Briefing
- OpenAI is moving to control its own silicon stack with Jalapeño, a custom inference accelerator designed specifically for ChatGPT and related LLM workloads, reducing dependence on third-party chip suppliers.
- The chip was produced in partnership with Broadcom and arrives eight months after OpenAI first announced the custom chip deal.
- OpenAI framed the project as part of an ambition to 'build the full stack', signalling that Jalapeño is the first in a planned series of proprietary hardware generations.
- Custom inference silicon typically lowers per-query compute costs and improves latency at scale, which matters directly to OpenAI's unit economics as ChatGPT usage grows.
- The immediate question for competitors and hyperscalers is whether OpenAI will seek to deploy Jalapeño exclusively internally or eventually offer it as part of a broader infrastructure service.
Analysis
- Broadcom's AVGO faces a structural revenue ceiling on its AI custom silicon business as OpenAI's in-house capability matures. Broadcom's hyperscaler ASIC franchise is valued partly on expanding wallet share from AI labs; OpenAI's stated ambition to 'build the full stack' signals future chip generations may reduce Broadcom's design and co-packaging revenue per chip, even if Broadcom retains a manufacturing role in Jalapeño's current generation. The risk compounds if other large AI labs interpret OpenAI's move as validation for similar vertical integration.
- Nvidia's inference GPU demand from OpenAI faces a long-term displacement risk, not an immediate one. Custom inference silicon optimised for specific LLM workloads, as Jalapeño is described, typically delivers better performance-per-watt and lower per-query cost than general-purpose GPUs for production inference at scale. If OpenAI scales Jalapeño broadly across ChatGPT inference, its incremental Nvidia H100 and Blackwell procurement for inference workloads slows, reinforcing a bifurcation where Nvidia retains training dominance but loses inference share to custom ASICs across multiple large customers.
- The Jalapeño announcement accelerates a structural re-rating question for pure-play AI infrastructure providers including HIVE and similar GPU-as-a-service operators: if the largest AI labs vertically integrate silicon, the total addressable market for third-party inference compute narrows over a 3-5 year horizon. HIVE's recently announced $70M ARR Blackwell deployment is insulated in the near term because it serves Cohere, not OpenAI, but the precedent OpenAI sets pressures the long-term assumption that hyperscale AI inference demand will continue to flow through third-party GPU operators.
Historical Context
- 2023-2024
Google's TPU program demonstrated that hyperscale custom inference silicon can reduce per-query costs by 30-50% versus general-purpose GPUs at scale, establishing the business case that OpenAI is now following and providing a precedent for how long the transition from external GPU dependence to internal silicon takes.
- 2020
Apple's M1 transition showed that vertically integrating chip design allows a platform company to rapidly obsolete third-party silicon suppliers for its primary workloads; the analogy is imperfect but the market re-rated Qualcomm and Intel exposure to Apple revenue the moment Apple announced the in-house program, not at volume deployment.
- 2016
Microsoft's Project Catapult and Amazon's Inferentia programs established that large cloud and AI platform operators use custom silicon primarily to reduce third-party chip dependence on inference at scale; both programs took 3-5 years to materially affect Nvidia's revenue mix from those customers.
Connected Events
Nvidia launches $20bn bond deal, its first debt sale since 2021
Nvidia's $20bn bond issuance, its first debt sale since 2021, is partly a capital raise to fund continued AI infrastructure buildout at a moment when its largest inference customer is now moving to reduce reliance on Nvidia silicon.
HIVE secures $220M sovereign AI cloud contract with Bell and Cohere
HIVE's $220M sovereign AI cloud contract deploying 2,304 Nvidia Blackwell GPUs for Cohere represents the type of third-party inference infrastructure that OpenAI's vertical integration strategy is designed to replace internally, illustrating the divergence between OpenAI's trajectory and the broader GPU-as-a-service market.
Further reading
See Indexa more often on Google
Mark Indexa as a preferred source — your Top Stories will surface more Indexa coverage.
