xAI Colossus 2 Trains 7 Massive AI Models: 10T Parameter Breakthrough by Elon Musk

On April 8, 2026, Elon Musk announced a major milestone for SpaceXAI Colossus 2, revealing that the world’s first gigawatt-scale AI training supercluster is now simultaneously training seven advanced AI models. The update underscores xAI’s aggressive scaling strategy following its integration with SpaceX under the combined SpaceXAI entity.

The seven models currently in training on Colossus 2 are:

• Imagine V2 — The next-generation multimodal image and video generation model powering Grok’s creative capabilities
• 2 variants of 1T-parameter models (1 trillion parameters each)
• 2 variants of 1.5T-parameter models (1.5 trillion parameters each)
• One 6T-parameter model (6 trillion parameters)
• One 10T-parameter model (10 trillion parameters)

Musk added the concise remark: “Some catching up to do.” This signals xAI’s determination to rapidly close any gaps with leading frontier models from competitors and accelerate progress toward more capable general intelligence.

Understanding Colossus 2

Colossus 2 represents the second phase of xAI’s ambitious supercomputer project in Memphis, Tennessee. After Colossus 1 reached 200,000 GPUs in record time (built in just 122 days for the initial phase), Colossus 2 has crossed the 1 gigawatt (GW) power threshold — equivalent to the peak electricity demand of an entire major city like San Francisco.

Key features of Colossus 2 include:

• Hundreds of thousands of next-generation NVIDIA GPUs (with the broader Colossus ecosystem targeting over 1 million GPUs)
• Exaflop-scale computational performance
• Advanced power infrastructure combining gas turbines, Tesla Megapacks for energy storage, and high-efficiency cooling
• Designed for parallel multi-model training, enabling simultaneous experiments across different model sizes and architectures rather than sequential runs

This infrastructure allows xAI to iterate faster by testing multiple training recipes, data mixtures, and optimizations at the same time.

Breakdown of the 7 Models in Training

1. Imagine V2 The successor to Grok’s current image and video generation tools. It is expected to deliver substantial improvements in photorealism, complex prompt adherence, artistic versatility, and potentially enhanced native video generation or deeper multimodal integration.
2. 1T and 1.5T Parameter Variants (Four Models Total) These mid-to-large frontier-scale models serve as testbeds for different architectural tweaks, training methodologies, and capability enhancements. Running multiple variants in parallel accelerates comparison across areas such as reasoning, coding, long-context handling, and agentic behaviors.
3. 6T and 10T Parameter Models These represent the next frontier in model scale. A 10-trillion-parameter model would rank among the largest ever trained, holding potential for breakthroughs in complex multi-step reasoning, sophisticated world modeling, and advanced multimodal intelligence. Training at this scale demands not only immense compute but also cutting-edge optimizations for training stability and efficiency.

The ability to train all seven models concurrently highlights Colossus 2’s massive parallel processing capacity and xAI’s philosophy of rapid, bold experimentation.

Context: The SpaceXAI Integration

The announcement arrives amid the recent SpaceX acquisition of xAI via a share-swap deal, forming SpaceXAI. This merger is unlocking synergies in high-power energy systems, rapid infrastructure deployment, Starlink connectivity, and real-world data streams from Tesla, X, and SpaceX operations. Colossus 2 benefits directly from this combined engineering expertise, particularly in managing gigawatt-level power demands and scaling hardware at unprecedented speed.

Why This Update Matters in the AI Race

• Compute Leadership: By operating at gigawatt scale and training models up to 10T parameters in parallel, xAI continues to push the limits of scaling laws while emphasizing execution speed.
• Parallel Experimentation: Instead of focusing on a single massive model, the cluster supports diverse research threads, potentially leading to faster discovery of optimal designs.
• Multimodal Advancement: Imagine V2’s development alongside massive language models positions future Grok versions as more complete systems capable of seamless text, image, and video understanding and generation.
• Competitive Edge: The “catching up” comment reflects a confident, competitive stance. While other labs train individual large models, xAI is running an entire portfolio of frontier-scale experiments simultaneously.

What Comes Next?

xAI has not disclosed exact timelines for completion or release, but successful training of these models could result in:

• Significant upgrades to the Grok family (potentially Grok 5 or Grok 6 series) later in 2026
• Major leaps in Grok’s image and video generation features via Imagine V2
• New performance benchmarks in reasoning, creativity, and real-world task handling

The parallel training strategy also allows xAI to evaluate and combine the strongest elements from different variants for final flagship models.

This development reinforces xAI’s (and now SpaceXAI’s) commitment to building the most powerful AI infrastructure on the planet at unmatched speed. The gigawatt era of AI training is not just arriving — it is already here and expanding rapidly.

Stay tuned to VFutureMedia.com for ongoing coverage of AI supercomputing, frontier model releases, and the evolving SpaceXAI ecosystem. The race toward more capable and useful AI continues to accelerate.

This is a developing story based on Elon Musk’s announcement on X. Follow official channels from @elonmusk and @xai for the latest updates.