SpaceX has built a custom AI training stack in C designed to run on 220,000 NVIDIA GB300 GPUs, claiming potential speed improvements of over 10 times compared to Google's JAX framework. The announcement, made by Elon Musk on May 28, 2026, represents a fundamental shift in how frontier AI labs approach the infrastructure behind training the world's most powerful language models.

Why Would SpaceX Build Its Own AI Training Software?

To understand the significance, you need to know what JAX is and why nearly every major AI lab uses it. JAX is a machine learning framework developed by Google that handles the complex math behind training AI models. It's flexible, fast, and widely trusted. OpenAI, DeepMind, Anthropic, and even xAI in its early days all relied on JAX or similar frameworks.

But here's the catch: frameworks like JAX are designed to work on any hardware configuration. That flexibility comes with a cost. Every layer of software between your training job and the actual GPU chips consumes computing power. When you're training on 10 GPUs, that overhead is negligible. When you're training on 220,000 GPUs, every percentage point of wasted efficiency translates into millions of dollars and weeks of lost time.

SpaceX took a radically different approach. Instead of using a general-purpose framework, they wrote their training stack directly in C, one of the oldest and lowest-level programming languages still in common use. C gives programmers direct control over memory, processor instructions, and how data flows to the hardware. There's no middleman, no abstraction layer between the code and the silicon.

What Makes This Training Stack Different From Everything Else?

The technical choices SpaceX made reveal a philosophy borrowed from aerospace engineering: vertical integration. If you want something done right, build it yourself. The training stack was written with precise knowledge of exactly how many GPUs exist in the cluster, how they're physically connected, and how data should flow between them.

The infrastructure supporting this stack includes several critical components:

• Direct Hardware Mapping: The software was written specifically for 220,000 NVIDIA GB300 GPUs, not designed to run on any hardware configuration. This exact-mapping approach eliminates the flexibility overhead that general frameworks carry.
• High-Speed Networking: The system uses 800 Gigabit Network Interface Cards (800G NICs) to move data between nodes. In distributed training, the speed at which GPUs communicate with each other often becomes the real bottleneck after individual GPU performance.
• Pipeline Parallelism: The training stack breaks up model layers into stages distributed across different GPUs, allowing multiple batches of data to flow through different stages simultaneously, like an assembly line.

For models with hundreds of billions of parameters, pipeline parallelism isn't a nice-to-have optimization. It's what makes training possible in the first place.

What Would a 10x Speed Improvement Actually Mean?

The claimed speedup sounds abstract until you translate it into real-world impact. Meta's Llama 3.1 405B model required approximately 3.8 × 10²⁵ floating point operations to train. Meta used 16,000 H100 GPUs to complete that training run in 54 days. A 10-fold improvement in training speed would compress similar-scale projects from months to weeks and slash training costs from hundreds of millions of dollars to tens of millions.

This isn't just about speed. Faster training cycles mean researchers can run more experiments, test different model architectures more frequently, and iterate on improvements monthly instead of quarterly. The compound effect on research velocity could be transformative for any organization that can harness it.

To be clear, independent verification of the 10x claim doesn't yet exist. Experts have noted that achieving consistent improvements at this scale is genuinely challenging. Amdahl's Law, which describes the limits of parallel speedup due to sequential bottlenecks, places hard constraints on what's achievable. Communication overhead between 220,000 GPUs doesn't vanish just because the code is well-written. But even achieving a fraction of the claimed gains, such as 3x, 4x, or 5x improvements, would be transformative.

Where Will This Training Stack Actually Be Used?

Musk confirmed that the new training stack will power Grok v5, xAI's next major model release. This isn't a roadmap item or a future possibility. It's a concrete commitment.

The hardware context for this announcement is crucial. Following SpaceX's acquisition of xAI in February 2026, the company now controls what may be the world's most powerful AI training complex: Colossus 2 in Memphis, Tennessee. Colossus 2 became operational at gigawatt-scale power in January 2026, making it the world's first coherent AI training cluster to reach that threshold. The facility houses approximately 550,000 to 555,000 NVIDIA Blackwell-series GPUs, primarily GB200 and GB300 chips, and operates at roughly 1 gigawatt of power. That's equivalent to the peak electricity demand of a city the size of San Francisco.

As of April 8, 2026, Musk confirmed that Colossus 2 is simultaneously running seven distinct model training jobs, including image generation models and language models scaling up to 10 trillion parameters. For context, GPT-4 is widely estimated to have approximately 1.8 trillion parameters. A 10 trillion parameter variant would represent a generational leap in scale.

How Does This Fit Into SpaceX's Broader Strategy?

SpaceX's S-1 registration statement, filed ahead of its planned initial public offering, explicitly reframes the company as a vertically integrated AI infrastructure platform. The filing argues that future AI competition will be decided by control of underlying physical systems: chips, energy, networking, manufacturing, and deployment capacity.

A proprietary training stack is a direct extension of that thesis. If SpaceX can train models meaningfully faster than competitors using the same NVIDIA hardware, it gains an asymmetric advantage. More model iterations per dollar, faster research cycles, and a competitive moat that doesn't depend on getting scarce next-generation chips first.

The broader Musk ecosystem stands to benefit too. Tesla's Dojo supercomputer has pursued a similar philosophy of custom silicon, custom interconnects, and custom software for autonomous driving training. A proven bare-metal training stack developed at SpaceX could inform or accelerate that work, though the two remain separate organizations with separate compute infrastructure.

Version 1.0 of the training stack isn't finished yet, and benchmark numbers haven't been published. But the direction is clear: SpaceX is betting that owning the full stack, from the C code to the GPU to the data center power supply, is the only way to compete at the frontier of AI. Whether the speed gains match the ambition is the question that V1.0 will have to answer.