🐦 推特Wall St EngineCodex · 实盘不做把握 4/10不做原因:证据不足
北京时间·🕒 发布 22:14:13时区推断·⟳ 抓取 22:14:23·⚡ 决策 08-25 23:02:39 +48m16s·62.7 天前
SemiAnalysis披露OpenAI Jalapeño推理芯片早期基准,显示能效优于部分英伟达、AMD和Google加速器
事件方向对博通偏正、对英伟达和AMD偏负,但公开事实主要是早期基准披露和2027年远期放量,缺少当前可量化收入增量、实时报价、内部去重和事件研究结果,不能形成可执行正期望订单。
📄 原文Wall St Engine@wallstengine↗ 原文
More details on OpenAI’s Jalapeño chip from SemiAnalysis: SemiAnalysis says Jalapeño is not narrowly optimized for Open…
不做 博通(AVGO.US)
方向偏多但不是首次披露合作,工程样品到2027年放量,缺少当前报价、订单金额和收入占比,不能给可执行LONG。
不做 英伟达(NVDA.US)
事件对推理GPU竞争偏负,但Jalapeño不覆盖训练且OpenAI仍继续使用外部芯片,未达到明确强空催化。
不做 AMD(AMD.US)
被基准对比削弱叙事,但没有客户流失、订单取消或业绩下修证据,做空强度不足。
不做 Celestica(CLS.US)
可能受益于机架和系统集成,但未取得量级、收入弹性和行情校验,暂不交易。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
结论摘要
Jalapeño证明OpenAI自研推理ASIC能力增强,直接合作方博通最相关,NVDA/AMD存在远期竞争压力;但该消息并非首次披露OpenAI-Broadcom合作,芯片仍处工程样品阶段,生产放量主要在2027年,且OpenAI仍会继续使用外部AI芯片,短线做多博通或做空英伟达都缺少量级和价格锚。
消息传播链
- OpenAI与Broadcom已在6月24日官方发布Jalapeño,确认其为双方共同构建的LLM推理加速器平台。
- TechCrunch报道OpenAI在Hot Chips披露Jalapeño更详细基准结果,称其在tokens/user和每千瓦吞吐方面领先当前推理处理器。
- The Verge称Jalapeño由OpenAI与Broadcom开发,InferenceX测试显示相对GB200/GB300有能效和延迟优势,但2026年部署有限、2027年放量。
- DCD报道OpenAI在Hot Chips披露700W TDP、128芯片机架、2048 ASIC pod和15.4TB/s HBM4带宽等硬件细节。
关键理由
- 博通是OpenAI Jalapeño官方合作方,若2027年量产兑现,将强化其定制AI ASIC收入叙事。
- Jalapeño只覆盖推理且仍为工程样片,短期对英伟达训练GPU和整体AI基础设施需求替代有限。
- 量级句:公开材料只给出约100MW下一部署目标和2027年逐步放量,未披露芯片ASP、博通收入份额或OpenAI采购替代比例,无法估算对博通收入、市值或英伟达收入的短线增量占比。
主要风险
- 若SemiAnalysis完整报告显示测试口径有限或AgentX长上下文负载表现弱于预期,博通受益逻辑会降级。
- 若市场已在6月OpenAI-Broadcom官方发布和此前Reuters/The Information报道中充分定价,8月基准披露可能成为见光死。
反向证据
- 做多前查反向 · 来源 web · 日期 2026-08-25 · 影响 weakens
公开报道显示Jalapeño仍为工程样品,2026年部署有限、2027年才逐步放量,削弱博通短线收入确认强度。
检索:OpenAI Jalapeño production ramp 2027 Broadcom revenue replace Nvidia 原文 ↗ - 做空前查反向 · 来源 web · 日期 2026-08-25 · 影响 weakens
Axios称Jalapeño只用于推理且OpenAI仍依赖外部硬件进行模型开发,不支持对英伟达作强空。
检索:OpenAI Jalapeño inference only Nvidia still depends external hardware training 原文 ↗ - 做多前查反向 · 来源 web · 日期 2026-06-24 · 影响 weakens
OpenAI-Broadcom合作在6月已官方披露,8月基准细节的预期差低于首次合作公告。
检索:OpenAI Broadcom Jalapeño chip official June 24 2026 原文 ↗
先例如何影响了这次决策
对本次决策的影响:先例把博通方向保留为正向观察,但不足以在无报价、无量级和无event_study结果下给ENTER;同时否决基于该消息直接做空NVDA/AMD。
方向:先例支持定制AI ASIC合作对ASIC供应商偏正,但对GPU龙头的负面更多是中长期竞争叙事而非强单日做空。
涨幅量级:缺少同市场、同机制、同相位且有实测日线的高相似样本;Marvell-Google有正式商业协议和权证,强度高于本次基准披露。
形态:可见样本中,正式商业协议反应更强,媒体线索或技术发布更依赖预期差,容易被既有叙事吸收。
历史先例(按相似度)
- 相似度 7/10 AVGO 2024-10-29 Reuters独家报道OpenAI与Broadcom、TSMC打造首款内部AI芯片,同时仍使用AMD和Nvidia芯片。 原文 ↗
事件后 +1日 -1.4% · +2日 -5.3% · +3日 -5.8% · +4日 -6.0% · +5日 -3.0%
⚠非首发(转载/聚合)·来源层级不符(声明wire/实为aggregator)·未核对目标相位 — 该行涨幅别用来定止盈止损
相似度依据:同为OpenAI、Broadcom、TSMC自研AI芯片合作确认,机制与本次最接近,但不是性能基准。
经验:合作确认强化博通定制AI芯片敞口,但对英伟达替代的短期交易结论不能自动外推。 - 相似度 6/10 AVGO 2024-07-18 The Information报道OpenAI与Broadcom讨论AI芯片后,Bloomberg称博通股价上涨约3%。 原文 ↗
事件后 +1日 -2.0% · +2日 +0.3% · +3日 +2.0% · +4日 -5.7% · +5日 -7.0%
⚠来源层级不符(声明media/实为issuer)·锚前1-2日已异动5.2%>锚定日2.0%·未核对目标相位 — 该行涨幅别用来定止盈止损
相似度依据:同为OpenAI与Broadcom定制AI芯片合作消息,但当时是媒体线报而非基准披露。
经验:定制ASIC客户线索能推升博通,但线报阶段更像叙事重估而非订单兑现。 - 相似度 6/10 MRVL 2026-08-20 Google与Marvell达成AI芯片合作并授予最高约122亿美元权证,Marvell涨约10%、Broadcom跌约5%。 原文 ↗
事件后 +1日 +5.8% · +2日 -0.1% · +3日 -3.4%
⚠锚前1-2日已异动9.8%>锚定日5.8% — 该行涨幅别用来定止盈止损
相似度依据:同属超大客户定制AI芯片合作,且涉及与Broadcom份额竞争,但不是OpenAI和Jalapeño。
经验:明确商业协议和股权激励可显著推升定制硅供应商,单纯性能披露力度弱于正式商业协议。 - 相似度 5/10 MSFT 2023-11-15 微软Ignite发布Maia 100 AI加速器,媒体称微软股价随后触及纪录高位。 原文 ↗
事件后 +1日 +0.0% · +2日 +1.8% · +3日 +0.1% · +4日 +2.1% · +5日 +1.0%
相似度依据:同为大型云/AI公司发布自研AI加速器,但交易标的是平台公司自身且微软市值逻辑不同。
经验:自研芯片发布通常更利于平台公司利润率叙事,对既有GPU供应商不是单日强空。
原文全文
展开原文 · 2,276 字
More details on OpenAI’s Jalapeño chip from SemiAnalysis:
SemiAnalysis says Jalapeño is not narrowly optimized for OpenAI models.
It is a general-purpose inference ASIC, and in the workloads they tested it outperformed $NVDA, $AMD and Google accelerators on several inference efficiency metrics.
The chip was designed from scratch for LLM inference and went from initial team buildout to tape-out in roughly 16 months.
Performance:
• DeepSeek R1 exceeded 700 tokens/sec/user at concurrency 1
• Kimi K2.5 and GPT-OSS reached roughly 1,400 tokens/sec/user in selected tests
• Jalapeño achieved these results using single-token prediction, without speculative decoding or prefill/decode disaggregation
• SemiAnalysis says its output-token throughput per MW already exceeds the published July Vera Rubin results, despite Rubin using multi-token prediction
• On performance per total cost, Jalapeño and Vera Rubin are currently roughly comparable based on available results
These are still early benchmarks.
SemiAnalysis tested 8k/1k workloads and has not yet run its more demanding AgentX benchmark, which stresses long-context, multi-turn agentic workloads. Rubin is also already beginning to ship, while Jalapeño remains at the engineering-sample stage.
Hardware:
• Built on TSMC N3P
• 700W TDP
• HBM4 with 15.4TB/s of memory bandwidth
• SemiAnalysis says that is currently the highest memory bandwidth per package among shipping/near-shipping accelerators
• A0 silicon produced the current benchmark results
• B0 is already in the fab and is expected to improve perf/W by roughly 25%
• B0 delivers 13.4 PFLOPs of MXFP4 versus 17.5 PFLOPs of dense NVFP4 for a similarly sized Rubin compute die, but at lower power
Scale:
• 128 Jalapeño ASICs per rack
• Up to 2,048 XPUs can be connected across a 16-rack scale-up domain
• A complete host + ASIC rack pair consumes roughly 160kW
• OpenAI’s next deployment target is around 100MW
OpenAI is also using Codex heavily in the chip’s development. SemiAnalysis says AI-assisted design reduced SIMD area by 8% and matrix-engine area by 10%, while Codex has been used to write and optimize kernels for new models.
Production is expected to ramp gradually through 2027, with most output currently planned for Q4 2027.
SemiAnalysis says Jalapeño is not narrowly optimized for OpenAI models.
It is a general-purpose inference ASIC, and in the workloads they tested it outperformed $NVDA, $AMD and Google accelerators on several inference efficiency metrics.
The chip was designed from scratch for LLM inference and went from initial team buildout to tape-out in roughly 16 months.
Performance:
• DeepSeek R1 exceeded 700 tokens/sec/user at concurrency 1
• Kimi K2.5 and GPT-OSS reached roughly 1,400 tokens/sec/user in selected tests
• Jalapeño achieved these results using single-token prediction, without speculative decoding or prefill/decode disaggregation
• SemiAnalysis says its output-token throughput per MW already exceeds the published July Vera Rubin results, despite Rubin using multi-token prediction
• On performance per total cost, Jalapeño and Vera Rubin are currently roughly comparable based on available results
These are still early benchmarks.
SemiAnalysis tested 8k/1k workloads and has not yet run its more demanding AgentX benchmark, which stresses long-context, multi-turn agentic workloads. Rubin is also already beginning to ship, while Jalapeño remains at the engineering-sample stage.
Hardware:
• Built on TSMC N3P
• 700W TDP
• HBM4 with 15.4TB/s of memory bandwidth
• SemiAnalysis says that is currently the highest memory bandwidth per package among shipping/near-shipping accelerators
• A0 silicon produced the current benchmark results
• B0 is already in the fab and is expected to improve perf/W by roughly 25%
• B0 delivers 13.4 PFLOPs of MXFP4 versus 17.5 PFLOPs of dense NVFP4 for a similarly sized Rubin compute die, but at lower power
Scale:
• 128 Jalapeño ASICs per rack
• Up to 2,048 XPUs can be connected across a 16-rack scale-up domain
• A complete host + ASIC rack pair consumes roughly 160kW
• OpenAI’s next deployment target is around 100MW
OpenAI is also using Codex heavily in the chip’s development. SemiAnalysis says AI-assisted design reduced SIMD area by 8% and matrix-engine area by 10%, while Codex has been used to write and optimize kernels for new models.
Production is expected to ramp gradually through 2027, with most output currently planned for Q4 2027.
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
这条决策早于决策账本(2026-10-08 起逐条追加),没有追加记录。