Grok This!

By Dr. Robert Castellano

For most of the past year, the world believed Elon Musk’s AI strategy had stalled. Tesla (NASDAQ: TSLA) stopped discussing Dojo, the company’s custom-built supercomputer, and xAI’s Grok model appeared sporadically with little technical transparency.

Meanwhile, Nvidia GPU orders soared, Tesla expanded its data center footprint at unprecedented speed, and Musk shifted public focus toward robotics and vehicle autonomy. Analysts concluded that Dojo was quietly shelved, Grok was a side project, and Tesla’s internal AI roadmap had become fragmented.

In a recent interview with Barron’s, Musk humbled the critics. For the first time, he described the internal evolution of Grok, clarified the difference between Grok and Dojo, and explained why one project ended while the other accelerated. Grok was not paused. It was rebuilt and scaled into a far more ambitious system in service of Musk’s long-standing goal: to merge real-time perception, digital cognition, and physical-world robotics.

What Grok is; what Dojo was

Grok and Dojo were often lumped together but served entirely different purposes. Grok is xAI’s frontier-scale multimodal intelligence model, designed to understand language, images, audio, and real-time video, and capable of dynamic reasoning and tool usage. It is intended to function as an early-stage general intelligence system, connecting conventional digital corpora with real-time data streams from X and, eventually, from Tesla’s robotics platforms.

Dojo was a highly specialized supercomputer built around Tesla’s D1 ASIC. It was designed to train Tesla’s full self-driving networks using massive volumes of driving video supplied by the global Tesla fleet. Many observers assumed Dojo was Tesla’s version of a training cluster for AI models like Grok. No, it was a dedicated accelerator for a narrow class of vision problems.

Grok’s hidden lineage 

The public saw Grok as a series of modest releases; its internal development trajectory was considerably more ambitious. Grok-1 appeared as a chatbot with a unique conversational tone. Grok-1.5 added multimodal capability through vision inputs and larger context windows. Still, these early public releases gave the impression that Grok was a relatively small model.

Because Musk did not release parameter counts, analysts were unaware that Grok-2, Grok-3, and Grok-4 were built on a three-trillion-parameter architecture, placing them in the frontier category. Grok Heavy added a coordinated multi-agent reasoning system that compared outputs and resolved disagreements before generating a final answer. As Musk later confirmed, this multi-agent structure was the foundation for Grok-5.

Grok progressed through substantial architectural upgrades while the market assumed Musk was falling behind. Compute capacity remained a constraint, addressed by the rapid construction of the Colossus GPU supercluster that would ultimately support Grok-5.

Grok-5 itself is the first model Musk believes might yield some early-stage version of general intelligence, due to its six-trillion-parameter scale, multimodal architecture, and emphasis on real-time video understanding. The table illustrates Grok’s mostly sub rosa evolution.

The significance of Grok-5

Grok-5 shifts frontier AI from language-first models to perception-based intelligence capable of real-time video interpretation. Its six-trillion-parameter design makes it one of the largest models ever attempted, and its multimodal nature positions it to analyze dynamic environments in ways closer to human cognition. The integration of Grok Heavy’s multi-agent reasoning transforms the model into a system capable of internal deliberation rather than simple pattern recognition.

Dojo and Grok diverged because they solved fundamentally different problems and required different infrastructures. Dojo’s custom ASIC and tile architecture excelled at processing and training automotive vision networks, but it lacked flexibility and depended on a custom compiler stack that could not keep pace with the rapid evolution of multimodal frontier AI.

Grok, by contrast, required a framework capable of scaling across modalities, expanding parameter counts, and iterating quickly. Nvidia’s ecosystem, with its optimized software libraries and growing GPU supply, offered exactly that environment. As Nvidia’s clusters became the standard for frontier AI, Grok’s GPU-driven framework enabled it to scale quickly, incorporate multi-agent structures, and support real-time video processing.

By consolidating around Nvidia’s GPU infrastructure, Musk unified Tesla’s autonomy efforts, xAI’s general intelligence ambitions, and the development of systems that can integrate real-time data streams with physical-world robotics. Tesla shifted away from experimental architectures—Dojo—toward a scalable, GPU-driven platform capable of competing directly with the world’s leading AI laboratories. Tesla’s Grok is back, not as a curiosity or a conversational experiment, but as a frontier-class intelligence system at the center of Musk’s next decade of AI development.

Investor takeaway

With the return of Grok, Musk has resolved the architectural fork that had divided his AI strategy. Grok, built on scalable infrastructure, is able to absorb rapid increases in compute, context, and modality. For investors, this consolidation simplifies the landscape: Tesla and xAI are now aligned behind a single model family supported by Nvidia’s accelerating hardware roadmap. This will accelerate Musk’s pace of innovation.

With Grok-5 moving to six trillion parameters and integrating real-time video understanding, xAI is entering the same strategic domain as OpenAI, Google DeepMind, and Anthropic. When paired with Tesla’s unique access to real-world driving, robotics, and sensor data, Grok becomes more than a model; it becomes a platform capable of spanning digital and physical intelligence. Investors following Tesla’s autonomy roadmap, the emerging robotics market, or Musk’s broader AI ecosystem should recognize that Grok’s maturation enhances the long-term option value embedded in Tesla’s non-automotive businesses.

The strategic consolidation around Nvidia GPU infrastructure also has implications for timing and competitiveness. By shifting away from custom silicon and fully embracing Nvidia’s accelerated cadence, Musk gains a platform that renews itself on an 18–24-month cycle. This allows xAI and Tesla to iterate faster and positions Grok to advance in parallel with Nvidia’s Blackwell and Rubin generations. Grok will lead Tesla’s AI-driven businesses to become more central to its valuation over time. The return of Grok is not just a technical milestone, but a strategic realignment. It reinforces the long-term investment case for Tesla and the broader Musk AI ecosystem.