If You Believe in AI, You Have to Bet on This
Speakers
Ejaaz Ahamadeen
— Co-Host, Limitless Podcast (Bankless)
Josh Kale
— Co-Host, Limitless Podcast (Bankless)
Summary
The episode argues that inference—serving AI models—is becoming the dominant use of compute, surpassing training. Custom accelerator chips like those from Etched, Cerebras, and others are critical for high-throughput, low-power inference. The hosts identify publicly traded plays on this trend, naming Cerebras, MediaTek, and Broadcom as asymmetric bets, while also advising not to overlook NVIDIA's ability to adapt.
- Inference demand has flipped from one-third to two-thirds of AI compute and is heading toward 80%.
- NVIDIA GPUs are underutilized (30-40%) for inference, creating an opening for specialized ASICs.
- Etched is a private startup building a whole-rack inference system with 10-50x efficiency over GPUs.
- Cerebras (CBRS) is a public inference chip company that the hosts view as deeply undervalued.
- MediaTek and Broadcom are identified as chip design companies benefiting from the custom inference boom.
- OpenAI's Jalapeno chip and other custom efforts highlight vertical integration as the endgame.
- Despite the rise of ASICs, NVIDIA's acquisition of Grok shows it will attempt to stay competitive.
- The hosts believe the market is underpricing the inference trend and are personally betting on some public names.
Ideas
Cerebras is undervalued inference play.
Cerebras is a publicly traded inference chip company that the market is currently mispricing. The stock is down 35.5% from its IPO, despite inference demand exploding as AI agents and autonomous workloads drive token consumption. Ejaaz believes this creates an asymmetric bet, because inference performance is the new moat and Cerebras's accelerator chips are designed specifically for high-throughput, low-latency inference.
MediaTek benefits from inference chip design.
MediaTek is helping design custom inference chips for AI workloads and has already returned 180% year-to-date. Ejaaz includes it as part of the asymmetric opportunity in inference, implying that MediaTek's role in chip design will benefit from the shift toward specialized inference accelerators.
Broadcom offers asymmetric inference exposure.
Broadcom is a major designer of custom AI chips, including OpenAI's Jalapeno inference chip. Despite this, Broadcom is up less than 10% year-to-date, making it an undervalued play on the inference chip trend. Ejaaz frames it as another asymmetric bet in the inference space.
NVIDIA will adapt inference chips.
NVIDIA is aware of the shift toward inference and has acquired Grok, a custom accelerator company. Josh believes NVIDIA will integrate inference-optimized capabilities into future GPUs, creating a new product line. He warns against counting NVIDIA out, implying it will remain a leader even as custom ASICs emerge.
