Building a frontier, top-tier A.I. next year will cost more than a billion dollars. Because A.I. is becoming increasingly powerful every day in regard to the economy, the question of who owns those costly A.I.s is somewhat comparable to the question of who owns the most concentrated capital in the world. There are practical questions of the social impact of large language models and how they are trained politically, for those outlooks embedded in software algorithms will influence billions of people on sundry issues. The normative A.I. training parameters will train humans indirectly too.
Questions of national sovereignty in relation to A.I. ownership and operator’s nationality are also meaningful. Few would be comfortable today in the United States if China were to be the sole provider of A.I. for the United States. For that matter at least half the country would not be comfortable with Democrats ruling A.I. development. Political and power orientation of A.I. might induce some creative thought about its social effect on the populus. Would one want the government or Wall Street to be the sole providers of A.I. access?
Real employee wages have increased about 30% since the end of the Cold War. The growth of Wall Street has increased about 3800%. Information concentration is going the same way in regard to corporate information vs that of ordinary people. Information about ordinary people obtained by A.I. compared to information ordinary people have about corporate proprietary information is also asymmetric. Concentrated wealth of such an extreme degree as concentrated A.I.seems unAmerican. A.I. power concentrated in a few corporations is entirely accepted though, and at least rarely considered by the majority of society. There may be consequences of that.
Here is some data on A.I. today.
https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance
Gemini- Training a top-tier "frontier" AI model (like GPT-4, Gemini Ultra, or Grok 3) costs over \(\$100\text{–}200\) million in computing power alone as of 2024–2025, with costs projected to exceed \(\$1\) billion by 2027. Very few exist—roughly 5 to 10 truly top-tier frontier models exist globally, dominated by companies like OpenAI, Google, Anthropic, and xAI. [1, 2, 3, 4, 5]
Cost of Frontier AI Setup
- Training Costs: Training a frontier model requires massive, specialized computing clusters. Estimated training costs for flagship models are immense: GPT-4 was estimated at \(\$78\text{–}100+\) million, while Google's Gemini Ultra was around \(\$191\) million.
- API Usage Costs (2026): For users accessing these models via API, pricing is structured by tokens.
- Claude 4 Opus: \(\$15.00\) input / \(\$75.00\) output per million tokens.
- Gemini 2.5 Pro: \(\$1.25\text{–}2.50\) input / \(\$10\text{–}15\) output per million tokens.
- Grok 3/4: \(\sim\$3.00\) input / \(\$15.00\) output per million tokens.
- Consumer Subscription Costs:
Number of Existing Frontier Models
The number of truly "frontier" models—those setting the state-of-the-art—is extremely low due to the high barrier to entry. [1, 2, 3]
- Core Actors: The main players are OpenAI (GPT-4o/5), Google (Gemini Ultra/Pro), Anthropic (Claude Opus/Sonnet), and xAI (Grok).
- Estimated Count: Only a handful of organizations currently possess the computational resources (\(\text{>10,000s}\) of GPUs) and capital to train these models, resulting in fewer than 10-15 distinct flagship, truly leading-edge models globally, though many more "near-frontier" models are emerging. [1, 2, 3, 4, 5]
Note: Cost and capability estimates are based on industry trends as of early 2026
As of mid-2026, the U.S. generally leads in top-tier, proprietary AI model performance, but China has rapidly closed the gap, nearly erasing the U.S. advantage through highly efficient, open-source models. While American models like Anthropic’s Claude Opus 4.6 maintain a narrow edge in advanced reasoning, Chinese models like DeepSeek R1 and Alibaba’s Qwen often provide 90% of the capability at 10% of the cost, making them highly competitive. [1, 2, 3, 4]
Key Differences in the AI Race:
- Performance vs. Efficiency: U.S. models often win on raw power and capability (e.g., GPT-4, Claude). China has shown incredible ability to create highly optimized models that are cheaper and, through open-source approaches, often more accessible for customization.
- Hardware and Compute: The U.S. retains a substantial advantage in total compute, backed by immense capital expenditures from firms like Microsoft, Alphabet, Amazon, and Meta. China faces constraints due to U.S. chip export controls but excels in utilizing “mature” chips for inference.
- Open Source Dominance: Chinese firms are dominating the open-weight model space, providing top-tier alternatives that are heavily used globally.
- Application Areas: The U.S. leads in software-based AI applications, while China often takes the edge in AI “bodies” (robotics, drones) and industrial application, supported by heavy government subsidies. [1, 2, 3, 4, 5, 6]
The Gap Is Closing
By March 2026, the performance gulf in chatbot “Arena scores” had shrunk to just 39 points, a significant drop from the vast, year-over-year lead previously held by the U.S.. The consensus is that while the U.S. holds a slight edge in foundational innovation, China is an equal, if not superior, competitor in efficiency and specialized implementation. [1, 2, 3]
| Rank [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12] | Model Name | Owner / Developer | Performance Score (%) |
|---|---|---|---|
| 1 | Gemini 3.1 Pro Preview | Google DeepMind | 96.1% |
| 2 | Claude Opus 4.7 | Anthropic | 94.2% |
| 3 | GPT-5.5 | OpenAI | 93.1% |
| 4 | GPT-5.4 Pro | OpenAI | 92.8% |
| 5 | Claude Mythos Preview | Anthropic | ~92.5%* |
| 6 | Gemini 3.1 Pro | Google DeepMind | 87.0% |
| 7 | Grok-4.20 Expert | xAI (Elon Musk) | ~86.5%* |
| 8 | Llama 4 Maverick | Meta | ~85.8%* |
| 9 | Claude Opus 4.6 | Anthropic | 85.0% |
| 10 | Qwen 3 | Alibaba Group | ~84.2%* |
| 11 | DeepSeek-V3 | DeepSeek (Liang Wenfeng) | 78.2% |
| 12 | Mistral 3 | Mistral AI | ~77.8%* |
| 13 | Ernie 5.0 | Baidu | ~77.1%* |
| 14 | Kimi-K2 Instruct | Moonshot AI | ~76.5%* |
| 15 | Gemma 3:12b | Google DeepMind | ~75.4%* |
| 16 | Command R4 | Cohere | ~74.2%* |
| 17 | Llama 3.1 405B | Meta | 75.4% |
| 18 | Codex 2 | OpenAI | ~72.1%* |
| 19 | Claude Code | Anthropic | ~71.8%* |
| 20 | Reflection-2 | Reflection AI | ~70.5%* |
