OpenAI published its first public benchmarks for the Jalapeño chip on August 26. On the InferenceX benchmark, Jalapeño posted 1.9x better efficiency than Nvidia's GB200 and GB300 systems, running at 700 watts versus 1,200-1,400 watts for the Blackwell hardware.
Same night: Nvidia reported $96.2 billion in Q2 revenue, up 106% year-over-year. The stock rose 4.7%.
The juxtaposition was not subtle.
The Hyperscaler Silicon Playbook
Google built its first Tensor Processing Unit in 2015. Meta's MTIA ships internally at scale and is never available to outside customers. Amazon's Trainium 3 runs inference workloads exclusively on AWS. Microsoft deployed Maia 200 across its Azure AI services. Every major hyperscaler has been building custom silicon for years.
The pattern is consistent: first, buy Nvidia to move fast. Then, once inference workloads are large enough and predictable enough, build a chip optimized for exactly those workloads. GPUs are general-purpose accelerators. Custom ASICs are purpose-built for specific inference architectures. The efficiency gap grows as models stabilize.
OpenAI's Jalapeño is built with Broadcom as the design partner, the same firm behind Google's TPU, Meta's MTIA, and Microsoft's Maia. That alignment is not coincidental. Broadcom has quietly become the engineering backbone of the hyperscaler silicon movement.
What This Means for Nvidia
Nvidia's gross margin averaged over 70% through the AI buildout. That margin reflects both monopoly pricing and genuine technical leadership. The H100 and GB200 are exceptional hardware. Jalapeño does not obsolete them. It competes with them in one specific workload class: inference.
Training remains Nvidia's strongest moat. Large model training requires the kind of dense matrix computation that Nvidia's GPU architecture handles better than any current alternative. But inference is where the volume is. Two-thirds of all AI compute is now inference. As hyperscalers shift those workloads to custom silicon, the addressable market for Nvidia's data center segment narrows at the margin.
Bearish risks remain real. Jalapeño's benchmarks are self-published by OpenAI. Independent verification is limited. First-generation custom chips historically underperform against the incumbent. SemiAnalysis reviewed the architecture and called it industry-leading across multiple model types, but broad production deployment is not the same as a benchmark win.
The Recalibrated Picks-and-Shovels Thesis
The AI infrastructure trade was simple: everything runs on Nvidia, so own Nvidia. That thesis worked from 2023 through mid-2025.
The updated version requires more precision. HBM memory suppliers (SK Hynix, Samsung, Micron) sell to both sides. Every GPU Nvidia ships needs HBM4. Every custom ASIC that displaces a GPU also needs HBM4. The memory suppliers do not care who wins the inference efficiency benchmark. They supply the winner either way.
SK Hynix holds approximately 50-55% of the HBM market. It has already secured 70% of Nvidia's HBM4 orders. As OpenAI, Google, Meta, and Amazon scale their custom chips, HBM demand does not decline. It redistributes.
The infrastructure picks-and-shovels play shifts from "own the GPU monopoly" to "own the memory stack that every accelerator requires." That is a more durable position.
Follow @UTXOMacro for daily Bitcoin, macro, and AI breakdowns.