SAN FRANCISCO — Nexus Labs, the Silicon Valley semiconductor startup that has operated in near-total secrecy since its $4.2 billion Series D round in 2024, stunned the technology industry Sunday by unveiling the Aurora X1 — a photonic AI processor that the company says outperforms the fastest existing graphics chips by a factor of one thousand on standard machine learning benchmarks. If independently verified, the announcement would represent the single largest leap in AI processing power since the invention of the GPU.
The Aurora X1 abandons traditional transistor-based computation in favor of photonic circuits, which use pulses of light rather than electrical signals to perform calculations. Because photons travel faster than electrons and generate virtually no heat, the architecture eliminates two of the most persistent bottlenecks in modern AI hardware: processing speed and thermal management. Nexus claims the chip can sustain peak performance indefinitely without the thermal throttling that limits conventional processors during extended inference tasks.
At a packed keynote address in San Francisco's Civic Auditorium, Nexus CEO Dr. Rina Osei demonstrated the chip running a trillion-parameter language model with a latency of under four milliseconds — a task that currently requires a rack of dozens of high-end GPUs and takes several seconds. The demonstration drew gasps from an audience that included representatives from every major technology company. "We are not iterating," Osei told the crowd. "We are beginning again."
Independent experts who reviewed Nexus's technical whitepaper said the claims, while extraordinary, are grounded in legitimate physics. "The principles here are sound," said Dr. Arun Patel, a photonic computing researcher at MIT who was given advance access to the documentation. "The question is whether they have actually solved the manufacturing challenges that have blocked photonic chips from commercial viability for thirty years." Those challenges include precisely aligning optical components at nanometer scale — a process that has historically been prohibitively expensive to replicate at volume.
Nexus says it has solved the manufacturing problem through a proprietary chip-packaging technique developed in partnership with TSMC, and that it is already producing Aurora X1 units at low volume at a facility in Taiwan. The company announced partnerships with three major cloud providers — none of which it named — who will offer Aurora-based compute capacity beginning in the third quarter of this year. Pricing was not disclosed, but analysts expect the initial cost per unit of compute to be substantially higher than conventional GPU-based alternatives.
The announcement sent shares of established chipmakers plunging in after-hours trading. Nvidia fell 11 percent, AMD dropped 9 percent, and Intel lost 7 percent of its market value within an hour of the keynote. Nexus itself is not publicly traded, but sources familiar with the company's cap table said private investors were being offered secondary shares at a valuation above $200 billion — which would make it the most valuable private technology company in history.
Not everyone in the industry is convinced. Several engineers who spoke anonymously said they remain skeptical until the chip can be evaluated in independent testing environments. "Extraordinary claims require extraordinary evidence," said one veteran chip architect. "We've seen companies demonstrate impressive benchmark numbers on carefully selected tasks before." Nexus says it will begin shipping developer kits to vetted research institutions next month, at which point independent verification will become possible.