Africa's Try-On Pitch Assumes the Megabytes Are Free
Augmented try-on is being sold into African e-commerce as the cure for buyer distrust, but a 4 MB garment model outweighs the entire median mobile page in the region with the world's least affordable data. The cost that decides these markets is the van, not the render.
Parallax Pincer
The dress falls correctly now. The shoulder seam sits where a shoulder is, the hem swings a half-beat behind the leg, and the fabric reads as fabric rather than the shrink-wrapped vinyl that virtual try-on was producing five years ago. Then weigh it. Shopify’s brief for the studios that build its merchants’ 3D product models says “the total file size should be about 4 MB”, which makes one garment heavier than the entire median mobile home page HTTP Archive measured in July 2025, at 2,362 KB.
That garment is the heaviest object on the page, and in African e-commerce the shopper pays for it by the megabyte. Augmented try-on is being sold into Nigeria, Kenya and South Africa as the fix for a trust problem: buyers who will not commit to clothes they cannot handle. What buyers there distrust is whether the box arrives at all. The picture got cheap years ago; the delivery never did.
Shopify runs the pitch on its Nigerian domain. The explainer there credits AR try-on with lifting browsing and sales “by almost 20%” and cutting returns “up to 64%”. Both figures come from Perfitly, a company that sells virtual fitting software. File size appears once in the copy, as a caution that the model “has to balance detail and a small file size to avoid lag”. Lag is the wrong worry.
Africa has the least affordable mobile data in the world. The ITU’s 2025 Africa report prices the median entry-level mobile broadband plan, 2 GB a month, at 4.2 per cent of gross national income per capita in 2024, against the UN Broadband Commission’s two per cent target and a global median of 1.4 per cent. Nigeria’s regulator cleared a 50 per cent tariff rise, after which MTN’s 15 GB weekly bundle went from ₦2,000 to ₦6,000, a tripling Africanews reported as exceeding the cap. A shopper who views forty garments in AR spends 160 megabytes deciding not to buy anything.
Speed is not the constraint, which is what makes the assumption careless rather than wrong. Ookla’s 2025 medians, reported in February by tech.africa, put Nigeria at 44.14 Mbps and Kenya at 45.37, with South Africa at 65.7. Those networks will carry four megabytes without complaint. They will also bill for it, in the region the ITU records as having the world’s lowest internet-use penetration, 36 per cent against a global average of 74 per cent.
The picture has never been the expensive part of selling clothes to people who cannot touch them.
Montgomery Ward tested that in 1872 with a single sheet listing 163 items, and the engravings that followed stayed crude for decades, incapable of showing how a cloth hung. What made the apparel pages work on farms a week’s post from Chicago was the money-back guarantee added in 1875. The trust was underwritten rather than rendered. Every mail-order house that came after copied the guarantee before it copied the artwork.
Jumia’s accounts show what that underwriting still costs. The platform closed 2025 with $818.6 million in GMV across 22.6 million physical-goods orders, a $60.1 million pre-tax loss, and fulfilment expense of $1.97 per order in the fourth quarter, down 12 per cent year on year and still the largest line item tied directly to physical delivery. Upcountry regions supplied 61 per cent of fourth-quarter orders, up from 56 per cent, which suggests the average delivery is lengthening rather than shortening. Jumia no longer publishes a returns rate; the last clean figure, in its 2019 IPO prospectus and revisited by TechCrunch, put failed deliveries and returns at 14.4 per cent of 2018 GMV.
A vendor claiming 64 per cent fewer returns is claiming to move the one line item that decides whether African e-commerce clears breakeven, and no platform operating in these markets has published that result. The render is close to free now, pulled as open weights off a hub that Nvidia has just bought, as we reported earlier today; in May we argued the try-on models were already good enough and the bottleneck had moved to catalogue metadata. What would earn the megabytes is a fit prediction accurate enough to send the van out once. Until someone publishes that number, the four-megabyte dress is a well-cut costume billed to the buyer.