Deep Investment ResearchAugust 14, 202611 min read

DeepSeek Raises Peak Cache Pricing More Than 12x. Open Weights Did Not Kill Pricing Power—They Moved the Moat to Deployment.

DeepSeek will introduce peak and off-peak V4 pricing at 16:00 UTC on August 16. The sharpest question is why an official API can raise prices when anyone can download the weights. Open weights weaken model monopoly; they do not give customers reliable capacity, low latency or zero-operations deployment for free.

One Announcement, Three Increases and Two Time Zones

The Information's public AM page listed DeepSeek's V4 price hikes. DeepSeek's official page says the new peak/off-peak schedule starts August 16, not today.

Peak cache-hit input rises from $0.003625 to $0.044 per million tokens, uncached input from $0.435 to $1.32 and output from $0.87 to $3.96. Off-peak is half the new peak rate.

V4-Pro / 每百万Token	8月14日现价	8月16日起低谷	8月16日起高峰
缓存命中输入	$0.003625	$0.022	$0.044
缓存未命中输入	$0.435	$0.66	$1.32
输出	$0.87	$1.98	$3.96
相对现价	基准	约6.1x / 1.5x / 2.3x	约12.1x / 3.0x / 4.6x

Data view from this article

缓存命中输入0.004

缓存未命中输入0.435

输出0.87

Open Weights Do Not Mean Free Production

The MIT-licensed model has 1.6T total parameters, 49B active parameters and an approximately 865GB repository. A vLLM production recipe uses eight GB200 GPUs for the full model.

Customers can exit, but must still pay for hardware, networking, utilization, reliability and operations. Open weights cap pricing power; they do not eliminate it.

Cost Per Solved Task Matters More Than Token Price

CAISI found V4 about eight months behind the frontier. It was cheaper than GPT-5.4 mini on five of seven comparable benchmarks, but ranged from 53% cheaper to 41% more expensive.

Those results used earlier prices. After the increase, DeepSeek must justify the difference with success rate, latency or integration cost.

AI算力时钟在蓝色低谷与金色高峰之间切换，表现DeepSeek即将采用的峰谷Token定价

The Countercase: Load Management, Not Monopoly Rent

DeepSeek launched V4 around cost-effective million-token context. Peak/off-peak pricing can shift batch workloads away from constrained hours.

The bullish case is sustainable service economics; the bearish case is accelerated migration to Flash, self-hosting or rivals. Retention decides.

Three Master-Investor Tests

An open-model blueprint separated from a managed AI factory by the deployment moat of GPUs, networking and reliability

Buffett, Munger and Howard MarksTurn pricing power into evidence

Buffett: a price increase only proves a moat when customers stay.
Munger: invert the failure path—cheaper hosting, automatic routers and higher self-host utilization all lower exit cost.
Howard Marks: a great open model and a good price are different propositions.

Conclusion: A Public Stress Test of the Moat

DeepSeek is testing whether customers value managed deployment more than the right to leave. Watch peak/off-peak mix, Pro retention versus Flash, third-party serving prices and true self-host cost.