DeepSeek: What Is This AI and How Does It Actually Work?
- Stéphane Guy
- May 21
- 7 min read
For the past several weeks, the tech world has had one name on its lips: DeepSeek. This Chinese AI model, which landed in front of global audiences with almost zero prior fanfare, has triggered a wave of fascination, scrutiny, and genuine alarm. Touting near-ChatGPT capabilities, an open-source architecture, and a training cost claim that rattled Silicon Valley to its core, DeepSeek is simultaneously impressive, disruptive, and controversial. So what exactly is this AI, and what can it actually do?
In Short
DeepSeek is a large language model developed by a Chinese startup founded in 2023. It stands out for its dramatically lower cost and open-source availability, unlike ChatGPT, which requires a paid subscription for full access.
The AI is accessible via web and mobile (Android and iOS) and functions as a conversational chatbot. It supports advanced features including live web search and document analysis.
DeepSeek generates text, assists with academic tasks via specialized sub-models (DeepSeek Math, DeepSeek Coder), proposes creative ideas, and can be integrated into external platforms via its API.
Certain politically sensitive topics, particularly those censored in China, such as the 1989 Tiananmen Square massacre, return blocked responses, marking a distinct form of content moderation compared to ChatGPT or Gemini.
Multiple countries, including Italy and Australia, have raised serious data privacy concerns, citing Chinese national security laws that can compel companies to share user data with the state.
DeepSeek: The Company Behind the AI
DeepSeek is a remarkably young Chinese company. It was founded in May 2023 by Liang Wenfeng. The company's mission is unambiguous: democratize generative AI and conversational agents, and position itself as the global reference in the field, if not the reference.
To get there, DeepSeek has leaned on a deceptively simple but potent strategy.
Its model is open-source, meaning the code is publicly accessible, auditable, and free to use. This puts it in direct contrast to ChatGPT, which gates its most powerful features behind a paid subscription. The second pillar of the strategy is cost efficiency. DeepSeek's creators claim to have built an AI with capabilities broadly comparable to OpenAI's flagship models at a fraction of the expense. According to DeepSeek's own technical documentation, “DeepSeek V3 was trained on 2,048 H800 GPUs for approximately two months. In total, the model required 2.79 million GPU hours at an estimated cost of $5.58 million.”*
However, that headline training figure has faced significant pushback. Bernstein, the global investment research firm, challenged the $5–6 million claim directly in a report, noting that the figure only accounts for GPU pre-training compute and excludes research and development expenses, infrastructure costs, prior experimentation, and dataset validation. The Bernstein analysis, published on their official site at bernstein.com, argues that these omitted costs are far from trivial, and that building a genuinely competitive frontier model from zero still requires enormous capital investment, regardless of how efficiently the final training run was executed.*
What Is DeepSeek?
At its core, DeepSeek is a large language model (LLM): it parses user inputs, interprets them, and generates a textual response. In this fundamental sense, it is closely analogous to ChatGPT. The chat interface, in fact, closely mirrors OpenAI's design, a deliberate choice that lowers the adoption barrier for users already comfortable with conversational AI.
DeepSeek is, in the most precise sense, a conversational agent. If you want a deeper dive into how these systems are built and how they process language, our piece on artificial neural networks covers the underlying architecture in detail.
Where Is DeepSeek Available?
DeepSeek is accessible via its official website as well as through dedicated apps for Android and iOS, the same distribution model used by ChatGPT. All you need is a connected device. No specialized hardware, no subscription.
How to Use DeepSeek
Getting started with DeepSeek is as frictionless as using any major chatbot. Landing on the homepage, users are presented with two options: Start Now (direct access to the chat window) and Get DeepSeek App (mobile download). It is worth noting that the platform currently operates in English and Chinese; language preferences can be adjusted in settings by clicking your profile icon at the bottom-left corner of the interface.
That same profile menu also lets you sign out, delete all conversation history, or contact the company. New conversations are initiated via the blue New Chat button at the top-left. Conversation history is accessible just below it.
Within the chat interface, three functional controls are worth knowing:
DeepThink, triggers a slower, more deliberate reasoning process for complex queries
Search, activates live web search and returns sourced links alongside responses
Paperclip icon, allows users to upload documents or images for analysis directly within the conversation
For developers, DeepSeek's API enables direct integration into websites and applications.
What Can You Do With DeepSeek?
Help With Academic and Professional Tasks
One of DeepSeek's structural advantages is its mixture-of-experts (MoE) architecture. Rather than activating the full model on every query, DeepSeek routes requests to specialized sub-models. DeepSeek Coder handles software development tasks; DeepSeek Math targets algebraic problem-solving. According to the company, when you ask a question on a specific subject, only the relevant expert model is activated, a design choice that reduces inference costs significantly.
In practice, this means DeepSeek can assist with everything from history essays to calculus problems. The standard caveats apply: like all large language models, DeepSeek can hallucinate or produce factual errors. It supplements critical thinking, it does not replace it. Always verify outputs before using them in any high-stakes context.
Generating Ideas for Personal and Professional Projects
Struggling to find a content angle, brainstorm a campaign concept, or draft a project brief? DeepSeek performs reliably as an ideation engine. The breadth of its training data means it can surface relevant references, frameworks, and creative directions quickly, making it a useful first-pass research and inspiration tool. For an overview of how AI is reshaping work like this across industries, see our deep-dive on AI and the future of work.
Integration via API
DeepSeek's API unlocks a wide range of enterprise and developer use cases: automated translation pipelines, customer-facing chatbots, content moderation tools, and more. At $2.19 per million output tokens*, it undercuts the major Western providers substantially, making it a genuinely attractive option for teams building at scale. This connects to a broader structural shift in the AI industry, one we explore in detail in our analysis of the jobs AI will kill, transform, or create.
Lists, Recommendations, and Curated Selections
Like ChatGPT, DeepSeek handles structured recommendation requests competently. Ask for restaurant suggestions in a given city, a curated reading list, or a ranked selection of software tools, and it delivers a formatted list within seconds. Enabling the Search function adds cited web sources to these outputs, letting you cross-reference and validate the AI's suggestions independently.
DeepSeek and User Data: What You Need to Know
Like all major AI platforms, DeepSeek collects usage data and user profile information. The critical distinction lies in what happens to that data once collected. Under Chinese national security law, companies operating in China can be legally compelled to provide government authorities with access to user data upon request, with no obligation to notify users that this has occurred.
A growing list of governments and agencies have restricted or banned DeepSeek access, citing these exact concerns.
Italy was among the first to act: the country's Data Protection Authority (Garante) launched a formal investigation into DeepSeek's compliance with GDPR, and imposed an immediate halt on the processing of Italian users' personal data : “The Italian DPA described DeepSeek as a serious privacy risk due to its lack of transparency, failure to cooperate with authorities, and potential exposure of user data. Investigators raised alarms about how personal data was being collected, where it was stored, and whether Italian users were properly informed.”*
Australia followed suit, with Home Affairs Minister Tony Burke declaring DeepSeek an "unacceptable risk" to national security and issuing a mandatory directive banning the app from all government devices.*
Taiwan, South Korea, and multiple U.S. congressional offices and federal agencies have implemented similar restrictions.
This intersects with a wider debate about AI, data privacy, and what you're actually handing over when you use these tools, a debate that DeepSeek has thrown into sharp relief.
DeepSeek is already carving out a legitimate position as a high-performance, cost-efficient AI solution. Its open-source model and API economics make it genuinely attractive to developers. Its conversational capabilities are competitive with established Western models across most mainstream use cases. But the platform carries unresolved liabilities: embedded political censorship, a data governance framework that conflicts with Western privacy law, and an origin-country context that gives regulators legitimate grounds for concern.
If DeepSeek wants to hold its ground globally, against Gemini, Claude, ChatGPT, and whatever comes next, it will need to demonstrate far greater transparency about how it handles user data and what relationship it maintains with the Chinese state. Until then, it will remain one of the most technically impressive, and politically complicated, AI products on the market.
FAQ
Can DeepSeek compete with ChatGPT or Gemini?
On core functionality, text generation, coding assistance, document analysis, reasoning tasks, DeepSeek is genuinely competitive. Its strongest differentiators are its open-source model and aggressively low API pricing. Where it still trails is in linguistic quality and nuance for languages beyond English and Chinese, ecosystem breadth, and the reliability of factual outputs.
Is DeepSeek actually free?
Consumer access via the website and mobile apps is free. Professional usage via the API is billed per token, though rates are substantially lower than major competitors. At $2.19 per million output tokens versus $60 for OpenAI's o1, the cost differential is not marginal, it is structural.
Why is DeepSeek generating so much controversy?
Because its infrastructure, data storage, and corporate entity are all based in China, where national security law mandates that companies cooperate with government data requests. This is not a theoretical risk. It is a documented legal obligation that multiple governments have cited as grounds for banning or restricting the platform.
Can DeepSeek replace ChatGPT entirely?
Not universally. DeepSeek is a strong general-purpose tool for cost-conscious users and developers, but ChatGPT retains advantages in Western-language processing, plugin and integration ecosystem maturity, and overall output consistency. The choice depends heavily on the specific use case and risk tolerance around data governance.
What are DeepSeek's most distinctive capabilities?
Beyond standard chat functionality, DeepSeek's standout features are:
Specialized sub-models for math and coding that activate automatically based on query context
Native document and image analysis within the chat interface
DeepThink mode for deeper, step-by-step reasoning on complex problems
A developer-grade API at a price point that undercuts every major Western competitor
Comments