Jalapeño
OpenAI's Jalapeño chip. OpenAI

OpenAI put a number on Nvidia's table less than a day before Nvidia had to answer for its own. On Tuesday, at the Hot Chips conference at Stanford, OpenAI released the first performance results for Jalapeño, its in-house inference chip built with Broadcom — and claimed it beats Nvidia's GB300 on efficiency and speed. Nvidia reports fiscal second-quarter results Wednesday after the close.

The timing is the story as much as the chip. Wall Street walks into tonight's print already holding a fresh data point suggesting Nvidia's biggest customer needs it a little less than it used to — not proof of anything, but a new variable in how investors read whatever guidance Nvidia gives.

The numbers themselves: OpenAI says Jalapeño delivered 1.5 to 1.9 times more AI work per watt and cut end-to-end latency by 1.7 to 3.6 times versus Nvidia's GB200 and GB300 systems, running on SemiAnalysis's public InferenceX benchmark across three models — OpenAI's own GPT-OSS 120B, DeepSeek's R1 and Moonshot AI's Kimi K2.5. SemiAnalysis said it witnessed selected runs inside OpenAI's own lab; OpenAI supplied the underlying figures. That's a meaningful step short of an outside lab running the full suite independently, and it matters for how much weight the numbers should carry.

Two caveats OpenAI built into its own release: Jalapeño was never tested against Vera Rubin, Nvidia's newer platform, and the chip only handles inference — it can't train models, which is still where Nvidia's advantage is least contested. OpenAI's Richard Ho, the company's hardware chief, said the chip's low power draw — a 700-watt rating against GB300's 1,400 watts — should cut data center costs. Volume deployment isn't expected until 2027.

Nvidia arrives at its own moment already unsettled. The stock had run a seven-session losing streak — its longest since 2022 — before steadying in the days ahead of the report. Management guided to roughly $91 billion in quarterly revenue; consensus sits a notch above that. Whatever Nvidia says about margins, China exposure and next quarter's demand will now be read partly against a competitor's chip that didn't exist as a public benchmark 24 hours ago.

Hong Kong's stake in the print

The Hong Kong angle runs through Semiconductor Manufacturing International Corp (0981.HK), the top-weighted name on the Hang Seng TECH Index and the closest regional proxy to the AI-chip trade. SMIC and the wider index have moved on AI-chip headlines before — the stock jumped in a rally tied to a domestic AI model release last week. Southbound flows into SMIC and its Hang Seng TECH peers on Thursday will carry some of whatever tone Nvidia's call sets tonight, particularly on China and margin commentary, even though Jalapeño itself has nothing to do with China's chip supply chain.

None of that resolves tonight's real test, which is Nvidia's own numbers against Nvidia's own guide. But for the first time, that test arrives with a competitor's benchmark already in the room.