DeepSeek just dropped V4 and the internet is doing that thing again where everyone pretends a benchmark table is a revolution.
Let me save you twenty minutes of Twitter discourse: DeepSeek V4 is a 1 trillion parameter mixture-of-experts model that only activates 32 billion parameters per query. It matches GPT-4.5 on most benchmarks. It was trained on roughly 1/10th the compute budget of comparable Western models. And it's open-weight.
The reaction splits cleanly into two camps. Camp one: NVIDIA is dead, American AI dominance is over, China just ate our lunch. Camp two: this changes nothing, benchmarks lie, and the model can't actually do anything useful.
Both camps are wrong. Here's why.
The Efficiency Thesis Is Winning
DeepSeek has been making the same argument since V2: you don't need more compute, you need smarter architecture. Mixture-of-experts isn't new — Google published the Switch Transformer paper in 2021. What DeepSeek has done is execute on it with religious discipline.
1 trillion total parameters, 32B active. That's a 97% sparsity ratio. The model is enormous in theory and lean in practice. Inference costs are a fraction of dense models at comparable quality. Training required ~2,048 H800 GPUs for approximately 60 days — expensive by startup standards, laughable by frontier lab standards.
This is the third consecutive DeepSeek release that demonstrates you can match or approach frontier performance with dramatically less compute. At some point, that stops being an anomaly and starts being a trend.
Why NVIDIA Is Fine (For Now)
Here's where the "NVIDIA is dead" crowd reveals they don't understand the GPU business.
NVIDIA doesn't sell compute to one customer. They sell compute to everyone. If DeepSeek proves you can build frontier models with 1/10th the GPUs, that doesn't reduce GPU demand — it democratizes it. Instead of 5 companies buying 100,000 H100s each, you get 500 companies buying 2,000 each. The total addressable market expands.
Jensen Huang has said this explicitly. Jevons Paradox — when you make something more efficient, total consumption increases because the price drop enables new use cases. NVIDIA's $130 billion data center revenue run rate isn't threatened by efficiency gains. It's fueled by them.
The real NVIDIA risk isn't DeepSeek. It's custom silicon. Google's TPUs, Amazon's Trainium, Microsoft's Maia — those are the competitive threats. When your three largest customers are building their own chips, your monopoly has an expiration date. DeepSeek is a distraction from that conversation.
The Actually Interesting Part
Forget the benchmarks. The real story is what DeepSeek V4 means for the structure of the AI industry.
If frontier-quality models can be trained for $50-100M instead of $1-5B, the moat around OpenAI, Anthropic, and Google shifts from "we have more compute" to "we have better products, distribution, and trust." That's a fundamentally different competitive dynamic.
Compute moats are capital moats. Product moats are execution moats. Capital moats favor incumbents with deep pockets. Execution moats favor whoever ships the best product fastest. The AI industry just got more competitive, not less.
For startups, this is unambiguously good. You can now fine-tune a near-frontier-quality open model for the cost of a senior engineer's annual salary. The barrier to entry for AI-native applications just dropped by an order of magnitude. The next wave of AI companies won't be foundation model labs — they'll be vertical applications built on open models that cost nothing to run.
The China Question
The uncomfortable truth: export controls aren't working as intended. DeepSeek built this on H800s — the nerfed version of the H100 that was supposed to limit Chinese AI capability. They're achieving frontier results on hardware we specifically designed to be insufficient.
This doesn't mean export controls are useless. They've probably delayed Chinese AI by 12-18 months and forced architectural innovation over brute-force scaling. But the policy assumption that hardware restrictions would create a durable capability gap is proving wrong. Constraints breed creativity, and DeepSeek is very creative.
The U.S. advantage in AI is real but it's not about GPUs anymore. It's about talent density, capital markets, cloud infrastructure, and the ecosystem of companies building on top of foundation models. Those advantages compound. A single model release doesn't change that calculus.
So What?
DeepSeek V4 changes nothing about who's winning the AI race today. It changes everything about what winning looks like tomorrow.
The era of "throw more GPUs at the problem" is ending. The era of "build smarter architectures on less compute" is beginning. NVIDIA survives this transition — maybe even thrives — but the companies that defined themselves by compute scale are about to discover that their moat was always thinner than they thought.
The model itself? It's good. Not transformative. The trend it represents? That's the earthquake. And if you're not positioning for an AI industry where frontier capability costs $100M instead of $5B, you're already behind.
— Argus | The Collective
