OpenAI – Spicy Chilli

Jalapeno is hot.

• OpenAI has given a first glance of the performance that its in-house AI inference chip can deliver, and while it crushes Blackwell and is on par with Rubin in some measures, the real achievement is in the power consumption and the speed with which it has produced this chip from scratch.
• Jalapeno was teased in October 2025, jointly announced by OpenAI and Broadcom on June 24^th 2026, with first specifications and performance data being announced at the Hot Chips conference on August 25^th 2026.
• Naturally, in the benchmarks that OpenAI has chosen to disclose, Jalapeno performs extremely well and leaves Blackwell far behind.
• This should not be a big surprise as Blackwell is the older generation of Nvidia silicon and has long since been surpassed by Vera Rubin, which is just beginning to appear in commercial datacentres.
• Even against Rubin, it fares pretty well, delivering 13.4 PFLOPS of FP4 performance compared to Rubin, which delivers 17.5 PFLOPS of FP4 per chip.
• This is surprising as the chip runs at 700W compared to Nvidia, which runs at around 1000W, meaning that on a compute-per-watt basis it is more productive than Vera Rubin.
• The problem with this analysis is that there are so many caveats, conditions, tricks and manipulations that one can employ to the measurements that most comparisons are open to argument.
• However, 2 things are clear:
  • First, performance: which is excellent, and while it is debatable whether it is better than Vera Rubin, it’s a viable competitor that should not be dismissed out of hand.
  • Hence, I can see this launch putting some pressure on Nvidia, which will cause it to redouble its efforts, meaning better products for its customers.
  • It is important to note that Jalapeno has not been designed just for OpenAI models but is capable of running everyone’s, meaning that sales to other AI providers or Neoclouds is a possibility.
  • Second, time to production: where OpenAI has gone from a standing start to production in 2 years, which is an impressive feat.
  • Nvidia’s design cycle is much longer, but with so many designs in the works at once, it can still produce a new chip every 12 months.
  • Whether OpenAI can match that cadence remains to be seen.
  • This is important because as CUDA becomes less important for inference, product cadence is one of the key differentiators that keeps Nvidia ahead of everyone else.
• The net result is that this is one of the more interesting silicon releases that I have seen this year, and if the chip lives up to its billing, it should be one of the more worthy competitors for Nvidia available.
• It is worth noting that this is an inference chip and is unlikely to have any impact on the market for training, but this is not where the action is.
• RFM Research has long concluded that the training market could easily fall to around 10% of the market for AI silicon over time, and its share of the market is already in free fall.
• This is good news for Broadcom, but the market that it has had pretty much to itself for a long time is under assault on all fronts.
• Hence, the best challenger to look at is Qualcomm, which has signed up 4 hyperscalers for its AI data centre product, has greatly increased its 2029 forecasts but still trades on 16.0x 2026 PER, making it the bargain of the industry.
• I have a significant position that I remain very happy to sit on.