How Did DeepSeek Grow 10X in One Year?
How Did DeepSeek Grow 10X in One Year?
The big story of the last few years in the world of artificial intelligence has been told mostly through the same companies: OpenAI, Google, Anthropic, Meta and Microsoft. Larger models, larger data centers, multibillion-dollar GPU investments and ever-increasing compute budgets have become almost the industry's constant formula.
Then DeepSeek came along and showed that this formula might not be the only valid way.
The company, based in Hangzhou, China, first attracted worldwide attention with its V3 and R1 models. Then, in 2026, it moved to another level not only technically but also commercially with the V4 family. DeepSeek's revenue in the first seven months of 2026 reached approximately 475 million yuan, or 70.7 million dollars. An important correction needs to be made here: This figure is approximately 10 times the revenue of DeepSeek in the whole of 2025, not the same period last year. Therefore, the company's growth rate is more remarkable than even the news headline implies.
But what makes DeepSeek interesting is not just its revenue growth. The company announced a net loss of 715 million yuan in the same period. So it's not profitable yet. However, there is a much more important sign behind the numbers: DeepSeek's total gross profit margin is 44.6 percent, and the gross profit margin in the API business alone has reached 82.9 percent. This is a remarkable rate considering how expensive it is to run models in the AI industry.
Therefore, the story of DeepSeek is much bigger than the headline “A rival to OpenAI emerged from China”.
The real story might be this:
The next phase of the artificial intelligence race will be determined not by who builds the biggest model, but by who can deliver intelligence at the lowest cost.
Why Is DeepSeek Growing So Fast?
An AI company's revenue model may seem complex at first glance, but on DeepSeek's side, the equation is pretty straightforward. The company doesn't just want people to use chatbots from its web or mobile app. The real big economic potential lies in developers and companies embedding DeepSeek models into their own products.
A software company may use DeepSeek in its customer support bot. A coding platform can generate code from the model. An agency can connect it to an automated content system. A SaaS company can have documents analyzed. A business can build its own AI agent systems on the DeepSeek API.
Every query, every token and every agent transaction means potential revenue for DeepSeek.
For this reason, the success of V4 Flash should not be evaluated solely on the basis of being a "good chatbot". The model is designed to go directly into developers' infrastructure.
DeepSeek's official API infrastructure supports both OpenAI format and Anthropic format. The practical meaning of this is significant: many applications already developed for other major model providers become much easier to migrate to DeepSeek.
This may seem like a technicality.
It's actually an extremely powerful growth strategy.
Because instead of telling the customer "rewrite your entire system", you say "change the model and try it".
Why is DeepSeek V4 Flash So Important?
At the center of DeepSeek's rise in 2026 is DeepSeek V4 Flash.
The V4 family is divided into two basic models: V4 Pro, which has higher capacity, and V4 Flash, which focuses on speed and cost efficiency. According to DeepSeek's own technical description, although V4 Flash has a total of 284 billion parameters, it uses approximately 13 billion active parameters in each process. V4 Pro, on the other hand, has 1.6 trillion total parameters and 49 billion active parameters.
The "active parameter" issue here is important to understand the economic advantage of DeepSeek.
A model may be huge, but the computational cost can be significantly reduced if the entire model does not need to run simultaneously for each query. DeepSeek's approach proceeds through more efficient architectures where only the relevant sections are activated when necessary.
The result?
Less calculation.
Less memory usage.
Lower inference cost.
Cheaper API.
And potentially higher profit margins.
DeepSeek's API gross margin, which reached 82.9 percent, shows that this technical architecture is starting to yield results not only on the benchmark chart but also on the balance sheet.
1 Million Tokens Now Standard
One of the remarkable features of the V4 family is that it offers a 1 million token context window.
Simply put, the context window determines how much information the model can “hold” at the same time. As the context grows, the model can evaluate much larger documents, long conversations or large code bases in a single session.
A novel.
A large archive of contracts.
Hundreds of pages of research.
A large software project.
Very long user history.
Context capacity is critical for all of these.
With V4, DeepSeek claims to standardize the 1 million token capacity in all official services, instead of being a premium feature. Behind this are architectural improvements such as DeepSeek Sparse Attention to reduce computation and memory costs.
This nicely summarizes DeepSeek's overall strategy:
Making expensive features cheaper instead of selling a more expensive AI.
Prices Reveal DeepSeek's Real Weapon
In the AI industry, it is easy to compare a model through a benchmark. In the real commercial world, another question is much more important:
“How much will it cost me to use this a million times?”
DeepSeek is being aggressive here.
As of August 2026, in official API pricing, 1 million cache-miss input tokens for V4 Flash are approximately $0.14, and 1 million output tokens are approximately $0.28. On the V4 Pro side, the same values are approximately $ 0.435 and $ 0.87, respectively. In case of a cache hit, the entry cost goes down even further.
These numbers may not have much meaning for the ordinary user.
But it is for a software company with millions of users.
Because as the artificial intelligence product scales, small token price differences can turn into cost differences of hundreds of thousands or even millions of dollars.
Therefore, explaining DeepSeek's growth solely by model quality would be incomplete.
The price/performance ratio is at the heart of the company's brand positioning.
New Competition in Artificial Intelligence: “Intelligence per Dollar”
For a long time, the most important benchmark in the AI world was “which model is smarter?” happened.
However, as the sector turns into a commercial structure, a different benchmark becomes important:
How much intelligence do I get for a dollar?
One model may be 5 percent better than another.
But what if it's 10 times more expensive?
For a company, the answer may not always be to use the model with the highest benchmark score.
The frontier model may not be required to print emails to the user.
A giant model may not be required for a simple coding task.
For millions of routine queries in the customer support system, a smaller but faster model may make much more economic sense.
DeepSeek V4 Flash fits exactly into this market.
That's why the name "Flash" doesn't just mean speed.
It describes a product strategy.
Coding and AI Agents are the New Growth Engine for DeepSeek
In the current version of V4 Flash, DeepSeek emphasizes that agentic capabilities, that is, AI agent capabilities, have been improved.
Using terminal.
Writing code.
Making changes in the repository.
Calling a ride.
Following a task through several steps.
Acting in web or software environments.
Next-generation AI systems no longer just produce answers; doing business.
In the V4 Flash update released by DeepSeek at the end of July, serious performance improvements were announced in Terminal Bench, NL2Repo, CyberGym and different coding-agent benchmarks. The model is also offered with Responses API support and optimizations for the use of Codex type agents.
The economic opportunity for the company here is extraordinary.
A user can ask the chatbot 10 questions per day.
But an AI coding agent can consume thousands or even tens of thousands of tokens for a single task.
As agents run longer, the amount of model usage can increase dramatically.
That's why in the future, one of the big revenues for AI companies may come not from chat, but from digital workers.
It is no coincidence that DeepSeek has aggressively prepared the V4 family for agent use.
Why Is DeepSeek Still Making a Loss?
At first glance, it may seem negative that a company earning 475 million yuan in revenue incurred a loss of 715 million yuan.
But this is not surprising in the current economic structure of the AI industry.
Model training is expensive.
GPU infrastructure is expensive.
Researchers are expensive.
Data centers are expensive.
Developing new models is expensive.
As DeepSeek grows, it invests more in infrastructure, and the company prefers capacity expansion over short-term profit.
What is more interesting is that the economy of generating income amidst the damage is improving.
The fact that the company's total gross margin in the first seven months of 2026 reached 44.6 percent and 82.9 percent on the API side indicates that the basic product economics of the business are strengthening.
In other words, it's not like DeepSeek can't make money.
He spends the money he earns on growth.
These two situations are not the same thing financially.
New Investment Round of 50 Billion Yuan
It is enough to look at the investment side to understand the extent of DeepSeek's growth plans.
Current news indicates that the company is preparing to raise approximately 50 billion yuan in new capital, or roughly $7.4-8 billion. The targeted valuation is approximately 500 billion yuan, or approximately 74 billion dollars. Reuters also confirmed that the company's second investment round is on the agenda again and targets a valuation of approximately 500 billion yuan.
This figure now moves DeepSeek out of the “promising Chinese AI startup” category.
The $74 billion valuation is at the level of technology companies that have reached significant scale worldwide.
Moreover, there are reports that the company is working on a potential Shanghai IPO.
DeepSeek's story is changing from a startup story to an infrastructure company story.
Why Did the World Start Talking About DeepSeek One Year Ago?
To understand today's V4 success, it is necessary to go back a little.
DeepSeek's recognition on a global scale started especially with the V3 and R1 models. The company questioned the perception created by large American AI laboratories that “huge budgets are mandatory for frontier AI”.
R1's reasoning abilities, V3's performance, and the company's clearer presentation of its models attracted serious attention in Silicon Valley.
Suddenly, DeepSeek was not just a company developing new models.
It was a question mark:
If it is possible to produce similar performance at lower cost, are the economic assumptions of the AI industry correct?
Today, with V4, the company takes the same question further.
This time, it is trying to reduce not only the research cost but also the usage cost.
Being “Cheap” Is Not a Weakness for DeepSeek
Cheapness in technology brands can sometimes be perceived as a quality problem.
On the contrary, DeepSeek turns its cost advantage into engineering superiority.
The message is not:
“We are cheaper because we are a smaller company.”
The message is:
“We are cheaper because we design the system more efficiently.”
This is very different in terms of brand perception.
The company's emphasis on technical publications, open weights, architectural descriptions, and developer community also supports this idea.
DeepSeek does not pretend to be a lifestyle AI brand.
It acts like an engineering brand.
This is perhaps one of the reasons why the company has built a strong brand value among developers despite not being very visually flashy.
Open Models Are Part of DeepSeek's Marketing Strategy
DeepSeek's release of the V4 Preview family as open-weight is not only for academic benefit.
Openness is also a distribution strategy.
Model goes into Hugging Face.
Researchers are testing.
Developers are experimenting.
Community integration is improving.
Companies evaluate it in their own infrastructure.
Then, it is possible to pass a certain part of it to DeepSeek's API service.
In other words, the open model is also a giant product demo.
This approach is very similar to open-source growth strategies in the software world.
Bring the product as close to the developer as possible.
Reduce the usage barrier.
Build community.
Then move the user that requires scale to the commercial infrastructure.
Considering that DeepSeek does not create its international impact solely with its advertising budget, the importance of this strategy becomes even more evident.
The Real Difference Between DeepSeek and OpenAI and Anthropic
The comparison here is only "which model is better?" It is not right to do it through.
Companies' strategies are different.
OpenAI is increasingly creating a vast ecosystem of consumer and enterprise products. ChatGPT is a huge user platform on its own.
Anthropic positions Claude strongly, especially in the security, corporate use and coding side.
DeepSeek, on the other hand, is growing aggressively through its cost efficiency, open model approach and developer infrastructure.
There is still a significant difference between them in terms of financial scale. According to data from The Information, DeepSeek's seven-month revenue of $70.7 million is well below the revenues of large US-based laboratories.
Therefore, it would be wrong to say “DeepSeek has surpassed OpenAI”.
But another sentence might be more accurate:
DeepSeek showed that OpenAI and Anthropic cannot define the AI economy on their own terms.
There is now a powerful third model:
high performance + aggressive price + light weight + efficient inference.
DeepSeek Means More for China
DeepSeek is not just a commercial company; It has become one of China's symbolic companies in the global AI competition.
Chinese companies face US export restrictions on access to advanced Nvidia GPUs. At first glance, this is a big disadvantage.
But the same limitation creates another engineering pressure:
Producing more results with less compute.
It is difficult to think of DeepSeek's obsession with efficiency without this context.
If you can't get unlimited GPU, you have two options.
You quit the race.
Or you look for a way to do the same job with less GPU.
DeepSeek seems to have chosen the second path.
And ironically, this obligation may turn into the company's global competitive advantage.
What Does DeepSeek Mean for Designers and Advertising Agencies?
For Voldi Creative, one of the most interesting points in the rise of DeepSeek is that artificial intelligence is increasingly turning into a commodity, that is, a standard infrastructure service, for the creative industry.
Today, an advertising agency does not have to choose a single model when using AI.
A model in text production.
Another model in coding.
Another model in research.
A different system in the visual.
A cheaper model can be used in background automation.
Powerful low-cost models like DeepSeek are accelerating this multi-model world.
For example, the most expensive frontier model may not be required for a system that classifies thousands of product descriptions within the agency. A more economical model can be used to categorize social media comments, clean catalog data, analyze customer documents, or use decision-making automation systems.
In this case, artificial intelligence is not just a creative tool.
Operations infrastructure.
And it is precisely in this second form of use that DeepSeek's pricing strategy is very powerful.
The Future of Big AI Models May Not Be a Single Giant Model
DeepSeek's success calls into question another assumption in the AI world.
Will one giant model do all the work in the future?
Probably not.
Maybe dozens of different models will run in a company's AI infrastructure.
Simple task → small model.
Code → coding model.
Long research → reasoning model.
Visual → image model.
Video → video model.
Sensitive corporate data → local model.
High volume automation → low cost Flash model.
This structure is very similar to cloud computing.
We do not use the most expensive server for every job.
We select resources according to need.
The AI economy is likely to advance to the same point.
DeepSeek V4 Flash is one of the best examples of this future.
Why is the 82.9 Percent API Margin More Important than the News Headline?
In the DeepSeek news, everyone's eyes are on the 10x revenue increase.
I think the more important figure is 82.9 percent.
Because that's the gross margin from API access.
One of the biggest problems of AI companies is the inference cost. Every time the user asks the model a question, the company incurs the cost of an actual GPU. While revenue increases for some AI businesses as the number of users increases, costs can also increase aggressively.
If DeepSeek can truly sustain this level of gross margin on the API side, the company's underlying economic model could be extremely strong.
Of course, we should not confuse gross profit with net profit here. When R&D, employees, model training, infrastructure investments and other expenses are added, the company still makes a loss.
But this figure tells us something else:
DeepSeek does not burn money while running models; spends money to grow.
This distinction is extremely important for investors.
Can There Be a 74 Billion Dollar Company?
The potential valuation of $74 billion against seven-month revenue of approximately $70 million seems extraordinarily high at first glance.
And it's already high.
Investors are not buying DeepSeek today here.
It is investing in the share it can get from the future AI infrastructure market.
If DeepSeek API can grow its revenue at the current rate for several years, gain enterprise customers, and launch a powerful new model family after V4, today's revenue numbers could become meaningless very quickly.
However, the opposite is also possible.
Competition in the AI industry is incredibly fast.
The model that is leading today may fall behind after a few months.
Google may lower the price.
OpenAI can produce more efficient models.
Anthropic API can reduce costs.
New Chinese competitors may emerge.
Therefore, although DeepSeek's current growth is impressive, the $74 billion expectation prices in serious future growth.
DeepSeek's Greatest Achievement Perhaps Changing Brand Perception
Until a few years ago, when global AI competition was discussed, China was mostly considered as the “market lagging behind American models”.
DeepSeek has dramatically changed this perception.
When the new model comes out today, people check the benchmark results.
Developers compare API price.
Rival laboratories read their technical reports.
Investors follow their financial results.
This is one of the strongest brand positions a tech company can achieve:
Your competitors are forced to follow you.
DeepSeek's most valuable asset today may not just be V4 Flash.
It may be this technical reputation that he has gained in the global AI world.
The Story of DeepSeek is Actually the Story of Efficiency
The motto of the first period of the artificial intelligence race was almost “more”.
More parameters.
More GPUs.
More data.
More capital.
Larger data center.
DeepSeek asks a different question:
Can we produce the same intelligence with fewer resources?
Today's financial results are starting to show that this problem has implications not only from an engineering perspective, but also from a commercial perspective.
The company's revenue has grown approximately 10 times compared to the whole of 2025. API margin is extremely high. V4 Flash is being aggressively optimized for developers and agent systems. One million token context is being standardized. The second major investment round is being prepared and the company's valuation is said to reach approximately 74 billion dollars.
It is too early to say that DeepSeek will win this race.
But the company achieved something much more important.
He changed the rules of the race.
Now an AI company can only ask “how powerful is your model?” It is not enough to ask.
“How fast?”
“How obvious?”
“How easily is it integrated?”
“How much calculation does it require?”
And perhaps most importantly of all:
How much does one million uses of this intelligence cost?
The rise of DeepSeek shows us that in the future of the AI economy, the biggest company may not necessarily be the company that makes the biggest model.
Maybe the winner will be the company that separates intelligence from electricity, server and token costs as much as possible and turns it into an infrastructure accessible to everyone.
And that's exactly what DeepSeek is pursuing right now.
Nur Oğuz