Google 75 Percent Code Written by AI TPU 8 Challenges Nvidia

Google just played its biggest hand. At Cloud Next 2026, the ai nsfw generator company revealed that 75 percent of all new code is now written by AI. The eighth generation TPU is here with dual-chip architecture. Performance jumped three times. Nvidia should be watching closely.

clothes remover

CEO Sundar Pichai took the stage with a simple message. Gemini is now everywhere. The keynote was not about future promises. It was about what Google is shipping today. And the numbers are staggering.

The headline act was the eighth generation TPU. For the first time, Google split the architecture into two chips. TPU 8t handles training. TPU 8i handles inference. Each is optimized for its specific job. Together, they cut agent serving costs by 3 times.

After ten years of development, the eighth generation TPU finally challenges Nvidia’s GPU dominance head-on. Beyond chips, Google announced a wave of software products. Workspace Intelligence brings Gemini to every business app. The AI app platform AI-APP lets anyone build intelligent applications. Workspace and Gemini are now deeply integrated.

Google also introduced a new distributed reasoning architecture. It enables global collaboration with no limits. Gemini now processes 160 billion tokens per minute, up from 100 billion. Over 350 thousand enterprise customers are using it. Ten billion lines of code have been migrated.

Most shockingly, 75 percent of new code inside Google is now AI-generated. Engineers review and approve it. Internal agent collaboration systems handle code migration 6 times faster than human teams alone.

Google is now competing with OpenAI and Anthropic on a full-stack level. Chips, models, cloud infrastructure, and enterprise software. Everything is connected.

The eighth generation TPU is Google’s declaration of war on Nvidia. For the first time, Google split TPU into two distinct chips. TPU 8t is built for training. TPU 8i is built for inference. This is the result of ten years of grinding. The gap is closing.

One analyst summarized it perfectly. The eighth generation TPU’s training plus inference dual-track strategy directly targets Nvidia’s core business.

The battle between Nvidia GPUs and Google TPUs is no longer theoretical. It is happening now.

TPU 8t is the training beast. It pushes the boundaries of scale-out architecture. When training large language models, scale is everything. Each pod’s computing power tripled with Ironwood. A single TPU 8t Superpod packs 9,600 chips delivering 121 ExaFlops of peak performance.

Through the new Virgo interconnect, combined with JAX and Pathways software stack, Google achieves near-linear scaling of TPU clusters. Google also optimized tail latency by 97 percent, ensuring that even the slowest parts of the distributed system stay fast.

TPU 8i is the inference specialist. This chip is designed specifically for serving agents at massive scale. Each TPU 8i pod contains 1,152 chips, achieving huge throughput with low latency.

At the chip level, through key optimizations, Google eliminated the queuing effects that slow down agent teams. The chip architecture uses a distributed Boardfly structure. Four chips form a base unit, which scales to 36 chips, then to a full TPU 8i pod.

Latency and throughput have always been conflicting goals in chip design. The memory wall is real. TPU 8i solves this with 288GB of high-bandwidth memory and 3 times more on-chip SRAM than previous generations. The working memory for models stays entirely on the chip.

For agent collaboration scenarios requiring high-level reasoning, TPU 8i delivers 80 percent better cost performance per dollar of compute.

Most importantly, the eighth generation TPU fully integrates Axion Arm CPUs. Through chip plus network plus liquid cooling plus power supply full-stack optimization, every watt of power becomes compute power. As the saying goes, Google’s TPU gets sharper with every generation. After ten years of grinding, it is indeed getting sharper.

But chips were just the opening act. The number that stunned the entire industry was the AI code generation statistic.

Google revealed that 75 percent of new code is AI-generated. Engineers review and approve it. This is up from just 50 percent last year. At the same time, Google is not fully AI-coding like Anthropic claims to be. Google is taking a more balanced approach.

Internally, Google has quietly entered the era of AI writing code. Human engineers no longer write code from scratch. Instead, they direct agent teams. One person manages multiple AI agents, each handling different parts of the codebase.

A particularly complex code migration project, using agent plus engineer collaboration, completed 6 times faster than human-only teams. The initial version of the Gemini Mac app was built entirely using Google’s internal agentic platform called Antigravity. The idea and prototype came from Swift code, but AI did the heavy lifting.

Interestingly, Google internally uses a new rule. If you can use Claude, use Claude. If you cannot use Claude, use Gemini. Engineers joke that Google employees are actually using Claude for coding. It turns out that even DeepMind researchers prefer Claude for their own work.

Another interesting detail leaked from internal Slack channels. In benchmark tests, Gemini actually outperformed Claude. Google has made AI coding tools a daily KPI for all employees. Using AI tools is now part of performance reviews.

In the office software battle, Google officially launched Workspace Intelligence. Gmail, Docs, Sheets, Slides, Drive, and Chat all get AI superpowers. Every app becomes intelligent. Every workflow gets automated.

Chat becomes the command center. Google Chat is now the unified inbox for everything. Instead of switching between ten browser tabs, checking emails, documents, and files, you just tell Chat your goal. Gemini works in the background and pushes results directly into the chat window.

Every morning, Chat’s AI assistant delivers a daily briefing. It summarizes key conversations you need to follow, identifies upcoming meetings, and flags action items. You can schedule meetings, create documents, and generate presentations with a single sentence.

Workspace Intelligence is Google’s nuclear weapon in the office software war. Sheets becomes fully automatic. Gemini in Sheets has reached enterprise-level intelligence. You can describe what you want in natural language, and Sheets fetches data from emails, files, and web pages. Humans become supervisors rather than spreadsheet builders.

Docs stops being a writing tool and becomes a document engine. Gemini in Docs can write, edit, and rewrite documents. It can summarize long reports, suggest improvements, and apply formatting changes directly. The killer feature is automatic information graph generation. One click creates a knowledge graph from your document, with images, tables, and charts added automatically.

Slides turns one sentence into a full presentation. You describe your goal. Gemini generates the outline, finds relevant data, creates charts, and builds the slides. Workspace Intelligence claims it can reduce presentation creation time by 90 percent.

Gmail gets an AI inbox that automatically sorts and summarizes emails. Drive Projects turns scattered files into organized knowledge bases. Everything is connected. Everything is intelligent.

Beyond the office, Google announced enterprise-grade AI platforms. Gemini Enterprise leads the way in enterprise AI deployment. API calls surged from 100 billion to 160 billion tokens per minute. The new version added 35 enterprise features.

Google also launched the Gemini Enterprise Agent Platform, an operating system for managing thousands of enterprise agents. Through Vertex AI, it provides a unified tool chain for building, deploying, and optimizing agents. The platform supports over 200 AI models including Claude and Gemini 3.1 Pro.

A key feature called Memory Bank gives agents persistent long-term memory. The more they process, the smarter they become. Nothing is forgotten.

In the AI platform war, the battlefield is crowded. OpenAI has ChatGPT with hundreds of millions of users. Anthropic wins with Claude’s coding abilities. Microsoft owns the enterprise channel.

Google chose a different path. Chips, models, office software, and agent platforms. Full stack, fully connected. The battle has just begun. But Google just showed the world that it is playing to win.