If you cancel a $20-a-month AI service and run a large language model (LLM) on your own PC, will you actually save money? In an account published by Lifehacker Japan, the author replaced Perplexity Pro with a local LLM and found it could handle most of their everyday work. The author also reports, however, that electricity costs and maintenance work increased, and that Perplexity still has an advantage for real-time web search.
If you already own a PC with a GPU, whether you come out ahead depends mainly on three conditions: how much the electricity costs, whether the model you want to use fits in GPU memory (VRAM), and whether the quality of its Japanese is good enough for your purposes. If you would need to buy a new PC, you also have to consider how many months it would take to recoup the upfront cost. Using actual model specifications and electricity rates, we examine where the line falls between keeping a flat-rate subscription and saving money by moving to a local LLM.
Subscriptions aren't the only thing to compare open-weight models against
Open-weight models are models whose trained weights are published, so users can download and run them. The terms of use vary by model, however. Google's Gemma 4 and Qwen3.5-27B are released under the Apache 2.0 license, while Llama-3.3-Swallow-70B, which has been enhanced for Japanese, must follow the Llama 3.3 license as well as the usage restrictions in the Gemma Terms of Use. If you are considering commercial use, you need to check the license for each model.
Even when a model's weights are public, though, you don't necessarily have to run it on your own PC. Cloud providers host the same models and offer them as APIs where you pay only for what you use. For example, according to the price comparison site pricepertoken, as of October 2026 the output price for Llama 3.3 70B Instruct was $0.32 per million tokens at DeepInfra and $1.04 at Together. We picked DeepInfra and Together as examples because they were the cheapest among the providers compared.
If you spent the entire $20 a month on output tokens, you could generate roughly 62.5 million tokens with DeepInfra or roughly 19.2 million with Together. With DeepInfra, that works out to an average of about 2.08 million tokens per day. Tokens are the units in which an AI processes text, and input tokens are billed separately in practice, so $20 a month would not actually buy this much output. Even so, if your main uses are text generation and summarization, it is worth considering pay-as-you-go APIs before buying an expensive PC.
$20 a month is the base price of ChatGPT Plus, and it is also what the author of the account paid for Perplexity Pro. Flat-rate services and APIs for open models differ in features, however. Flat-rate services offer things like web search, image understanding and the latest models, whereas the features available through an API depend on the model and provider you choose. The two can't be compared on price alone.
| Option | Monthly cost | Upfront investment | Where your data goes |
|---|---|---|---|
| Flat-rate AI subscription | Around $20 | None | The service provider |
| Open-weight model API | Pay as you go (e.g., $0.32–$1.04 per million output tokens; input billed separately) | None | The API provider |
| Local LLM | Electricity and maintenance effort | GPU, memory, etc. | No external transmission if run fully locally |
The big advantage of running locally is that you can process input data without sending it outside. In exchange, you bear the hardware, electricity and maintenance burdens yourself. It is important to compare not just cost but also privacy and the effort of operating the setup.
An RTX 4090 costs roughly ¥14 per hour in electricity
The Lifehacker Japan author says they bought a laptop for about $1,600 (roughly ¥240,000). Converting at $1 = ¥150, a $20 monthly fee comes to ¥3,000, so ¥240,000 equals 80 months of subscription fees. However, a PC can also be used for work other than AI, so it isn't appropriate to count the entire purchase price as an AI cost. The actual payback period has to be calculated by dividing the portion of the purchase price you allocate to AI by the monthly difference between the subscription fee and the cost of running locally.
Beyond the upfront cost, electricity is easy to overlook. The account also names rising electricity bills as a hidden cost but gives no specific figure. So we estimate using the guideline unit price of ¥31/kWh (including tax), revised on July 22, 2022 by the Fair Trade Conference for Home Electric Appliances.
According to NVIDIA's official specifications, the GeForce RTX 4090 has a Total Graphics Power (TGP) of 450W. If the GPU runs at that power for one hour, the electricity cost is as follows.
\[ 0.45\text{ kW}\times31\text{円/kWh}=13.95\text{円} \]
In other words, if an RTX 4090 runs continuously at its rated 450W, the electricity cost for the GPU alone is about ¥14 per hour. That is about ¥420 for 30 hours a month and about ¥1,400 for 100 hours. 100 hours a month is an average of about 3.3 hours a day, and compared with a ¥3,000 monthly AI subscription, electricity alone would account for nearly half.
If you use two RTX 4090s to run a 70B-class model and both operate at 450W, the total is 900W. Under those conditions the cost is about ¥28 per hour, or ¥2,790 for 100 hours a month, which is 93% of the ¥3,000 monthly fee. In a two-GPU setup, electricity alone could come close to the cost of a flat-rate subscription.
This is only an estimate that assumes the GPU runs continuously at its rated 450W, however, and it does not represent actual power draw during inference. Power consumption varies with the model and the workload, and the estimate excludes power used by the CPU, main memory, cooling and so on. It can serve as a rough upper-end guide for the GPU alone, but it is not the electricity cost of the whole PC. To learn the actual cost, you need to measure the whole PC's power consumption with a power meter and multiply by your actual hours of use.
Which models fit in the GPU memory you have?
As important as electricity is the amount of GPU memory (VRAM) needed to run a model. Using 4-bit quantization, which converts model weights from 16 bits to 4 bits, reduces the weight data to roughly a quarter in theory. Actual model files also contain additional information, however, and the loss of precision from quantization can affect answer quality in some cases.
For example, according to Qwen's published model information, the 4-bit quantized (Q4_K_M) version of Qwen2.5-Coder-32B-Instruct is about 19.9GB. At run time, memory is needed not only for the model itself but also for things like the KV cache, which holds the input and information generated so far. The model information site llmrun estimates that the required memory is around 22–28GB depending on context length.
Here, the Lifehacker Japan account contains a statement worth treating with caution. The author reports that on a laptop with an RTX 4060 and 8GB of VRAM, a 4-bit 32B model fit in VRAM and ran at 25–30 tokens per second. But a model of about 20GB cannot fit in its entirety in 8GB of VRAM. They may have used offloading, which places part of the model in main memory, but the article doesn't let us confirm how memory was allocated at run time. Therefore, the reported generation speed can't necessarily be reproduced on a PC with the same amount of VRAM.
With an RTX 4090's 24GB of VRAM, the weights of a 4-bit 32B-class model can be loaded into GPU memory. When handling long contexts, however, additional memory is required and capacity may fall short. For 70B-class models, the conditions get tougher. According to the GPU information site Spheron, even a 4-bit quantized Llama 3.3 70B needs about 38GB, so it does not fit on a single 24GB GPU.
To handle a 70B-class model, you can use a GPU with 48GB or more of VRAM, spread the model across multiple GPUs, or place some of the weights in main memory. Distributing across multiple GPUs requires a compatible runtime environment, though, and offloading to main memory tends to slow generation. Mini PCs with large amounts of memory and the MacBook Pro, which Gizmodo Japan has covered, are also options, but they raise the cost of adoption. If you buy additional hardware to use a higher-performance model, it could take a long time to recoup the cost relative to a monthly subscription.
On the other hand, in 2026 there are more options that deliver high performance even with relatively small models. Gizmodo Japan also notes that Qwen3.5 and Gemma 4 come in multiple model sizes aimed at a wide range of devices, including ordinary laptops. Gemma 4 comes in a 26B mixture-of-experts (MoE) model and a 31B Dense model (which uses all parameters in computation), and Google's official specifications estimate the memory needed to load them at 4-bit quantization at about 14.4GB and 17.5GB respectively. These figures include 20% overhead but do not include additional memory such as the KV cache. Even so, they show that if you don't insist on 70B-class models, you can run a local LLM with less VRAM.
If you'll use it in Japanese, check performance first
When using a local LLM in Japanese, you need to consider multilingual support and the quality of Japanese prose separately. According to Google, Gemma 4 is trained on more than 140 languages, and the Qwen3.5-27B model card describes support for 201 languages and dialects. But supporting many languages doesn't mean the Japanese text will necessarily read naturally. Especially for writing articles or technical documents, you need to check actual output for handling of specialized terminology, understanding of context and naturalness of prose.
Another approach is to choose a model that has been additionally trained for Japanese. On March 10, 2025, a research team from Institute of Science Tokyo and the National Institute of Advanced Industrial Science and Technology released Llama-3.3-Swallow-70B, a model based on Llama 3.3 with enhanced Japanese ability. In the research team's evaluation, its average score on Japanese MT-Bench, which measures Japanese dialogue performance, was 0.772, below GPT-4o's 0.848 and Qwen2.5-72B-Instruct's 0.835. Even a model tuned specifically for Japanese does not necessarily match the performance of the cloud models or other open models it was compared with.
Llama-3.3-Swallow-70B is also a 70B-class model, so even with 4-bit quantization it needs about as much memory as the original Llama 3.3 70B. Based on Spheron's estimate, you should expect to need about 38GB or more, and it won't fit on a single 24GB GPU. Using two GPUs increases power consumption and purchase costs, so if you choose a large model to improve Japanese quality, you need to consider whether the benefit justifies the burden.
The Lifehacker Japan author mainly uses their setup for code review, documentation, data analysis and technical troubleshooting. The account contains no detailed evaluation of Japanese prose quality, and conclusions drawn from code-centered work can't simply be applied to writing and editing Japanese articles. The surest basis for a decision is to generate the kinds of documents you handle every day with several models and compare the quality.
Who comes out ahead by canceling, and who should keep subscribing
The author of the account isn't giving up Perplexity entirely either; they plan to move to the free plan. They were able to replace things like code review and documentation with a local LLM, but they rate Perplexity's convenience for real-time web search as higher. A local LLM can also use web search if combined with something like Open WebUI, but the setup is complicated, so the same ease of use can't be easily reproduced. For image understanding and checking the latest information, too, you need to confirm what the local model and surrounding tools can do.
On cost, the more likely you are to already own a PC with a GPU and be able to do your work with models of 32B class or smaller, the greater the benefit of adopting a local LLM. For example, if you use a single RTX 4090 at its rated 450W for 30 hours a month, the electricity cost for the GPU alone is about ¥420. Setting aside the PC's own power draw and maintenance effort, that is far cheaper than a ¥3,000 monthly subscription. Conversely, if you buy additional GPUs to run a 70B-class model, or run several GPUs for long hours, the cost advantage shrinks.
If you'd use a 70B-class model, comparing against pay-as-you-go APIs also matters. If you ran two RTX 4090s at their rated 450W for 100 hours a month, the electricity cost for the GPUs alone would be ¥2,790. Meanwhile, if you put ¥3,000 a month ($20) toward the Llama 3.3 70B API, you could buy about 19.2 million to 62.5 million output tokens, ignoring input fees.
How much text a local LLM can generate in the same time, however, depends on inference speed. 100 hours is 360,000 seconds, so the number of output tokens you can generate is found by multiplying tokens per second by 360,000. The 25–30 tokens per second reported by Lifehacker Japan relates to a different model and hardware and can't be used for comparison with 70B-class models. You need to measure generation speed and power consumption in your own environment and compare them including the API's input and output fees.
Ultimately, whether you should cancel a $20-a-month AI subscription depends not only on model performance but also on how much and for what you use it. If you already own a sufficiently powerful PC and mostly write text or process code offline, there's a good chance a local LLM can cut your fixed costs. On the other hand, if you frequently use searches for up-to-date information, image understanding or advanced reasoning, keeping the subscription can be the more rational choice.
To decide, first measure the whole PC's power consumption and calculate the monthly electricity cost with "power consumption (W) ÷ 1,000 × hours of use × electricity rate." Then check whether the model you want fits in your VRAM and whether its Japanese quality meets your needs. If a local LLM can replace most of the features you need, you could cancel the subscription and cover only the missing features with an API or a free service. Rather than moving everything to local, using each option for the purpose it suits best is what balances cost and convenience.