Google's Gemini 4 Argon sees 3x surge in early usage despite compute limits

Surging demand for Gemini 4

Google's newest AI model, Gemini 4 Argon, has seen a significant spike in adoption since its release. According to Google CEO Sundar Pichai, the model processed almost 25 trillion tokens in its first 72 hours of availability. This figure represents a massive jump in scale for the company's AI infrastructure.

When compared to the previous iteration, Gemini 3.0, the new model's usage is about 3x higher over the same three-day period. This growth suggests that the market is eager for the advanced capabilities of the Argon model, particularly in areas like coding and knowledge work. The data highlights a rapid transition toward more complex AI interactions.

Despite this high demand, Pichai noted that Google remains capacity constrained as it works through the launch. The company is currently balancing the need to provide high-quality responses with the physical limits of its computing power. This tension between user demand and hardware availability is a central challenge for the company's current growth strategy.

Model efficiency and access

Because Argon is a large and expensive model to serve, Google is employing specific strategies to manage its reach. Currently, the model is available to Google AI Pro and Ultra subscribers, as well as through the company's API. This gated approach allows Google to prioritize high-value users while managing the heavy computational load required for frontier models.

To reach a broader audience, Google is using a process called distillation to create smaller, more efficient versions of the model. Distillation is a method where a large model helps train a smaller one to be nearly as capable but much faster and cheaper to run. Pichai stated that the company will continuously expand access as these more efficient versions become available.

The company also recently restructured how free users interact with its AI tools. Free Gemini users currently only have access to the Flash-Lite model, while the more powerful Argon-derived features are reserved for paid tiers. This tiered system ensures that the most resource-intensive capabilities are balanced against the company's infrastructure limits.

Infrastructure and capital spend

The rapid growth of Gemini 4 Argon comes at a significant financial cost for Alphabet. The company recently reported negative free cash flow for the first time since its IPO, driven by massive capital expenditures. Google is currently guiding for capital spending between $195 billion and $205 billion for 2026, with further significant increases expected in 2027.

Pichai defended this aggressive spending by pointing to durable long-term contracts and strong demand signals across consumer and enterprise products. He noted that the company is making long-term compute decisions that are already reaching into 2028 and 2029. This long-term planning is necessary because building the custom chips required for these models takes several years of development.

The company is also navigating a volatile market where many players are competing for AI dominance. While there are concerns about a potential bubble, Pichai focused on the signals Google can control, such as backlogs and contracts. He compared the current situation to Google's late entry into the cloud market, which required a period of aggressive investment to achieve scale.

The shift toward agents

A major theme in the launch of Gemini 4 Argon is the move toward agentic tasks, where AI can perform actions rather than just generating text. Pichai described Argon as a frontier model with strengths in coding and knowledge work. He highlighted how the model can persistently work through complex codebases, such as Chromium, to build new features without human intervention at every step.

This shift toward agents has significant implications for the web and the advertising industry. As agents begin to perform tasks like navigating browsers or completing workflows, the way people reach websites and interact with ads will change. Google is working on interoperable standards, such as the Universal Commerce Protocol, to help different parts of the economy work together with these agents.

To ensure these agents behave safely, Google has increased its investment in infrastructure around the models. This includes sandboxing, real-time monitoring, and telemetry to provide early signals if a model behaves unexpectedly. Pichai emphasized that this real-time monitoring is critical for maintaining safety as models are deployed to enterprise customers and government partners.