The slowest and most capital-hungry thing a Singapore founder could build has just raised another US$130 million.
Acrab, headquartered in Singapore and founded in 2024, made the Series B public on 6 August, with Vertex Ventures SEA & India and Vertex Growth returning alongside new institutional money from Europe and the region. The round takes what the company has raised past US$480 million, US$350 million of it in June alone.
It expects its first revenue some time this year.
What the money buys is a chip called GΞLIX 1 and a device called Agent Box, a personal edge system that runs large-model inference, persistent memory and agent orchestration on the machine in front of the user rather than in somebody else’s data centre. Purpose-built silicon, edge models, and the software that schedules them.
Three layers, all of them hard, none of them fast.
That is a strange place for the region’s biggest recent cheque to land. Southeast Asian venture funding has stayed thin through 2026, with investors turning selective after three years of contraction, and the money that does move has gravitated to agentic AI and generative AI rather than to fabrication.
The lesson sitting inside the round is about moats, and it belongs to founders who will never etch anything.
Consider what happened to the layer most operators actually built on. Chinese model providers now price their APIs at 10 to 20 percent of what foreign alternatives charge, on UBS estimates, and train comparable systems for roughly a tenth of the cost. They hold API gross margins of 20 to 40 percent while doing it, on GPU utilisation above 70 percent against an industry figure closer to half.
The rented layer gets cheaper every quarter, and nobody has yet found the floor.
The rented layer gets cheaper every quarter, and nobody has yet found the floor.
A company whose entire product is a prompt, a login and a monthly invoice sits in the path of that curve. Each price cut in Hangzhou removes a slice of its margin, and the founder finds out from a changelog.
So the question worth putting to each line of a business this month is narrow, and it answers in one sentence. If the model underneath went to zero tomorrow, what remains that a competitor cannot buy in an afternoon?
For some operators the honest answer is a supplier list built over four monsoons, or a Johor customs lane with the officers’ names in it, or three years of who-clicked-with-whom that no one else has collected. For others the honest answer is nothing, which is a useful thing to learn on a Tuesday rather than in a board meeting.
Acrab’s answer is the most expensive one on the menu and the most fragile in the near term, because a company with a chip still to qualify and no revenue booked has a burn rate its backers will keep having to feed.1
The shape of that answer, though, costs nothing to copy.
Vertex paid US$130 million for the one layer of the stack that nobody else can re-price by press release.
Vertex paid US$130 million for the one layer of the stack that nobody else can re-price by press release.
Footnotes
-
The June round was US$350 million. The August round was US$130 million. The direction of that sequence is the thing to watch, not the sum. ↩