Overview
- Perplexity confirmed Tuesday that it plans to use Nvidia's Vera processors after internal testing found roughly 1.5x faster performance on continuous AI‑agent coding tasks compared with traditional CPUs.
- Perplexity's practical tests included copying code repositories and running software tests and in some sandboxes reported up to 1.9x speed gains, though the company declined to disclose purchase volumes.
- Vera is built around 88 custom 'Olympus' cores and is designed for strong single‑core performance, high memory bandwidth and lower energy use to support long‑running agent loops that repeatedly fetch model results and act.
- Nvidia has already placed Vera systems with major cloud and AI customers and is pitching the chip as a challenger to Intel and AMD, but wider adoption depends on independent benchmarks and software integration work.
- If Vera scales as Nvidia hopes, it could shift server buying toward agent‑optimized CPUs and change cloud cost and performance tradeoffs, but supply, competition from x86 and custom chips, and undisclosed orders pose clear risks.