Staff Reporter | Cheng Lu
News Editor | Wen Shuqi
On August 12th, the Google I/O Connect China developer conference kicked off in Shanghai. Right after Typhoon “White Dolphin” swept through, thousands of developers queued up in the rain, flying in from all over the country and even from overseas. Their expectations were largely the same: to catch a glimpse of Google’s latest AI breakthroughs.
No brand-new flagship model? That was hardly a shocker.
Back at the U.S. I/O conference in May, Google rolled out updates like Gemini 3.5 Flash and Gemini Omni. But the much-anticipated Gemini 3.5 Pro? It’s been delayed for three months now. At the Shanghai event, management revealed that Gemini 4 is already in its pre-training phase, so they only dropped a few incremental updates like Gemini 3.6 Flash.
Beyond the models themselves, Google dedicated most of its time to TPUs, developer tools, Agent workflows, Cloud, and hyper-specific use cases across film, gaming, retail, going-global strategies, and smart hardware.
“There was no real ‘wow’ moment,” said Zhang Xiang, a developer from Canada attending for the second year in a row. In his view, the Transformer architecture that Google pioneered is a classic case of “the flower blooming outside the wall.” OpenAI, Anthropic, and even domestic players have been catching up and overtaking at a blistering pace—that’s just the hard truth now.
This isn’t just a vibe from the conference floor. In early August, Google’s parent company Alphabet went through its own personnel “typhoon”: Chief Scientist Jeff Dean announced he’s leaving the company after 27 years to start a new venture with three core researchers, while DeepMind founder Demis Hassabis stepped down as CEO to become chairman. On the day of that announcement, Alphabet’s market cap evaporated by roughly $190 billion. Product pressure and talent drain—both hitting Google at the exact same moment.
That might be the most significant backdrop to this China edition of Google I/O. That said, some developers told us that the model race is inherently back-and-forth. Seats at the top table keep shuffling, and the sheer number of new models released this year proves it. “Being behind temporarily doesn’t mean it’s permanent.” In his view, Google still has a lot of cards up its sleeve, and catching up is very much on the table.
What we noticed this year is that the on-site story focused much more on the compute power beneath the models and the applications layered on top of them.
Compute is the heavy hitter. This year, Google split its TPU line into two distinct series: TPU 8t, built specifically for training, and TPU 8i, optimized for agent reasoning—all to squeeze more efficiency out of the entire compute stack.
Application companies need models, model companies need compute, and compute has become the scarcest resource in the AI arms race. How it gets allocated sits at the core of so many issues. In July, Google—with its massive compute resources—posted a stellar earnings report: Q2 2026 cloud revenue hit $24.768 billion, up a whopping 82% year-over-year. But the market still wasn’t impressed. CEO Sundar Pichai has repeatedly pushed back against the “Google is falling behind” narrative, recently touting that Gemini has crossed the 1 billion user mark.
Walking the floor at Google I/O Shanghai, you could see the company stacking more and more on top of its models, leveraging its entire ecosystem.
YouTube, Google Maps, Gmail, Android—Google holds a treasure trove of legacy ecosystem assets, along with the APIs and data capabilities that come with them. Even though there’s nothing new to say on the model front, Google is hammering home its ecosystem strength when it comes to application deployment.
We spotted one booth surrounded by a crowd of brands: a cross-border e-commerce Agentic AI solution built around UCP (Universal Commerce Platform).
In the old days, online shoppers had to hop between different platforms and merchant pages, then handle search, selection, and payment separately. Now UCP works like a unified protocol layer: users just give AI a command, and the Agent handles understanding the request, hunting down products, and even completing the purchase. This is the kind of shift e-commerce competition might see in the AI era—products need to be “seen” by AI. Nike has already announced it’s plugging into Google’s UCP.
Booth staff told us that merchants and brands can now tap into YouTube’s social trend data for viral product selection, use Google Maps heatmap data to pick store locations in new cities, and link ad campaigns to Google Ads—all stitched together by AI into one seamless workflow.
Of course, competitors are running the same playbook. Meta has Instagram and Facebook, and Microsoft has Office and Windows.
For the enterprise customers actually writing the checks, the China event zeroed in on the going-global play.
Focusing on production-grade deployments of Google’s AI tech stack, Chen Junting, President of Google’s Greater China and Korea region, shared some Chinese enterprise success stories. vivo is using TPU as its overseas AI compute foundation, cutting operational costs by 48%. Baidu’s no-code platform Miaoda leveraged Gemini to speed up app delivery tenfold. XPENG became the first automaker in Asia-Pacific to ship new cars with Google Maps Auto SDK natively built in. The same day, Google also announced a “Startup Going Global” initiative, expanding its cross-border e-commerce acceleration centers from Xiamen and Shenzhen to Guangzhou and Hangzhou.
Hardware was another highlight. Over at the expo area, the XREAL Aura—the first wired glasses running Android XR—had a line that never seemed to shrink. This device, co-developed with Qualcomm and XREAL, weighs under 95 grams and comes with deep Gemini integration. Chinese attendees rarely get the chance to try cutting-edge hardware like this, and the spatial computing and immersive feel left a strong impression on many who tested it.
“Google’s Notebook LM, which is genuinely one of their best products this year, didn’t even get stage time at I/O—yet developer feedback on it has been fantastic,” Zhang Xiang said. “I’m hoping that next year, or the year after, Google I/O will bring more of those genuinely jaw-dropping moments.”
As for whether the next-gen Gemini can put Google back in the top tier of the model race—that’s the question this Shanghai I/O left hanging, a cliffhanger for the rest of us to chew on.
