There is a company in Singapore that has raised close to half a billion US dollars and has never sold a single thing. Not a small amount of revenue. None.
It has a chip. It has a box. It has a founder story, a serious set of engineers, and one of the bigger names in Singaporean venture writing the cheque. What it does not have is a customer you can name or a price you can look up.
And here is the part that should make you sit up. That is not a scandal. In 2026, that is a business model.
So this one is about a Singapore chip company called Acrab. But Acrab is really just the door. Behind it sits the question every founder and every investor in this region needs an answer to. When a company raises hundreds of millions before it has a product in the market, is that smart money seeing the future early, or is it the dumbest money in history dressed in a lab coat? And underneath that, something more practical. This whole idea of running powerful AI on a small box on your desk instead of in the cloud. Is it real, or is it a slide in a deck?
The box
On 6 August, Acrab closed a $130 million Series B led by Vertex Ventures SEA and India, together with Vertex Growth. Hold that name, Vertex, in your head. We are coming back to it, and it turns out to be the most interesting thing in the whole story.
That round took Acrab past $480 million in total funding. The company was founded in 2024 and only came out of stealth this year. By its own description it has visible paths to industrial deployments and expects revenue within 2026. Read that last part again. Expects revenue within 2026 is a very polite, very well lawyered way of saying we have not made any money yet.
So what did they build? This is where it gets genuinely interesting, and I do not want to be cynical about the engineering, because the engineering looks real. The chip is called GELIX 1. It is a system on a chip built on a 5 nanometre process, which is a serious modern manufacturing node, pairing a 20 core Arm CPU with a dedicated neural processing unit and 273 gigabytes per second of unified memory bandwidth. If none of those words mean anything to you, here is the translation. This is a real piece of high end silicon, not a science project.
Around it they built a product called Agent Box. Think of it as a small computer that sits on your desk and runs AI models locally, on the device, without sending your data off to a data centre somewhere. Acrab says it is designed to support models in the 100 billion parameter class, with persistent memory, multimodal input, and agents that actually do things, all running on the box in front of you. That is the pitch.
Now let me do the part I actually get paid to do, which is not believe the pitch.
The headline number is a benchmark. Acrab claims its chip processed a prompt at 1,416.8 tokens per second, against 188.9 on a Mac Mini with an M4 Pro. Roughly seven and a half times faster than a high end Apple machine. Sounds incredible. It might be. But let me tell you what that number actually is, because this is where you earn the right to have an opinion.
That measurement is something called prefill. Prefill is the speed at which the machine reads and digests your question before it starts answering. It is the reading phase, not the writing phase. It is genuinely useful, and it is exactly one half of the job. The other half, where the machine generates the answer token by token, is called decode.
Acrab has not published its decode rate. It has not published time to first token. It has not published power consumption. It has not published thermal data. And crucially, nobody independent has verified any of it. Every one of these numbers comes from Acrab. Even the tech press covering the launch flagged this openly. The claim is prefill only. It is not proof the chip is seven and a half times faster at everything. It is proof it is fast at one specific thing, in one specific test, that the company designed.
And the 100 billion parameter claim, the one doing the work in every headline, notice the wording. Designed to support. Not runs. Not we measured it running a 100 billion parameter model at this speed while drawing this many watts. Designed to support is a target, not a result.
I am not saying anyone is lying, and I want to be clear about that. I am saying that as of today, GELIX 1 is in validation. It is chasing its first industrial adoption. The first market is the smart cockpit inside cars. No named customer, no public price, revenue as a forecast.
So here is your honest scorecard on the product. Real chip, real team, one genuinely fast benchmark on the reading half of the task, and the entire set of numbers that would tell you whether this actually works quietly missing.
Half a billion for nothing is now normal
Before you laugh at half a billion dollars for a company with no revenue, look at what is happening everywhere else. Acrab is not the strange one. Acrab is the small one.
San Francisco. Safe Superintelligence, founded by Ilya Sutskever, formerly chief scientist at OpenAI, raised $2 billion at a $32 billion valuation. Around 50 employees. No product. No published research. And it has said out loud, on purpose, that it does not intend to ship anything until its mission is complete. Nvidia is one of its backers. Two billion dollars, fifty people, nothing to sell, by design.
Thinking Machines, founded by Mira Murati, formerly chief technology officer at OpenAI, raised roughly $2 billion at a $12 billion valuation, then another $5 billion at $50 billion. To its credit it did finally ship a model in July. But it raised the first two billion on a founder name and a thesis, not on a product.
And this is not fringe. In 2025, OpenAI and Anthropic between them absorbed around 14 percent of every venture dollar invested anywhere on earth. Anthropic was valued at $965 billion by May this year. The money is not spread out. It is piling into a handful of bets on faith.
That is America. Here is where I actually wanted to take this, because Asia is running the same playbook with local accents, and almost nobody here describes it that way.
Tokyo. Sakana AI raised $135 million at a $2.65 billion valuation, making it the most valuable startup in Japan. And here is the elegant part. Sakana’s entire philosophy is small. Cheap, efficient models trained on small datasets, tuned to the Japanese language and Japanese business. One of its founders, Llion Jones, co-wrote the original research paper that created the modern AI boom. Rich valuation, thin revenue, sold on pedigree and a thesis that says bigger is not always better. Sit with that, because it is the same bet Acrab is making. The anti-giant bet. Do more with less.
Beijing. Moonshot AI raised $2 billion at a $20 billion valuation and is reportedly chasing $30 billion. But here is the twist, and this is the one that should sting. Moonshot is not pre-revenue. Its annualised recurring revenue passed $200 million in April. Its rivals Zhipu and MiniMax already listed in Hong Kong in January. The Chinese champions are not raising on faith. They are raising, making money, and ringing the opening bell.
Bangalore. India has decided AI is a sovereignty question. Sarvam raised $234 million at a $1.5 billion valuation, led by the Indian IT giant HCLTech. And Ola’s founder Bhavish Aggarwal is building Krutrim, which is not just making models but building its own chips and its own computing infrastructure. That is a direct echo of what Acrab is doing, wrapped in the flag of national self-reliance.
Put the map together. Silicon Valley is doing pre-product on pure faith. Japan is pricing efficiency on pedigree. China is already making money and ringing the bell. India is treating this as national security. And Southeast Asia’s entry into the contest is a box from Singapore that runs AI models on your desk.
Pre-revenue is not pre-product
One distinction, and it is the difference between a lazy take on this and a correct one.
Pre-revenue is not the same as pre-product. Safe Superintelligence has literally nothing to ship, deliberately. Acrab has a chip and it has a box. And a chip company is not a model company. Silicon costs a fortune before you make a single dollar, because you have to design it and manufacture it long before anyone can buy it. That is the nature of hardware, not a warning sign.
So if you are going to judge Acrab, do not judge it against a software lab in California. Judge it against Sakana in Tokyo and Krutrim in Bangalore. Judge it against the other people trying to build hardware and sovereignty in Asia. That is the fairer fight.
Where the engine was built
So who runs it? The chief executive is Dr Ken Phua, the former co-chief executive of Arm China. He spent two decades at Arm and stepped in to co-run its China business in 2020, during a very messy boardroom fight.
Let that land. The technical spine of this Singaporean deep tech contender was built at the top of Arm China.
I want to be careful and fair here, because this is exactly the kind of thing this show exists to say out loud without turning it into a smear. That is a real, world class credential. It is not a red flag on its own. But it is the same thread we pulled on with Manus and with DayOne. When something is wearing a Singapore jersey, it is always worth asking where the engine was actually built. Here, the engine has an Arm China lineage.
There is also a small mystery I could not fully close. A couple of outlets name someone called Tian Qi as the 2024 founder, and that name appears almost nowhere else, while Phua is the public face. I am flagging it, not asserting it. If you know who else is in there, tell me.
So is edge AI real, or is it a slide?
Now the question the whole thing hangs on, in the plainest language I can manage.
Edge AI just means running the model on the device in front of you, instead of sending your request over the internet to a giant data centre and waiting for the answer to bounce back. On the box, not in the cloud. That is the whole idea.
Why would you want that? Four reasons, and they are all real. Speed, because there is no round trip to a server. Privacy, because your data never leaves the building. Cost, because you are not paying a cloud bill every time you ask a question. And it works when the internet does not.
This is a real market, not a fantasy. Edge AI chips are worth somewhere around $4.4 billion this year, growing more than 20 percent a year. Nvidia already holds close to 40 percent of edge AI compute revenue, and it went and bought Groq, one of the most interesting inference chip companies going, to own more of it. Serious people are putting serious money here.
But here is the crack in the Acrab story that no press release is going to point at.
The sweet spot for running AI on a device is small models. The industry even has a nickname for it, the Goldilocks zone: roughly seven to thirteen billion parameters, with a bit of clever compression. With four bit quantisation, a four billion parameter model fits in under a couple of gigabytes and runs on your phone. And these small models are shockingly good. There is a documented case of a fine tuned model with under three billion parameters beating a model more than two hundred times its size on a narrow enterprise task. Gartner expects that within a couple of years companies will use small, task specific models around three times more than giant general purpose ones.
So the direction of travel is clear. On device, small and specialised beats big and general.
Now look back at Acrab’s headline. One hundred billion parameters, on a box. That is an order of magnitude beyond where the rest of the field says the sweet spot sits.
So one of two things is true. Either Acrab has a genuine architectural edge that lets it do something the rest of the industry cannot, in which case show me the decode rate, show me the watts, let someone independent test it, and I will be the first one standing up and clapping. Or the 100 billion number is marketing running out ahead of the physics. Right now, on the numbers they have released, I cannot tell you which. And neither can they.
Why this region is the right market for it
Let me make the constructive case, because I do not want to leave you with only the skepticism.
Southeast Asia might be the best market on the planet for edge AI, for reasons that have nothing to do with Silicon Valley.
Data sovereignty is becoming law here. 89 percent of Indonesian enterprises now say keeping their data in their own country is a key factor in technology decisions, and you hear the same thing in Malaysia, the Philippines and Thailand. ASEAN is actively pushing on it. Meanwhile the region runs on patchy connectivity and expensive cloud bills.
Put those together. Data that must stay onshore, internet you cannot always trust, and a cloud invoice that hurts. Running the model locally on a box stops being a gimmick and starts being the obvious answer. That is not a California idea imported here. That is a genuine regional fit.
The tell is on the cap table
Which brings me back to the name I told you to hold on to. Vertex.
Vertex, the venture firm leading this round, is a wholly owned subsidiary of Temasek. And Temasek is owned by the Government of Singapore. Peel back one layer and the marquee backer of this deal is effectively the Singapore state.
And the state is not being shy about it. In this year’s budget Singapore committed S$800 million specifically to semiconductor research, sitting inside a S$37 billion national research plan, and it is merging its investment arms to fund deep tech.
So let me reframe the whole thing as bluntly as I can. Acrab is not really a startup story. It is industrial policy wearing a startup jersey.
Singapore has looked at a world where AI is a matter of national power and chips are the new oil, and it has decided it wants to own a piece of the silicon. So it is putting state money, through a state owned fund, into a chip company run by the former co-chief executive of Arm China, at the exact moment it is pouring billions into semiconductors nationally. This is a country building a strategic asset and calling it a Series B round.
And honestly, once you see it that way, it is a better story and a more defensible one. This is not foreign intellectual property quietly routing through Singapore to dodge the geopolitics and sell to American buyers, which is a pattern we have criticised here before. This is the opposite. This is Singapore deliberately hiring Arm China talent to build something it intends to keep.
So the honest question stops being is Acrab a real startup. It clearly is a real engineering effort. The honest question becomes this. Is it a smart sovereign bet, or is the state paying startup prices and startup multiples for a national champion it wants for reasons that are not purely financial?
My actual take
For whatever it is worth, from someone who writes small cheques for a living and not state sized ones.
The category is very, very real. Small models on real devices work today, and Southeast Asia has better reasons to want them than almost anywhere else. So I am not bearish on the idea.
What I am cautious about is the specific claim and the specific price. Half a billion dollars, no revenue, a benchmark that measures half the job, a flagship number sitting well outside where the rest of the field says the physics work, and a cap table that traces back to a government. None of that is disqualifying on its own. All of it together means you keep your applause ready and your hands still until the independent numbers show up.
Applaud that Southeast Asia finally has a serious hardware bet with real talent and real ambition behind it. Then ask them for the decode rate, the power draw, and the name of one paying customer. If those show up and they are good, this is one of the most important companies in the region. If they keep not showing up, what you have is a very expensive, very well credentialed box that a government wanted to exist.
Here is the one line I want you to take into your next investor meeting or your next board call. Pre-revenue is not pre-product, and a benchmark you cannot reproduce is not a benchmark. It is a poster.
The whole world, from San Francisco to Tokyo to Beijing to Bangalore to right here in Singapore, is spending enormous amounts of money on artificial intelligence before the revenue shows up. Some of that will look like genius in five years. A lot of it will look like the most expensive lesson a generation of investors ever paid for. Your job, and my job, is to tell the difference before the money moves, not after.
Acrab might be the real thing. The Singapore government clearly thinks the category is worth owning. I just want to see the other half of the numbers before I call it a winner.
If this was useful, send it to one founder or one investor who needs it. And if you know who actually founded this company, tell me. I read every reply.
A benchmark you cannot reproduce is a poster.
Real. Raw. Relatable.
... --- ...
