Introduction

Fermat's Last Theorem, which has baffled humanity for over three hundred years, has actually been completely cracked by Anthropic's Claude in just eleven days! I'm Mu-Yen. Today on "Mark's Tech Insights," besides talking about this mathematical feat that involved writing tens of millions of lines of code, we'll also explore the shocking 10,000x surge in electricity bills as AI gets smarter, and Meta's bizarre strategy of publicly "putting a price tag" on everyone's private code. Want to know how much your code is worth? I'll tell you in a bit!

Today's Top Stories

1. Anthropic's Claude Cracks 358-Year-Old Math Riddle, Completing the Largest Lean Mathematical Proof in History

• Source: Anthropic (https://www.anthropic.com/research/formalizing-fermats-last-theorem）
• Summary: Anthropic announced today that its AI model, Claude, operated autonomously for eleven days on the Prove2Me platform, successfully formalizing and computer-verifying "Fermat's Last Theorem" using the Lean programming language. This project involved writing nearly 13 million lines of code, proving close to 30,000 intermediate theorems, and consuming about 6 billion tokens, making it the largest Lean proof in history.
• Most Surprising Aspect: A super-puzzle that baffled humanity for nearly 400 years—which took mathematicians years just to verify the solution back then—has now been solved by AI on its own in less than two weeks, passing verification with the strictest computer logic.
• Taiwan Perspective: Taiwan's academia and tech industry have always highly valued formal software verification. Claude's breakthrough means that in the future, not just mathematics, but also code verification for Taiwan's semiconductor chip designs could enter a golden age of full automation.
• Discussion Points:
  1. How do we build trust in AI when it can autonomously prove complex mathematics that humans cannot easily verify?
  2. Will this accelerate the adoption of formal verification in Taiwan's IC design workflows?
• Podcast Script Suggestion: Folks, this is a massive earthquake in the mathematics community! Fermat's Last Theorem baffled humanity for over three centuries. When mathematician Andrew Wiles finally cracked it, his handwritten manuscript was incredibly thick, and it took the community years to confirm there were no mistakes. Now, Claude spent just eleven days, working tirelessly to write 13 million lines of code, patching all the logical loopholes itself, and passing the strictest computer verification. It's like hiring a super math professor who never sleeps, writing a lifetime's worth of papers in eleven days. That's insane!

2. The Smarter the AI, the Hotter the Earth? Study Shows Agent Task Power Consumption Surges 10,000x

• Source: Bloomberg (https://www.bloomberg.com/news/articles/2026-09-03/ai-s-environmental-impact-per-task-balloons-with-more-complexity）
• Summary: Independent evaluation agency Vals AI assessed 16 AI models using EcoLogits and found that when AI transitions from simple Q&A to complex "AI Agent" tasks, its carbon emissions, water usage, and power consumption skyrocket by nearly 10,000 times. For example, asking an AI to write a complete web application for you consumes as much energy as an average household uses in two and a half hours.
• Most Surprising Aspect: We think of AI as just thinking in the cloud, but having it write a small web page for you consumes enough electricity to run your home's air conditioning for two and a half hours.
• Taiwan Perspective: For Taiwan, which is in a critical phase of energy transition and serves as the heart of AI computing power, this "10,000x power consumption" Agent trend will pose a massive electricity challenge for Taipower and the green energy sector.
• Discussion Points:
  1. When enterprises deploy Agents on a large scale, will carbon tariffs (like CBAM) become an unexpected cost bomb?
  2. Can Taiwan's server ODMs seize this opportunity to promote higher-efficiency liquid cooling solutions?
• Podcast Script Suggestion: Weren't you all excited hearing about Claude proving math theorems just now? But don't forget, that burned a whopping 6 billion tokens! Today, Bloomberg reported that when we upgrade AI from "chatbots" to "intelligent agents" that handle end-to-end tasks for us, power consumption immediately skyrockets by 10,000 times! Asking it to write a web page consumes enough electricity to run your living room's AC for two and a half hours. This isn't just artificial intelligence; it's an absolute power hog. It looks like Taiwan's energy-saving and cooling technologies are truly going to become the world's lifesavers next.

3. McKinsey 2026 AI Report: Over 30% of Enterprises Choose to Build Software with AI Instead of Buying

• Source: McKinsey (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai）
• Summary: McKinsey's latest global survey reveals that up to 32% of enterprises, empowered by powerful AI coding agent tools, have decided to forgo purchasing off-the-shelf software packages in favor of in-house development. In the tech industry, this figure is as high as 41%, with 40% of giant corporations deploying AI agents at scale.
• Most Surprising Aspect: The biggest competitor for traditional Software-as-a-Service (SaaS) providers has now become the customer's own in-house AI programmers.
• Taiwan Perspective: Many small and medium-sized software distributors or system integrators (SIs) in Taiwan need to be highly alert. When clients realize they can use AI to build custom ERP or CRM systems themselves, traditional software outsourcing and licensing models will face a major shake-up.
• Discussion Points:
  1. How should SaaS companies pivot to avoid being replaced by software that clients "build themselves with AI"?
  2. When Taiwanese enterprises use AI to build their own software, how should they address subsequent maintenance and cybersecurity compliance issues?
• Podcast Script Suggestion: In the past, when a company wanted a management system, they either spent a fortune on Microsoft or Oracle licenses or hired external contractors. But McKinsey's report today reveals a shocking trend: one-third of companies are now telling software vendors, "No thanks," because they are just having their own AI Agents build it! This is a massive blow to the traditional software industry. The products you worked so hard to develop can now be generated by clients with a few clicks using AI. This is a wake-up call for software service providers in Taiwan—if you're still selling rigid, off-the-shelf software packages, you'll soon be left behind.

4. Meta Offers Insane 90% Discount, but the Catch Is Using Your Private Code to Train AI

• Source: TechCrunch (https://techcrunch.com/2026/09/03/meta-is-paying-to-peek-at-how-you-use-their-latest-ai-model/）
• Summary: To promote its latest Muse Spark code model, Meta has launched an astonishing discount of up to 90% on average, slashing input token prices from $1.25permillionto1.25\; per\; million\; to$0.10. The catch is that developers must agree to let Meta read their prompts and outputted code to train future AI models.
• Most Surprising Aspect: Your trade secrets and proprietary code now have a clear price tag, and Meta loudly announced that price to the world today.
• Taiwan Perspective: Many high-tech manufacturing and chip design firms in Taiwan are extremely sensitive about intellectual property (IP). This "privacy-for-discount" scheme will likely trigger immediate red flags for cybersecurity departments in Taiwanese enterprises.
• Discussion Points:
  1. Would you be willing to hand over your company's core code to Meta as training material just to save over 90% on API costs?
  2. Will this "privacy monetization" pricing model become the new normal for future AI services?
• Podcast Script Suggestion: Meta's move today is absolutely brilliant—they are putting "privacy" and "money" on opposite sides of the scale for you to choose. Want to use their latest Muse Spark to write code? Sure, they'll give you a 90% discount. What used to cost over a dollar per million tokens is now just ten cents—it's so cheap it's practically free! But the only catch is that your prompts and the code you write will all be "glanced at" by Meta to train their next-generation models. This deal might look incredibly sweet to cash-strapped indie developers, but for Taiwan's semiconductor or hardware giants holding valuable trade secrets, cybersecurity heads will probably break into a cold sweat just looking at this contract!

5. Tesla Cybercab Hits the Streets of Austin in Style, Only to Face Federal Investigation Hours Later

• Source: TechCrunch (https://techcrunch.com/2026/09/04/feds-launch-investigation-into-teslas-cybercab-deployment/）
• Summary: Just hours after Tesla hosted its Cybercab launch event in Austin, the National Highway Traffic Safety Administration (NHTSA) launched an investigation into approximately 1,000 of its autonomous robotaxis on the road. The vehicle completely lacks a steering wheel, brake pedals, and rearview mirrors. Tesla had previously deployed them based on "self-certification" of compliance with federal safety standards, and now regulators are strictly reviewing whether this self-certification complies with regulations.
• Most Surprising Aspect: The champagne from the new car launch had barely been popped when the federal investigation order arrived. The crux of this dispute is whether Tesla is essentially "grading its own final exam."
• Taiwan Perspective: Taiwan is also actively promoting autonomous driving and smart transportation testing. This "self-certification" battle between Tesla and US regulators will serve as an important reference model for Taiwan's future autonomous vehicle regulations.
• Discussion Points:
  1. Can Tesla's "hit the road first, clash with regulators later" strategy actually accelerate the adoption of autonomous driving?
  2. For fully autonomous vehicles without steering wheels or pedals, who should legally bear the ultimate safety responsibility?
• Podcast Script Suggestion: Tesla's Elon Musk never runs out of headlines! The Cybercab launch event in Austin had just wrapped up, and everyone was still marveling at that ultra-futuristic design with no steering wheel, no brake pedals, and not even rearview mirrors. Yet, just a few hours later, the NHTSA's investigation order slammed right down. The funniest part is that Tesla previously told the government, "We tested it ourselves, it's absolutely safe." Isn't that just grading your own final exam? Now the government is coming to regrade the paper. This head-on clash of "tech giant vs. national regulator" is honestly more exciting than a Hollywood movie!

6. SEMICON Taiwan 2026 Debuts with Record Numbers! Taiwan's Chip Output Surges 40%, but Is AI's Bottleneck Now "This Wire"?

• Source: The Manila Times / AFP (https://www.manilatimes.net/2026/09/03/business/foreign-business/taiwan-chip-industry-sees-revenue-soaring-40/2417319）
• Summary: SEMICON Taiwan 2026 opened grandly in Taipei, hitting record-breaking scales and attracting over 100,000 professionals. As Taiwan's semiconductor output is projected to surge 40% this year to approach $300 billion, the exhibition sent out a critical signal: the performance bottleneck of AI systems has shifted from "chip computing" to "inter-chip transmission lines," with the energy consumed by data movement now surpassing that of computation itself.
• Most Surprising Aspect: While Taiwan's chip output is set to hit $300 billion this year, experts warn that the bottleneck for AI computing power is no longer how fast the chips are, but rather the "traffic congestion and power consumption" between them.
• Taiwan Perspective: This means TSMC's advanced packaging technologies (such as CoWoS) and Silicon Photonics will be the ultimate remedies for Taiwan to maintain its crown as the "Silicon Island" and solve global AI transmission bottlenecks.
• Discussion Points:
  1. When transmission energy consumption exceeds computing energy consumption, what new opportunities does this bring to Taiwan's packaging and communication chip industries?
  2. How can Taiwan leverage this 40% surge in semiconductor output to further consolidate its bargaining power in the global AI supply chain?
• Podcast Script Suggestion: Speaking of AI, how could we leave out Taiwan? SEMICON Taiwan 2026, held in Taipei over the past few days, was absolutely packed. Output is projected to skyrocket to $300 billion, with a year-on-year growth rate of 40%—that's just insane! However, the hottest topic at this year's exhibition wasn't how much smaller chips have gotten, but rather the "transmission bottleneck." Simply put, chips are computing too fast now, but the speed of data "queuing up for transmission" between chips is too slow, and the power consumed during transmission is actually more than the chip computation itself! It's like building an ultra-high-speed bullet train, but having only a single ticket gate at the station. It looks like TSMC's advanced packaging and Silicon Photonics are once again the only hope for saving the world's AI!

Outro

Looking at everything today, whether it's Claude solving a centuries-old math puzzle or the new chip bottlenecks revealed at SEMICON Taiwan, we can see that AI is developing at a breathtaking pace. However, the accompanying surge in electricity bills, privacy controversies, and regulatory challenges remind us that while we pursue technology, the real-world costs are not to be ignored. Thank you for listening to today's "Mark's Tech Insights." I'm Mu-Yen, and I'll see you next time!