AGIBOT Did Not Beat Unitree. Read the Table Again.

The headline everyone ran in August was "AGIBOT overtakes Unitree." Two trackers published H1 2026 humanoid shipment tables within days of each other, and both put AGIBOT first, Unitree second. The overtake became the story.

The overtake is the least interesting thing in either table. Smart Analytics Global and Counterpoint Research used different methods and different form-factor rules, and they still landed on the same five companies in the same order. That agreement is the durable output. Everything below the rank column is where the two tables come apart, and that is where the reading skill lives.

A shipment table is one ledger. It records how many units left a vendor's building, as estimated by an outside analyst. Two more ledgers sit underneath it: where those units are working, and who paid what for them. The scoreboard shows rank. The ledgers show what a "unit" actually is, and for the H1 2026 tables the three ledgers disagree with each other more than the trackers disagree among themselves.

Two trackers, one order

Smart Analytics Global (SAG) counts about 19,100 humanoids shipped worldwide in the first half of 2026, up 272 percent year on year (https://www.techrepublic.com/article/news-humanoid-robot-shipments-chinese-vendors/). Counterpoint counts more than 22,000, up about 300 percent (https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026). That is roughly a 15 percent gap in absolute units. On AGIBOT specifically the gap is about 8,400 (SAG) versus about 9,700 (Counterpoint); on Unitree it is about 5,900 versus more than 7,000.

The order does not move.

*SAG remainder figures via Humanoids Daily, not a primary SAG release (https://www.humanoidsdaily.com/news/global-humanoid-shipments-surge-272-in-1h-2026-as-agibot-overtakes-unitree).

Don't blend the two columns. Galbot at 900 and Galbot at 1,100 are two estimates of the same unknown, and neither vendor has filed a number that would settle it. Axis Intelligence notes that no top-five vendor has published H1 2026 unit counts; every number in the table is an outside estimate (https://axis-intelligence.com/china-humanoid-robot-companies/).

A 15 percent spread between trackers is normal for a market with no filed unit data. Forbes ran the range wider still, framing the half at "up to 30,000" (https://www.forbes.com/sites/johnkoetsier/2026/08/20/humanoid-robot-shipments-up-300-up-to-30000-so-far-in-2026/). Forbes published no methodology behind that figure, so it doesn't belong in the table. It does tell you how wide the range gets once you step away from the two firms that showed their work.

Both tables are right about the order and neither can be right about the total, because neither is counting the same object.

The first ledger: shipped, and what counts as a unit

"Shipped" here means a tracker's estimate of units that left a vendor. Not customer-confirmed, and not paid for.

The bigger problem is the object being counted. AGIBOT runs at least three product lines that get lumped under "humanoid": the A-series biped, the X-series compact platform that feeds rental and education channels, and the G-series wheeled base. Unitree says it counts bipeds only, and has claimed about 18,000 cumulative bipeds excluding wheeled units (https://www.humanoidsdaily.com/news/global-humanoid-shipments-surpass-22-000-in-h1-2026-as-agibot-leads-and-form-factor-debates-intensify). Galbot, third on both tables, ships mostly wheeled platforms.

So the number-one vendor's count includes categories the number-two vendor excludes by its own rule, and the number-three vendor sits almost entirely in a category number two doesn't count at all. Whether that makes AGIBOT's lead larger or smaller depends on which rule you adopt, and the trackers don't publish the split.

Reading rule: before you compare two vendors' unit counts, find out which form factors the tracker included and whether the vendor's own counting convention differs. If you can't find out, the comparison is between two definitions rather than two companies.

The second ledger: deployed, and the mix problem

Here the trackers stop disagreeing at the margin and start disagreeing about what kind of market this is.

SAG says industrial and commercial applications took more than 70 percent of H1 shipments (https://www.techrepublic.com/article/news-humanoid-robot-shipments-chinese-vendors/). Counterpoint's breakdown: entertainment 33.6 percent, data collection and research 27.0 percent, service 19.0 percent, manufacturing 12.8 percent, warehousing 4.9 percent (https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026). On Counterpoint's ledger, entertainment plus research comes to over 60 percent of the half while manufacturing plus warehousing comes to under 18, which reads as a different market from the one SAG describes.

Those two pictures cannot describe the same 20,000 robots under the same definitions. The likelier explanation is definitional rather than factual. "Commercial" in SAG's taxonomy probably absorbs much of what Counterpoint files under entertainment and service; a robot doing a trade-show demo is a commercial sale to somebody. And a research lab buying twenty units for data collection is a commercial customer too. I'd rather say the two firms drew their category lines in different places than declare one of them wrong.

But that leaves a reader with no reliable answer to the question that matters, which is how many of these are doing work someone would pay for again. Inside the top five, the clearest deployment stories are UBTECH's Walker S line going into automotive plants and Leju's push into data-center operations. For AGIBOT and Unitree, the volume leaders, the trackers give shares and no site-level detail.

Reading rule: a shipped unit is not a working unit until someone says where it works, and the two trackers disagree on that more than on anything else.

The third ledger: revenue, and the 4.8x spread

Only one ledger has audited numbers, and they're for FY2025, not H1 2026. Use them anyway. They are the only place a reader can see what a "unit" is worth.

UBTECH reported FY2025 humanoid revenue of RMB 820.557 million on 1,079 units (https://axis-intelligence.com/china-humanoid-robot-companies/). That works out to about RMB 760,000 per robot. Axis puts UBTECH's stated capacity above 6,000 units a year, which implies utilization somewhere around 18 percent. That utilization figure is derived from those two numbers, not reported by the company, so hold it loosely.

Unitree reported FY2025 revenue of RMB 868 million on more than 5,500 units, was profitable, and has a STAR Market IPO in motion. Because the unit figure is a floor, per-unit revenue is a ceiling: at most about RMB 158,000 per robot, and likely less.

Two companies, same top five, and a spread of roughly 4.8x in revenue per unit. On the shipment ledger UBTECH's 700 to 1,000 units look like a rounding error next to Unitree's roughly 6,000, yet on the revenue ledger their FY2025 totals sit within 6 percent of each other, so the table most readers see is the one that hides the difference.

Reading rule: identical unit counts can represent businesses that differ by a factor of five in money. The shipment table hides this entirely, and it hides it most where the gap is largest.

AGIBOT vs Unitree through all three ledgers

Run AGIBOT and Unitree through the three ledgers and the same two names come out ranked differently each time.

Shipped ledger: AGIBOT leads on both trackers, 44 percent to about 31 on SAG, more than 43 to about 31 on Counterpoint. The SCMP framed this as AGIBOT taking the top spot ahead of its own IPO (https://www.scmp.com/tech/tech-trends/article/3363544/agibot-overtakes-unitree-top-global-humanoid-robot-vendor-first-half-amid-ipo-push). But AGIBOT's total includes wheeled G-series and compact X-series units that Unitree, by its stated convention, wouldn't count. Strip those and the gap narrows by an amount nobody outside the two trackers can compute.

Deployed ledger: unknown for both at H1 2026, with only circumstantial evidence to go on. AGIBOT's rental and compact X-series volume sits in exactly the segments Counterpoint's application split assigns to entertainment and research rather than industrial or commercial work (https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026). Unitree's price point leads to the same reading, since a per-unit ceiling around RMB 158,000 is a university-lab and content-studio number, not a factory-floor number. Neither company publishes a site list, and neither tracker fills the gap. The honest entry for both is a blank.

Revenue ledger: Unitree has audited FY2025 revenue, a stated unit floor, and reported profitability. AGIBOT has an IPO filing in progress, which means audited figures are coming, but they are not in front of us today (https://www.scmp.com/tech/tech-trends/article/3363544/agibot-overtakes-unitree-top-global-humanoid-robot-vendor-first-half-amid-ipo-push). On this ledger Unitree is the only one of the two with a number, and a company with a number outranks a company without one by default.

So the headline holds up in a narrow sense. "AGIBOT beat Unitree in H1 2026" is true on the shipped ledger, unverifiable on the deployed ledger, and unknowable on the revenue ledger until the prospectus lands. That takes nothing away from AGIBOT. It describes how much a single scoreboard can carry.

Reading rule: when a ranking flips depending on which ledger you open, the ranking is telling you about the ledger, not the companies.

What the full-year guesses are worth

Two trackers, two full-year numbers, and one piece of company guidance. SAG projects roughly 60,000 humanoids shipped in 2026 at about $1.6 billion in value (https://www.techrepublic.com/article/news-humanoid-robot-shipments-chinese-vendors/). Counterpoint says more than 50,000 (https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026). UBTECH has guided to about 5,000 units of its own for the year (https://axis-intelligence.com/china-humanoid-robot-companies/).

Divide SAG's two figures and the market averages roughly $27,000 per robot. That is an inference from a projection, not a reported figure, and it deserves the same caution as the utilization number earlier. But it is instructive. A $27,000 average is a Unitree-shaped market rather than a UBTECH-shaped one, where FY2025 units cleared RMB 760,000 each. If SAG's value estimate is right, the 2026 fleet is dominated by low-priced units going to buyers who are not automotive plants.

UBTECH's 5,000-unit guidance cuts the other way. Hit it, and UBTECH alone would be 8 to 10 percent of either tracker's full-year total at a per-unit value several times the market average. Miss it, and the H1 pace of 700 to 1,000 units suggests a full year closer to 2,000. The company's own number is the most testable forecast on the page, and the one most likely to be checked against reality by next spring.

Three tests before you quote a shipment number

First, check the rank before the number. SAG and Counterpoint disagree on total volume by a wide margin and agree on order at the top. If a claim survives the swap between trackers it is about the market, and if it collapses when you switch from SAG's table to Counterpoint's it is about the methodology, which is worth knowing before it goes in a headline.

Second, ask what a unit is. Wheeled, compact, kit, rental. The word "humanoid" is doing category work that the shipment tables do not show. Unitree's convention would shrink AGIBOT's total by an unknown amount. Anyone quoting the 44 percent should be able to say whether they mean the AGIBOT definition or the Unitree one.

Third, divide revenue by units wherever audited revenue exists. It exists for two of the top five today, and it will exist for a third once AGIBOT files. The 4.8x spread between UBTECH and Unitree is the single most useful figure in this piece, and it appears nowhere in either tracker.

The H1 2026 scoreboard was never wrong. AGIBOT shipped the most. Unitree shipped the second most. Chinese vendors took the top five. All of that is likely true. The scoreboard was just one ledger of three, and the other two are where the money and the work are recorded.