AI gets all the headlines, but the real magic happens in the hardware. Every breakthrough we see – from chatbots to image generators – runs on an intricate network of chips and memory systems processing unfathomable amounts of data. There’s just one problem: as AI models balloon in size, we’re hitting a wall. The traditional way of moving data between processors and memory can’t keep up anymore. That’s where Celestial AI comes in. This Santa Clara startup is rethinking the entire approach to data movement in AI systems, and they’re doing it quietly, away from the spotlight.
The story of Celestial AI starts with a simple but crucial insight. Back in 2020, David Lazovsky and Preet Virk saw something that others had missed: AI’s biggest hurdle wouldn’t be raw computing power – it would be moving data around efficiently. Both founders knew their stuff. Having spent years in semiconductors and photonics, they’d watched AI models grow so massive that no single memory chip could handle them. Their big idea? Use light instead of electricity to move data between chips.
The Photonic Fabric
The heart of the innovation at Celestial AI would be its Photonic Fabric, a technology that addresses inefficiencies in electronic interconnects. Here, data is transferred by light between compute and memory for high speed and low latency. Unlike traditional optical interconnects, which often have to convert optical signals to electrical signals at the edge of a chip, Celestial’s technology delivers data directly to the point of compute, anywhere on the silicon. This removes the so-called “beachfront problem,” where optical links are limited to the perimeter of a chip, and opens up new possibilities for bandwidth and efficiency.
Phil Winterbottom, Celestial’s Chief Technology Officer and a former Bell Labs researcher, is the architect behind this technology. Winterbottom’s expertise in photonics and his experience in founding successful startups like Entrisphere and Gainspeed have been instrumental in shaping Celestial’s vision. The Photonic Fabric is a practical solution that has already attracted significant interest from hyperscalers, semiconductor companies, and memory vendors.
Strategic funding and partnerships have driven Celestial AI from a fledgling startup to a major player in the AI hardware space. In March 2024, the company announced a $175 million Series C funding round, led by the US Innovative Technology Fund (USIT) and supported by a dozen other investors. This latest round brings Celestial’s total funding to over $340 million.
The funds will be used to speed up the commercialization of the Photonic Fabric platform, to grow the engineering and technical teams, and to further develop collaborations with hyperscalers and semiconductor partners. Lazovsky highlighted that the company is experiencing “broad customer adoption” and works closely with “some of the largest hyperscalers” to address data-memory bottlenecks. These collaborations are not just about integrating Celestial’s technology into existing systems but also about co-designing solutions that can future-proof data center architectures.
It allows Celestial to work within industry standards and protocols, such as PCIe, UCIe, and CXL, and its ability to be directly integrated into today’s data center infrastructures without making substantial changes in the software. Our photonics technology is really differentiated and unique with superior characteristics,” says Lazovsky. “With the explosion of generative AI workloads based on large language models and the strain that is causing to current data center architectures, the demand for optical connectivity is growing exponentially.”
Memory Wall
One of the most critical challenges in AI hardware is the so-called “memory wall,” where the growth of on-chip memory capabilities lags far behind the exponential increase in data requirements for advanced AI workloads. This mismatch often leads to stranded memory, where a significant portion of memory resources remains underutilized. This can account for as much as 25% of the memory in Azure that is stranded and hence results in inefficiencies to the tune of billions of dollars that could be costed by the data centers.
Celestial AI Photonic Fabric can help to achieve this, first through memory disaggregation. That is the separation of memory from compute, hence making it pooled across hardware within the data center. This approach not only improves memory utilization but also reduces power consumption and latency. Lazovsky explains, “Today, compute and memory are closely coupled. The only way to add more high-bandwidth memory is to add more compute, whether the additional compute is required or not. Celestial’s tech enables memory disaggregation.”
The company’s technology also brings significant energy savings, a critical factor as data centers grapple with rising power demands. Reducing the amount of electricity required for data movement, Celestial’s photonics allow chips to allocate more power to computing tasks, effectively boosting performance. “Energy is becoming the gate for more AI, and data transmission consumes 55 to 70% of power consumption—it’s not the compute power,” Lazovsky notes.
Celestial AI is gaining attention with its Photonic Fabric, but it operates in a less crowded space than photonic integrated circuit players like Ayar Labs and Renovus, which focus on optical interconnect solutions. The market for these technologies is projected to reach $26.42 billion by 2027.
On the other hand, Celestial AI competes with some of the biggest players in the industry. Google Cloud and AWS leverage their cloud dominance to integrate AI/ML services into enterprise ecosystems. NVIDIA has remained the benchmark for GPU-accelerated computing, while Intel is continuously advancing its portfolio of AI-focused semiconductors. IBM Watson, with deep AI research, remains one of the strong players in specialized areas, NLP and computer vision.
What sets Celestial AI apart is its unique approach: it brings data directly to the point of compute while supporting industry-standard architectures.
The company’s recent acquisition of silicon photonics intellectual property from Rockley Photonics further strengthens its position. This acquisition, which includes over 200 patents, enhances Celestial’s portfolio in areas like optoelectronic systems-in-package, electro-absorption modulators, and optical switch technology. “The addition of Rockley’s IP into our technology platform further accelerates the growth of Celestial AI’s highly valuable intellectual property portfolio and amplifies the strength of our position,” Lazovsky said.
As Celestial AI continues to scale, its focus remains on delivering scalable, energy-efficient solutions tailored to the demands of next-generation AI compute and networking. The company is actively engaged with multiple Tier-1 partners, including hyperscalers, processor vendors, and memory companies, to bring its Photonic Fabric products to market. With plans to grow its workforce to around 160 employees by the end of 2024, Celestial is aiming to be an enabler of the transition from general computing to accelerated computing.
Yesterday, Celestial AI announced its Board of Directors appointment of Lip-Bu Tan, a highly regarded leader in the semiconductor industry with over 40 years experience in deep-tech expertise coupled with a history of scaling transformative technologies. As founder of Walden International and key investor in Celesta Capital and Walden Catalyst Ventures, he pushed the frontiers of semiconductor design, electronic design automation, and AI solutions.
The photonics industry is still in its early stages, and engineering hurdles remain, particularly in areas like data-to-analog conversion and signal regeneration. However, Celestial’s ability to attract top talent, secure significant funding, and forge strategic partnerships suggests that it is well-equipped to navigate these challenges. In the words of Thomas Tull, chairman of USIT, Celestial AI’s work on connectivity is “critical to addressing key concerns across performance and energy efficiency.”