Why some chip technologies make it to production - and others don't

Artificial intelligence (AI) has increased demand for chips across computing, power systems, connectivity and edge devices, noted Aditya Mathur, Founder and Managing Director of venture capital firm Elev8.vc. “Almost every single country wants to have its own technology stack,” he said.

The rise in demand has led to duplicated supply chains as countries try to secure access to fabrication and key materials, instead of using global networks. This makes it harder for new technologies to fit into existing production systems, as companies now need to account for different supply arrangements, which can slow adoption.

The growing complexity also changes how technologies are built and assessed before they reach production. As Carlos Mazure, Chief Strategy Officer at A*STAR Institute of Microelectronics, pointed out, developing new systems now requires a much more interconnected view.

“You can no longer design a new generation of data centres without taking into account the racks, the GPU stacks, the memories, the packaging, the interconnects, the photonics, everything,” he said. 
 

Energy consumption is now one of the major factors that decides if new technologies can be deployed widely. The semiconductor industry has spent decades improving compute performance, but power consumption now limits how far that can go. 

“If you can save just 10% to 20% of a photonic interconnect chip, that would reduce about 4 gigawatts of power, which equates to billions [of dollars in operating costs] in a 20-gigawatt AI data centre,” said Benjamin. 

Energy use at this scale directly affects whether new infrastructure can be built and operated. “You will need new materials, and then you need to have new ways to connect these materials to the devices,” he added. 

This has led companies to prioritise integrating different materials and components within the same chip system to manage energy and performance together. 
 

With energy now a serious constraint on the industry, attention is turning to where power is being lost and how to reduce those losses.

According to Dr Jekaterina Viktorova, CEO and Co-Founder of semiconductor startup Syenta, a key inefficiency is how chips operate when running AI workloads, where they spend much of their time waiting instead of computing.

“During inference, chips are idle 80 per cent of the time,” she said. “Most of the energy of the chips goes to moving data around, and that’s extremely power inefficient.”

This inefficiency is likely to be a source of the next phase of growth, because solving the problem would improve performance and reduce energy use. 

Syenta is tackling this challenge at the manufacturing level through its localised electrochemical manufacturing technology. This process builds interconnects at micron resolution without using lithography at every step. Its goal is to reduce the time chips spend waiting for data. 

The company is also exploring how to support quantum systems and neuromorphic computing, where the way components are linked plays a bigger role than the number of transistors. 

Jekaterina stressed the importance of sustained funding and industry backing to move from a working idea to production. “You need to think through your go-to-market strategy like it’s a war plan,” she said.

Few startup ecosystems provide sufficient risk capital at the earliest stages, and many technologies fail before proving themselves. 

Support at this stage often depends on individuals. “You need friends, you need believers, but you also need a champion,” said Carlos.  

A champion pushes a project forward and gets it tested despite internal resistance. 

Jekaterina credited early backing as a major factor in Syenta’s survival. “When we started, we got extremely lucky meeting a few people like SGInnovate’s Benjamin Lim early on, who believed in our big vision,” she said. 
 

From an investor’s perspective, Aditya said that large semiconductor companies already invest heavily in their own research & development and consider external partners only when they clearly lack internal capabilities. 

“If a large corporate is going to work with a startup, it needs to be in a few areas,” he said. “One is a technology or material that they can’t develop and manufacture themselves.” 

Another is applying AI to chip design and manufacturing, and a third involves improving how high-speed chips communicate, particularly at the edge, where data must move quickly. 

“You need to be 100x better than anyone else and have a completely different approach to be considered relevant,” added Aditya. This sets a high benchmark for adoption. 

Learn more about Singapore’s fast-growing deep tech community and the startups that SGInnovate supports.