Hardware sensitive algorithms redefine ai
They also take away traditional logic
Last month everything changed.
November and December have been disruptive months for the ai industry.
First was the inflective entry of China and India. MetaX the Nvidia competitor out of China had a blockbuster ipo listing. And nearer home, BharatGen – India’s answer to ChatGPT and Deepseek started operations with seed stage funding of USD 150 million. Both are important events – but not something that will disrupt the status quo. MetaX will make cheaper GpU and BharatGen will do a better job of multilingual and smaller LLMs. Both have value.
But the real disruption is happening in the back lanes of Korea, Taiwan, Singapore and India. Away from the bright lights and showbiz of Silicon Valley. The long-anticipated “interconnect bottleneck,” where electrical copper links could not keep up with the data demands of trillion-parameter models, has been broken. The industry has shifted from electrons to photons by embracing photonic chipsets, fundamentally changing how AI systems communicate and scale. They are no longer in silicon and silicon fabs don’t make them.
Nvidia hasn’t done this.
Nor has Open AI.
By replacing copper traces with lasers and fiber optics, AI clusters can now achieve orders-of-magnitude improvements in bandwidth density and energy efficiency. These chips are made on diamond substrates.
Fibre optics first connected continents. Then it connected telephone exchanges. It finally became fibre to the home. And now it has become “fibre to the chip”. Copper is being replaced by glass fibre. These approaches 3D-stack photonic integrated circuits directly onto GPUs or switches, enabling “edgeless I/O” where data exits the chip as light from its interior.
Startups have gone further, introducing photonic interposers with more than 100 Tbps of bandwidth, effectively allowing thousands of accelerators to behave like a single processor. This represents the most significant hardware advance since high-bandwidth memory (HBM).
The moat in AI is shifting from algorithms alone to the photonic efficiency of underlying hardware. By cutting energy consumption by 95%, photonic chipsets decouple AI progress from runaway power consumption, easing environmental concerns. It also enables disaggregated, memory-centric computing, where large pools of memory can be dynamically shared across processors—critical for future multimodal models requiring petabyte-scale active memory. The memory itself will connect to the optical fabric directly.
The interconnect bottleneck will soon vanish, marking the beginning of a global, light-based computational fabric that underpins the next era of artificial intelligence.
Open source
Not proprietary
And that is the end of a monopoly.
It is now a new industry. One that will had new leaders. The incumbents will fight to survive. But they are unlikely to thrive.
