Why AI is so expensive (and American AI even more so)

You pay $20 a month for ChatGPT or Claude, and now you’re being offered $100 or $200 plans. Meanwhile, DeepSeek charges a fraction of that for its best model. Why such a gap? Because behind every answer there are wildly expensive chips, power plants’ worth of electricity, and hundreds of billions of dollars in investments waiting to be paid back. Here’s where the bill comes from, with sources for every number.

In short: AI is expensive because training a frontier model now costs hundreds of millions of dollars, Nvidia’s chips are scarce and sold at huge margins, and Microsoft, Meta and Amazon each spend more than $100 billion a year on data centers. American AI costs more because it chases maximum performance and has to earn back those colossal investments. Chinese models are cheaper thanks to leaner architectures and an open-weights strategy, even if their headline costs don’t tell the whole story. For students, the good news is that “good enough” models are becoming very affordable, even free.

Training a model: hundreds of millions of dollars

The research institute Epoch AI has tracked these costs for years. Its finding: the cost of the final training run for frontier models (amortized hardware plus energy) has grown 2.4x per year since 2016. If the trend holds, the largest training runs will cost more than a billion dollars by 2027. Hardware accounts for 47 to 67% of total development costs, research staff for 29 to 49%, and energy for just 2 to 6%.

A concrete example: Epoch estimates that training Grok 4, xAI’s model, cost around $490 million and 310 million kilowatt-hours. That’s an outside estimate, not a company figure. And the race isn’t slowing down: according to Stanford’s 2026 AI Index, global AI compute capacity has grown 3.3x per year since 2022.

GPUs: scarce, expensive, and one supplier cashing in

Models are trained on thousands of GPUs, specialized chips dominated by Nvidia. In early 2024, Raymond James analysts estimated Nvidia was selling its H100 for $25,000 to $30,000 apiece. A lab like DeepSeek reportedly has around 50,000 GPUs from this Hopper generation, according to SemiAnalysis.

The best measure of scarcity is Nvidia’s own accounts. In the quarter ending in late July 2026, the company posted $96.2 billion in revenue, up 106% year over year, including $89 billion from data centers, with a 75% gross margin. In other words, out of every $100 of sales, Nvidia keeps about $75 after the cost of the products sold. At least part of that margin ends up in the price of your subscription.

Big Tech is spending like never before

Running those chips takes data centers. The figures in the latest quarterly results are staggering:

Company	Capital spending	Source
Microsoft	$115.9 billion in the fiscal year ending June 2026	Microsoft
Meta	$130 to $145 billion expected for 2026	Meta
Amazon	$54.2 billion in Q2 2026 alone, up from $32.2 billion a year earlier	Amazon

At Amazon, the effect is dramatic: free cash flow turned negative (a $7.6 billion outflow over twelve months), driven by higher capital spending. The company says plainly: “This increase primarily reflects investments in artificial intelligence.”

OpenAI plays in the same league. The Stargate project, launched on January 21, 2025 with SoftBank, Oracle and MGX, planned $100 billion right away and up to $500 billion by 2029. OpenAI also signed a deal with Oracle for $300 billion of computing power over five years and committed to $250 billion of Azure services. Its 2025 revenue, by comparison, was $13.1 billion.

Energy: a growing bill

Electricity is a small share of a training run’s cost, but a big deal at global scale. According to the International Energy Agency, data centers used around 415 terawatt-hours in 2024, about 1.5% of the world’s electricity. That’s set to more than double, to around 945 TWh by 2030, slightly more than Japan’s total consumption today. AI is the main driver of that growth.

Inference: every answer has a cost

Inference is the moment the model answers you. Here the trend is actually good: according to the 2025 AI Index, the cost of GPT-3.5-level AI dropped more than 280-fold between November 2022 and October 2024. Epoch AI measures price declines of 9x to 900x per year at constant performance, depending on the task, while warning that the fastest drops may not last.

So why isn’t your bill going down? Three reasons.

1. The best models keep getting pricier. OpenAI’s GPT-6-Astra and Anthropic’s Claude Fable 5.1 both cost $10 per million input tokens and $50 per million output tokens.
2. Usage is exploding. A coding agent running for an hour burns far more than a quick question.
3. Some prices are going back up. Google says Gemini 3.8 Flash will rise from $0.75 to $1.50 per million input tokens on January 1, 2027, and from $3.75 to $7.50 for output.

Why subscriptions keep climbing: 20, 100, 200

The $20 tier is still the norm, but “heavy use” plans are multiplying:

Service	Standard plan	Higher tiers
ChatGPT	Plus: $20/month	Pro x5 at $100 and Pro x20 at $200 since April 2026 (Wikipédia)
Claude	Pro: $20/month, $17 billed annually	Max from $100 (Claude)
Gemini (France)	AI Pro: €21.99/month	AI Ultra at €99.99 or €219.99 (Google)
Le Chat (Mistral)	Pro: $14.99/month	$5.99 for verified students (Mistral)

The logic: a flat fee can’t cover someone running agents all day. So companies sell usage multipliers (5x, 20x). And they’re looking for other revenue: on January 17, 2026, OpenAI announced it would start testing ads in the free version of ChatGPT, starting in the US.

Why Chinese models are cheaper

The contrast is stark. Here are some public prices per million tokens, input then output:

• GPT-6-Astra (OpenAI): $10 and $50.
• Claude Opus 5.5 (Anthropic): $4 and $20.
• DeepSeek V4-Pro: $0.66 to $1.32 and $1.98 to $3.96 depending on the time of day, half price off-peak.
• MiMo-V2.6-Pro (Xiaomi): $0.435 and $0.87.

Several things explain it.

A lean architecture. DeepSeek V4-Pro has 1.6 trillion parameters, but only 49 billion are active for each token. This “mixture of experts” design cuts compute, and with it the cost of every answer.

The chip squeeze. US export controls cut China off from the best GPUs: the H800, a throttled H100 designed for that market, was itself restricted in late 2023. Chinese labs had to do more with less.

Headline costs to take with a grain of salt. DeepSeek V3’s famous number comes from its technical report: 2.788 million H800 GPU hours, or $5.576 million at $2 per hour. The report itself says this excludes prior research and experiments. SemiAnalysis believes DeepSeek’s hardware spend is “well higher than $500M” over the company’s history. Anthropic also accuses DeepSeek, Moonshot and MiniMax of training their models on millions of Claude answers obtained through around 24,000 fraudulent accounts, an accusation no court has ruled on.

Cheaper per token doesn’t always mean cheaper per task. In September 2025, the US NIST evaluation found that one US reference model cost 35% less than the best DeepSeek model for similar performance. In April 2026, on the other hand, running Artificial Analysis’s benchmarks on DeepSeek V4 Pro cost $1,071, more than four times less than Claude Opus 4.7. The same organization flags very high hallucination rates for V4 Pro and V4 Flash on its dedicated test. Price is only one criterion.

Bubble or not?

Facts first. In late July 2026, South Korea’s stock market, until then carried by chipmakers riding the AI boom, fell nearly 11% and then 6% in two sessions, wiping out $2.18 trillion in market value amid fears of weaker AI demand and heavy leverage. It was still up 41.5% for the year in dollar terms.

On the worried side, OpenAI’s Sam Altman said back in August 2025 that investors as a whole were “overexcited about AI” and that some would get burned. In October 2025, the Bank of England warned that “the risk of a sharp market correction has increased”, given tech valuations.

On the confident side, Nvidia’s Jensen Huang says in his latest results release that “AI has reached its inflection point” and that compute now turns into revenue. The numbers partly back him up: Azure revenue topped $100 billion in Microsoft’s fiscal year, and investors value OpenAI at $852 billion and Anthropic at $965 billion.

Our take: both camps are right at the same time. Demand is real, but spending is running ahead of revenue. As long as that gap lasts, US companies have every reason to keep their best models expensive, while Chinese and open-source competition drags down the price of “good enough” models. That’s mostly good news for your budget.

What this means for you

1. Start free. The free versions of ChatGPT, Claude, Gemini and Le Chat are enough to revise, understand a lecture or quiz yourself. Our guide to free and student AI offers sorts them out.
2. Use student pricing. Mistral offers Le Chat Pro for $5.99 a month to verified students, and in France Google lists its AI Plus plan as free for a year for students: check the terms in your country before signing up.
3. Don’t pay more than $20 unless you use coding agents heavily. With Claude, annual billing brings Pro down to $17 a month.
4. Try open models locally. With Ollama, ollama run qwen3.8 installs Qwen3.8-27B (18 GB): no cost per message, but you need a machine with plenty of memory.
5. For coding projects, pick the smallest model that does the job. GPT-6-Luna costs $0.10 per million input tokens, a hundred times less than GPT-6-Astra. Batch processing halves prices at Mistral, and DeepSeek charges half price off-peak.
6. Use AI as a coach, not a contractor. Precise questions and short answers cost less and teach you more.

FAQ

Why does ChatGPT Plus cost $20 when DeepSeek is free?

Both have a free version. The difference is at the high end: OpenAI has to earn back hundreds of billions of dollars in data center commitments, while DeepSeek relies on leaner models and open weights. And the DeepSeek app raises other questions, about your data.

How much does it cost to train an AI like ChatGPT?

Companies don’t publish these numbers. Epoch AI estimates Grok 4’s training at around $490 million and expects the largest training runs to pass a billion dollars by 2027.

Did DeepSeek really cost $6 million?

Not in the way people think. The $5.6 million covers only DeepSeek V3’s final training run, priced at $2 per GPU hour. The report itself excludes prior research, and SemiAnalysis puts the company’s hardware spend well above $500 million.

Will AI prices go down?

At equal performance, yes: the cost of GPT-3.5-level AI fell more than 280-fold in two years. But the most powerful models stay expensive, and some prices are rising, like Gemini 3.8 Flash, which will double in January 2027.

Further reading

• Free AI and student offers: every plan that lets you pay nothing, or close to it.
• Chinese AI: why Europe is wary: what to know before using DeepSeek or Qwen.
• The AI race: pessimists vs optimists: the big-picture debate behind the bubble talk.
• The budget-etudiant skill: fit your subscriptions into a budget that holds up.