Training frontier AI models requires computing power that only a handful of organizations can afford. This concentration has consequences that extend far beyond who can build the biggest model.
The Numbers
Training GPT-4-class models costs somewhere between $50 million and $100 million in compute alone — not including researcher salaries, data acquisition, or failed experiments (Epoch AI, 2024). Claude 3 Opus likely cost similar amounts. Google's Gemini Ultra probably cost more.
These are estimates because the companies don't disclose exact figures. The opacity itself is revealing.
The compute comes from specialized chips (GPUs, TPUs) manufactured by a tiny number of companies, assembled into clusters by an even tinier number of cloud providers, and rented at rates that price out all but the wealthiest organizations.
Who Has Access
The organizations that can train frontier models:
- OpenAI (backed by Microsoft)
- Anthropic (backed by Google and Amazon)
- Google DeepMind
- Meta AI
- A few others
That's roughly it. There are well-funded startups training smaller models, and academic labs doing important research at smaller scales, but the compute-intensive frontier belongs to a handful of players.
This is a structural feature, not a temporary accident. Training costs have grown faster than algorithmic efficiency improvements have reduced them. Each new generation of model requires more compute than the last.
The Infrastructure Layer
Below the model providers sits another layer of concentration:
Chip manufacturing is dominated by Nvidia (GPUs), with AMD distant second and Google's TPUs only available on Google Cloud. TSMC in Taiwan fabricates the advanced chips for everyone. A single company, ASML in the Netherlands, makes the lithography machines TSMC needs. As Kate Crawford documents in Atlas of AI, this material infrastructure — the mines, factories, and supply chains — shapes AI development in ways that pure software analysis misses (Crawford, 2021).
Cloud providers — Amazon Web Services, Microsoft Azure, Google Cloud — own the data centers where these chips run. They're also the primary customers for chips and the primary investors in the model providers.
Energy infrastructure determines where data centers can be built. Training runs consume megawatts continuously for months. Few grids can handle the load, and those that can are increasingly contested.
Why This Matters
The oil analogy breaks down in some ways — compute is renewable, oil isn't — but captures something important: a small number of entities control a critical resource, and that control shapes what gets built and for whom.
Research agendas get set by funders. Academic AI researchers increasingly depend on industry compute grants. The research that gets done is the research that can get compute. Questions that don't interest deep-pocketed sponsors don't get studied.
Deployment patterns follow capital. Applications that generate revenue get prioritized. Healthcare in wealthy countries attracts investment. Agriculture in developing countries doesn't. The technology theoretically works for both; the infrastructure favors one.
Competition policy struggles to keep up. Traditional antitrust focuses on consumer prices and market share. But when the market is "who can train the next generation of foundational AI," there are maybe five competitors globally, and they're all entangled through investment, cloud services, and talent movement.
Alternative Futures
This concentration is not inevitable. And history offers reasons for cautious optimism. Different choices could distribute compute more widely — and some of those choices are already being made:
Public compute infrastructure — national or international facilities that provide training resources to researchers, startups, and civil society organizations. The EU is experimenting with this through EuroHPC. The model has precedent: ARPANET was public infrastructure that seeded the internet; the National Science Foundation's NSFNET established open protocols that enabled an innovation ecosystem no private network could have planned. Public compute for AI follows the same logic. India's IndiaAI Mission and the UK's AI Research Resource are early steps in this direction.
Open model ecosystems — if training is too expensive to democratize, fine-tuning and deployment might not be. Open-weight models like Llama and Mistral let smaller organizations build on top of frontier capabilities without bearing frontier training costs. Projects like EleutherAI, BLOOM, and AI2's OLMo demonstrate that genuinely open development is possible — and the rise of small language models means useful AI no longer requires frontier-scale compute.
Algorithmic breakthroughs — training efficiency could improve faster than model size grows. This has happened in other fields. It's not a certainty but a possibility worth investing in.
Governance of the chokepoints — if the chip supply chain runs through specific companies and countries, those points become sites for regulation. Export controls on advanced chips are already a tool of geopolitics (Sastry et al., 2024).
The Question
How compute is distributed will determine who shapes AI development. Right now, a few companies in a few countries control the resource. That's a choice, embedded in policy and investment decisions, not a fact of nature.
The infrastructure question isn't separate from the social question. It's the foundation.
Related
- Open Source vs. Open Weights
- API Economy
- Who Regulates Foundation Models
- Hardware Bottleneck
- Frontier Lab Oligopoly
- AI Chip Export Controls
- Data Center Energy
- Inference Tax
- Public AI
Sources
- Epoch AI. "Trends in Machine Learning Hardware." 2024. Epoch
- Stanford HAI. "AI Index Report." 2024. Stanford HAI
- Benaich, Nathan & Hogarth, Ian. "State of AI Report." 2024. State of AI
- Crawford, Kate. Atlas of AI. Yale University Press, 2021. Publisher
- Sastry et al. "Computing Power and the Governance of AI." GovAI, 2024. GovAI