Written by｜Chen Cong

Produced by｜Shi Tianhao Observation

DeepSeek has adjusted its prices twice within a span of seven days.

The first adjustment took place on the evening of August 13. Alongside the announcement of the official version of V4 Pro, DeepSeek introduced new pricing. Starting at midnight on August 17, the output price for the flagship model V4-Pro during peak hours surged from 6 yuan to 27 yuan per million Tokens, marking a 350% increase. Meanwhile, the input price for cache hits rose from 0.025 yuan to 0.30 yuan, 12 times the original rate. Peak-valley pricing was also implemented, with peak hours defined as 9 AM to 12 PM and 2 PM to 6 PM daily, and all other times at half price.

The second adjustment occurred six days later. Starting at midnight on August 23, Saturdays and Sundays were no longer subject to peak-valley pricing, with a uniform low-price rate applied throughout the day.

Two adjustments in a single week—one price hike and one de facto price cut—may seem contradictory, but they serve a unified purpose. Computing power is in short supply during peak hours, and price adjustments are employed to shift time-insensitive demand to nights and weekends.

The market, highly sensitive to price changes, initially responded to the initial hike. On August 17, the day the new prices took effect, the A-share computing power sector experienced a surge, with T Time hitting a 20% daily limit.

The market perceived not just that Tokens had become "more expensive," but that computing power was "scarce."

However, on the first trading day following the second price adjustment, August 24, the market cooled down. By the close, the Shanghai Composite Index fell by 0.61%, the Shenzhen Component Index dropped by 2.11%, and the ChiNext Index tumbled by 3.14%. Computing hardware led the decline: Zhongji Xunchuang's A-shares closed down 7.44% at 872.88 yuan, with trading volume exceeding 30 billion yuan and net selling by major funds surpassing 3.9 billion yuan. Its Hong Kong shares fell over 12% to 1002 HKD, resulting in its combined A+H market cap falling below 1 trillion yuan. Tianfu Communication dropped over 11%, New Easycom over 9%, and T Time, which had hit a daily limit a week earlier, fell over 8%.

On the same day, Alibaba announced the placement of 710 million new shares at 112.70 HKD each, raising 80 billion HKD, all earmarked for full-stack AI capabilities and infrastructure. The subscription was oversubscribed within an hour and ultimately nearly tripled, with sovereign wealth funds and other long-term investors accounting for over 40% of the allocation. This marked Alibaba's first placement since its 2019 Hong Kong listing and the largest primary follow-on offering in Hong Kong's history.

On one hand, the secondary market was selling optical modules; on the other, long-term capital was queuing up to fund AI infrastructure.

These two opposing capital flows made choices on the same day, raising the question: what are they pricing?

I. Price Hikes as a Supply-Demand Report

To understand pricing, consider three sets of figures.

From August 3 to 9, the weekly call volume for DeepSeek-V4-Flash's official version on the OpenRouter leaderboard reached 8.83 trillion Tokens, up 570% week-over-week, ranking first globally and nearly 800 billion Tokens ahead of second-place Tencent Hy3. On August 1, it processed a record 8 trillion Tokens in a single day; just three days later, on August 4, its API began frequently returning "capacity insufficient" errors.

At the time, the market was unaware that four days later, DeepSeek would announce a price hike and that nine days later, the new prices would become official.

Nationally, the gap is structural.

According to the National Data Bureau, China's daily Token call volume surged from 100 billion in early 2024 to 100 trillion by the end of 2025 and 140 trillion by March 2026, a thousandfold increase in two years. Yet supply-side mismatches persist: roughly 80% of real-time inference demand is concentrated in the east, while about 80% of training and batch tasks are in the west. Data center delivery cycles have shrunk to around 100 days, but matching (supporting) power infrastructure takes two to three years.

In other words, the bottleneck isn't just chips—it's also electricity and land.

Peak-valley pricing emerged in this context. Its essence isn't promotion but scheduling. By offering half-price lows, batch tasks are shifted from peak daytime hours, reserving scarce peak periods for real-time demands willing to pay 27 yuan. All-day weekend lows extend this logic.

Demand is so intense that capacity must be sliced by the hour—a testament to scarcity.

II. Price Hikes as a Business Model Announcement

For the past two years, the industry has been engaged in a price war.

Why a price war? Its subtext is that the product isn't valuable enough to charge for upfront. The computing power chain was burdened by this logic for two years, with markets reflecting that smarter models and cheaper calls made upstream orders feel like one-time deals. The turning point emerged from pricing.

On August 9, Morgan Stanley released a report titled "Farewell to Price Wars, Hello Intelligence Wars." It analyzed official pricing from ByteDance, Alibaba, Baidu, Tencent, MiniMax, Zhipu, Yuezhi Dark Side, and DeepSeek: in Q2 2026, the average API input price for Chinese large models was 4.9 yuan per million Tokens, with output at 21.9 yuan, up 48% and 80%, respectively, from Q1 2025. The report concluded that model intelligence, not price, determines long-term competitiveness; low-price market share crushes margins, and thin margins can't fund next-gen model training.

DeepSeek's pricing rhythm has also been dissected by securities firms.

For example, Hualong Securities divided its actions this year into three steps: first a price cut, then peak-valley pricing, and finally an overall hike. The conclusion: the year-long low-price competition among domestic large models has officially ended, and pricing benchmarks are entering a correction phase. Indeed, DeepSeek isn't the only one raising prices this year. Zhipu has increased API prices three times, Tencent Cloud twice, and Alibaba Cloud and Baidu Intelligent Cloud have followed suit.

Interestingly, DeepSeek was the one that drove Token prices to rock-bottom two years ago, forcing the industry to follow; now, its 350% price hike tests demand elasticity for the entire sector.

When the initiator of a price war retreats, the impact is different. Businesses must see viable returns on computing power investments.

Stock markets buy expectations, so they reacted first.

III. Where Does the Money from Price Hikes Flow?

1. Up the Chain, Every Link Has Financial Statements to Capture It

In simple terms, the extra revenue collected by model companies becomes purchases of computing power, flowing from model vendors to cloud providers, then to servers, optical modules, chip foundries, and equipment.

Over the past two weeks, every link in this chain has released financial results.

Optical modules, closest to models, began delivering first. On the evening of August 23, Zhongji Xunchuang disclosed its Interim Report (semiannual report): revenue reached 41.778 billion yuan, up 182% YoY; net profit attributable to shareholders hit 13.651 billion yuan, up 242%. Management said on the earnings call that optical module orders, previously rolling on three-month cycles, now see many clients signing through 2027. Further upstream, semiconductor foundry HuaHong Hongli reported Q2 revenue of $717.5 million, a record high, with capacity utilization at 102.8%; management attributed 60% of this growth to price hikes and 40% to expansion. Equipment player AMEC reported H1 revenue of 6.691 billion yuan, up 34.89% YoY; net profit attributable to shareholders surged over 300% to 2.825 billion yuan.

The logical chain is in motion, and top-tier investments are accelerating. According to reports, TrendForce on August 7 raised its 2026 global AI server shipment growth forecast from 28% to nearly 31% and projected that the nine largest cloud vendors globally would increase capital expenditures by about 90% this year; Microsoft, Amazon, Alphabet, and Meta alone plan to spend $735–760 billion in 2026.

2. Deeper Changes in Monetization Models

Previously, computing power providers charged rent, billing by card and hour for fixed revenue.

On July 29, Xingyun Technology announced a new approach in an Announcement (announcement). Its subsidiary revised a five-year contract with a leading large model client (widely believed to be Yuezhi Dark Side) from 1.014 billion yuan to 3.053 billion yuan, expanding computing units from 128 to 256; monthly service fees for the eight already-delivered units rose 21.21%, with new units up 36.36%. Reportedly, the new terms are called Token revenue-linked fixed service fees.

Computing power providers no longer charge fixed rent but instead take a share of clients' future model call revenue—the first time A-shares have included Token revenue sharing in a major contract.

This shift is worth noting. Charging rent treats computing power as a cost; sharing revenue treats it as an asset. The same GPU represents two different businesses. After Token price hikes, revenue-sharing terms let computing power providers directly benefit from Token appreciation—the higher the client's selling price, the larger their share. CITIC Securities describes this as a shift from fixed monthly rents to usage-based billing tied to actual Token volumes; Galaxy Securities puts it more bluntly: from "selling resources" to "selling output."

Demand structures are also evolving to support this model. BOC International notes that inference demand is spreading from internet applications to office, finance, industrial, and government sectors, with computing power investment shifting from training-focused heavy assets to balanced training and inference. Training is a phase-specific heavy investment; inference is daily recurring revenue—the latter is the premise for revenue-sharing models. Servers, switches, optical modules, PCBs, and liquid cooling will all expand accordingly.

After Token price hikes, the industrial chain is revaluing itself.

IV. How Far Can This Rally Go?

The first constraint is demand elasticity.

For price hikes to hold, users must stay. Around the time DeepSeek announced its hikes, Alibaba open-sourced its 2.4-trillion-parameter flagship model Qwen3.8-Max—the first time a Max-tier flagship was open-sourced.

Opening the door to private deployment means every cent of premium from price hikes lowers substitution costs for competitors. To gauge whether hikes can stick, watch if call volumes drop, if low-period utilization rises, and if cache hit price differentials continue to narrow.

The second constraint is rising standards. On August 24, Zhongji Xunchuang's semiannual report showed revenue doubling, net profit up 242%, and orders signed through 2027—yet its stock still fell. Zheshang Securities' observation proved true: the market has shifted from buying grand narratives to buying evidence of delivery. Good financials aren't enough; firms must also address whether their technical routes can be bypassed. Zhongji Xunchuang's revenue is heavily overseas-dependent (94.8%). When news broke on August 4 that the U.S. FCC was drafting an import ban on 800G and 1.6T optical modules, its stock nearly hit a daily limit.

Technical routes and trade policies—any shift here could override orders signed through 2027 and trigger revaluation.

In gold rushes, shovel sellers profit first.

But that's only half true. Shovel prices ultimately depend on gold.

Tokens must be valuable for computing power to be valuable; computing power must be valuable for chips, servers, and optical modules to secure orders.

This week, DeepSeek's price hike put the industrial chain's account books to the test. It asks whether demand is real and whether what you hold today will still be needed tomorrow.

Standard answers don't exist yet.

But until they do, price hikes will serve as the industrial chain's thermometer—measuring both demand heat and route viability.