Redistributed

Training a frontier AI model now costs roughly what it takes to build a skyscraper. That number keeps climbing. Understanding where the money goes reveals who gets to build these systems, who profits from them, and who gets locked out.

The Cost Escalation

The trajectory is steep enough to be disorienting. OpenAI's GPT-2, released in 2019, cost an estimated $50,000 to train. GPT-3, a year later, ran to roughly $4.6 million. GPT-4, released in 2023, reportedly cost over $100 million. By 2025, credible estimates for frontier training runs had crossed $1 billion, with some projections for next-generation models approaching $10 billion (Cottier et al., 2024).

That is a five-order-of-magnitude increase in six years. No other technology sector has seen its entry costs escalate this quickly. To understand why, you need to follow the money into four buckets: compute, energy, data, and labor.

What the Money Buys

Compute: The Dominant Cost

The largest share of a frontier training budget goes to GPU clusters. A state-of-the-art training run in 2025 requires tens of thousands of NVIDIA H100 GPUs — or their successors — running continuously for months. A single H100 costs roughly $30,000–$40,000. A cluster of 25,000 of them represents close to $1 billion in hardware alone, before you pay for the data center to house them, the networking to connect them, or the engineers to keep them running.

The engineering challenge is formidable. Distributed training across thousands of GPUs demands sophisticated parallelism strategies — splitting the model, splitting the data, managing communication overhead between chips. When you are running 25,000 GPUs for three months, even a 1% failure rate means hundreds of hardware failures that must be handled without losing training progress. Checkpoint systems, fault tolerance, and cluster management are entire subfields of engineering that only a handful of teams have mastered.

This is not just expensive. It is expensive in a way that demands specific, concentrated infrastructure — exactly the kind of capital expenditure that creates natural monopolies.

Energy: The Hidden Bill

A large GPU cluster draws tens of megawatts of power continuously. Training a frontier model can consume as much electricity as a small city uses in a month. At industrial electricity rates, the energy bill for a single training run can reach tens of millions of dollars.

But the real cost is not just the electricity price — it is securing the supply. AI labs are now competing with each other, and with the rest of the economy, for access to reliable, large-scale power. Microsoft has restarted a reactor at Three Mile Island. Amazon has purchased a nuclear-powered data center campus. Google has signed deals for small modular reactors that do not yet exist. The energy demands of frontier training are reshaping power markets and infrastructure investment decisions (see The Environmental Cost of AI and The Data Center Boom for the downstream consequences).

Data: Increasingly Contested

Pre-training data was once treated as essentially free — scrape the web, process it, train on it. That era is ending. A wave of copyright lawsuits from publishers, artists, and media companies is forcing labs to negotiate licensing deals. The New York Times lawsuit against OpenAI, filed in late 2023, signaled that major content producers would not accept unlicensed use of their work (Grynbaum & Mac, 2023).

The result is a new cost category. Data licensing deals now run to hundreds of millions of dollars collectively. Reddit's deal with Google, reportedly worth $60 million per year, provides a template: platforms with large amounts of human-generated content can charge for access. As more content owners assert their rights, the cost of assembling a high-quality, legally defensible training dataset will continue to rise.

Beyond licensing, there is the cost of data curation and cleaning. Raw web scrapes are full of duplicates, spam, toxic content, and low-quality text. Filtering and processing this into a useful training set requires significant engineering effort and compute.

Human Labor: The Invisible Foundation

Training a frontier model requires substantial human labor that is often invisible in cost discussions. RLHF (reinforcement learning from human feedback) depends on thousands of annotators who rate model outputs, write demonstrations, and provide comparative judgments. Red teamers probe the model for harmful behaviors. Evaluators assess capabilities and safety properties.

Much of this work is performed by contract workers, often in lower-income countries, paid wages that would be considered low even by local standards. An investigation by Time magazine found that Kenyan workers employed through a contractor for OpenAI were paid less than $2 per hour to label content that included graphic descriptions of violence and abuse (Perrigo, 2023). This is a pattern, not an exception — one documented extensively in Ghost Work and in reporting on The RLHF Workforce and Data Labeling in the Global South. The economics of RLHF depend on cheap human judgment at scale.

The engineering talent, by contrast, commands the opposite end of the pay spectrum. Senior ML researchers at frontier labs earn $1–5 million per year in total compensation. The labor market for people who can actually architect and execute a frontier training run is perhaps a few hundred people worldwide. Their scarcity is itself a barrier to entry.

Who Funds This

The capital required to build frontier models comes from a narrow set of sources. Big Tech companies — Google, Meta, Microsoft, Amazon — can self-fund from their existing cash flows. Google's parent Alphabet spent over $30 billion on capital expenditure in 2024, much of it AI-related. Meta has committed to spending $60–65 billion on AI infrastructure in 2025.

Startups like OpenAI and Anthropic rely on massive venture capital rounds and strategic investments. OpenAI raised $6.6 billion in late 2024 at a $157 billion valuation. Anthropic has raised over $10 billion, with significant backing from Google and Amazon.

A newer category of funder is the sovereign wealth fund. Saudi Arabia's Public Investment Fund and the UAE's investment vehicles have poured billions into AI companies and infrastructure. The geopolitics of AI training are now inseparable from its economics.

This funding structure means that frontier AI development is controlled by a handful of corporations and the investors behind them (The Venture Capital Trap explores how this funding model shapes what gets built). There is no public equivalent. No government-funded lab operates at the frontier. The closest analogy — and it is an imperfect one — would be if the internet had been developed entirely by three or four private companies with no DARPA, no NSF, no public university research. Some initiatives point toward alternatives: the EU's sovereign AI investments, national compute programs, and projects like BLOOM (which trained a frontier-scale multilingual model through international academic collaboration) demonstrate that public and cooperative paths are technically feasible, even if they remain underfunded relative to private efforts.

The Return on Investment Question

When does a $1 billion training run pay for itself? The honest answer is that nobody knows with certainty. The economics depend on how many users the resulting model attracts, what they will pay, and whether the model retains its competitive advantage long enough to recoup costs before the next generation arrives.

OpenAI reportedly reached $3.4 billion in annualized revenue by late 2024, but was still losing money — spending more on compute, talent, and operations than it earned. The bet is that revenue will grow faster than costs, but that bet assumes continued model superiority in a market where competitors are spending just as aggressively.

This creates a peculiar dynamic: companies are investing billions on the assumption that their models will generate returns that justify the investment, while simultaneously racing against competitors making the same bet. It looks less like a business model and more like an arms race.

The DeepSeek Disruption

In January 2025, the Chinese lab DeepSeek released its V3 model, claiming training costs of approximately $5.6 million — a fraction of what Western labs spend on comparable models (DeepSeek, 2024). The model performed competitively with systems that reportedly cost 20 to 50 times more to train.

The claim deserves scrutiny. The $5.6 million figure likely represents only the final training run's GPU-hours and excludes research costs, failed experiments, data preparation, and the extensive pre-existing infrastructure DeepSeek had built. Still, even accounting for those caveats, DeepSeek demonstrated that architectural innovation — particularly mixture-of-experts architectures and efficient training techniques — can substitute for brute-force compute scaling.

This matters for the redistribution question. If training costs can be reduced by an order of magnitude through better methods, the barrier to entry drops. More organizations can participate. The natural monopoly weakens. DeepSeek's efficiency innovations, along with techniques like distillation (training smaller models to mimic larger ones), suggest that the "spend more, get more" paradigm has limits — a possibility also examined in The Scaling Laws Plateau. The Rise of Small Language Models trained through distillation represents one practical consequence: capable models that run on consumer hardware, partially decoupling AI capability from billion-dollar budgets.

The Redistribution Question

The economics of training frontier models are not just a technical curiosity. They determine the structure of the AI industry, which in turn determines who benefits from the technology.

When training costs are $1 billion and rising, only a few organizations can play. Those organizations — predominantly American and Chinese tech companies — become the gatekeepers for a technology that will reshape the economy. Everyone else becomes a customer, a dependent, or an afterthought.

There is an alternative. Public investment in AI infrastructure, modeled on CERN or the Human Genome Project, could ensure that frontier capabilities are not exclusively controlled by private capital. The EU's proposed investments in sovereign AI, and various national AI compute initiatives, gesture in this direction — but at funding levels that remain a fraction of what private companies spend.

The question is whether the current cost trajectory makes public alternatives feasible, or whether the economics have already locked in a structure where a few corporations control the foundation layer of an increasingly AI-dependent economy. The answer depends partly on whether DeepSeek-style efficiency gains continue, making frontier training accessible to more organizations, or whether the next generation of models requires resources that only the largest companies can muster.

Either way, the cost structure is the power structure. Follow the money, and you find who decides what AI can do, who it serves, and on whose terms.

Sources

Cottier, Ben et al. "The Rising Costs of Training Frontier AI Models." Epoch AI, 2024. Epoch AI
Grynbaum, Michael M. and Ryan Mac. "The Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted Work." The New York Times, December 2023. NYT
Perrigo, Billy. "OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic." Time, January 2023. Time
DeepSeek. "DeepSeek-V3 Technical Report." 2024. GitHub
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher

General

The Economics of Training a Frontier Model