What does it really cost to train a language model in 2026?

Training a language model has one main bill: renting graphics processors (GPUs), the chips that do the arithmetic, for as long as the run lasts. The unit that matters is the GPU-hour: one GPU for one hour. Eight GPUs for two hours is 16 GPU-hours. Multiply GPU-hours by the hourly rate and you have the core of the budget.

The range is enormous, which is why headlines about billion-dollar training and hobby projects at under $100 can both be accurate.

What GPUs rent for today

These are public list prices in US dollars per GPU-hour, read on November 2, 2026. Taxes are extra, and prices move, so check before you budget. Reserved contracts can differ.

Runpod sells two tiers at different prices, so read who runs the machines in each before you put sensitive data on one. Marketplaces and spot offers can go much lower, with less certainty that the machine stays yours.

What real runs cost at those prices

The hours come from the projects' own reports. The dollar column multiplies them by the prices above, unless noted.

Three notes on reading this table. First, nanochat's README quotes about $48, using roughly $3 per GPU-hour. At Lambda's listed $31.92 an hour for the 8-GPU machine, the training alone (1.65 hours on the current leaderboard) is about $53, and a full run with setup and evaluation is closer to $64. Its author also mentions about $15 on spot machines, whose prices change constantly. Second, the same "GPT-2 grade" took about $43,000 in 2019, by the author's account. Third, only DeepSeek's cost is a company's own number, and DeepSeek says it leaves out earlier research and experiments.

The costs that are not on the rate card

Failed and experimental runs. The published number is normally the final successful run. Serious projects try many versions first. For small models this is your main cost after the first success.

People. A 2024 analysis of frontier models by Cottier and colleagues finds that for models such as GPT-4 and Gemini, staff cost tens of millions of dollars, about as much as the chips, while energy is only 2 to 6%. It estimates that the most compute-intensive runs have grown about 2.4 times a year since 2016, and that if the trend continues the largest will pass a billion dollars by 2027.

Storage and data. A filesystem you forget keeps billing. Lambda, for example, charges for a filesystem as long as it exists but not for moving data in or out. Other clouds may charge for outbound data, so read the terms.

Forgotten machines. On-demand instances usually bill until you terminate them, even when idle. An 8-GPU node left running over a weekend, about 60 hours, costs about $1,900 at Lambda's rate, far more than the training itself.

If you are in Canada

The prices above are in US dollars, so the exchange rate adds to a Canadian bill. University researchers have another route: the Digital Research Alliance of Canada gives account holders CPU and GPU time for opportunistic use, without priority and without guarantees. More priority requires its Resource Allocation Competition, open to faculty, adjunct faculty and librarians. The 2027 round closes on November 3, 2026.

How do you estimate your bill before you start?

Use one formula: tokens to train on, divided by tokens processed per GPU-hour, multiplied by the price. A token is the piece of text the model reads, roughly three to four characters of English.

SmolLM3 gives a real throughput for reference. It processed 11.2 trillion tokens in 221,184 GPU-hours, which works out to about 50 million tokens per GPU-hour, or about 14,000 per second per GPU. Suppose you want to pretrain a similar model on 500 billion tokens:

• 500 billion ÷ 50.6 million = about 9,900 GPU-hours
• 9,900 × $3.99 = about $39,500
• Double it for experiments, as a planning assumption, and budget about $80,000

That throughput depends on the model size, the batch settings and the machine, so measure your own. Run for ten minutes, read the tokens per second from the training log (nanochat reports it as train/tok_per_sec), and redo the arithmetic with your number. Those ten minutes on an 8-GPU node cost about $5, and they can stop you from starting a run you cannot finish.