North American robot orders reach 17,995 units worth $1.17bn in H1 2026, A3 data show

The Analysis

A robot order book is an arrival stream, and I read it the way I read any queue: what is arriving, how fast, how heavy is each job, and how long before it turns into capacity at the other end. The Association for Advancing Automation put out its first-half 2026 numbers for North America, and the headline pair looks dull until you hold the two halves against each other. Companies ordered 17,995 robots worth 1,166 million dollars in the first half of 2026. Units up 2.0 percent year on year, value up 6.6 percent.

Two percent is nearly flat. Six point six percent is not. When units barely move and dollars move three times as fast, the queue is not getting longer, each job inside it is getting heavier. Divide one by the other and the implied average price per unit is up roughly 4.5 percent year on year. That number is the most useful thing I can pull out of the release and it is not printed in the release; it is arithmetic on two numbers that are.

I care about this as an importer and exporter because robots are not a technology story to me. They are a capacity story with a lag. Somebody ordered 17,995 machines in six months. Those machines get uncrated, installed, integrated and brought up to rate somewhere between two and four quarters from now, and when they come up they change the volumes on my lanes, they change the lead times at my suppliers, and they change what I pay if I want to buy some myself. Three channels, three different clocks.

Start with the quarter split, because the half-year hides it. I have the first-half totals and the second-quarter figures, so the first quarter is whatever is left: 17,995 minus 8,940 is 9,055 units, and 1,166 million minus 622 million is 544 million dollars. Put the two quarters side by side and something odd shows up. Units went from 9,055 to 8,940, down about 1.3 percent. Dollars went from 544 million to 622 million, up about 14.3 percent. Average price per unit works out to roughly 60,100 dollars in the first quarter and roughly 69,600 dollars in the second, a jump of close to 16 percent in three months.

I do not extrapolate a one-quarter move like that. What I do is ask which of three things caused it, because the answer changes what I tell my sourcing team. It could be mix: more expensive machines, bigger payloads, more axes, vision packages bolted on. It could be price: list increases sticking. Or it could be that the cheap end of the market stopped ordering while the expensive end kept going. Those three have completely different consequences for a buyer. Mix means your budget for the same machine is unchanged and you can relax. Price means waiting genuinely costs money. A collapse at the low end means the entry-level segment is in trouble and that is a demand signal about small manufacturers, which is a different story again. The release does not split it out, so I treat 16 percent as a question to ask my suppliers, not as a forecast.

The sector numbers are the part with real freight meaning. Semiconductor, electronics and photonics buying rose 35 percent. Life sciences, pharma and biomedical rose 32 percent. Automotive OEM orders fell 25 percent. Read carefully: those are growth rates, not shares. The release I have does not give me the absolute units behind each sector, so I refuse to convert a growth rate into a share of 17,995 in my head. This is where people get sloppy. A sector growing 35 percent off a small base can still be a rounding error next to a sector falling 25 percent off a large one. If your planning depends on which of those is true, go and get the denominator before you move a dollar. Quantify first, then argue about what it means.

What I can say without the denominator is directional and it is still useful. Two sectors are pulling hard in the same half-year and one is retreating. Semiconductor and electronics capacity is being built. Life sciences capacity is being built. Automotive OEM capital equipment is being cut by a quarter. If your book of business leans into automotive OEM, this is a leading indicator that somebody upstream of you is preparing for less volume, and capital equipment orders lead physical output by quarters. If you lean into electronics or pharma, somebody upstream is preparing for more, which means tighter capacity later and a stronger argument for booking space earlier rather than later.

Then the cobot line. Cobots took 15.4 percent of units in the first half, which on 17,995 orders is somewhere around 2,770 machines. I read that the other way round: about 84.6 percent of what got ordered is not a cobot. Conventional automation is what is actually being bought, and conventional automation comes with fencing, floor anchors, safety validation, integration to upstream and downstream equipment, and a commissioning window measured in weeks. That matters for two reasons. It means the two-to-four quarter lag I assumed is closer to the long end for most of these orders, and it means the bottleneck in most projects is not the robot, it is the integration and the floor. I have never seen a project slip because the arm arrived late. I have seen plenty slip because the safety sign-off took five weeks longer than the plan.

Now the arithmetic on my own money, with the assumptions nailed down. Assume I am budgeting a cell of six units. Assume I pay something near the half-year blended average of about 64,800 dollars per unit, which is 1,166 million divided by 17,995. Six units is 388,800 dollars of hardware. Assume integration, installation, safety and commissioning run at forty percent of hardware, which is 155,500 dollars, for an all-in of about 544,300 dollars. Assume each unit displaces one and a half people across a two-shift operation, so nine people. Assume a loaded cost of 34 dollars an hour at 2,080 hours, which is 70,720 dollars a head, so 636,500 dollars a year. Payback lands around 10.3 months.

Now run the same sheet six months later with prices up 4.5 percent, the half-year implied figure. Hardware becomes 406,300 dollars, all-in 568,800 dollars, payback about 10.7 months. Waiting half a year costs me roughly 24,500 dollars in capital and about half a month of payback. That is a real number and it is smaller than it sounds. Which is my point: a 4.5 percent price drift is not a reason to rush a capital decision. It is a reason to stop treating the decision as reversible and start the paperwork, because the thing that will actually hurt me is not the price, it is discovering in March that the lead time was thirty weeks and I needed it in eighteen. So the discipline is this: get lead time in writing from three sources, take the median, and let the median set your order date. Price is the second-order term.

For the inbound side, treat the machine the way you treat any other capital import. The robot price is one line in a landed cost that also carries freight, duty, installation and the internal hours of everyone who touches the project. I am not going to give you a classification or a duty answer, because that is not my lane and I would rather you ask your broker than trust a column. What I will say is that the landed-cost model should be built before anyone signs, with a written quote validity window on every line, because a machine quote that is open for thirty days and a project that takes ninety days to approve are two different risks. Hold the supplier to a stated validity period and re-price on expiry rather than discovering the change at invoice.

Then the part everyone skips: what does success look like after the machines are in. Baseline before, measure after. Take throughput per labour hour on the cell today, by shift, for four weeks. Install. Measure the same thing for four weeks after a ninety-day settling period. Threshold: I want at least a twelve percent improvement in throughput per labour hour, and I want the 90th-percentile cycle time to improve and not just the average, because the average hides the stalls. Guardrails: scrap or error rate must not rise more than one percentage point, unplanned downtime on the cell must stay under five percent of scheduled hours, and maintenance hours per week must not exceed what the plan said. Stop condition: if throughput is under the twelve percent mark at day one hundred and eighty, freeze further cells and go looking at the process upstream, because at that point the constraint is almost never the robot.

That last sentence is the whole column, really. Buying machines does not fix a process you have not fixed. If the reason your line is slow is that material arrives late, or the inbound dock has two doors and needs four, or the changeover eats ninety minutes, then a robot sits in the middle of that waiting. It will wait faster than a person waits, which is a nice trick and worth nothing. Model the line first, find the constraint, and only then decide whether the constraint is one that a machine addresses. Sometimes the answer is a second dock door and a scheduling rule, at a fraction of the capital.

Where does this leave the reading of the A3 numbers? As a leading indicator with a stated lag, not as a verdict. Write the lag down in whatever planning document you keep, so that the next person who opens the file knows that two-to-four quarters is an assumption of mine and not a published figure. Set a review cadence on it: I re-read the order data every quarter, and I only act on a sector move when I have seen it in two consecutive quarters, because a single half-year print is one observation and one observation is noise.

So, the calendar, starting from where we are in late August 2026. By Friday 4 September, the commercial team tags the top twenty lanes by end sector and computes what share of our volume sits with automotive OEM customers or with electronics and life sciences customers. That share, not the national number, is what drives our exposure. In the week of 7 September, procurement writes to our top ten suppliers and asks two questions only: what is your current quoted lead time in weeks, and how long is that quote valid. Median the answers, put the number in the plan, and re-ask every quarter. By 30 November, re-run the exposure split and compare it to the September baseline. And on 31 January 2027, do the honest check: has the capacity anybody ordered in the first half of 2026 actually shown up as volume on our lanes. If the answer is no after three quarters, my lag assumption was wrong and I say so in writing rather than quietly extending it.

Alternatives, because buying is not the only response to any of this. If you are exposed to automotive OEM and the equipment orders say volumes soften, the response is not necessarily to chase replacement business; it may be to shorten committed volumes, renegotiate minimums, and hold your own capex. If you are exposed to electronics and the orders say capacity is coming, the response is to book space earlier and lock rates for longer, not to add trucks you cannot fill yet. If you want the productivity without the capital, there is usually a process answer first: changeover reduction, better sequencing, a second shift on the constraint. And if what you actually need is flexibility rather than throughput, the cobot share of fifteen percent is telling you that most buyers in this market are still choosing fixed automation, which is worth asking about before you copy them.

The traps are mostly statistical. A 2.0 percent unit move over a half-year is inside the noise band for a series like this, and building a network plan on it is a mistake I have made and will not repeat. Orders are not installations and installations are not productive capacity; do not collapse three stages into one number. These are North American orders, so do not read them as a global figure, and do not read them as a proxy for your own region unless you have checked. Growth rates are not shares, and I have already said why that one bites. And the biggest trap of all is counting machines instead of counting output: units ordered tells you what somebody spent, not what they will make, and the gap between those two is where every bad automation business case I have ever reviewed went wrong.

I will finish on the queue. Arrivals into the North American capital stock ran at 17,995 machines in the first half of 2026, the average job got about 4.5 percent more expensive, and the mix is drifting toward electronics and life sciences and away from automotive. None of that tells you what to do on its own. It tells you where to put your instruments. Measure your own exposure, measure your own lead times, write down the lag you are assuming, and hold your own projects to a threshold you set before the first machine is uncrated. Quantify first. Then optimise, if there is anything left to optimise.

Let me make the exposure tracker concrete, because watching your sector mix is not actionable until somebody builds the table. One row per lane or per customer group. Columns: end sector, annualised units or containers, revenue, committed volume under contract, and the quarter-over-quarter change in each. Five fields, updated monthly, owned by one person in commercial. Then the thresholds: any row where the end sector is automotive OEM and revenue share exceeds thirty percent gets pulled into a volume review that quarter; any row in electronics or life sciences growing faster than fifteen percent year on year gets a capacity conversation with your carriers before the next tender cycle. The point of the table is not the reporting. It is that when the A3 numbers move, you already know which rows of your own book move with them, and you can act in days rather than discovering it in the next quarter's volume variance.

A caution about what an order is. An order is a signed intention, not a shipped machine, and it can be pushed. Order books get renegotiated when conditions turn, and the customers who ordered in the first half can defer delivery in the second. That cuts both ways for me. It means the capacity signal is softer than the headline 17,995 suggests, and it means I should be looking for confirmation in a second series instead of reading one release. Confirmation I can actually get: my own suppliers' order books, my own customers' production schedules, and the lead times quoted to me, which lengthen when capacity is genuinely tight and shorten when it is not. Three independent readings before I touch a network plan. One release is a hypothesis.

Run the downside as a number rather than carrying it around as a worry. Assume our book has 28 percent of revenue with automotive OEM customers. Assume the 25 percent decline in automotive OEM robot orders passes through to softer physical output at some fraction, and assume the pass-through to our volumes is one third of that, because capital equipment cuts do not map one-for-one onto freight. That is 28 percent times 25 percent times one third, roughly 2.3 percent of revenue exposed over the horizon. Assume a 14 percent contribution margin on that revenue, so margin at risk is about 0.33 percent of total margin.

Small. Now run the electronics side with the sign flipped: assume 22 percent of revenue facing not lost volume but tighter capacity, assume three percent rate inflation on that book, which is about 0.66 percent of revenue in added cost. Both are survivable and both are worth insuring against, and that is exactly what the arithmetic is for. It converts a frightening sector headline into two small exposures with owners, dates and hedges.

There is a physical read in the mix too, which matters if you run a warehouse rather than just moving boxes. About 84.6 percent of what was ordered was not a cobot, and conventional automation arrives with fencing, floor anchors and a commissioning window measured in weeks. Somebody's floor is going to be closed for a while and their inbound schedule will be disrupted inside that window, which is a freight event you can plan for if you know it is coming. If you have customers mid-automation, ask when their commissioning window falls and treat it the way you would treat a plant shutdown: expect appointment slots to change, expect weeks with no receipts at all, and expect a catch-up surge afterwards. Two phone calls in advance and you have saved yourself a month of missed appointments.

One last piece of discipline, because I have watched this go wrong. Do not let a supplier's automation announcement become a negotiating event. If a vendor tells you they are installing robots and lead times will improve as a result, ask for the number and the date, put it in the contract as a service commitment, and keep buying exactly as you did until it shows up in your own receipt data. Announcements are not capacity. The 17,995 figure is somebody's spending, and spending becomes throughput only after installation, commissioning, staffing and a great deal of mundane work that never appears in a press release.

↑ Back to the key points

— By Ravi Chandran