Conclusion First: Cracks Have Appeared in Nvidia’s ‘Monopoly’

China’s Meituan has trained a 1.6 trillion parameter AI model without using any Nvidia GPUs.

1.6 trillion parameters. That’s on par with GPT-4. The mere fact that this was achieved without Nvidia could fundamentally shake the cost structure of the AI industry.

Why is this important? The answer is simple: A significant portion of the API fees you pay each month is made up of the cost of Nvidia chips.

The Nature of Nvidia’s ‘Monopoly’

First, let’s clarify the structure.

Currently, Nvidia holds about 80-90% of the market share for GPUs used in AI development. The top model, the H100, costs around 4 to 5 million yen per unit. Training large-scale models requires thousands to tens of thousands of these units. In other words, training just one model can incur hardware costs in the billions to tens of billions of yen.

Where do these costs get passed on? To the API fees.

OpenAI’s GPT-4o charges about $2.5 (approximately 375 yen) for every million tokens of input and $10 (approximately 1,500 yen) for output. Claude 3.5 Sonnet charges $3 for input and $15 for output. When companies start using these services seriously, monthly costs can easily reach tens of thousands to hundreds of thousands of yen. For small and medium-sized enterprises in rural areas, these amounts are not at the level of “let’s give it a try.”

In other words, there is a direct line connecting Nvidia’s GPU prices → Cloud inference costs → API fees → Small and medium-sized enterprises’ adoption decisions. If Nvidia’s monopoly crumbles, the upstream of this line will change, affecting the costs for small and medium-sized enterprises downstream.

What Meituan Has Proven

Meituan reportedly used domestically produced chips. Chinese companies, unable to obtain Nvidia’s latest GPUs due to U.S. export restrictions, have been compelled to explore paths without Nvidia.

What’s crucial here is the fact that “huge models can be trained with alternative chips.”

The prevailing wisdom until now has been: “The Nvidia CUDA ecosystem is essential for training large-scale models. Other chips cannot deliver the performance. Software optimization cannot keep up.”

Meituan has overturned that conventional wisdom at a scale of 1.6 trillion parameters, where excuses cannot be made.

Of course, whether the training efficiency and cost efficiency are on par with Nvidia is another matter. Even if training that takes 100 days on Nvidia’s H100 takes 150 days on domestically produced chips, if the chip cost is less than half, the total cost will decrease. What matters is not whether it can be done, but how the costs will change.

Alibaba’s Ban on Claude Code Indicates ‘Another Division’

Around the same time, news emerged that Alibaba had banned the internal use of Claude Code as “high-risk software.”

This is not just a matter of security policy. It signifies that the “division” in the AI supply chain is spreading beyond chips to the software layer as well.

U.S.-made AI tools are being shut out of the Chinese market. Chinese-made AI chips cannot enter the U.S. market. As a result, the structure of two independent AI ecosystems is accelerating.

This division will cause confusion in the short term. However, in the medium to long term, it means intensified competition. Both camps will independently develop chips, train models, and provide APIs. As the sources of supply increase, prices will decrease. This is basic economics.

So, When Will API Fees Drop?

This is the crux of the matter. I want to answer the question, “This is interesting, but when will our API costs go down?”

What Is Already Happening

In fact, API fees have dramatically decreased over the past year.

• The cost per input token for GPT-4 has dropped by about 85% from its announcement in March 2023 to the end of 2024.
• Google’s Gemini 1.5 Flash charges $0.075 per million tokens. Compared to the cost per token at the time of GPT-4’s announcement, this is about 99% off.
• DeepSeek’s API offers prices that are less than one-tenth of those of U.S. competitors for comparable performance.

The main reason for this downward trend is the efficiency of models and optimization of inference. However, if chip supply diversifies, further structural price declines will occur.

Future Scenarios

Short-term (6-12 months): Meituan’s achievements will not directly reflect in API fees. The ecosystem for domestically produced chips is still maturing and is not yet at a stage where it can be commercially offered as an external API. However, low-cost APIs from Chinese companies like DeepSeek and Qwen will increasingly become viable options.

Medium-term (1-3 years): The performance of AMD, Intel, Google TPU, and domestically produced Chinese chips will improve, weakening Nvidia’s price dominance. As cloud providers diversify their chip procurement sources, inference costs could potentially drop to one-third to one-fifth of current levels. A monthly API usage cost that was previously 100,000 yen could be reduced to 20,000 to 30,000 yen.

Long-term (3-5 years): In addition to chip diversification, edge inference (running AI on local devices) will become widespread. There will be more cases where there is no need to call an API at all. Costs will approach nearly zero in some areas.

What Should Small and Medium-sized Enterprises Do ‘Now’?

If you’ve read this far and thought, “Then I’ll just wait until it gets cheaper,” that’s a mistake.

There are two reasons.

First. There are areas where API fees are already ‘sufficiently’ low.

With Gemini 1.5 Flash, the cost is about 11 yen per million tokens. Automating internal inquiries, summarizing daily reports, generating drafts for estimates — for these applications, you can keep costs to just a few thousand yen per month. If you think it’s “expensive,” it’s because you’re only looking at the price list for the top models. If you choose a model that fits your needs, you can start using it today.

Second. The value of AI lies not in ‘cheapness’ but in ‘systematization.’

Even if API fees are halved, if you don’t know how to use it, zero is still zero. Conversely, if you establish a system now that identifies “this task can be automated with AI” or “this data can be analyzed by AI,” when the fees drop, your profits will increase accordingly.

What small and medium-sized enterprises in rural areas should do is not to chase the trends of the latest chips. They should find one task within their operations that can be replaced by AI and try it out today.

What Will Really Change Is Not ‘Cost’ But ‘Assumptions’

What Meituan’s 1.6 trillion parameter model has shown is the fact that “huge AIs can be created without Nvidia.” This is not just about chip prices; it also indicates that the prerequisites for AI development have changed.

“AI is expensive,” “only large companies can use it,” “it doesn’t work without the latest GPUs” — these assumptions are being dismantled one by one.

The development of a business system that used to cost 3 million yen can now be done for 300,000 yen with AI coding tools. An outsourcing cost of 500,000 yen per month can be reduced to 5,000 yen with an AI chatbot. Such reversals will accelerate each time the chip monopoly crumbles.

I want to ask: Is the AI utilization that your company has postponed because it’s ‘expensive’ or ‘difficult’ really still expensive? Is it truly still difficult?

The answer has probably already changed. The only way to find out is to try.