In April, when DeepSeek launched its V4 series models, the company lit a fuse under the global AI industry. Its pricing was already aggressive: V4-Pro started at 12 yuan per million input tokens, while V4-Flash undercut nearly every Western competitor. But there was a catch: the discounts were temporary, set to expire at the end of May.

On May 23, DeepSeek removed the catch. Permanently.

The company announced that the 75 percent discount on its flagship V4-Pro model would become permanent, keeping prices at a quarter of their original level. API costs now range from just 0.025 to 6 yuan per million tokens—approximately 0.0035 to 0.83—depending on usage type.

This is not a limited-time promotion. This is a structural shift. DeepSeek is betting that it can sustain these prices indefinitely, and in doing so, it is forcing every other AI model provider — OpenAI, Anthropic, Google, and the rest — to either follow suit or cede the cost-sensitive segment of the market.

This article explains why DeepSeek can afford to slash prices permanently, what it means for the AI industry’s profit margins, and why this price war may be the most consequential development in enterprise AI since the launch of ChatGPT.

The Numbers — How Cheap Is DeepSeek, Really?

Before the permanent cut, DeepSeek’s V4-Pro pricing was already disruptive. The new permanent rates are even more aggressive.

The New Pricing Structure

In dollar terms, the new rates range from 0.0035 to 0.87 per million tokens, depending on usage type. The lowest tier—cache input at 0.025 yuan per million tokens— is effectively free for any organization processing repetitive queries.

How This Compares to Western Models

The gap between DeepSeek and its Western rivals is now measured in orders of magnitude:

For developers comparing DeepSeek V4 Flash—the lighter, faster version—the disparity is even sharper. DeepSeek V4 Flash costs 0.18 per million input tokens and 0.25 per million output tokens, making it up to 40 times cheaper than GPT-4o.

At these prices, the economic case for switching to DeepSeek — for any non-regulated, non-US enterprise — becomes overwhelming.

How DeepSeek Can Afford This — The Huawei Factor

The most obvious question from Western observers is: how can DeepSeek sustain prices this low without bleeding cash?

The answer lies not in DeepSeek’s software, but in China’s hardware — specifically, Huawei’s Ascend 950 AI chips.

From Nvidia Dependency to Domestic Supply

When DeepSeek launched V4 in April 2026, it explicitly stated that the Pro version’s higher pricing was due to “constraints in high-end compute capacity,” limiting availability. At the time, it also noted that Pro pricing was expected to fall sharply once Huawei Ascend 950 supernodes were launched in large quantities in the second half of the year.

DeepSeek did not explicitly attribute the permanent price cut to increased supply of Ascend 950 chips. But the timeline is unmistakable. The company’s ability to lock in permanent 75 percent discounts aligns almost perfectly with the expected ramp-up of Huawei’s domestic AI silicon.

The US Export Control Paradox

Ironically, Huawei’s Ascend chip business has benefited directly from the US export controls that were designed to cripple China’s AI ambitions. Restrictions prevent Nvidia from selling its most advanced semiconductors in China, creating a captive market for domestic alternatives. While separate curbs on chipmaking equipment exports have limited Huawei’s ability to scale up Ascend production, the company is still ramping capacity faster than many analysts predicted.

Architectural Efficiency

DeepSeek has also optimized its V4 architecture to run efficiently on Huawei’s chips. The company confirmed at launch that V4’s fine‑grained expert parallelism scheme had been validated on both Nvidia GPUs and Huawei Ascend NPUs. This cross-platform compatibility allows DeepSeek to source compute from whichever supply chain is most cost-effective — and right now, that is domestic Chinese hardware.

The result is a cost structure that Western providers, locked into Nvidia’s premium GPU ecosystem and facing higher labor and energy costs, simply cannot match.

Performance — Is 75% Cheaper Also 75% Worse?

Price cuts mean nothing if the underlying model is unusable. DeepSeek’s performance has been the subject of intense scrutiny, and the picture is more nuanced than either detractors or boosters admit.

The CAISI Evaluation

A recent evaluation by the US National Institute of Standards and Technology’s Center for AI Standards and Innovation (CAISI) tested DeepSeek V4 Pro against leading American models. The report concluded that V4 Pro is “the most powerful Chinese AI model” CAISI has evaluated to date. However, it also found that V4 Pro’s capabilities lag top US frontier models by approximately eight months, with performance roughly equivalent to GPT-5 from the previous generation.

On CAISI’s scoring system, DeepSeek V4 Pro achieved around 800 (±28), which is very close to GPT-5.4 Mini at 749. The model was tested across cybersecurity, software engineering, natural sciences, abstract reasoning, and mathematics — all critical domains for enterprise AI deployment.

The Cost-Performance Tradeoff

The CAISI evaluation also noted that V4 Pro’s cost advantage is significantly better than its US counterparts. This is the heart of DeepSeek’s strategy: it does not need to be the smartest model in the room. It only needs to be smart enough at a fraction of the price.

For the vast majority of enterprise tasks — document summarization, customer support, data extraction, content generation — the eight-month capability gap is irrelevant. At 75 percent of the cost, V4 Pro becomes the default choice for price-sensitive developers.

The CAISI Debate

Not everyone accepts CAISI’s findings at face value. Industry observers have questioned whether a single overseas evaluation can definitively define the capabilities of Chinese AI models, pointing to methodological differences and the rapid pace of model updates. Nonetheless, even critics acknowledge DeepSeek’s commanding cost advantage.

The Industry Response — Will Others Follow?

DeepSeek’s permanent price cut puts Western AI providers in an uncomfortable position.

The Margin Squeeze

OpenAI, Anthropic, and Google have built their business models on premium pricing. GPT-4o output costs 10 per million tokens; 10 per million tokens; Claude Opus 4.5 costs Claude Opus 4.5 costs 15. If DeepSeek can offer comparable utility at a fraction of that price, the entire value proposition of Western models comes into question — especially for non‑US enterprises that are not subject to export controls or data sovereignty concerns.

The Response So Far

As of May 2026, none of the major Western providers have announced matching price cuts. Some have taken a different tack: a Stanford University report published in late 2025 found that Chinese AI models “seem to have caught up or even pulled ahead of their global counterparts” in certain benchmarks. Rather than compete on price, US firms may lean into their perceived advantages in safety, reliability, and integration with existing enterprise software stacks.

However, history suggests that when one player in a commodity market slashes prices permanently, the others are forced to respond. The question is not if the price war will spread, but how quickly and how deeply.

The Two-Tier Market

Some analysts predict that the AI model market is already bifurcating into two tiers: a high‑margin, high‑performance tier for enterprises willing to pay a premium for the absolute best models, and a low‑cost, high‑volume tier for price‑sensitive applications where “good enough” is sufficient.

DeepSeek is aggressively claiming the low‑cost tier. Whether Western providers can hold the line in the premium tier — or whether they will be forced to compete on both fronts — remains to be seen.

Why This Price War Is Different — Structural, Not Promotional

Not all price wars are created equal. Most are temporary promotions designed to grab market share, after which prices return to normal. DeepSeek’s permanent cut is different.

The Supply Chain Advantage

DeepSeek’s cost advantage is not based on promotional subsidies — it is based on a fundamentally cheaper supply chain. By running on Huawei’s Ascend chips rather than Nvidia’s premium GPUs, DeepSeek’s cost per inference is structurally lower. This is not a discount that can be matched by simply slashing prices; it requires a complete re‑engineering of the underlying hardware stack.

The Scale Advantage

China’s domestic AI market is enormous. As Chinese enterprises shift from experimentation to production deployment, the volume of inference tokens processed on domestic hardware will explode. That volume creates a virtuous cycle: more usage → more data for optimization → lower unit costs → more usage. Western providers, locked out of the Chinese market by export controls, cannot access this scale advantage.

The Geopolitical Dimension

The US export controls that were supposed to cripple China’s AI progress have, paradoxically, given Chinese firms like DeepSeek a competitive advantage in cost. By cutting off access to Nvidia’s premium chips, Washington forced Chinese AI labs to optimize for efficiency rather than raw power. That efficiency is now paying dividends in the form of dramatically lower operating costs.

Some analysts have gone further, noting that DeepSeek’s aggressive pricing “directly swings the knife at the profit margins of the entire generative AI market”. Whether Western providers can respond without gutting their own profitability is an open question.

What This Means for Enterprises and Developers

For organizations that have been priced out of the AI revolution — or that have been forced to ration usage due to high costs — DeepSeek’s price cut is a game-changer.

Instant Cost Savings

For any enterprise running inference workloads at scale, switching to DeepSeek could reduce cloud AI spending by 90 percent or more. A company spending 100,000 per month on GPT−4o API calls could cut that to approximately 8,700 using DeepSeek V4 Pro, or even less using the Flash variant.

The Caching Advantage

DeepSeek’s aggressive caching discounts — input tokens with cache hit cost just 0.025 yuan per million tokens, or $0.0035 — are particularly attractive for applications that process repetitive queries, such as customer support chatbots, document processing pipelines, or internal knowledge bases. At these prices, the marginal cost of each additional query approaches zero.

The Compliance Caveat

Not every organization can switch to a Chinese AI model. US government contractors, financial institutions subject to strict data sovereignty rules, and companies handling sensitive customer data may be unable to use DeepSeek for regulatory or compliance reasons. For these organizations, Western providers will retain pricing power.

The Developer Opportunity

For independent developers and startups, DeepSeek’s permanent price cut lowers the barrier to entry for AI‑powered applications. Small teams can now build and deploy AI features that would have been prohibitively expensive just months ago. The long tail of AI development — niche applications, experimental projects, and creative tools — will be the primary beneficiary of this price war.

Frequently Asked Questions (FAQ)

Q1: Is the 75% discount really permanent?
A: Yes. DeepSeek announced on May 23, 2026, that the 75 percent discount on its V4-Pro model will be permanent, keeping prices at a quarter of their original level. The discount was originally set to expire at the end of May.

Q2: How much does DeepSeek V4-Pro actually cost?
A: API costs range from 0.025 to 6 yuan per million tokens (approximately $0.0035to$0.0035to0.83), depending on usage type. Input with cache hit is cheapest at 0.025 yuan, output is 6 yuan, and input with cache miss is 3 yuan.

Q3: How does DeepSeek compare to GPT-4o in quality?
A: The US government’s CAISI evaluation found DeepSeek V4 Pro lags top US frontier models by about eight months, performing roughly at the level of GPT-5 from the previous generation. However, the evaluation also noted that V4 Pro is “the most powerful Chinese AI model” tested to date.

Q4: Why is DeepSeek so much cheaper than Western models?
A: DeepSeek runs on Huawei’s Ascend 950 AI chips, which are cheaper than Nvidia’s premium GPUs. The company also benefits from a lower-cost domestic supply chain and architectural optimizations that improve efficiency.

Q5: Can I use DeepSeek V4-Pro from outside China?
A: Yes. DeepSeek offers API access to developers outside China, though pricing and availability may vary by region. The company has a global API endpoint for international customers.

Q6: Will OpenAI and Anthropic lower their prices in response?
A: As of May 2026, no major Western provider has announced matching price cuts. Some analysts predict that the AI model market will bifurcate into a premium tier (Western models) and a value tier (DeepSeek and others), rather than a straight price war.

Q7: Is the performance gap worth the price savings?
A: It depends on your use case. For tasks that require the absolute highest performance — complex reasoning, advanced coding, cutting-edge research — Western models may still be worth the premium. For the vast majority of enterprise applications (summarization, extraction, customer support), DeepSeek is likely more than sufficient at a fraction of the cost.

Q8: How does this connect to your earlier coverage of US export controls?
A: Directly. US restrictions on Nvidia chip sales to China created a captive market for Huawei’s Ascend chips. DeepSeek optimized V4 to run efficiently on those chips, and the resulting cost savings are now being passed to customers. The policy intended to slow China’s AI progress may have inadvertently accelerated its cost advantage.

Conclusion — The New Normal

DeepSeek’s permanent 75 percent price cut is not a temporary promotion. It is a structural shift in the economics of AI inference, driven by a fundamentally cheaper supply chain, architectural efficiency, and a domestic market that Western providers cannot access.

For developers and enterprises, the implications are enormous. AI-powered applications that were once prohibitively expensive are now within reach. The long tail of AI development — niche tools, experimental projects, and creative applications — will flourish in this low-cost environment.

For Western providers, the challenge is equally significant. They can no longer ignore DeepSeek as a fringe competitor. The price war has entered a new phase, and it might not stop until the entire industry has been reshaped.

References & Further Reading

• Reuters – “China’s DeepSeek to make permanent 75% price cut on flagship V4‑Pro AI model” (May 23, 2026)
• CAISI (US NIST) – DeepSeek V4 evaluation report (May 2026)
• CoinMarketCap – “DeepSeek makes 75% price cut permanent on V4-Pro as AI price war enters new phase” (May 24, 2026)
• Engadget – “DeepSeek permanently reduces the price of its flagship V4 model by 75 percent” (May 23, 2026)
• CNBC – “AI’s widening pricing divide, plus Big Tech earnings” (April 30, 2026)
• Stanford University – AI Index Report (late 2025)

The post The AI Price War Just Entered a New Phase — And It Might Not Stop appeared first on Explain This Tech.