NVIDIA's largest customer took out a benchmark test of a competitor's chip on its earnings day, and Iran lost at the negotiating table the moment it lost control of the Strait of Hormuz.
1｜OpenAI Unveils Jalapeño Chip on NVIDIA Earnings Day, Officially Starting the Inference Battlefield
OpenAI released the first benchmark test of its in-house inference chip Jalapeño at Hot Chips 2026. According to SemiAnalysis's InferenceX test, Jalapeño achieved 1.5 to 1.9 times the throughput per kilowatt on three large language models compared to NVIDIA's GB300, with end-to-end latency reduced by 1.7 to 3.6 times. The performance gap widens under interactive loads, being nearly 2.1 to 4.1 times faster. Drawing 700 watts of power, it competes against a flagship chip drawing 1,400 watts.
Timing is the signal itself. NVIDIA released its largest-ever single-quarter earnings report tonight after hours (with a consensus revenue of $92 billion), while OpenAI chose the same day to announce its benchmark. Jalapeño, developed in collaboration with Broadcom, went from team formation to silicon in just 16 months. As reported by Tom's Hardware, the deployment schedule targets small batches by the end of 2026 and mass production in 2027. However, by then NVIDIA's product lineup will also have iterated. What truly changes is not who is faster but the inference chip market shifting from "NVIDIA exclusive" to a dual-track pattern of "customer in-house development + NVIDIA."
(Source: SemiAnalysis / TechCrunch / Tom's Hardware)
2｜US Notifies Allies of Shift from Strike to Sanctions as the Hormuz Mine Clearance is Completed
According to an exclusive report by Axios, US Secretary of State Pompeo recently informed foreign ministers of multiple countries that the US will "temporarily" refrain from launching a new round of strikes against Iran, turning to economic pressure and ensuring navigation through the Strait of Hormuz. On the same day, Trump announced that the US Navy had cleared all mines from the main shipping channel in the strait.
The underlying logic of this shift is leverage reversal. US officials told Axios that the mine clearance marks a "turning point" in this conflict, neutralizing Iran's primary leverage over the global energy market. In the past two weeks, Iran's main oil export terminal on Khark Island saw nearly no tankers docking. Crude oil fell over 5% this week. A US official was quoted as saying, "The Iranians have lost control of the strait, and now the US is controlling it." This policy is expected to continue at least until after the midterm elections. Pakistani Army Chief General Munir visited Tehran on Sunday for mediation, but the White House stated there are currently no plans for negotiations with Iran. (Continued from yesterday's report)
(Source: Axios / ABC News / CNBC / Times of Israel)
3｜Anthropic Aiming for $2 Trillion IPO in October, to Become Largest Ever
According to reports from Fortune and Yahoo Finance, Anthropic's investors are pushing for an IPO in October with a target valuation of $2 trillion or more, surpassing SpaceX's $1.77 trillion valuation during its June listing, making it the largest IPO in history. Goldman Sachs, JPMorgan, and Morgan Stanley are the joint bookrunners, with the public S-1 filing expected by the end of August. CNBC reports that the prospectus will explicitly list "AI backlash" as a risk factor.
Supporting this valuation is revenue growth. As reported by Bloomberg on August 17, Anthropic's annualized revenue has exceeded $650 billion, growing over 7 times since the end of 2025. The company's API market share reached 32% in Q2, surpassing OpenAI's 25% for the first time. From a security lab to the world's most valuable IPO, Anthropic has completed this journey in less than four years. Valuation is a signal of power, indicating how much the market believes there is still room for growth in the AI infrastructure layer.
(Source: Fortune / Yahoo Finance / Bloomberg / CNBC)
4｜NVIDIA Reports $920 Billion Earnings Tonight, Surprise Margin Narrows to 4.6%
NVIDIA released its Q2 FY2027 earnings report after Wednesday's market close, with Wall Street consensus revenue of $920 billion, representing a growth of approximately 96% year-over-year. The company's own guidance was $910 billion plus or minus 2%, with the consensus exceeding the midpoint by 0.9%. According to Yahoo Finance, NVIDIA has surpassed its own guidance for 13 consecutive quarters, but the extent of these beats has steadily narrowed from 22.8% in Q2 FY2024 to 4.6% in the last quarter.
The narrowing itself is a signal. It signifies that the market has already priced in NVIDIA's consistent outperformance, amplifying reactions to any numbers below expectations. On the same day, OpenAI released benchmarks for its competitor inference chip (see Item 1), and Google is reportedly spending $900 million a month to rent 110,000 GPUs. NVIDIA's moat is not just quarterly revenue but the depth of the entire AI compute infrastructure cycle's reliance on it. Tonight's numbers are not answering "how much was earned" but rather "is the reliance deepening or starting to loosen?" (Continuation of yesterday's report)
(Source: Investing.com / Yahoo Finance / IG International)
Also worth noting ↓
Dwarkesh Patel's July post argued that AI hash rate prices could skyrocket 10 to 15 times. While the top labs saw a 10x increase in revenue, the hash rate capacity only grew by 3x. The profit margin cannot reach 95% to bridge this gap. SemiAnalysis founder Dylan Patel predicted in the same podcast episode that Anthropic and OpenAI will control the majority of global available hash rate by 2028. (Source: Dwarkesh Podcast / SemiAnalysis)
Stability AI raised $76 million in funding, with the image generation track continuing to attract investment in consolidation. In a new round of funding following a significant drop in valuation, the market is betting on survivors. (Source: TechCrunch)
Perplexity partners with NVIDIA to develop a portable AI computing device. The search engine company focuses on hardware, while NVIDIA expands its ecosystem on the inference side. (Source: TechCrunch)
Chris Malone, Head of Data Center Operations at OpenAI, resigns during a period of accelerated AI infrastructure development. Data centers are the AI company's most significant assets, so the departure of the operations executive during an expansion phase is worth noting. (Source: TechCrunch)
Apple's M6 chip will be the first 2nm processor, but it will skip the Pro and Max versions directly to the M7 series. TSMC's N2 process is the first customer to adopt it. The base model MacBook Pro will be updated by the end of this year, while professional users will have to wait until 2027. (Source: MacRumors / Cult of Mac)
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
