the direction is the point: the flown kilo carries an enormous amount of value and an almost trivial share of the spend. Which means an AI spend forecast is a bad proxy for air demand, and the 33% rise in spend next year can coexist with flat tonnage without anyone lying. The marginal dollar went into electricity and real estate, not into a pallet.

The second thing nobody says out loud is that 85,000 tonnes is a global monthly aggregate over a bucket that is not one commodity. Servers, networking gear, cooling hardware, test equipment, spare parts and returned units all sit in it, and they do not move on the same lanes or on the same clock. Ömer Kaya would take the same slide and go straight to the lane question, which city pairs carried the April peak and which ones are still climbing, and I would not argue with him for a second, because the aggregate is exactly where the answer hides. A global number that is down 15% can contain a lane that is up 30% and a lane that is down 40%, and your contract is signed on one lane, not on the world.

Think about what actually decides whether a box flies. It is never invoice value on its own, it is value density against the cost of time. A pallet of accelerators is worth two or three orders of magnitude more than the space it occupies, and the person buying it is a project manager with a commissioning date, so it flies. A pallet of rack frames is the reverse and it sails. Most of the AI bill of materials sits between those extremes, and where each item lands decides whether it ever appears in Rotate's 85,000 tonnes. That is why the curve is lumpy, and why it can fall while the project pipeline is full: the pipeline moved from the items that fly to the items that do not.

Capacity does not respond to that the way a spreadsheet would. A freighter pulled into an AI lane in 2025 carries a wet lease, a crew rotation, a maintenance slot and a handler contract, and none of those unwind inside a quarter. So when tonnage steps down 15%, the aircraft stay and load factor eats the difference. The first signal is therefore not a headline rate. It is quieter: carriers quote shorter validity, forwarders get offered space they did not ask for, and the charter desk becomes unusually willing to negotiate. If you are buying, those three tells reach you four to six weeks before any published index shows the turn.

The information asymmetry runs the other way too, and importers underuse it. Rotate's number is a market aggregate published monthly. Your own booking pattern is a sample of one company, available daily, and it turns earlier because it sits next to the order book. If your AI-related bookings in September and October came in under plan while the sales team still forecasts growth, you already know something the Miami panel cannot tell you. The trick is having both numbers in the same room. Most companies keep the market forecast in procurement and the booking actuals in operations, and the conversation that matters never happens.

The timing is what makes this urgent rather than interesting. 2027 capacity gets priced between now and Chinese New Year, and the data being used to price it is August data discussed on 28 October. Two months of lag in the number and another two months of negotiation in front of you. Spot reacts to a change in the curve within four to six weeks. Contract rates react on 1 January, all at once, whether or not the curve cooperated. So you are deciding in November with a photograph from August about a year that starts in January, and the only sane response to that is to build a contract that survives being wrong instead of a forecast that insists on being right.

Do the exposure math on a realistic book. Suppose a mid-size forwarder moves 40 tonnes a week of AI-related electronics, which is about 173 tonnes a month. Suppose they sign 2027 at 1.4 times their 2026 average to protect service levels, so 242 tonnes a month. In a flat market their share probably delivers 190. The shortfall is 52 tonnes, or 52,000 kilos, and at $4.20 a kilo that is $218,400 a month, $2.6 million across the year, before anyone counts the space they could have bought cheaper on the spot. Take-or-pay clauses usually bill 80% to 100% of the rate on the shortfall, so the number is not a worst case, it is the standard case for a book that grew its commitment faster than its demand.

Translate the slide into metal and it gets easier to feel. Fifteen percent of 100,000 tonnes is 15,000 tonnes a month, 180,000 tonnes a year. At roughly 100 tonnes of payload on a widebody freighter, that is 1,800 departures a year, about five a day. Five widebody departures a day that somebody scheduled, sold, and then did not fill. For a carrier that is the difference between a good year and a poor one. For a shipper it is the reason capacity on the AI lanes can loosen faster than a headline about $3.6 trillion suggests, and the reason waiting until February to book is not obviously worse than signing in November.

There is a basis problem in how these lanes get priced, and it is worse than usual here. Spot indices on AI-heavy routes are thin, because most of the volume moves on bilateral agreements rather than in the public market. An index can move 5% while the transacted rate on a 200-tonne annual commitment moves 15%. Anyone signing an index-linked contract without checking how that index is built is buying protection against a number that may not describe their lane. Ask what the panel is, how many contributors, what volume it covers, and whether it includes the fuel and security surcharges that make up a third of the invoice. If the answers are vague, index the contract to your own basket rather than someone else's.

When the curve turns, the loss does not land evenly, and this is where a lot of buying teams misread their own P and L. On an AI lane the forwarder's margin is thin in percentage terms but meaningful in dollars per kilo, so a volume shortfall hits contribution faster than an equal rate cut. Losing 15% of the volume on a 6% gross margin costs more profit than a 5% rate decline across the same book. That is the arithmetic reason the correct response to a flattening curve is to defend volume flexibility rather than chase the last point of rate. Most teams do the opposite, because rate is the number printed on the page and volume is the number buried in a clause nobody read.

Plenty of 2025 contracts are still running and are now underwater, and that changes the negotiation in a way almost nobody discusses. A forwarder holding space bought at peak rates against a flat market has two options: eat the difference, or go back to the carrier and re-cut the deal. Carriers can see the same curve, and in my experience they are far more flexible on rate than on volume, because a half-empty aircraft hurts them more than a lower yield. The opening is not to demand a cut. It is to offer a firmer commitment on one lane in exchange for a re-opener on another. That trade is on the table right now, and it will not be if the vertical turns back up in the first quarter.

Now the counter-case, because I do not want to be the person who calls a peak off four months of data. August is a slow month in almost every air market, and a summer trough is not a trend. If the vertical prints above 95,000 tonnes in November or December, the flattening story is dead and the April number was a seasonal high inside a climbing year. If it prints near 80,000, then 2025 was the peak and 2027 has to be priced down. There is a third path that would fool everybody: spend keeps rising and the hardware keeps getting denser per kilo, so the same dollars ship in fewer pallets.

That alone would produce exactly the chart we are looking at while demand is perfectly healthy. And a fourth: the non-urgent tranche starts sailing, because a two-week delay on a rack that will sit in a warehouse for a month is not worth $4 a kilo. Any of those four worlds is consistent with the data, and they call for different contracts.

Notice how thin the evidence actually is underneath the headline, because this matters if you are about to act on it. What we have is two anchors, April and August, plus a growth series that peaked in mid-2025. Nobody on that stage published a monthly line from January to October, and a step down from 100,000 to 85,000 is consistent with three very different shapes: a plateau that settled at 85,000, a slide that is still going, or a trough that recovers in the fourth quarter. The difference between those three is worth millions on a 2027 contract, and it cannot be resolved by quoting two numbers harder. So treat the eight-five as a level, not a trend, until a third and fourth print land in the same range.

There is one more read available that costs nothing, and it comes from the other side of the table. Ask the carrier or forwarder you are negotiating with what length of commitment they want. If they are pushing three-year deals at 2025 pricing, their own book is telling them the vertical is coming back and they want to lock the rate before it does. If they insist on twelve months and refuse a re-opener, they are telling you the opposite, whatever their slide deck says about long-term AI demand. Counterparty behavior on term length is the cheapest forecast in the market, and almost nobody asks for it directly.

If I had to weight them, I would put the summer-trough explanation near 40%, the denser-hardware explanation near 25%, the shift to ocean near 20%, and a genuine peak near 15%, and I would hold all four loosely until the fourth-quarter print. The weights matter less than the habit of writing them down, because the value of a weight is that you can change it when evidence arrives instead of rewriting the story afterwards and calling it foresight.

What I would actually do sits in a narrow band. Re-base the forecast on the last six months of real numbers and stop extrapolating 2025. Cap the growth assumption in single digits until the Q4 print confirms a direction. Split the AI book by lane and report the top three lanes monthly, because the aggregate will keep lying to you. Build the 2027 contract with a volume band of plus or minus 15%, a quarterly re-opener tied to a published index, and a step-down that triggers automatically if the vertical prints below 90,000 tonnes for two consecutive months. And price the downside at 85,000 flat before signing anything, so that the plan works in the world we can see rather than the one the conference slide described.

The trap on the other side is just as expensive. When a curve turns, the instinct is to wait, and waiting is fine until carriers start pulling freighters off the lane and the capacity that looked loose in November is gone in February. That is the asymmetry nobody prices: soft demand does not guarantee soft capacity, because capacity leaves faster than it arrives. Keep a floor of committed space on the two lanes you cannot afford to lose, index the rest, and treat the remainder as spot. That structure costs a little in a rising market and saves the year in a falling one, which is exactly the trade a buyer should be making with two-month-old data.

↑ Back to the key points

— By Grace Yang