Sarah Chen
The Sputnik Moment of Artificial Intelligence
Let me break this down: imagine a startup with a fraction of OpenAI's budget managing to create an AI that competes head-to-head with GPT-5. That trains its models for $5.6 million when OpenAI spends $100 million or more. That publishes model weights for free so anyone can use them.
That's exactly what DeepSeek is doing.
On January 27, 2025, when DeepSeek released its R1 model, NVIDIA lost $600 billion in market value in a single day. The largest single-day loss for any company in Wall Street history. Microsoft lost another $150 billion. Analysts called it the "Sputnik moment" of AI.
And now, in February 2026, DeepSeek-V4 is coming. Internal leaks suggest it could surpass Claude Opus 4.5 in programming with a 1 trillion parameter model and a 1 million token context window.
The trick is an innovation called mHC (Manifold-Constrained Hyper-Connections) that allows training larger models with less hardware. And what most guides won't tell you is that this fundamentally changes the economics of artificial intelligence.
What Is DeepSeek and Why Should You Care
DeepSeek isn't a typical Silicon Valley startup. It's a Chinese company founded in Hangzhou by a former quantitative trader who decided to use his hedge fund profits to build AGI.
The Numbers That Matter
| Metric | DeepSeek | OpenAI | Anthropic |
|---|---|---|---|
| Flagship model training cost | $5.6M | $100M+ | Not disclosed |
| API price (input/1M tokens) | $0.28 | $2.50-$3.00 | $3.00 |
| API price (output/1M tokens) | $0.42 | $10.00 | $15.00 |
| Open source model | Yes (public weights) | No | No |
| Valuation | ~$3.4B | $300B+ | $60B+ |
What most guides won't tell you is the magnitude of this difference: for the same money it costs to use GPT-4o, you can make 10-40 times more queries with DeepSeek. A 100K token context that costs $5.50 on GPT-4 costs $0.90 on DeepSeek.
DeepSeek V3 vs the Giants (Current Benchmarks)
| Benchmark | DeepSeek V3 | GPT-4o | Claude 3.5 Sonnet |
|---|---|---|---|
| MMLU-Pro (general knowledge) | 75.9 | ~77 | ~76 |
| GPQA (advanced science) | 59.1 | ~58 | ~59 |
| MATH-500 (mathematics) | 90.2% | ~85% | ~88% |
| LiveCodeBench (programming) | #1 | Top 5 | Top 3 |
DeepSeek already leads in mathematics and programming. With V4, they're aiming to completely dominate the coding niche.
The mHC Architecture: The Secret to Efficiency
This is where things get technical, but let me break it down so it makes sense.
The Problem mHC Solves
Think of it like training a very large neural network. As you add more layers, information flowing through the network tends to amplify uncontrollably. It's like a microphone producing feedback: a small signal gets amplified until it becomes unbearable noise.
This phenomenon is called gradient explosion and it's one of the main obstacles to training larger models. The traditional solution is to use more computing power to stabilize training. More GPUs, more money, more time.
mHC proposes something different: instead of letting information amplify without control, it projects it onto a mathematical structure called a manifold that guarantees the total amount of information is conserved.
How It Works (Simplified)
-
Multiple parallel streams: Instead of a single path for information, mHC uses 4 "streams" that process data in parallel
-
Sinkhorn-Knopp constraint: A mathematical algorithm that ensures when information mixes between streams, it doesn't amplify. It can redistribute, but never grow uncontrollably
-
Controlled maximum gain: While traditional architectures can amplify signals up to 3000 times, mHC limits amplification to ~1.6 times
The Concrete Results
| Metric | With mHC | Without mHC |
|---|---|---|
| Training overhead | +6.7% | Baseline |
| Stability | High | Low in large models |
| BIG-Bench Hard (reasoning) | +2.1% improvement | Baseline |
| Scalability | Tested up to 27B params | Limited by gradient explosion |
The trick is that 6.7% extra compute during training saves you from having to use 10x more GPUs to stabilize the process. It's a small investment with an enormous return.
Liang Wenfeng: The Founder Nobody Knows
If you ask me who's the most important person in AI that almost nobody knows, I'd say it's Liang Wenfeng.
From Rural Village to Billionaire Hedge Fund Manager
| Fact | Detail |
|---|---|
| Birth | 1985, Mililing village, Wuchuan, Guangdong |
| Parents | Both primary school teachers |
| Education | Bachelor (2007) and Master (2010) in electronic engineering, Zhejiang University |
| Estimated net worth | ~$4.5 billion (2025) |
Liang founded High-Flyer, a quantitative trading hedge fund, in 2016. By 2021, it managed over 100 billion yuan (~$14 billion). The strategy: use machine learning algorithms to predict market movements.
From Hedge Fund to AI Lab
In April 2023, Liang announced he would convert part of High-Flyer into an AGI laboratory. In July, that lab spun off as DeepSeek.
What most guides won't tell you is how he funded it: 100% with his own money. When Silicon Valley VCs offered investment, Liang rejected them. He didn't want the pressure of quick "exits" or investor interference.
The Master Move: 10,000 A100 GPUs
Before the U.S. restricted exports of advanced chips to China, Liang acquired 10,000 NVIDIA A100 GPUs. Those chips, now banned, are the infrastructure that allows DeepSeek to compete with OpenAI.
According to some reports, DeepSeek may have up to 50,000 Hopper GPUs including some H100s obtained through intermediaries. NVIDIA denies this, claiming DeepSeek only uses "legally acquired H800s" (a limited version allowed for China).
DeepSeek-V4: What We Know (and What We Don't)
DeepSeek-V4 hasn't been officially released yet. It's expected in mid-February 2026. But leaks and GitHub updates reveal a lot.
Leaked Specifications
| Feature | DeepSeek-V4 (leaked) |
|---|---|
| Parameters | ~1 trillion (MoE model) |
| Context window | 1 million tokens (with DSA) |
| Architecture | MoE + mHC + Engram |
| Consumer hardware | 2x RTX 4090 or 1x RTX 5090 |
| Efficiency vs Transformers | -50% overhead |
| Open weights | Yes (probable) |
The "Reasoning Core"
One of the most interesting innovations is what DeepSeek calls Reasoning Core: a separate module within the model specialized in step-by-step reasoning. Think of it like the model has a "deep thinking mode" it can activate for complex problems.
This is similar to what OpenAI did with o1/o3, but integrated directly into the base architecture.
Projected Comparison with Competitors
| Aspect | DeepSeek V4 | Claude Opus 4.5 | GPT-5 |
|---|---|---|---|
| SWE-bench (coding) | >80% (leak) | 80.9% | ~78% |
| Context | 1M tokens | 200K | 128K |
| Expected price | 10-40x cheaper | Premium | Premium |
| Open weights | Yes | No | No |
| Availability | Global (with restrictions) | Global | Global |
Why 15 Countries Have Banned DeepSeek
This is where geopolitics comes into play. DeepSeek isn't just a technical threat to Silicon Valley; it's a national security threat according to several governments.
Countries That Have Banned or Restricted DeepSeek
| Country/Region | Date | Scope | Official Reason |
|---|---|---|---|
| Italy | January 2025 | Total | GDPR violation |
| Australia | February 2025 | Government | National security risk |
| Taiwan | February 2025 | Government + schools | National information danger |
| South Korea | February 2025 | Government | Data collection |
| Czech Republic | July 2025 | Public administration | Servers in China/Russia |
| India | 2025 | Finance Ministry | Data vulnerability |
| USA | Various | NASA, Navy, Congress, Texas | Foreign access |
Companies like Microsoft, Mitsubishi Heavy Industries, and Toyota have also banned internal use.
The Legal Problem
The underlying reason is simple: according to DeepSeek's privacy policy, all user data is stored on servers in China. And according to China's National Intelligence Law (2017), any organization must "support, assist, and cooperate with national intelligence efforts."
In other words: the Chinese government can legally demand access to your DeepSeek conversations without notifying you.
The Impact on the Industry: End of OpenAI's Business Model?
What most guides won't tell you is that DeepSeek doesn't just threaten Silicon Valley's technology. It threatens their business model.
The Math That Scares OpenAI
If DeepSeek can train comparable models for $5.6 million instead of $100 million, and then publishes them for free with open weights, why would you pay $20 per month for ChatGPT Plus?
Jack Clark, Anthropic co-founder, said it clearly:
"DeepSeek means AI proliferation is guaranteed."
Developers are already running DeepSeek models locally on their own servers. No monthly subscriptions. No usage limits. Without their data going to any company.
Silicon Valley's Response
U.S. companies are responding in several ways:
-
Price reductions: OpenAI and Anthropic have significantly lowered API prices over the past year
-
Emphasis on security: Positioning themselves as the "safe" choice against Chinese models
-
Vertical integration: Microsoft, Google, and Amazon use their models in their own products where AI cost is secondary
-
Differentiation: Focusing on enterprise use cases where trust and support matter more than price
How to Use DeepSeek (If You Decide To)
If after reading all this you want to try DeepSeek, here are your options.
Option 1: Official API
- Site: platform.deepseek.com
- Price: $0.28/1M tokens input, $0.42/1M tokens output
- Warning: Your data goes to servers in China
Option 2: Run It Locally
DeepSeek publishes model weights on Hugging Face and GitHub. You can download them and run them on your own hardware.
Requirements for V3 (current version):
- Minimum: 2x RTX 4090 (24GB VRAM each)
- Recommended: GPU with 80GB+ VRAM or distributed cluster
Expected requirements for V4:
- Minimum: 1x RTX 5090 or 2x RTX 4090
- For full 1M token context: significantly more
Option 3: Intermediary Providers
Companies like Together AI, Fireworks, and others offer access to DeepSeek models from U.S. infrastructure. You pay a bit more but your data doesn't go directly to China.
Frequently Asked Questions
Is DeepSeek safe to use?
Depends on your definition of "safe." Technically it works well. But if you're concerned about the Chinese government accessing your conversations, the answer is no: according to their own terms of service, they store data in China under Chinese law. For casual personal use it's probably irrelevant. For sensitive company information, clearly not.
Is DeepSeek better than ChatGPT?
In some things yes, in others no. DeepSeek V3 leads in mathematics (MATH-500: 90.2%) and programming (LiveCodeBench: #1). GPT-4o has better general knowledge (MMLU-Pro: 77 vs 75.9). For most users, the practical difference is minimal; the price difference is enormous.
Can I run DeepSeek on my computer?
Yes, if you have enough hardware. Smaller models (7B, 16B parameters) run on consumer GPUs. The full V3 model needs minimum 2x RTX 4090. V4 will probably need more, though they promise optimizations for consumer hardware.
Why is DeepSeek so cheap?
Three reasons: (1) Architecture innovations like mHC that reduce compute requirements, (2) Lower labor costs in China, (3) They don't need to generate returns for investors because Liang funds it with his own money. Basically, DeepSeek operates more like a research project than a business.
When does DeepSeek-V4 come out?
Expected mid-February 2026. No exact date confirmed. GitHub updates and internal leaks suggest development is advanced.
Conclusion: The Future of AI Is No Longer Decided Only in Silicon Valley
DeepSeek represents something bigger than a company or an AI model. It represents the end of American hegemony in artificial intelligence.
For decades, the world's most advanced technologies were developed in U.S. labs with budgets no other country could match. DeepSeek proves that efficiency and innovation can compensate for lack of raw resources.
The trick is this has enormous implications:
For developers: Access to frontier AI no longer requires paying premium subscriptions. Models comparable to GPT-4 are available for free with open weights.
For companies: The cost of integrating AI into products drops 10-40x. What was previously viable only for Big Tech is now within reach of startups.
For the industry: The AI race is no longer won with more money. It's won with better research. OpenAI, Anthropic, and Google will have to innovate faster, not just spend more.
For geopolitics: China is no longer "years behind" the U.S. in AI. They're competing head-to-head, and on some benchmarks, winning.
When DeepSeek-V4 launches in February, we'll probably see another "DeepSeek Monday" in the markets. But the real impact won't be on NVIDIA's stock. It'll be on how we think about who controls the future of artificial intelligence.
Would you use a Chinese AI knowing your data goes to servers under Beijing's jurisdiction? Or would you rather pay 10x more for the "security" of Western models? The answer to that question will define the AI market for years to come.
