OpenAI's Jalapeño Chip Challenges Nvidia's AI Margins

A New Contender in the AI Silicon Race

OpenAI has reportedly developed an in-house AI chip codenamed Jalapeño that outperformed Nvidia's Blackwell-based systems on key inference-efficiency benchmarks, signaling that even the world's most prominent AI developer is no longer content to rely solely on Nvidia's GPUs 1. The disclosure adds OpenAI to a growing list of major technology companies pursuing custom silicon, a trend that analysts say could gradually erode Nvidia's dominant margins in the AI hardware market 1.

Custom Silicon Momentum Builds Across the Industry

OpenAI's move is not happening in isolation. Google is reportedly working with AMD on a next-generation Tensor Processing Unit that would integrate on-package CPU cores directly alongside the TPU, according to multiple reports describing the arrangement as a hybrid AI ASIC design 23. This chip is said to be aimed specifically at agentic AI and reinforcement-learning workloads, a class of applications that increasingly demands tighter integration between processing and memory to run efficiently 3. If accurate, the partnership would mark a notable shift for Google, which has historically designed its TPUs largely in-house, and would also represent a significant new business line for AMD beyond its existing GPU competition with Nvidia 2.

The custom-silicon push extends further down the supply chain as well. Marvell, a key player in AI networking and custom chip design for hyperscale customers, is set to report quarterly earnings with investors watching closely for signs of continued momentum in its custom silicon and AI networking segments 4. Marvell's business model, which involves designing tailored chips for large cloud companies, stands to benefit directly from the broader shift away from off-the-shelf GPU purchases toward bespoke hardware built for specific workloads 4.

Beyond the Data Center

The custom-chip trend is not confined to cloud computing and large language models. Alphabet's Waymo unit has unveiled its own custom 5-nanometer AI chips designed to process sensor data onboard its autonomous vehicles, underscoring how specialized silicon is becoming central to scaling real-world AI applications like robotaxis 5. As Waymo expands its autonomous fleet, in-house chip design offers the company greater control over performance, power efficiency, and cost compared with relying on general-purpose hardware 5.

Why It Matters

Taken together, these developments illustrate a broader industry pattern: the largest AI developers and their infrastructure partners are increasingly building or co-designing their own chips rather than depending entirely on Nvidia. For Nvidia, whose outsized margins have been built on GPU scarcity and near-universal demand, the rise of credible in-house alternatives from OpenAI, Google, AMD, Marvell's customers, and even Waymo suggests inference costs and hardware sourcing strategies across the AI sector are entering a more competitive and fragmented phase.