API价格降到打印纸的十分之一：DeepSeek连续降价背后的算盘 (English)

Generated: 2026-06-22 05:33:19

---

I Almost Couldn't Sleep Last Night—DeepSeek Has Truly Lost Its Mind

At 2 AM, I was curled up on the sofa writing code with Cursor when my phone suddenly buzzed. I glanced at it—a push notification from DeepSeek: V4 Pro API permanently reduced by 75%, cache hits at just 0.025 yuan per million tokens.

I froze for a second, then instinctively opened my API billing page.

Last month: over 3,800 yuan.

Guess what my first reaction was? Not joy. A little panic. Seriously, panic.

---

It All Started in May 2024

Back then, DeepSeek V2 had just launched, slashing API prices to 1 yuan per million input tokens and 2 yuan per million output tokens. The whole community exploded—some called it a loss-leader stunt, others said Liang Wenfeng had lost his mind.

Me? I was still using Claude at the time, spending about $300 a month. When I saw DeepSeek's prices, my first thought was the same as yours: It's this cheap—can it even work?

I gave it a try. V2 wasn't as good as Claude at code generation, but it was genuinely impressive at math reasoning and logic puzzles. And the key thing—it was cheap! For the same tasks, the cost was only a tenth of Claude's.

Then what? In September 2025, V3.2-Exp dropped prices again, cutting output costs by 75%. By then I'd already started migrating some of my workflows to DeepSeek. Honestly, V3.2's coding ability was already competitive—though it occasionally wrote some weird logic, but hey, it was cheap enough that you could just tweak it a few times.

This April, the V4 series launched, with prices slashed to a tenth of the original launch price. I posted on my Moments: "DeepSeek is about to strip every competitor bare."

Now, V4 Pro has gone permanently discounted, turning a limited-time offer into the new normal. Cache hits at 0.025 yuan per million tokens—what does that mean? Let me do the math for you: writing a moderately complex Python function consumes about 2,000 tokens, costing less than 0.00005 yuan. That's cheaper than printing a piece of paper!

---

What's Liang Wenfeng Really Up To?

Some people call him a cyberspace bodhisattva. I don't see it that way.

Look at his stance during fundraising: personally investing 20 billion yuan, accounting for 40% of total funding. That money didn't fall from the sky. He was very clear with investors: the goal is AGI, not profit.

Sounds idealistic, right? But if you look closely at his business model, the logic chain is crystal clear.

First, DeepSeek's costs are genuinely dropping.

They've developed KV Cache compression technology that allows model inference to run on cheap storage like SSDs, NAND flash, and LPDDR memory. Before, you had to buy HBM—now you don't. How severe is the HBM shortage? GB200s ordered in November 2025 won't ship until Q2 2026—that's longer than the longest delivery cycle for a Tesla Model Y!

Second, price cuts are essentially about capturing the ecosystem.

Think about it: when an API is so cheap that trial and error costs just a few dozen yuan, how will companies react? They'll try it first. Once they try it, they'll adapt their systems and invest resources. When enough developers and enterprises build around DeepSeek's technology stack, DeepSeek stops being just a foundation model company—it becomes an AI infrastructure company.

Third, this move hurts competitors.

Many rivals still rely on external funding. On the revenue side, they're being squeezed by DeepSeek's low-price, high-volume strategy—ARR growth slows, margins get thinner, but burn rate doesn't drop. Investors start reassessing. Capital markets were already losing patience with AI.

---

The Funny Thing Is, AI Prices Are Rising Everywhere Else

Microsoft canceled its internal Claude Code license because token costs were too high. Uber's CTO reported in April this year that their entire 2026 AI budget was spent within four months. 95% of engineers use AI coding tools monthly, and 70% of committed code is AI-generated.

The overseas big three—OpenAI, Google, Anthropic—are all raising prices. GPT-5.5 long-context version costs 374 yuan per million tokens; Claude Opus 4.7 costs 204 yuan. Meanwhile, DeepSeek V4 Pro cache hits cost just 0.025 yuan.

The gap? 15,000 times.

I've tested it: on code generation tasks, V4 Pro is within about 10% of GPT-5.5's performance, but at 1/15,000th the price. What do you think companies will choose?

---

What Fascinates Me Most Is the Opportunity for Domestic Chips

Before, if you wanted to enter the AI game, you had to spend millions on NVIDIA GPUs. Now DeepSeek has made models more memory-efficient, better at long contexts, and easier to split across different hardware—effectively creating interfaces for domestic GPU, storage, and server manufacturers.

Huawei's Ascend 950 already uses a dual-mode architecture combining SIMD and SIMT, and code generated by TileLang can migrate seamlessly from the 910 to the 950. DeepSeek has been using TileLang extensively for over a year, iterating in real-world scenarios from training to inference.

This isn't a lab demo—it's industrial-grade validation.

Previously, companies like NVIDIA, SK Hynix, Samsung, and Micron saw revenue surges because all AI hardware pressure was concentrated on GPUs and HBM. Now DeepSeek has brought domestic manufacturers into the game—they can participate in every link: inference, caching, storage, scheduling. The model no longer serves only the most powerful hardware, and hardware doesn't have to be the most powerful to be valuable.

---

But How Long Can This Last?

I'm a little worried.

DeepSeek raised 70 billion yuan, with Liang Wenfeng personally contributing 20 billion. That money will last a while, but AI costs are skyrocketing. The compute needed to train the next generation of models grows exponentially, and electricity costs are rising too.

CATL has entered the game; JD.com and NetEase have joined as well. But at the end of the day, this is still a money-burning game.

But look at it from another angle—if DeepSeek can truly build an ecosystem through price cuts, getting enough companies to adapt around its technology stack, then it gains pricing power. At that point, making money becomes a natural outcome.

Liang Wenfeng says the goal is AGI, not profit. But AGI needs money—a lot of it. So cutting prices isn't about losing money; it's about growing the ecosystem at the lowest cost, and using that ecosystem to fund R&D.

That's a ruthless strategy.

---

What I'm Doing Now

For daily development, I use DeepSeek V4 Pro. For complex problems, I fall back on GPT-5.5. My monthly cost has dropped from over 3,000 yuan to less than 500. Honestly, the experience isn't much different.

This might be exactly what DeepSeek wants—to get you using it until you can't do without it.

In the end, you'll realize that the most expensive thing isn't compute—it's missing the ticket to an era.