OpenAI's 1.9x Per-Watt Chip Gain Puts Worldcoin (WLD) in AI-Crypto Focus

OpenAI's Jalapeño Chip: First Benchmarks

Worldcoin (WLD), the AI-linked altcoin, traded down 5.8% over the past 24 hours even as OpenAI's first custom inference chip, Jalapeño, dominated the AI-infrastructure conversation. OpenAI, working with Broadcom, said the ASIC-style processor beat Nvidia's current GB300-class systems on per-watt efficiency and latency in its own tests. The chip is built on TSMC's N3P process and uses six HBM4 memory stacks per device, for 216 GiB of capacity and 15.4 TB/s of bandwidth at a 700W thermal design power, according to the spec breakdown shared alongside the results. OpenAI says the part delivered a 1.5–1.9x improvement in per-kilowatt peak throughput and 1.7–3.6x lower end-to-end latency against a comparison system built around Nvidia GB200 while running GPT-OSS 120B, DeepSeek R1 and Kimi K2. The company also says its internal models, GPT-Astra and Codex, helped compress the design-to-tapeout cycle to nine months. The initial results were published on August 25 alongside a multi-generation roadmap for the chip family. OpenAI framed the device as an inference chip, meaning it is tuned for running trained models rather than training new ones. The design follows the same hyperscaler playbook as Google's TPU and Amazon's Trainium, making OpenAI the third cloud-scale builder to bring a custom inference chip through Broadcom.

Independent assessments, however, have pointed to several caveats in the public results. Jalapeño was not tested against Nvidia's upcoming Vera Rubin platform, and OpenAI's benchmark excluded speculative decoding, an acceleration technique that uses a small draft model to speed up inference. On raw compute, a 128-chip Jalapeño system is rated at 1.7 exaFLOPS of 4-bit MXFP4 performance, while Nvidia's GB200 and GB300 NVL72 racks carry roughly 1.46–2x more aggregate compute and about 10% more memory; Jalapeño's edge is memory bandwidth, which is close to 20% higher. OpenAI has also not disclosed actual power draw, leaving external estimates of 40–60% power savings unconfirmed. An analyst note on the HBM4 supply chain flagged Samsung's exclusive role as a risk: if OpenAI continues to rely on a single HBM4 supplier, output scaling could be capped until SK hynix or Micron parts are qualified. Richard Ho, OpenAI's chip chief, told the Hot Chips conference that Nvidia remains a very good partner and that OpenAI will still need a large volume of Nvidia silicon while production ramps. Volume production is expected in 2027, with limited supply in 2026. This leaves the latest result closer to a lab win than a deployment proof point. OpenAI also continues to use Cerebras Systems for some model workloads, citing its large compute demand.

Separately, OpenAI's infrastructure organization has lost its data-center chief. Chris Malone, who joined the company in March 2025 from Meta and Google, has left, becoming the latest executive departure as OpenAI prepares for a potential initial public offering. The exit, first reported by one US newspaper and confirmed by another outlet, was followed by a company statement that the infrastructure unit had been reorganized to meet the pace of expansion. OpenAI is targeting roughly $600 billion in compute spending by 2030, and its data-center buildout has become politically sensitive in the United States. Malone's departure follows the exits of revenue lead Denise Dresser and eight-year veteran Brad Lightcap this month, with product and business lead Fidji Simo also stepping back last month. The company insisted the data-center team remains strong and deeply experienced, with clear leadership and technical capacity to execute its infrastructure plans. OpenAI, valued at about $852 billion, filed confidentially with the SEC in June; CFO Sarah Friar has told staff the company expects to become public in 2027. Four other executives left in April, and the turnover is drawing investor attention because OpenAI is simultaneously expanding infrastructure spending and preparing for one of the largest technology IPOs on record.

Together, the three developments show an AI-infrastructure cycle advancing faster on design speed than on public proof of execution. Jalapeño's nine-month tapeout is a genuine milestone, but the missing power figure, single-source HBM4 supply and leadership churn all temper the bullish read. For Worldcoin (WLD), which trades as a proxy for AI-crypto sentiment, the near-term takeaway is not a change in fundamentals but a wider focus on how AI hardware is built. The official benchmark release puts AI trading narratives at the center of token attention; as long as OpenAI keeps shipping milestones, AI-linked assets such as WLD are likely to stay in play without a broader bear market signal.