Recently, DeepSeek announced that it will adjust API prices in mid-July, doubling the call price from 9:00 to 12:00 and 14:00 to 18:00 on weekdays. The output price of the V4-Pro model has increased from 6 yuan to 12 yuan per million tokens. At the same time, Zhejiang Province's new industrial and commercial time-of-use electricity prices officially came into effect on July 1, with 16:00 to 23:00 designated as the peak period, and an additional "peak" period from 18:00 to 22:00 in summer and winter, with the peak-to-valley price ratio widening to 2.05 to 0.2.

An AI company doubled the call price during the day, and a provincial power grid designated after 4 p.m. as a "expensive electricity period." The two industries used the same logic to adjust prices in the same month - both relied on price to drive away peak demand. DeepSeek is managing the time distribution of computing power, and Zhejiang is managing the time distribution of electricity.

The time distribution of computing power is essentially the time distribution of electricity.

AI's "electricity bill" is starting to get expensive

Let’s take a look at global bills first.

According to data from Gartner, global data center electricity consumption will reach 565TWh in 2026, a net increase of 118TWh from 447TWh in 2025. The annual increase of 118TWh is almost equivalent to the annual electricity consumption of the Netherlands. Among them, the power consumption of AI optimized servers skyrocketed from 95TWh to 175TWh, an increase of 84%. By 2027, the power consumption of AI servers will officially exceed that of traditional servers.

Gartner research director Linglan Wang’s judgment is: “Today’s AI computing power is limited by power supply.”

Microsoft CEO Nadella said in the podcast: "The biggest problem we face now is no longer a surplus of computing resources, but whether power can be quickly installed where there are data facilities. If we cannot do this, we may have a bunch of chips lying in the warehouse but not plugged in."

The power consumption of a GPU server is 3 to 5 times that of a traditional server, and the power consumption of a single NVIDIA Blackwell architecture GPU is 1kW, which is 40% more than the 700W of the previous generation Hopper. The "Stargate" data center built by OpenAI in Texas has a maximum power demand of 1.4GW and deploys more than 400,000 GPUs. What is the concept of 1.4GW? Enough for a small city.

The result is that in the end, AI companies spend billions to buy chips, only to find that the power may not be enough.

DeepSeek's peak and valley pricing is essentially the same operation in the power industry - it's more expensive during the day and cheaper at night, using prices to drive demand toward low periods. In fact, DeepSeek has moved the rules of the power grid to the computing power market.

The output price of the V4-Pro model is 6 yuan per million tokens during peak periods and 12 yuan per million tokens during peak periods. The input price of cache hits increased from 0.025 yuan to 0.05 yuan. In May this year, DeepSeek permanently lowered its price by 75%, triggering a price war for large models. A month ago we were still charging "floor price", but now we are charging "peak surcharge".

Why does it turn so quickly? The bottleneck on the supply side of computing power is greater than anyone expected. Industry public monitoring data shows that the demand for computing power has increased by several times in the past year, while the expansion rate of supply has been far from keeping up, with the gap between supply and demand being about half. Tencent Cloud rose by 5% in April, Alibaba Cloud rose by 5% to 34% in March, Baidu Smart Cloud rose by 5% to 30%, AWS rose by 15%, and some Google Cloud services rose by up to 100% in May. Increasing prices is not a competitive strategy, it’s because the supply side really can’t bear it.

DeepSeek's peak and valley pricing shows that its GPU cluster, like the power grid, has experienced peak congestion.

Zhejiang has designated the period after 4 p.m. as "expensive electricity period".What is it for?

The new industrial and commercial time-of-use electricity price in Zhejiang Province, which took effect on July 1, cuts the electricity price throughout the day into five levels: peak, peak, flat section, trough, and deep valley. The ratio is 2.05 to 1.85 to 1 to 0.4 to 0.2. The peaks are 10 times more expensive than the deep valleys. The original morning peak hours have been cancelled, and most hours during the day have been changed to flat rates. All peak hours are from 4 pm to 11 pm, and an additional peak is from 6 pm to 10 pm in summer and winter.

In recent years, Zhejiang's photovoltaic installed capacity has been expanding. Photovoltaic power generation is large at noon, and electricity is cheap or even surplus. But in the evening, the photovoltaic output drops to zero. Going home from get off work, turning on the air conditioner, cooking, and charging are all crowded together, and the load on the power grid rises sharply. This is what the power industry calls the "duck curve" - ​​there is too much electricity to use during the day, and too expensive to use at night.

Zhejiang’s adjustment is to allow factories and data centers to move work to noon. The document mentioned that if a steel company with an annual electricity consumption of 1 billion kilowatt-hours shifts 10% of its peak electricity consumption to trough periods, it can save more than 20 million yuan in electricity bills every year, which is enough to build an energy storage system.

The document also clearly mentioned that industries such as warehousing logistics and data centers with "stable loads and large adjustable space" can fully enjoy the dividends of low electricity prices. By moving computing tasks to noon or late at night, you can save a lot on electricity bills.

DeepSeek is managing the time distribution of computing power, and Zhejiang is managing the time distribution of electricity. The time distribution of computing power is essentially the time distribution of electricity - behind every Token is electricity.

According to data from Gartner, the power consumption of AI-optimized servers has increased by 84% in one year. The greater the computing power demand, the greater the power consumption. The tighter the electricity supply, the more like electricity the cost structure of computing power becomes.

The 2026 government work report includes "computing and computing collaboration" for the first time, revealing the country's top-level design plan for computing power. The computing power center and the power grid should be planned together. The computing power will follow the power, and the power dispatching depends on the computing power. This is not a vision, it is forced out by reality. Gartner predicts that data center power consumption will exceed 1,200TWh by 2030, and the grid power supply will not be able to meet demand. If computing power and electricity are not tied together at the planning level, things will be more troublesome later.

After the free lunch,Let’s start to calculate the details

Although DeepSeek has adjusted the price of the API, the price during peak periods is actually exactly the same as the price after the permanent price reduction in May. DeepSeek did not "increase the price", it just added a surcharge for daytime calls.

But for most developers, daytime is work time. 9 o'clock to 12 o'clock and 14 o'clock to 18 o'clock are the prime time for writing code, adjusting interfaces, and running tests. Doubling the price during this period is equivalent to adding more to most people’s daily costs.

Foresight Energy understands that under these conditions, the cost of real-time interactive applications may surge, and offline computing tasks can only be moved to night. In the past, hundreds of thousands of small teams used almost free APIs for prototype verification, and the cost was close to zero. After peak and valley pricing, if the business model relies on real-time calls, such as customer service, search, and code completion, the cost structure will need to be recalculated.

OpenRouter's report shows that the gap between open source and closed source models has remained stable between three and six months over the past 18 months. The cost-effectiveness of the open source model is eating up the market for closed source APIs. If DeepSeek continues to engage in price wars, it will not be able to achieve a closed-loop business. The price increase is not "no more money", it is the settlement after "no more money".

Zhejiang’s electricity price adjustment and DeepSeek’s peak and valley pricing occurred in the same week. One is in charge of electricity and the other is in charge of computing power. Computing power is becoming a new way of using electricity - not air conditioners, factories, or electric vehicles. It is the GPU cluster roaring late at night, tokens flowing between servers, and a kilowatt-hour of electricity burning behind every piece of code.

Gartner says AI computing power is limited by power supply. Nadella said the chip might not plug in. DeepSeek said that calling during the day will cost extra. Zhejiang said that you need to pay extra for using electricity at night. Together, these voices point in the same direction: computing power is no longer a matter of "you can have as much as you want". It has costs, it has bottlenecks, it has peaks and troughs, it requires scheduling, it requires planning.

The "free lunch" of the AI ​​industry is turning into an electricity bill that requires careful calculation.