SourceThe Guardian Technology•May 26, 2026Stalecollected in 2h

Musk and Altman AI Rivalry Intensifies Amid IPO Race

Read original on The Guardian Technology

#ipo#market-trends#ai-strategyopenai-/-spacexopenaispacexmetanvidiaelon musk

Understand how the IPO race between SpaceX and OpenAI is shifting the AI landscape and capital flow.

30-Second TL;DR

What Changed

SpaceX is planning a $1.75 trillion stock market debut.

Why It Matters

The potential IPOs of these AI-adjacent giants will likely reshape capital allocation in the AI sector, forcing smaller startups to compete for dwindling venture capital.

What To Do Next

Monitor the financial disclosures of these companies to identify shifts in AI infrastructure spending and R&D priorities.

Who should care:Founders & Product Leaders

Key Points

•SpaceX is planning a $1.75 trillion stock market debut.
•Meta is mandating internal job transfers to prioritize AI development.
•Nvidia reported revenue growth exceeding Wall Street expectations due to AI demand.
•OpenAI has achieved a breakthrough in solving an 80-year-old mathematics problem.

Key numbers$40 billion$80 billion$1.75$2

Deep Insight

Background and context from public sources — not the original article. 36 sources cited.

Enhanced Key Takeaways

•SpaceX's anticipated IPO, expected around June 12, 2026, aims to raise between $40 billion and $80 billion, potentially valuing the company at $1.75 trillion to $2 trillion, which would make it the largest IPO in history.
•OpenAI is also laying the groundwork for a public listing in the fourth quarter of 2026, with a target valuation exceeding $1 trillion, despite internal projections indicating losses of $14 billion in 2026.
•Elon Musk's AI venture, xAI, founded in March 2023, merged with SpaceX in February 2026, contributing an approximate $80 billion to the combined entity's $1.25 trillion valuation. xAI also acquired X (formerly Twitter) in March 2025 for $33 billion.
•Nvidia reported record Q1 fiscal 2026 revenue of $81.62 billion, tripling its profit year-over-year, primarily driven by the surging demand for AI training and inference hardware. The company also announced an $80 billion stock repurchase program and a 25-fold increase in its quarterly dividend.
•OpenAI achieved a significant milestone in May 2026, with a general-purpose reasoning model successfully disproving a long-standing conjecture in discrete geometry, known as the planar unit distance problem, originally posed by Paul Erdős in 1946.

Competitor Analysis

Architecture

OpenAI (GPT-4/5): Mixture-of-Experts (MoE) Transformer (GPT-4: ~1.8T parameters, 16 experts, 2 active per pass; GPT-5: capability-focused)
xAI (Grok 4): Decoder-only Transformer, MoE design (Grok-1: 314B parameters, 2 of 8 experts active per token)
Meta (Llama 3): Optimized Transformer architecture (Largest model: 405B parameters, 126 layers)

Key Differentiators

OpenAI (GPT-4/5): Multimodal (text & image input), advanced reasoning, dynamic mode switching (GPT-5)
xAI (Grok 4): Real-time access to X (Twitter) firehose, looser content moderation, "truth-seeking" stance
Meta (Llama 3): Open-weight models, multilingual support, Grouped Query Attention (GQA) for efficiency

Training Infrastructure

OpenAI (GPT-4/5): Not fully disclosed, but implies massive compute for 1.8T+ parameters
xAI (Grok 4): Colossus supercomputer (100,000+ NVIDIA H100 GPUs)
Meta (Llama 3): Meta's Grand Teton AI servers with H100 GPUs, 4D parallelism across 16,000 GPUs

Context Window

OpenAI (GPT-4/5): Up to 128,000 tokens (GPT-4 Turbo)
xAI (Grok 4): Up to 2,000,000 tokens (Grok 4.1 Fast)
Meta (Llama 3): Up to 128,000 tokens

Performance (Benchmarks)

OpenAI (GPT-4/5): GPT-5: 94.6% on AIME 2025 (math), 74.9% on SWE-bench Verified (coding)
xAI (Grok 4): Grok 4 (2025): Closed most benchmark gaps with GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Pro
Meta (Llama 3): Rivals top models like GPT-4 in benchmark tests

Feature/Model	OpenAI (GPT-4/5)	xAI (Grok 4)	Meta (Llama 3)
Architecture	Mixture-of-Experts (MoE) Transformer (GPT-4: ~1.8T parameters, 16 experts, 2 active per pass; GPT-5: capability-focused)	Decoder-only Transformer, MoE design (Grok-1: 314B parameters, 2 of 8 experts active per token)	Optimized Transformer architecture (Largest model: 405B parameters, 126 layers)
Key Differentiators	Multimodal (text & image input), advanced reasoning, dynamic mode switching (GPT-5)	Real-time access to X (Twitter) firehose, looser content moderation, "truth-seeking" stance	Open-weight models, multilingual support, Grouped Query Attention (GQA) for efficiency
Training Infrastructure	Not fully disclosed, but implies massive compute for 1.8T+ parameters	Colossus supercomputer (100,000+ NVIDIA H100 GPUs)	Meta's Grand Teton AI servers with H100 GPUs, 4D parallelism across 16,000 GPUs
Context Window	Up to 128,000 tokens (GPT-4 Turbo)	Up to 2,000,000 tokens (Grok 4.1 Fast)	Up to 128,000 tokens
Performance (Benchmarks)	GPT-5: 94.6% on AIME 2025 (math), 74.9% on SWE-bench Verified (coding)	Grok 4 (2025): Closed most benchmark gaps with GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Pro	Rivals top models like GPT-4 in benchmark tests

Technical Deep Dive

OpenAI GPT-4/5: GPT-4 utilizes a Mixture-of-Experts (MoE) architecture, reportedly featuring approximately 1.8 trillion parameters distributed across 120 layers. It employs 16 expert networks, with two experts routed per forward pass to manage computational costs. The model supports multimodal inputs, accepting both text and images, and offers an extended context window of up to 128,000 tokens. GPT-5, released in August 2025, emphasizes enhanced capabilities over raw parameter count, demonstrating strong performance in mathematical reasoning (94.6% on AIME 2025) and coding (74.9% on SWE-bench Verified). It incorporates a real-time router for dynamic switching between a fast mode for simple queries and a 'thinking' mode for complex reasoning.
Meta Llama 3: Llama 3 is built upon an optimized dense Transformer architecture. The largest model in the series boasts 405 billion parameters across 126 layers, with 16,384 dimensions and 128 attention heads. Key architectural features include pre-normalization (RMSNorm), SwiGLU activations, and Rotary Positional Embeddings (RoPE). It also incorporates Grouped Query Attention (GQA) for models with 70 billion parameters or more, enhancing efficiency during inference. Llama 3 uses a tokenizer with an expanded vocabulary of 128,000 tokens and was trained on sequences of 8,192 tokens.
xAI Grok: Grok models are based on a decoder-only Transformer architecture, with Grok-1 being a 314-billion-parameter Mixture-of-Experts (MoE) model where only two of eight experts are active per token. Subsequent versions, such as Grok 3 and Grok 4, continue to leverage MoE designs and are trained on xAI's Colossus supercomputer, which utilizes over 100,000 NVIDIA H100 GPUs. A distinctive feature is Grok's real-time access to the X (formerly Twitter) firehose, enabling it to incorporate current information. Grok 4, refined through 2026, includes features like a 'Think' mode for chain-of-thought reasoning and a 'DeepSearch' agent mode for multi-hop web research, with context windows extending up to 2,000,000 tokens in some variants.

Future ImplicationsAI analysis grounded in cited sources

The intense rivalry between AI leaders will accelerate the development of increasingly powerful and specialized AI models.

Competition for market dominance and IPO success drives rapid innovation in model architecture, training scale, and unique feature sets like real-time data access or advanced reasoning, as seen with OpenAI's mathematical breakthrough and xAI's integration with X data.

The massive capital requirements for advanced AI development will lead to further consolidation and strategic mergers in the tech industry.

The multi-billion dollar investments in AI infrastructure and training, exemplified by OpenAI's projected losses and xAI's merger with SpaceX, necessitate significant funding and strategic alliances to sustain growth and competitiveness.

Timeline

2015-12

OpenAI founded as a non-profit by Elon Musk, Sam Altman, and others.

2018-02

Elon Musk resigns from OpenAI's board of directors due to potential conflicts of interest with Tesla's AI efforts.

2023-03

Elon Musk founds xAI with the mission to understand the true nature of the universe.

2025-03

xAI acquires X (formerly Twitter) in an all-stock deal valued at $33 billion.

2026-02

xAI merges with SpaceX, valuing the combined entity at $1.25 trillion.

2026-05

OpenAI's general-purpose reasoning model disproves an 80-year-old mathematics conjecture by Paul Erdős.

2015-12
OpenAI founded as a non-profit by Elon Musk, Sam Altman, and others.
2018-02
Elon Musk resigns from OpenAI's board of directors due to potential conflicts of interest with Tesla's AI efforts.
2023-03
Elon Musk founds xAI with the mission to understand the true nature of the universe.
2025-03
xAI acquires X (formerly Twitter) in an all-stock deal valued at $33 billion.
2026-02
xAI merges with SpaceX, valuing the combined entity at $1.25 trillion.
2026-05
OpenAI's general-purpose reasoning model disproves an 80-year-old mathematics conjecture by Paul Erdős.