Semiconductors, just like the AI workloads they power, are exponentially evolving. Take Blackwell, Nvidia’s highly anticipated next-generation GPU, which is marketed as being able to run AI models at 25 times lower costs than the previous Nvidia H100 series.
But advancements in technology are happening at such a rate that current semiconductors could become obsolete in 3-5 years, that’s the view of Dr. Seth Dobrin, former IBM Chief AI Officer, and now founder and CEO of Qantm AI and General Partner at 1infinity Ventures.
Subscribe today for free
Speaking at Metro Connect 2025 in Fort Lauderdale, Dobrin highlighted that several innovative startups are creating technologies not only set to rival semiconductor stalwarts like Nvidia, AMD, and Intel but also consign their current technology to the annals of history.
Dorbin highlighted emerging technologies ranging from photonics, which uses electrons to process and transmit data at far higher speeds, to plasmonics, which uses surface plasmons (oscillations of free electrons at the interface between materials) to manipulate and transmit data at nanoscales. These technologies are set to provide even greater miniaturisation and efficiency in powering both computing and communications.
Another potentially game-changing concept Dobrin discussed was neuromorphic chips, where a neural network is integrated directly into a semiconductor, effectively mimicking the potential cognitive action of the human brain.
Dobrin outlined that he was aware of a handful of startups that were headway with spiking neural network technologies, with their added benefit being that they don’t require the two to three-nanometer fabs coveted by current hardware players that take considerable levels of intricacy to produce.
“They can use fabs that everyone else is abandoning, so that's why I say three to five years, as they don't need to build new fabs. They can take over fabs that other people are abandoning,” Dobrin said.
These concepts are still emerging at varying degrees. Photonics alone has already seen millions of dollars pumped into startups like Oriole Networks, Lightmatter, and Ephos as they work to develop next-gen chips that leverage light.
One startup Dobrin referenced on the neuromorphic side was Conscium, which is led by Daniel Hulme of Satalia-fame. The team there are attempting to create neuromorphic computing solutions using both hardware and software to simulate the brain's neural and synaptic structures.
But it’s not just startups looking to shake things up. Just last week, Microsoft unveiled Majorana 1, a quantum computing chip that leverages a new kind of architecture that could potentially unlock actual applications far sooner than the previously anticipated quantum timelines.
DeepSeek considerations: New technologies bring advantages
The recent rise of DeepSeek and the subsequent stock sell-off put current hardware solutions into stark contrast.
In the wake of DeepSeek, Dobrin said data centre operators and investors need to monitor these new technologies as they ramp up to beat the curve and get ahead.
“This isn’t going to be an overnight change,” Dobrin said. “You're going to have three or four emerging technologies accelerating this shift. Instead of one new technology gradually replacing the old, these advancements—like photonics, plasmonics, and neuromorphic hardware—will all develop at a similar pace, making the decline of current semiconductors happen much faster.”
Following the “two-week panic” brought on by DeepSeek, Dobrin said the general view was that demand for data centres and related hardware will still stay, owing to more of a temporary blip rather than a major disruption.
He added that the resulting concern about the Chinese startup purportedly copying technology from OpenAI was more akin to fair use — arguing that it's no different than what companies like OpenAI and Anthropic have done in terms of using publicly available data.
“Weeks before DeepSeek released their model, some graduate students at Berkeley built Sky-T1, but no one complained about that because it wasn't in Beijing. Everyone thought it was great, and while it wasn’t quite as good as DeepSeek, fair use is fair use.”
Already the overreliance on GPUs is being challenged by a hardware landscape that’s slowly becoming more diverse, with operators increasingly turning to other solutions like CPUs to power their workloads.
CPUs like AMD’s EPYC series are helping to power AI workloads in data centres. Dobrin highlighted databiome, which is developing small-scale language models, as an example of a firm actively running AI training and inference on CPUs.
“[Data Biome] work with telcos and gaming companies, replacing large language models and GPUs, which helps businesses save millions of dollars each month by reducing GPU usage. It has a huge impact on GPU capacity. They're also partnering with AMD and Intel.”
Dobrin said the divergence away from GPUs, along with the rise of emerging hardware solutions will ultimately help prevent operators from being solely reliant on Nvidia's ecosystem.
“AMD recently launched a great GPU, and there are GPU virtualisation platforms like Mako, which allow you to virtualise across any GPU using any language,” Dobrin said. “This creates an abstraction layer, so you're no longer locked into Nvidia's CUDA — you can run CUDA on AMD, Intel, or any other GPU. All of these developments are breaking Nvidia's previous lock-in.”
RELATED STORIES
DigitalBridge CEO on the AI revolution’s impact on investments
DeepSeek: Safety questions arise amid Australia ban & EU compliance uncertainty
