SAN FRANCISCO, April 3, 2026, 04:05 PDT
Nvidia said on Thursday it had optimized Google’s newly released Gemma 4 artificial intelligence models to run across RTX PCs and workstations, DGX Spark systems and Jetson edge modules, widening its bid to keep developers on its chips as AI workloads move beyond giant cloud clusters. Nvidia disclosed the rollout in a blog post, while Google introduced Gemma 4 as its newest open model family. NVIDIA Blog
The timing matters. Nvidia is trying to defend growth as the AI market shifts from training large models to “inference” — the step where trained systems generate answers — and to agentic systems that can plan tasks and use tools on a user’s behalf. At Nvidia’s GTC developer conference last month, Chief Executive Jensen Huang said “The inference inflection has arrived,” and eMarketer analyst Jacob Bourne said Nvidia’s $1 trillion revenue-opportunity forecast “underscores the durable demand” for its AI infrastructure despite investor worries over the payoff from heavy spending. Reuters
In a launch post, Google DeepMind executives Clement Farabet and Olivier Lacombe wrote that Gemma 4 is Google’s most capable open model family to date. Google said developers have downloaded Gemma more than 400 million times and built over 100,000 variants since the first release, and that the new family is being offered under Apache 2.0, a permissive license that allows commercial use and modification. blog.google
Google said the family includes four model sizes. The larger 26B and 31B versions can fit on a single 80GB Nvidia H100 data-center GPU and, in compressed form, run on consumer GPUs, while the smaller E2B and E4B models are built to run fully offline on phones, Raspberry Pi boards and Nvidia Jetson Orin Nano devices; the models were trained on more than 140 languages and support audio, image and video input. blog.google
Nvidia said the models had been tuned across hardware from Blackwell data-center systems to Jetson edge devices. In a technical post, the company said local deployment would suit customers that want on-premises systems, tighter control of data and faster response times for uses in industries such as healthcare and finance. NVIDIA Developer
For all the talk of PCs and edge devices, Nvidia is still a data-center company first. Its latest quarterly report showed data-center sales of $62.3 billion out of $68.1 billion in total revenue, while gaming and AI PC revenue was $3.7 billion. Huang said in that report that computing demand was rising as enterprise adoption of AI agents accelerated. NVIDIA Newsroom
That also brings risk. Reuters reported last month that inference is drawing more competition from central processors and custom chips built for specific workloads by companies such as Google and Meta. “Nvidia is definitely going to see more competition compared to a year ago,” KinNgai Chan, a managing director at Summit Insights Group, told Reuters. Reuters
Google also made clear Gemma 4 is not exclusive to Nvidia. The company said the models are optimized not only for Nvidia systems but also for AMD GPUs and Google’s own TPUs, or tensor processing units, making Nvidia one of several hardware options for Gemma 4 rather than the only one. blog.google
