🎙️ OpenAI 辣椒晶片挑戰 NVIDIA！人形機器人跑贏波特背後秘密｜AI 日報 Podcast

Intro

Today, Mu-Yen, host of "Mark's Tech Insights," brings you today's most mind-blowing tech news! A humanoid robot in Beijing actually ran the 100-meter dash in 9.32 seconds, shattering the 17-year-old world record held by Jamaican lightning bolt Usain Bolt. Later, we'll also tell you how OpenAI's latest self-developed "Jalapeño" chip is threatening NVIDIA's dominance, and why a pioneer platform for AI data labeling ultimately died in an absurd loop of "humans pretending to be AI"!

Today's Top News

1. Amazon Announces Shutdown of Mechanical Turk, the Service Jeff Bezos Once Called "Artificial Artificial Intelligence"

Source: CNBC (https://www.cnbc.com/2026/08/25/amazon-service-that-jeff-bezos-called-artificial-ai-is-shutting-down.html)
Summary: Amazon has announced it will shut down its Mechanical Turk platform on September 30, ending 21 years of operation. The platform once gathered 500,000 human workers to help label training data for machine learning at extremely low pay.
Most Surprising Aspect: The platform ultimately collapsed because up to 46% of human workers were secretly using AI to generate answers. This meant a platform designed for "humans acting as AI" ended up dying at the hands of "humans pretending to be AI."
Taiwan Perspective: Taiwan previously had many local startups focusing on data labeling. This news reflects that the era of relying solely on cheap manual labor for labeling has come to an end. In the future, high-end data refinement combined with professional fields like medicine and law will be the new opportunity for Taiwanese teams.
Discussion Points:
1. When the data used to train AI is itself generated by AI, will it lead to model collapse?
2. How should the gig economy transform in the AI era?
Script Suggestion: This is truly an epic piece of dark humor: Amazon is shutting down Mechanical Turk after 21 years of operation. This platform used to hire hundreds of thousands of humans to label data for machine learning for just a few cents per task, which Jeff Bezos jokingly called "artificial artificial intelligence." The ultimate irony is that researchers found that in the platform's later stages, up to 46% of workers were secretly using AI to write their answers. This means the humans we thought were training AI were actually using AI to pretend to be human, just to half-ass tasks for another AI that needed training. The platform basically played itself to death, signaling the end of the cheap manual labeling era.

2. OpenAI Unveils First Custom Chip "Jalapeño," Beating NVIDIA in Inference Performance Per Watt

Source: CNBC (https://www.cnbc.com/2026/08/26/openai-jalapeno-ai-chip-nvidia.html)
Summary: OpenAI has released performance benchmarks for its first custom inference chip, "Jalapeño." Its inference performance per watt is 1.5 to 1.9 times higher than NVIDIA's Blackwell, with significantly reduced end-to-end latency. These figures have been verified by a third-party organization.
Most Surprising Aspect: Although Jalapeño uses newer HBM4 memory, giving it a slight hardware advantage in comparisons, its energy efficiency still wins out even when compared to NVIDIA's next-generation Vera Rubin, which is also equipped with HBM4.
Taiwan Perspective: While OpenAI is flexing its custom chip muscles, this chip utilizes cutting-edge HBM4 technology. Behind the scenes, whether it's wafer fabrication or advanced packaging, it absolutely relies on the full support of TSMC and Taiwan's semiconductor supply chain.
Discussion Points:
1. Will OpenAI's custom chips accelerate Big Tech's push to break their sole reliance on NVIDIA?
2. Will application-specific integrated circuits (ASICs) for inference completely replace general-purpose GPUs in the commercial market?
Script Suggestion: OpenAI is really bringing the heat this time. Their custom inference chip, Jalapeño, has some seriously spicy specs—its inference performance per watt is nearly double that of NVIDIA's Blackwell. While some critics argue that using the newer HBM4 memory is a bit like using a cheat code, microprocessor research firms found in testing that even when compared to NVIDIA's next-gen Vera Rubin (which also features HBM4), this Jalapeño chip's token generation speed is still mind-blowing. This is fantastic news for Taiwan's supply chain because whether NVIDIA or OpenAI wins in the end, the advanced nodes and advanced packaging behind it all will still be outsourced to TSMC.

3. Anthropic IPO Imminent, Targeting a $2 Trillion Valuation to Challenge Wall Street Records

Source: Euronews Business (https://www.euronews.com/business/2026/08/26/anthropic-ipo-five-things-to-know-before-its-wall-street-debut)
Summary: Anthropic, the developer of the Claude model, is expected to launch its IPO within weeks, targeting a valuation of up to $2 trillion. Founded by former OpenAI employees, the company currently generates up to 80% of its revenue from enterprise customers and achieved operating profitability for the first time in Q2.
Most Surprising Aspect: These founders originally left because they felt OpenAI didn't prioritize safety enough. Now, their annualized revenue has skyrocketed from $9 b i l l i o n l a s t y e a r t o$ 47 billion as of May this year.
Taiwan Perspective: When adopting generative AI, Taiwanese enterprises highly prefer Anthropic's Claude model due to geopolitical and data security concerns. If this IPO succeeds, it will significantly boost the confidence of Taiwanese enterprise clients in long-term partnerships.
Discussion Points:
1. Why does a "safety-first" AI approach actually win the favor of enterprise clients?
2. Does Anthropic's $2 trillion valuation raise concerns about a new tech stock bubble?
Script Suggestion: The "traitors" who originally walked out of OpenAI might be about to make history. Anthropic, the parent company of Claude, is expected to go public on the US stock market within weeks, targeting a staggering valuation of $2 trillion—a scale that directly challenges SpaceX's previous records. What's most impressive is that they built their brand on "safety first," and now 80% of their revenue comes from enterprise clients, even turning a profit in Q2 of this year. This teaches us a lesson: in the tech world, being the "safety-first" good kid can still make you a fortune as long as your product is solid, and it can even win over enterprises that were originally skeptical about AI.

4. Beijing Humanoid Robot Games: "Lightning" Runs 100m in 9.32 Seconds, Breaking the Human World Record

Source: NBC News (https://www.nbcnews.com/tech/tech-news/chinese-humanoid-robot-lightning-beats-human-100m-world-record-rcna593869)
Summary: At the 2nd World Humanoid Robot Games held in Beijing, multiple robots broke the 100-meter world record held by Usain Bolt for 17 years, with Honor's "Lightning" robot running a 9.32-second race.
Most Surprising Aspect: The competing robots did not rely entirely on hard-coded programming by engineers. Instead, through reinforcement learning, they "invented" a brand-new sprinting posture with arms tucked close to the face and driving movement from the hips.
Taiwan Perspective: For humanoid robots to run fast and stand stable, precision motors, reducers, and sensors at the joints are critical. This is precisely the strength of Taiwan's precision machinery and electronic component manufacturers. The Taiwanese supply chain will play a core role in this robot wave.
Discussion Points:
1. What does it mean when AI evolves physical movements through reinforcement learning that humans never taught it?
2. When robot athletic performance completely surpasses humans, will the definition of sports events change?
Script Suggestion: Usain Bolt's 17-year-old 100-meter world record was actually shattered by a tin-can robot. At the games in Beijing, Honor's Lightning robot clocked an incredible 9.32 seconds. But what really gave me goosebumps wasn't its speed—it was that one of the robots, given only the objective to "run as fast as possible," used reinforcement learning to invent a bizarre sprinting posture with its arms tucked close to its face and driving forward from the hips. Humans didn't teach it this; it calculated this optimal solution on its own. It's honestly a bit chilling—AI is redefining physical limits in ways we can't even comprehend.

5. Stanford Study: AI Companions May Exacerbate Loneliness in Vulnerable Users

Source: Stanford University (https://news.stanford.edu/stories/2026/08/ai-companions-chatbots-loneliness-research)
Summary: A Stanford research team analyzed conversation data from over a thousand Character.AI users and found that users who frequently confide in AI chatbots actually became lonelier in real life.
Most Surprising Aspect: This negative impact is selective: those with smaller real-world social circles who rely more heavily on AI for emotional companionship suffer the most, creating a vicious cycle of "the lonelier you get, the more you chat; the more you chat, the lonelier you get."
Taiwan Perspective: Taiwan is facing a declining birthrate and a rising proportion of single people, making AI companionship services highly accepted in the local market. This study serves as a timely reminder that when developing social AI products, we must incorporate mechanisms that guide users back to real-world social interactions.
Discussion Points:
1. Why does an AI that is on-call 24/7 and perfectly caters to the user fail to solve human loneliness?
2. Should tech companies bear corresponding ethical responsibilities for mental health when developing AI companions?
Script Suggestion: Do you chat with Character.AI when you're lonely? A study from Stanford University, just published in a Nature portfolio journal, has thrown some cold water on that. They analyzed thousands of users and nearly 500,000 conversations, finding that people who try to find warmth in AI actually end up feeling lonelier in real life. It's a brutal, vicious cycle: the lonelier you are, the more you want to talk to a perfect, on-call AI. But after chatting, you find real humans too troublesome and imperfect, causing you to withdraw even further. Technology gives us meticulous virtual companionship, but it also acts like a wall, completely isolating us from the real world.

6. NVIDIA Q2 Earnings Out Tonight, Market Predicts Record-High Data Center Revenue

Source: S&P Global Market Intelligence (https://www.spglobal.com/market-intelligence/en/news-insights/research/2026/08/nvidia-earnings-preview-q2-2027)
Summary: NVIDIA is set to release its Q2 FY2027 financial results after the bell today. The consensus market expectation for revenue is a staggering $92.2 b i l l i o n, w i t h t h e d a t a c e n t e r s e g m e n t e x p e c t e d t o c o n t r i b u t e$ 85.7 billion—an upward revision of over 50% compared to expectations from a year ago.
Most Surprising Aspect: The projected revenue for the data center segment alone was collectively revised upward by analysts by over 50% in just one year, showing that the global AI compute arms race shows no signs of cooling down.
Taiwan Perspective: NVIDIA's earnings report is a major bellwether for Taiwanese tech stocks. Everything from TSMC's advanced capacity to second-half orders for server ODMs like Quanta, Wistron, and Inventec depends on the scorecard Jensen Huang presents tonight.
Discussion Points:
1. With data center revenue revised upward by over 50%, does this mean there is truly no ceiling for AI compute demand?
2. If NVIDIA's earnings fall even slightly short of expectations, how big of an impact will it have on the Taiwanese supply chain?
Script Suggestion: Tonight, retail investors and tech bosses across Taiwan will probably lose sleep because NVIDIA is about to drop its latest quarterly earnings. This isn't just any earnings report; it's the thermometer for the entire AI industry. The market currently expects their single-quarter revenue to reach a mind-blowing $92.2 b i l l i o n, w i t h o v e r$ 80 billion coming from selling chips to data centers alone—a figure over 50% higher than what everyone anticipated a year ago. If the scorecard Jensen Huang delivers tonight is even slightly less than perfect, Taiwan stocks might experience a massive earthquake tomorrow. Get your popcorn ready; let's watch the results unfold tonight.

Outro

Today, we covered everything from robots outrunning humans to AI companions making people lonelier—the pace of tech development is truly dizzying. I hope today's content has fully charged your tech knowledge. I'm Mu-Yen, host of "Mark's Tech Insights." See you next time!