Conference Presentation, Product Demonstration
Zero Flaps, Zero Excuses: Network Reliability at AI Scale | Credo x TensorWave | RAISE 2026
- Credo is projected to grow revenue from approximately $1.3 billion to $2.4 billion based on analyst expectations.
- The company is transitioning to zero-flap optical technology to overcome the seven-meter distance constraint of copper cabling, specifically targeting customers like Oracle with fabric reliability needs exceeding current third-party options by a factor of 1,000.
- Credo intends to deploy intelligent transceivers capable of proactively fixing marginal link issues and removing units before disruptions, aiming to reduce Oracle's "no trouble found" (NTF) metric, which currently stands at 95%.
- TensorWave currently operates the world's largest AMD GPU deployment with 8,192 live MI325s and has secured up to two gigawatts of long-term capacity in North America, scaling operations across Arizona, Pennsylvania, and Florida.
- TensorWave aims to reduce network deployment time from six weeks to less than one week to accelerate revenue generation for customers, addressing the financial impact of idle GPU factories where four weeks of downtime can result in billions of dollars in lost revenue.
- Credo plans to build optical zero-flap transceivers that match the reliability of active electrical cables, utilizing iterative firmware fixes derived from data collected in 20 thermal chambers located in Taiwan.
- The company anticipates that for rack structures exceeding seven meters, optical technology is the only viable solution, and that standardized, high-reliability optical technology will become a hard requirement for AI networks over the next few years.
- Credo is co-chairing the OCP Optics Reliability Workstream with Oracle to establish a standard for zero-flap (ZF) optics specifications, while also planning to pre-qualify these optics with various switch vendors to support NeoCloud customers.
- Credo expects to maintain an 88% market share as it continues to provide active electrical cables used at hyperscalers and neoclouds, noting that these cables typically operate at 120 million hours mean time between failures with over 32 billion field hours of demonstrated reliability.
- The strategy involves moving beyond traditional cable-based networking to system performance technology, where predictive telemetry detects degradation before failure to prevent link flaps and reduce false positives, thereby saving costs and enabling efficient resource deployment.
- Credo intends to deliver telemetry data to orchestration layers to allow operators to distinguish specific failure modes, such as dust in patch panels or crimped fiber, and to alert operators to take links out of service before issues like ESD damage manifest over weeks or months.
- The company views the shift to proactive reliability as essential for NeoCloud operators facing challenges in raising funds, securing GPU allocations, power, and permits, while also needing to avoid restarting large-scale LLM training sessions due to link instability.
- Credo encourages competitors to join the OCP workstream to contribute to new reliability standards and anticipates that in-band 1 PPS signal capabilities will be integrated into standardization efforts for network synchronization.
- The outlook identifies that lower reliability directly correlates to higher costs for data center operators, with downtime creating SLA issues and revenue losses, making reliability as critical as GPU performance for AI infrastructure.