DeepSeek’s 75% price cut and what it means for AI valuations

AI is getting dramatically cheaper, and China’s DeepSeek just pushed that trend into overdrive. The company has made a massive 75% price cut on its flagship V4 Pro model permanent, turning high-end AI into something that costs fractions of a cent per million tokens. That’s great news for users—but it raises serious questions about how today’s AI giants can justify their sky-high valuations.

What DeepSeek just announced

DeepSeek, a fast-rising Chinese AI startup, offers large language models you can access via API, similar to OpenAI, Anthropic, and other providers. It has already been competing aggressively on price, but its latest move goes much further.

The company is making a 75% price cut on its V4 Pro model permanent. In practical terms, that means its API prices will stay at just a quarter of their original level instead of being a temporary promotion.

According to the company, V4 Pro API costs now range roughly between $0.00025 and $0.61 per million tokens for different usage types, which works out to about $0.00035 to $0.83 per million tokens after all tiers are considered. Previously, prices were in the $0.01 to $2.41 per million token range.

How cheap is a million tokens, really?

To understand how extreme this is, it helps to know what a “token” is. A token is a small chunk of text—on average, about 3–4 characters in English. Roughly speaking, 1,000 tokens is about 750 words.

That means 1 million tokens is enough to process something like the entire Lord of the Rings trilogy plus a couple of extra books. DeepSeek is effectively saying: you can run that much text through some of its models for around a third of a penny.

Even at the high end of its new pricing—around $0.83 per million tokens—that’s still incredibly cheap for serious AI workloads. The lowest tiers will typically apply to smaller, faster models (similar to “flash” or “nano” models from other providers), while the more capable, frontier-style models sit closer to the top of the range.

How this compares to OpenAI and other U.S. players

DeepSeek isn’t the only one driving prices down. OpenAI and other U.S. providers have also been cutting costs while improving model quality. For some OpenAI models, input tokens are now around $0.40 per million tokens.

At that rate, you can feed in book-length content for well under a dollar. For most everyday use cases—short prompts, emails, summaries, coding help—the cost per request is just a tiny fraction of a cent.

DeepSeek’s latest move undercuts even those already low prices, especially at the smaller-model end of its lineup. This intensifies a race to the bottom on token pricing, where providers compete on both quality and cost per million tokens.

China’s hardware advantage: Huawei and local chips

One open question is why DeepSeek can afford to slash prices this aggressively. The company did not explicitly say whether the permanent price cut is tied to better access to Huawei’s Ascend 950 chips, which it uses to run V4 Pro efficiently.

Because of U.S. export controls, Chinese AI companies have limited access to Nvidia’s top-end GPUs. In response, domestic players like Huawei have been building their own AI accelerators and scaling them across Chinese data centers.

As more of this local hardware comes online, Chinese AI providers can train and serve models at lower cost, which makes deep price cuts like this more feasible. It also strengthens China’s position in the global AI race, as companies like DeepSeek can move faster, iterate more, and offer cheaper access than many Western rivals.

For more background on DeepSeek’s strategy and pricing philosophy, see our earlier breakdown: DeepSeek’s $10 billion valuation and rational AI pricing.

The economics problem: can cheap tokens support trillion-dollar valuations?

All of this leads to a bigger question: how do AI companies get to trillion-dollar valuations when their core product—tokens—is rapidly approaching near-zero cost?

If a provider charges $0.40 per million tokens, they need an astonishing volume of usage to generate the kind of revenue that would justify a trillion-dollar market cap. At $0.0003 per million tokens, the math becomes even more extreme. You would need almost unimaginable token throughput to make the numbers work.

Public market investors typically expect that a company’s valuation at IPO is the lowest it will ever be, with growth and profits ahead. But if token prices keep falling while competition increases, it becomes harder to see how pure API usage alone can support those expectations.

Why AI agents became part of the business story

One way U.S. companies have tried to square this circle is by betting on AI agents. Instead of users occasionally sending prompts, the vision is that autonomous agents will run constantly in the background—researching, coding, scheduling, analyzing data—and in the process, burn through huge numbers of tokens.

In theory, this always-on usage could dramatically increase token consumption, and therefore revenue, even if the price per token is low.

In practice, many current agent systems are highly inefficient. When compute feels “almost free,” developers tend to build workflows that call models repeatedly, chain them together, and run complex loops. That can spike token usage, but it doesn’t always translate into real business value.

We’ve already seen examples of this backfiring. Some providers have had to clamp down on certain agent frameworks and “harnesses” because they were burning so many tokens that they threatened the provider’s own margins.

Why a race to the bottom on tokens is risky

As more companies join the price war, token costs may continue to plummet. That’s great for developers, startups, and enterprises experimenting with AI. But it also creates several risks:

1. Margin compression. If everyone competes primarily on price, profit margins on raw API calls shrink. That makes it harder to fund the massive compute, research, and talent needed to stay at the frontier.

2. Overreliance on volume fantasies. Business models start to rely on speculative scenarios where agents or automated systems consume unimaginable numbers of tokens. If real-world adoption is more measured and efficient, projected revenues won’t materialize.

3. Bubble-style valuations. When valuations assume both falling unit prices and explosive volume growth, the math can quietly stop making sense. As the transcript put it: you can’t make “two plus two equal a trillion.”

Where the real value in AI might come from

All of this suggests that the long-term value in AI may not come from selling raw tokens alone. Instead, the most durable business models are likely to come from:

Integrated products. Tools that bundle AI into workflows—CRM systems, productivity suites, vertical SaaS—can charge for outcomes (better sales, faster support, fewer errors) rather than just tokens.

Domain-specific solutions. Healthcare, finance, law, and other regulated industries will pay for reliability, compliance, and integration, not just cheap text generation.

Proprietary data and distribution. Companies with unique data, strong brands, or deep customer relationships can build AI features that are hard to commoditize, even if the underlying tokens are cheap.

These dynamics also feed into the broader geopolitical competition over AI leadership, as explored in our look at Elon Musk, OpenAI, and the global AI power struggle.

What this means for developers and businesses today

For now, the takeaway is simple: building with AI is getting cheaper, fast. Whether you’re a solo developer or an enterprise team, you can process huge amounts of text for pennies, especially with providers like DeepSeek pushing prices down.

That creates a window of opportunity to experiment aggressively, prototype new products, and integrate AI into existing workflows without worrying too much about per-request costs.

At the same time, it’s worth being skeptical of business models that assume infinite token consumption or rely entirely on raw API resale. As token prices race toward the bottom, the real winners will likely be those who turn cheap AI into valuable, efficient, and defensible products.