Who Pays for Intelligence — The $191 Million Question

Article 4 of 5 · AI & Technology Series

Who Pays for Intelligence — The $191 Million Question The most powerful AI models in the world did not emerge from a clever algorithm alone. They emerged from something far more material: an almost incomprehensible quantity of electricity, silicon, water, money, and engineering time. Training a frontier AI model in 2026 is less like writing software and more like constructing a skyscraper — an industrial undertaking that requires the coordination of thousands of specialised components, a sustained energy supply measured in megawatts, and a financial commitment that only a handful of organisations on earth can make. Understanding what model training actually costs — and what it is buying — is essential to understanding who will hold the levers of AI power for the next decade. That is no longer a technical question. It is a geopolitical one.

What It Actually Costs to Train a Frontier Model The numbers are staggering, and they are only heading in one direction.

Those figures cover compute alone. The full cost of developing a frontier model is considerably higher. R&D staff costs — salaries, benefits, and equity compensation — account for between 29% and 49% of the total amortised development cost for models like GPT-4 and Gemini Ultra. Senior ML researchers at top labs command total compensation north of half a million dollars a year. Before a single GPU spins up, the payroll alone is staggering. There are other invisible line items too. The networking cables connecting GPU nodes — InfiniBand, the high-bandwidth interconnect standard for training clusters — account for roughly 9–13% of total training hardware costs alone, comparable to the entire energy spend. Training data is another cost that rarely gets quantified publicly, because most labs don't disclose what they pay to acquire, clean, and license it. And then there are the failures — failed training runs, hyperparameter experiments, restarts — which can multiply total compute spend by a factor of 1.2× to 4× beyond the final training run alone. What You're Actually Paying For: Breaking Down the Bill The GPU bill is the obvious one. Training compute for GPT-4 consumed an estimated 21 billion petaFLOPs of computation, and models like Gemini Ultra are estimated at approximately 50 billion petaFLOPs — driving those nine-figure compute bills. To put that in physical terms: training a GPT-MoE-1.8T model required approximately 25,000 Ampere-based GPUs running for three to five months, or alternatively 8,000 H100s running for 90 days. Whether you own or rent that infrastructure, the numbers are eye-watering either way.

The DeepSeek Disruption: Efficiency as a Weapon In January 2025, a Chinese AI lab called DeepSeek published results that sent shockwaves through the industry — and briefly wiped $600 billion from NVIDIA's market capitalisation in a single day.

DeepSeek trained a competitive frontier model for a reported $5.6 million in compute. The US labs had been spending hundreds of millions. The world wanted to know: how?

DeepSeek V3 trained on 2.79 million GPU hours at a reported compute cost of $5.6 million — a figure that, taken at face value, represented an order-of-magnitude reduction in training cost compared to US frontier models. The reaction in Silicon Valley was a mixture of alarm and scepticism. The scepticism was warranted. According to TechCrunch, DeepSeek's $5.6 million figure "excluded" infrastructure, experimentation, and failed training runs. CNBC reported that OpenAI is investigating whether DeepSeek used "distillation" from existing models — a fundamentally different and less expensive approach than training from scratch. But even with those caveats, DeepSeek's results validated something important: raw spending is not the only path to capable models. The lab deployed several genuine efficiency innovations:

Mixture of Experts (MoE): Only a subset of model parameters activate per input, reducing compute per forward pass by 4–8× Quantisation-aware training: Training in lower precision (BF16, FP8) reduces memory and compute requirements by 2–4× with minimal quality impact Data quality over quantity: Smaller, higher-quality datasets can match models trained on 10× more lower-quality data

DeepSeek R1 was trained for only $294,000 using efficiency optimisations — proving that brute-force spending is not the only path to capable models. The implication is significant: efficiency techniques are rapidly democratising the ability to build highly capable models, even as frontier training costs continue climbing at the top.

The Paradox: Costs Are Falling and Rising Simultaneously This is the most important and least understood dynamic in AI model training today.

The Infrastructure Arms Race: Chips, Power, and Water Behind every training run is a physical infrastructure that is becoming one of the most strategically contested resources on the planet.

The grid was not built for this. It was built for a world where electricity demand grew at 1–2% per year, predictably, with decades of warning. Now, hyperscalers are showing up at utility offices asking for hundreds of megawatts on three-year timelines. More than 70% of grid interconnection requests in the United States are ultimately withdrawn because the grid simply cannot accommodate them. The response from the largest AI labs has been to go directly to the source. Microsoft and Google have both announced plans to restart US nuclear power plants to support AI infrastructure. OpenAI is reportedly in talks to lease a planned 10-gigawatt data centre campus on federal land in Ohio — a campus measured in the output of multiple power plants. Broadcom has teamed up with Apollo and Blackstone on a $35 billion first-phase plan to build and finance compute capacity for top AI labs.

Who Is Winning — and What "Winning" Actually Means

As of mid-2025, the geopolitics of AI stands at a crossroads. On one path, the world slides further into fragmentation, with a digital iron curtain separating US-led and China-led tech spheres — data centres, networks, and AI ecosystems divided by incompatible standards and mutual suspicion. On another path, increased dialogue prevents the worst outcomes. The semiconductor export controls tell the story clearly. By mid-2025, US authorities had banned even specialised AI chips designed to meet earlier export rules, effectively closing the last major chokepoint for top-tier AI hardware. The message was unambiguous: the US intends to maintain a hardware advantage in AI at all costs. China is responding with state-directed investment in domestic chip production, though it remains years behind at the cutting edge. The result is a bifurcating global AI ecosystem — US-allied and China-aligned — with the rest of the world navigating between them.

The Cost Paradox — and What It Means for Everyone Else The most important insight from all of this data is the paradox at its centre: as frontier training becomes more expensive and more exclusive, the ability to use AI capability becomes cheaper and more democratic every month. In 2023, a customer-facing chatbot on GPT-4 cost a mid-size company $8,000 to $15,000 per month. Today, the same volume of requests on GPT-4.1 Nano or Gemini 2.5 Flash runs for under $200. Same capability tier. Same query load. Two orders of magnitude cheaper. What was frontier-expensive two years ago becomes achievable for well-funded startups. The implication: both trends sustain GPU demand. Frontier labs need more GPUs for larger models. Smaller organisations need GPUs to train what was recently impossible. The total addressable market for training compute expands in both directions. For most businesses, individuals, and developers, the training arms race among hyperscalers is largely irrelevant to their immediate decisions. The capability those training runs produce flows downstream through APIs at prices that are collapsing toward zero. The question is not whether you can afford to use AI. The question is whether you are using it — because your competitors almost certainly are. For governments and policymakers, the picture is different. The geopolitical competition for AI supremacy between the United States and China is far more than a race about algorithm quality — it is a multifaceted struggle relying on the physical foundations of energy infrastructure and secure supply chains for microchips and critical minerals. The nation that most effectively secures these resources will unlock AI's full potential. This struggle will also define the future global balance of power.

The chatbot era made AI feel like magic. The AI infrastructure era is a reminder that even magic needs a power cord.

The race is not just happening in server rooms and research labs. It is happening in power grids, chip fabrication plants, mining operations, and diplomatic negotiations. It is, in the most literal sense, a race to build the infrastructure of the future. And it is already well underway.