Google just did something that sounds backwards at first. Its engineers spent months making a rival’s AI model run faster, then published the entire recipe for free. The model is Qwen 3.5, built by Alibaba in China, and it now runs nearly five times faster on Google’s own chips than it did in April.
To understand why that is strange, think of it like this. Google sells its own AI, Gemini, so helping a competitor’s model seems self-defeating. Yet Google also sells the chips that AI runs on, called TPUs, and that is the real business here. So the message to companies is simple, since Google wants them to know its chips will run any AI they choose, not just Google’s own.
The problem the engineers solved is worth understanding. Qwen 3.5 is enormous, holding 397 billion settings that shape how it thinks. Cleverly, it only switches on 17 billion of them for each word it produces, which is like a huge library where you only walk to one shelf. That keeps it fast, yet the whole library still has to sit ready in memory, and that memory is expensive and limited.
What the Numbers Mean
| What Google measured | The result |
|---|---|
| Speed on long questions | About 4.7 times faster than before |
| Speed on long answers | About 3.1 times faster than before |
| How close to the chip’s limit | 82% on questions, 80% on answers |
| Model size | 397 billion settings, only 17 billion used per word |
| Memory per chip | 192GB, against 288GB on NVIDIA’s rival chip |
One fix shows how small the details get. The chip was fetching information in tiny chunks of 16 at a time, wasting effort on constant trips. Google’s team enlarged those chunks to 256, so the chip made fewer, bigger trips. That single change made the work 33.8% faster.
The honest part is what Google admitted about its own hardware. Each Ironwood chip carries 192GB of memory, while NVIDIA’s rival chip carries 288GB, a gap of roughly 50%. So Google is using smarter software to make up for having less room, which works, though it proves NVIDIA still leads on raw hardware.
Alibaba is not that of a hot welcome in tech landscape, most notably after it landed in hot waters with Anthropic. Anthropic, the US lab behind Claude, told American senators in June that operators tied to Alibaba’s Qwan lab generated 28.8 million exchanges with Claude through nearly 25,000 fake accounts. The alleged goal was distillation, meaning squeezing a rival’s model for answers and using them to train your own. Alibaba denies it, and none of it is independently verified, yet Alibaba responded by banning Claude Code outright.
