What’s stopping Asia’s chipmakers from taking on Nvidia? The answer lies in this bottleneck

NVIDIA’S rise from video game industry stalwart to the world’s most valuable public company has had some unintended consequences. For one, the semiconductor industry now seems to generate almost as much Fomo as it does revenue.

“There is a lot of buzz, and everyone wants a share of it,” says Ankush Wadhera, managing director, partner, and India leader of the semiconductor practice at Boston Consulting Group (BCG).

This transformation has led to thousands of artificial intelligence chip startups across Asia, many founded by veterans of older, pre-AI companies such as Intel and AMD.

Like Nvidia, all of these startups are fabless, meaning they design chips but leave the manufacturing to specialist foundries.

But while there are thousands of startups each looking to build the next breakthrough chip in AI, there are only a few major fabrication facilities, such as TSMC and Samsung Foundry, the chipmaking division of Samsung Electronics.

For a startup trying to build at the frontier, a relationship with one of these foundries can matter as much as their chips, and these companies are now fighting to prove they have what it takes to get into these fabs.

A path beyond Nvidia

For firms looking at fabs with more advanced nodes – a term that describes a generation of chipmaking technology – TSMC is still pretty much the market leader in that space, says Chan Yip Pang, executive director at VC firm Vertex Ventures. Getting an allocation at TSMC is also typically not so easy, adds Pang.

To put nodes in perspective, a 3 nm node is considered more advanced compared to a 15 nm node, as the former fits transistors closer together, which usually makes a chip faster and more power-efficient.

This also makes them much harder and more expensive to manufacture.

Despite the bottleneck caused by the lack of foundries, investors are still backing AI chip startups. Pang, for one, has invested in three semiconductor startups in South-east Asia.

South Korea’s FuriosaAI, founded in 2017, is one of the few startups in Asia to have secured production for its advanced AI chips.

Alex Liu, senior vice-president of product and business at Furiosa, admits to Tech in Asia that the entire process was “very complicated,” requiring the startup to obtain TSMC capacity and gain access to high-bandwidth memory as well as packaging partners and customer validation all at the same time.

So far, the company has raised US$246 million and has shipped products to customers and partners including Samsung and LG.

Furiosa’s flagship AI inference chip, called RNGD, entered mass production in January 2026. Fabricated on TSMC’s 5 nm process, its initial batch of 4,000 units was delivered with assembly partner Asus, a Taiwanese manufacturer that now makes AI hardware.

Furiosa plans to produce another 16,000 units this year, taking its total 2026 output to 20,000 chips. The cost of a single RNGD chip is estimated at US$10,000.

Being at the frontier

Bengaluru-based Agrani Labs, led by former Intel and AMD executives, is among the few startups in India and South-east Asia trying to build cutting-edge AI chips like Furiosa.

The company has raised US$8 million from Peak XV Partners. Agrani’s CEO, Dheemanth Nagaraj, is reportedly in talks to raise more than US$100 million. Nagaraj did not respond to multiple requests for comment.

Investors tell Tech in Asia that the CEO and his team are among the strongest chip design talent in Bengaluru, where an estimated 150,000 engineers work for US semiconductor companies such as Intel and AMD.

But even with US$100 million and a strong chip design, it may not be enough for the company to secure TSMC.

“At the leading edge nodes, even a small tape-out and validated initial volume cost a lot,” says Arjun Rao, general partner at deeptech VC firm Speciale Invest, which backed chip design company Morphing Machines. “Capacity access is a barrier for young startups given the large AI compute build out globally.”

Tape-out is when a finished chip design is sent to the factory to be made.

For its part, TSMC has “dedicated resources to work with emerging customers,” helping startups turn their chip designs into products that can be manufactured at volume, it told Tech in Asia in a statement. It also provides technical and business guidance, including on production schedules and technology road maps.

Need for powerful friends

With all that said, foundry access is just one constraint that startups are facing. AI chips need high-bandwidth memory, or HBM, so that data can be moved quickly enough through the chip to run large AI models.

A bottleneck in HBM chips – also due to a shortage of manufacturing facilities – has caused prices for memory chips to spike.

Startups cannot solve this memory constraint alone, and many have been seeking to build partnerships with memory suppliers.

For AI chip startups, a relationship with an HBM supplier can be crucial. SK Hynix, Samsung, and Micron dominate the market for the high-bandwidth memory used in AI chips.

Accessing “HBM is so important,” says Furiosa’s Liu. “Those are the conditions that you need to win,” though he says it is not about how much money you have. SK Hynix, for instance, makes it very clear it will not just support any startup.

That is a problem especially since that support is a prerequisite for many of the next steps to getting a company’s chips to clients or on the cloud.

TSMC’s CoWoS, the advanced packaging that puts AI chips and high-bandwidth memory together into one chip package, is the best-known example of such advanced packaging. It pairs Nvidia’s chips with HBM.

But TSMC can only produce so many CoWoS packaging at a time, another bottleneck for companies trying to build high-end AI systems.

Singapore-based Silicon Box aims to ease this obstacle. It uses advanced panel-level packaging to connect the AI processor, high-bandwidth memory and other components into a single chip package.

But before Silicon Box will engage with AI-chip startups, Mike Han, chief revenue officer of Silicon Box, says they need to fulfil several criteria.

According to Han, Silicon Box assesses a startup’s technical team, its founders’ track record, whether it has support from foundry, memory, and other supply-chain partners, as well as credible customer demand.

“We evaluate the technical capability of their team first and then the track record of their founding team,” he notes. “Whether they’ve been able to be successful in the past is a key factor.”

For Furiosa, courting customers did not begin after the engineering was complete; it began alongside this process.

“When you start to design the chip, you need to start to talk to the potential customers to understand the market trends,” says senior vice-president Liu. Those conversations went through multiple rounds as the company moved toward production, he says.

Creating national champions

There is also a difficult commercial question that startups need to address when gauging demand: Why would an AI lab or cloud provider buy a startup’s chip over Nvidia’s? Liu says cost and sovereignty issues are among the most important factors.

“As Furiosa, we are non-US, non-China technology, so we are neutral to everyone,” he says. “That reduces a lot of the political risk as well.”

Since chips have become central to geopolitical competition, governments are increasingly seeking their own alternatives to US and Chinese suppliers, says Rao of Speciale Invest.

“India should have a few strategic AI chip alternatives, not just one,” he says. “We’re building data centres, we’ll need the compute, so having homegrown alternatives is a good idea in the medium to long term.”

Pang of Vertex Ventures says that in South Korea, government support has complemented an established memory and semiconductor ecosystem to help produce companies such as Furiosa and Rebellions.

AI chip startups in India and South-east Asia, many of whom have been boxed out of the competition with government backed entities, are taking a less capital-intensive route.

As they struggle to access high-end fabs, many are focusing on designing ASICs, application-specific integrated circuits, for edge AI or particular workloads.

Edge AI chips can power real-time tasks for drones, robots, CCTV cameras, industrial machines, EVs and autonomous vehicles without sending data to a distant data centre.

Such chips may not need the most advanced manufacturing capacity or the same volumes of scarce high-bandwidth memory as a frontier AI data-centre chip. The talent needed to design such chips is more widely available and less expensive. Startups also have more ways to differentiate by building for specific uses, unlike in the crowded market for data-centre inference chips.

From edge to cutting-edge

Singapore-headquartered OptoML is building such chips for CCTV cameras, drones and robots. They are deployed on devices, on premises or at the network gateway, where they can reduce latency and compute costs for enterprises.

Saravana Maruthamuthu, CEO and co-founder of OptoML, says the company is currently piloting its chip with manufacturers, including Chennai-based Murugappa Group and Wheels India.

OptoML’s current chips use a 12 nm process node, an older manufacturing technology that has been used at scale for years. It is generally cheaper and easier to access than the 5 nm process that Furiosa uses.

Startups working with older chip generations have more manufacturing options. “If you are talking about more mature nodes, actually in Malaysia, there’s a foundry called Silterra,” says Pang of Vertex Ventures. “They are able to do the more mature nodes.”

South-east Asia is looking at six new foundries using these types of older technology by 2029.

India is also building out parts of its semiconductor supply chain, including fabrication, assembly, and testing.

The government’s India Semiconductor Mission, for instance, has backed several projects, while Tata Electronics is developing a fab in Dholera, Gujarat, that is expected to manufacture chips at nodes ranging from 28 to 110 nm.

Pang says it is unlikely that South-east Asia will produce a Groq- or Cerebras-style company in the near term, given the region’s funding constraints. “I’d expect South-east Asian chip startups to cluster around smaller, cheaper, more niche categories instead, which can still be sizeable,” he notes.

For some, a niche is only the start. OptoML is working toward a 5 nm data-centre inference chip.

“The hyperscaler is the ultimate goal that we want to go into,” CEO Maruthamuthu says. “But we are starting with where we can also deploy, build commercial traction and then move and reach it.” TECH IN ASIA