Abishek D Praphullalumar
DeepSeek’s launch in January 2025 was nothing short of a phenomenon. The app skyrocketed to the top of Apple and Google app stores with over three million downloads in just days. The company proudly claimed to have built a ChatGPT-level AI for just $6 million a fraction of OpenAI’s reported $100 million investment in GPT-4.
Sounds impressive, right? Too good to be true.
Governments aren’t buying it. The Pentagon and NASA blocked DeepSeek almost instantly. Italy and Taiwan went further and outright banned it. It didn’t take long for their concerns to be validated a massive cyberattack hit the platform, leaking user data, API secrets, and internal logs.
So what’s really going on here? Experts have been hyping DeepSeek as an “AI breakthrough,” but when you strip away the marketing, its technical limitations, security risks, and misleading cost claims paint a different picture.
DeepSeek’s Technical Claims vs. Reality
“DeepSeek’s breakthrough comes from cutting edge AI performance, open-source innovation, and dramatically reduced computational costs.”
Bold claims. But do they hold up? Let’s break it down.
1. Processing Power Is Nowhere Near What They Claim
DeepSeek R1 is supposed to handle 275 tokens per second, but real-world testing shows it barely reaches 32 tokens per second. The app takes an average of 60.51 seconds to respond far slower than its competitors.
For complex queries, responses can drag on for minutes, making it impractical for anything beyond simple conversations.
2. Severe Memory Limitations
DeepSeek boasts a 64,000-token context window, but that’s still leagues behind ChatGPT-4o’s 200,000 tokens. Worse, its memory management is inefficient, forcing the system to store excess key value pairs leading to slower processing times and limited recall ability.
3. “Efficiency” at the Cost of Performance
DeepSeek only uses 37 billion parameters per query a deliberate design choice to save computing power. But that comes with a tradeoff: less detail, more generic answers, and a lack of nuanced reasoning compared to ChatGPT or Claude.
DeepSeek performs well in math (90% accuracy) and coding (97% success rate), but when it comes to general reasoning, it’s GPT-4o level slow so bring patience.
DeepSeek’s Security Nightmare: A Hacker’s Playground
Security experts took a closer look at DeepSeek’s infrastructure. What they found? A disaster waiting to happen.
1. Hard Coded Encryption Keys (Yes, Really.)
DeepSeek’s iOS app still relies on outdated Triple DES encryption a relic from the late ‘90s. Worse, the encryption keys are hard coded into the app. This means if one key is exposed, hackers can unlock every user’s data.
2. Unencrypted Data Transmission
DeepSeek disabled Apple’s App Transport Security (ATS) a security feature designed to protect data in transit. This means:
- Chat histories can be intercepted
- Device and user info are sent in plaintext
- API secrets are exposed in transmission
The platform also transmits user data to ByteDance servers in China, raising serious concerns about data privacy and government access.
With no end-to-end encryption, anyone with network access can read, modify, or inject malicious data into DeepSeek’s system.
The $6M Myth: What DeepSeek Isn’t Telling You
“DeepSeek proves that China is innovating AI faster and cheaper. If true, it challenges the massive spending by U.S. tech companies.”
— Angelo Zino, Senior Equity Analyst, CFRA Research
$6 million to train an AI rivaling OpenAI? Sounds like a miracle or a lie.
Here’s what DeepSeek’s budget doesn’t include:
- Massive GPU Costs: DeepSeek operates ~50,000 Nvidia Hopper GPUs in data centers worldwide. Each H100 GPU costs $25,000 – $40,000 putting their hardware costs between $1.25 and $2 billion.
- Power and Infrastructure: AI models consume huge amounts of electricity. Maintaining DeepSeek’s server farms, cooling systems, and data pipelines is estimated at $944 million per year.
- AI Talent and R&D: AI engineers don’t come cheap. DeepSeek’s model required years of research, synthetic data generation, and training refinement.
Real Cost Breakdown
📌 Reported Cost: $6 million (GPU pre-training only)
📌 Actual Cost Estimate: $1.6 billion+ (infrastructure + operations)
DeepSeek’s own documentation admits their cost estimates leave out “prior research, experiments, and data acquisition.”
So no, DeepSeek isn’t a budget friendly miracle it’s just creative accounting.
Performance Testing: Where DeepSeek Stands
How does DeepSeek actually compare? Recent tests from the Knowledge Observation Group (KOG) paint a mixed picture:
MMLU (General Knowledge Test): 88.5% accuracy (competitive)
MMLU-Pro (Advanced Reasoning): 75.9%
GPQA (General Purpose QA): 59.1%
MATH-500 (Math Benchmark): Better than GPT-4o preview
DeepSeek performs well in structured logic and math tasks, but when it comes to complex reasoning and agentic tasks, it falls behind leading closed source models by 24.2%.
This performance gap gets even worse for agentic AI tasks, with DeepSeek lagging 317.7% behind top tier models.
The Verdict: Overhyped, Underwhelming, and Full of Risks
DeepSeek took the world by storm, but the reality doesn’t match the hype.
🔻 It’s too slow. Response times drag, making it impractical for high speed interactions.
🔻 It’s insecure. Hard coded encryption keys, plaintext data transmission, and China based servers are major red flags.
🔻 The $6M claim is misleading. The real investment is closer to $1.6 billion, raising questions about transparency.
DeepSeek shines in math and coding benchmarks, but for everyday users? It’s nowhere near the game changer it claims to be. If you’re considering DeepSeek for serious AI work, think twice. The risks far outweigh the benefits.
Final Thoughts: A Paper Tiger?
DeepSeek is riding the AI hype wave with big promises, but reality tells a different story. It’s slow, riddled with security flaws, hampered by processing limits, and its $6M development cost claim doesn’t add up.
It shines in math and coding, but for real-world AI tasks, it falls flat. Governments don’t trust it. Security experts are raising red flags. And if you’re looking for a serious AI tool, this isn’t it.
It may have millions of downloads today but when the hype dies down, will anyone still trust it?
