Why Google Made a Rival's AI 5x Faster
Google recently did something that sounds entirely backward: its engineers spent months optimizing a rival’s AI model to run faster, and then published the entire recipe for free. The model in question is Qwen 3.5-397B, an enormous open-source AI built by China's Alibaba. Thanks to Google's tweaks, Qwen 3.5 now runs nearly five times faster on Google’s own TPU hardware than it did in April 2026.
Why would Google help a competitor when they sell their own Gemini AI? The answer lies in hardware. Google wants enterprise customers to know its custom chips (TPUs) can efficiently run any AI model they choose, not just Google's proprietary ecosystem.
The Engineering Breakdown: How Qwen Got Faster
Qwen 3.5 is massive, containing 397 billion parameters (settings) that shape its logic. However, it uses a "Mixture of Experts" (MoE) architecture, meaning it only activates about 17 billion of those parameters for each word it generates. Think of it like a massive library where you only visit one specific shelf at a time. While this keeps the model fast, the entire library still has to sit ready in expensive, limited memory.
Here is a breakdown of what Google achieved:
| Metric | Result |
| Speed on long questions (Prefill) | ~4.7x faster |
| Speed on long answers (Decode) | ~3.1x faster |
| Hardware ceiling reached | ~82% (questions), ~80% (answers) |
| Model Size | 397 billion parameters (17B active) |
| Memory per chip | 192GB (Google) vs. 288GB (NVIDIA) |
One specific fix shows just how granular this engineering gets. Originally, the chip was fetching key-value data in tiny chunks of 16 tokens at a time. Google’s team enlarged those chunks to 256, allowing the chip to make fewer, larger data trips. This single adjustment drastically cut down processing time and memory bottlenecks.
Google Ironwood TPUs vs. NVIDIA GPUs
In its technical report, Google was highly transparent about its own hardware limitations. Each of its Ironwood (TPU v7x) chips carries 192GB of memory. In contrast, NVIDIA’s competing chips pack 288GB—a roughly 50% advantage.
Google is relying on smarter software co-design to make up for having less raw memory space. While their optimization playbook is incredibly effective, it also highlights that NVIDIA still holds the crown when it comes to brute-force hardware capacity.
Alibaba’s Hot Water With Anthropic
Despite this technical triumph, Alibaba is currently facing intense scrutiny in the broader tech landscape. In late June 2026, US-based AI lab Anthropic sent a letter to American senators accusing operators linked to Alibaba’s Qwen lab of "illicitly" accessing its Claude AI model.
Anthropic claims that between April and June, these operators used nearly 25,000 fraudulent proxy accounts to generate 28.8 million exchanges with Claude. The alleged goal was adversarial distillation—a process where a company squeezes a cutting-edge rival's model for answers and uses that high-quality data to train its own AI at a fraction of the cost.
(Note for your draft: While your original text mentioned Alibaba responded by banning Claude Code, the reality is that Anthropic and the US Commerce Department swiftly cut off foreign access to the Claude models to prevent further data harvesting.)
This clash highlights the escalating tension between AI labs fighting to protect their multi-million dollar intellectual property from being silently siphoned by global competitors.