TPUs vs GPUs and why Google is positioned to win AI race

The Byte’s Bits

🧠 Why Google's Stack Makes TPUs Unstoppable in the AI Race

💥 Holy S***! Salesforce CEO Marc Benioff Dumps ChatGPT for Google's Gemini 3

🚨 X's Location Feature Exposes Foreign-Run Political Accounts

✨ "Vibe Coding" Is Making Software Development Fun Again, Says Sundar Pichai

🔔 How a $200 Doorbell became a $4B Business. The Story of the Ring Doorbell

While the early years of AI were dominated by the race to train the largest models, the industry is now entering the "Age of Inference" the phase where those models are actually deployed and run. Google's custom-built Tensor Processing Unit (TPU) is designed specifically for this massive, cost-sensitive market, giving the company a key strategic advantage through deep vertical integration.

🔍 The Inference Shift: The AI computation market is rapidly shifting from the expensive process of training models (Nvidia's GPU stronghold) to inference (running the models). This market is expected to be orders of magnitude larger, as models are used billions of times daily, creating a massive, cost-sensitive demand.

🧠 ASIC's Time to Shine: Unlike general-purpose GPUs, Google’s TPUs are Application-Specific Integrated Circuits (ASICs) custom-designed for machine learning tasks. This specialization allows them to deliver superior performance per watt/dollar for inference workloads, offering a structural cost advantage over competitors.

🔗 The True Moat is Scale: The TPU's real power is its massive parallel scale, enabled by Google's unique Optical Circuit Switch (OCS) interconnects. This custom networking hardware links thousands of TPUs together, allowing Google to run large-scale distributed inference that rivals even the largest single-chip Nvidia systems.

💰 Vertical Integration Wins: By owning the entire stack, the chip (TPU), the interconnects (OCS), the software framework (JAX/XLA), and the cloud, Google can provide AI services at scale far more cost-effectively than competitors, positioning itself to capture the long-term, high-volume inference market.

I’m a firm believer in whoever owns the entire stack will win the race regardless of what race we are in…

Salesforce CEO Marc Benioff, a long-time daily user of ChatGPT, made a highly public declaration that Google's new Gemini 3.0 has permanently changed his preference, calling the model's performance "insane" and signaling a major shift in the competitive landscape of frontier AI.

⚡️ The 2-Hour Conversion: Benioff stated he had used ChatGPT every day for three years, but after spending just two hours with Gemini 3, he was convinced, declaring, "I'm not going back. The leap is insane."

🚀 Multimodal Superiority: The CEO specifically praised Gemini 3's superior performance across multiple modalities, noting improvements in reasoning, speed, image quality, and video processing—key areas that showcase the model's advanced, unified architecture.

🔗 Unprecedented Endorsement: Benioff's verdict carries immense weight as the head of Salesforce, a major enterprise platform with extensive AI partnerships (including with OpenAI), making his public switch a clear sign of Google's technological lead in the latest AI model generation.

🎯 The Race Heats Up: The viral announcement underscores the mounting pressure on OpenAI as rival tech giants like Google make significant advances, with the competitive landscape rapidly shifting based on the capabilities of the newest large language models.

I’m in the same boat. I noticed I am spending more time in Gemini than ChatGPT or Claude. I find the integration with the Google ecosystem and AI Studio purely outperforms the competition. Have an opinio,n drop a comment at the end of the newsletter.

A new transparency tool rolled out by X (formerly Twitter) last week, which displays an account's country of origin, instantly revealed that dozens of influential accounts claiming to represent U.S. political voices are actually operated from overseas, sparking a fresh debate on foreign influence.

🌍 Unmasking Foreign Operators: The feature allows users to see the country or region where an account is based by clicking on its signup date. Almost immediately, users and researchers discovered numerous popular accounts—many pushing pro- or anti-MAGA and pro-Democrat narratives—were based in places like Eastern Europe, Nigeria, India, and Bangladesh, contradicting their claims of being U.S.-based patriots.

🛡️ Tool's Purpose: X's head of product, Nikita Bier, stated the "About This Account" tool was introduced as "an important first step to securing the integrity of the global town square" and combating foreign troll farms by verifying content authenticity.

⚠️ Immediate Inaccuracies: The rollout was marred by issues, leading to the feature being briefly disabled. Some accounts of verified users, including NBC News journalists, incorrectly displayed locations matching recent travel rather than their actual home bases. The platform acknowledged the data "may not be accurate" due to VPN usage or changing IP ranges.

⚖️ Mixed Motivations: While some exposed foreign-run accounts may be tied to state-sponsored influence operations, experts noted that many are likely financially motivated grifters who simply post polarizing content, memes, and misinformation to drive engagement and profit.

Google CEO Sundar Pichai says a quiet revolution is taking place in software development, dubbed "vibe coding." This AI-assisted trend allows people without traditional engineering backgrounds to build functional apps simply by describing their desired outcome, making the creation process more enjoyable and accessible.

🎨 Low-Friction Prototyping: Vibe coding, a concept popularized by Google's Gemini team, refers to using advanced AI models like Gemini 3 to generate code and user interfaces based on high-level, natural language prompts (or "vibes"). This allows users—from HR staff to product managers—to prototype apps and visualize ideas in minutes, skipping the manual syntax work.

🚪 Lowering the Barrier to Entry: Pichai compared this trend to the early days of blogging and YouTube, noting that AI tools are opening a fresh path for creativity and job opportunities by dramatically lowering the barrier to software creation. Inside Google, the trend is evidenced by a sharp rise in first-time developers submitting changelists (small code updates).

⚠️ Warning: Not for Critical Systems: Despite the excitement, Pichai cautioned that vibe coding is not yet suitable for all applications. He stressed that the technology is not ready for complex, large-scale, or highly sensitive software where rigorous security and correctness are paramount, emphasizing the continued necessity of experienced engineers for mission-critical systems.

📈 The Beginning of a Shift: The CEO noted that the current state of AI-assisted development is the "worst it'll ever be," hinting that the next few years will bring even more transformative changes to how software is designed and built.

Since I commute every day by car, I have a lot of time to listen to tons of podcasts.

This week, I listened to a great episode of My First Million, one of my favorite podcasts, where they interviewed the Ring founder, James Siminoff. James talks about inventing the Ring doorbell camera, scaling, Amazon's acquisition, and rebuilding a town…

Great listen on Spotify or Youtube below.