BNP Paribas: Hot Chips 2026 Takeaways Covering Jalapeño and NVIDIA Networking
Nashnova编辑部
OpenAI's in-house inference chip Jalapeño debuted benchmark results at Hot Chips 2026, beating Nvidia's Blackwell on performance per watt in nearly every test; this signals that hyperscaler-designed silicon is moving from concept to combat, putting direct pressure on Nvidia's inference-market margins.
What did Jalapeño actually score?
SemiAnalysis tested Jalapeño on-site and found it outperformed Blackwell on performance per watt in nearly every scenario.
But the firm cautioned the comparison is "not entirely complete or fair" — Jalapeño uses next-generation HBM4 memory (the high-bandwidth memory that feeds data to the chip), while Blackwell runs on the prior generation.
This means → a fairer benchmark target is Nvidia's Rubin platform, which also uses HBM4. Rubin is already shipping to customers; Jalapeño remains an engineering sample.
In plain terms = Jalapeño is genuinely fast, but it ran with better ammunition. Once Nvidia loads the same ammunition, the gap may narrow.
What does this mean for Nvidia?
Inference is Nvidia's fastest-growing business segment. Multiple analysts told CNBC that Jalapeño poses a "material threat" to margins in that segment.
Yole Group analyst Adrien Sanchez noted that OpenAI had been one of Nvidia's largest single customers. Jalapeño's launch "directly raises competitive risk in Nvidia's most important customer relationship."
This reflects a structural shift: the biggest buyer is becoming a potential competitor.
The custom-chip wave — is OpenAI alone?
Far from it. Google unveiled a new TPU generation — tensor processing units purpose-built for AI. Meta partnered with Broadcom to deploy 1 gigawatt of custom AI silicon. Anthropic committed over $100 billion to AWS over the next decade, covering Amazon's in-house Trainium chips.
Omdia analyst Alexander Harrowell called this Nvidia's "greatest competitive threat" — roughly half of AI infrastructure capex comes from hyperscale cloud providers, and those same customers are now building their own chips.
He forecast that custom ASIC — application-specific integrated circuit, a chip designed for one task — shipments will surpass GPUs by 2028, though GPU revenue will take longer to overtake because of higher per-unit prices.
Where does Nvidia's moat still hold?
TrendForce analyst Fion Chiu argued that for compute-intensive workloads such as large-scale model training, Nvidia GPUs remain vital thanks to their programmability, software ecosystem, and ability to handle diverse workloads.
In plain terms = custom chips excel at running their maker's own models, but for tackling a wide range of new tasks, Nvidia's general-purpose advantage is hard to replace in the near term.
BNP Paribas flagged Nvidia's networking architecture as the conference's other key takeaway. This means → the market is watching not just how fast a single chip runs, but how efficiently chips talk to each other — and that is Nvidia's traditional stronghold.
How far is Jalapeño from real deployment?
Jalapeño was co-developed with Broadcom by a team of roughly 100 engineers; critical packaging tape-out took about nine months from project start.
OpenAI said the chip will be deployed into its compute infrastructure before year-end, with second- and third-generation designs already under way.
This means → in the short term Jalapeño is an internal cost-reduction tool for OpenAI. Whether it sustains its efficiency edge at scale will be the key proof point for whether this competitive shift is real.
市场有风险，内容仅供研究参考，不构成投资建议。