OpenAI Jalapeno Chip: Broadcom Wins, NVIDIA Loses Inference Crown
OpenAI's Jalapeno chip with Broadcom threatens NVIDIA's inference monopoly. SK Hynix's $29B US listing and Cerebras's earnings add context. This article analyzes who wins, who loses, and what happens next.
On June 24, 2026, Bloomberg's Ed Ludlow broke the news: OpenAI's first custom AI chip, codenamed Jalapeno, is here, built with Broadcom. This is not a test chip—it's a production inference processor designed to replace NVIDIA's H100 and B200 in OpenAI's data centers. The AI hardware landscape just shifted.
- OpenAI and Broadcom unveiled Jalapeno, a custom AI inference chip, on June 24, 2026, per Bloomberg's Ed Ludlow.
- Jalapeno is designed to reduce OpenAI's reliance on NVIDIA GPUs for inference, potentially cutting costs by 40-60%.
- SK Hynix announced a $29 billion US listing, the largest ever for a Korean company, to fund AI memory production.
- Cerebras reported its first quarterly earnings as a public company, showing revenue growth but ongoing losses in wafer-scale AI.
Why Did OpenAI Choose Broadcom Over NVIDIA for Jalapeno?
According to Bloomberg's Ed Ludlow, OpenAI partnered with Broadcom for chip design and manufacturing, leveraging Broadcom's expertise in networking and custom ASICs. The Jalapeno chip is an inference-specific processor, meaning it handles model execution, not training. This is a direct challenge to NVIDIA, which dominates inference with its H100 and B200 GPUs. OpenAI's reasoning is clear: NVIDIA's margins are high, and OpenAI's inference costs are ballooning as ChatGPT and GPT-5 usage grows. By owning the chip, OpenAI can optimize for its own models and cut costs. Broadcom, meanwhile, gains a marquee AI customer without competing with its own networking chips. The deal is non-exclusive, but Broadcom's design win is a significant revenue stream.
How Does Jalapeno Compare to NVIDIA's B200 in Inference Performance?
While Bloomberg did not provide benchmark numbers, industry sources suggest Jalapeno is optimized for transformer-based models, the architecture behind GPT-5. It likely uses high-bandwidth memory from SK Hynix—tying the two stories together. In contrast, NVIDIA's B200 is a general-purpose GPU that excels at both training and inference. The trade-off is clear: Jalapeno may be 2-3x more power-efficient for inference on OpenAI's models, but it lacks the flexibility to run other AI workloads. For OpenAI, that's a feature, not a bug. For the broader market, it means NVIDIA retains the training crown, but inference is now contested.
| Feature | OpenAI Jalapeno (Broadcom) | NVIDIA B200 |
|---|---|---|
| Primary Use | Inference (GPT-5, ChatGPT) | Training + Inference |
| Design Partner | Broadcom | In-house (NVIDIA) |
| Memory | Likely SK Hynix HBM4 | HBM3e |
| Efficiency (est.) | 2-3x B200 for inference | Baseline |
| Flexibility | Low (OpenAI models only) | High (any model) |
| Availability | Late 2026 (est.) | Now |
| Verdict | Winner: Broadcom/OpenAI for inference cost; NVIDIA for versatility | |
What Does SK Hynix's $29 Billion US Listing Mean for AI Memory?
Bloomberg reported that SK Hynix is planning to raise $29 billion in a landmark US listing. This is the largest IPO ever for a South Korean company and the biggest tech listing of 2026. SK Hynix is the dominant supplier of HBM (High Bandwidth Memory) for AI accelerators, including NVIDIA's H100 and B200. The listing is a bet that AI memory demand will continue to grow exponentially. According to SK Hynix's CFO, the funds will be used to build new fabrication facilities in the US, likely in partnership with Intel or TSMC. This is a direct response to the US CHIPS Act and a hedge against geopolitical risk. For OpenAI's Jalapeno, SK Hynix's US capacity ensures a stable memory supply.
Can Cerebras Survive the Custom Chip Wave?
Cerebras CEO Andrew Feldman, speaking on Bloomberg, reported the company's first quarterly earnings since going public. Revenue grew 35% year-over-year, but the company remains unprofitable. Cerebras's wafer-scale chips are designed for training, not inference, putting them in direct competition with NVIDIA, not OpenAI. Feldman argued that custom chips like Jalapeno are good for the ecosystem but that Cerebras's advantage is in training speed for large models. However, the market is skeptical: Cerebras's stock dropped 8% after earnings. The question is whether Cerebras can pivot to inference or find a niche in scientific computing. My bet: Cerebras will be acquired within 18 months by a hyperscaler like Microsoft or Google.
My thesis: The Jalapeno chip is a tactical win for OpenAI and Broadcom, but a strategic loss for NVIDIA's inference monopoly. In the short term, OpenAI will reduce its inference costs by 40-60%, improving margins on ChatGPT subscriptions. Broadcom gains a new revenue stream and a reference design for other AI customers. NVIDIA loses a marquee customer for inference, but its training business is safe for now. In the long term, the real winner is the custom chip ecosystem: Broadcom, Marvell, and even AMD will benefit as hyperscalers build their own silicon. The loser is NVIDIA's data center revenue growth, which will slow as inference shifts to custom chips. SK Hynix's US listing is a bet that this trend accelerates. My concrete prediction: By Q3 2027, at least three major hyperscalers (Google, Amazon, Meta) will announce custom inference chips, following OpenAI's lead.
- Prediction 1: By Q3 2027, OpenAI will deploy Jalapeno in at least 50% of its inference data centers, cutting its NVIDIA GPU dependency by 30%.
- Prediction 2: SK Hynix's US listing will close at a valuation exceeding $150 billion, making it the largest memory company by market cap.
- Prediction 3: Cerebras will be acquired by a hyperscaler (most likely Microsoft) within 18 months of this article's publication.
- June 2026OpenAI Jalapeno chip unveiled
Bloomberg's Ed Ludlow reports OpenAI's first custom AI inference chip, built with Broadcom.
- June 2026SK Hynix $29B US listing announced
SK Hynix plans the largest Korean IPO in history to fund US AI memory production.
- June 2026Cerebras first quarterly earnings
Cerebras reports 35% revenue growth but remains unprofitable; stock drops 8%.
Estimated Inference Chip Market Share by 2027
- Insight 1: OpenAI's Jalapeno chip is a defensive move against NVIDIA's pricing power, not an offensive one—it's about cost control, not performance supremacy.
- Insight 2: Broadcom's design win is more valuable than the chip itself; it positions Broadcom as the go-to custom silicon partner for AI.
- Insight 3: SK Hynix's US listing is a geopolitical hedge; it signals that Korean memory makers see the US as the primary market for AI hardware.
- Insight 4: Cerebras's struggles show that wafer-scale chips are a niche; custom chips like Jalapeno are the future of inference.
- Insight 5: The inference chip market will fragment by 2027, with hyperscalers owning their own silicon and NVIDIA dominating training.
Source and attribution
Bloomberg Technology
OpenAI Unveils First Custom AI Chip With Broadcom | Bloomberg Tech 6/24/2026
Discussion
Add a comment