DeepSeek V4 trained a 1-trillion-parameter model for just $5.6M, challenging GPT-5 and signaling a major shift in AI compute economics.

$5.6M, 1 Trillion Parameters: DeepSeek V4 Drops Nvidia

by
💡 Key Takeaways

  • DeepSeek V4 is a Mixture-of-Experts model — 1.6 trillion total parameters with 49B active per token, per official launch specs — reportedly trained for just $5.6M — a fraction of what U.S. frontier labs spend.
  • The model is optimized for Huawei Ascend 910C and Cambricon chips, with DeepSeek denying early access to Nvidia and AMD — a deliberate break from the U.S. semiconductor supply chain.
  • Released under Apache 2.0 with a 1-million-token context window and native multimodal capabilities, DeepSeek V4 accelerates the bifurcation of the global AI stack into competing U.S. and Chinese ecosystems.

$5.6 million. That is what it reportedly cost to train a trillion-parameter AI model capable of competing with GPT-5 and Claude Opus. The claim comes from DeepSeek, a Chinese AI startup that has turned efficiency into a geopolitical weapon.

But cost is only part of the story. DeepSeek V4 was purpose-built to run on Huawei’s Ascend chips — no Nvidia GPUs required. In short, this is not just another model release; it is the clearest signal yet that the global AI industry is splitting into two parallel technology stacks.

The model ships with a 1-million-token context window, native multimodal generation across text, image, video, and audio, and a fully open Apache 2.0 license. Jimmy Goodrich, a senior associate at CSIS’s Wadhwani AI Center, put it bluntly: “It’s been long known that DeepSeek has a really good team, and if they had access to even more compute, God knows how capable they would be.”

Cinematic illustration of a neural network architecture symbolizing the U.S.-China AI technology split
A trillion-parameter MoE architecture is splitting the global AI supply chain in two — Generated by OpenAI DALL-E 3

Inside DeepSeek V4: What a Trillion Parameters Actually Means

The headline number — one trillion parameters — is deliberately misleading if taken at face value. DeepSeek V4 uses a Mixture-of-Experts (MoE) architecture, which means only a fraction of the model activates for any given token. In this case, roughly 37 billion parameters fire per token, making inference dramatically more efficient than a dense model of equivalent scale.

The architecture also includes a 1-million-token context window, powered by a proprietary system called Engram Conditional Memory. Put simply, Engram uses hash-based constant-time lookups to retrieve relevant information from extremely long inputs, achieving a 97% accuracy rate on needle-in-a-haystack tests at the million-token scale. That is a meaningful jump over standard attention mechanisms.

On the multimodal front, DeepSeek V4 was designed from the ground up to handle text, image, video, and audio simultaneously. This is not a text model with vision bolted on afterward — multimodal capabilities were baked into the training process from inception. The model is expected to ship under an Apache 2.0 license with no MAU caps, no use restrictions, and full commercial freedom.

Key point: DeepSeek V4-Pro activates 49B of its 1.6 trillion parameters per token (official launch figures; pre-launch leaks had suggested ~1T/37B), pairs a 1-million-token context window with 97% retrieval accuracy, and delivers native multimodal processing — all under an Apache 2.0 license with no commercial restrictions.

The $5.6M Question: How Three Architectural Breakthroughs Slashed Training Costs

According to a CSIS report drawing on SemiAnalysis estimates, DeepSeek V4’s training consumed approximately 2.788 million GPU-hours at $2 per GPU-hour, totaling roughly $5.6 million. For context, training costs for U.S. frontier models routinely exceed $100 million. The key here is that this cost advantage is not the product of a single trick but three distinct architectural innovations working in concert.

The first is Manifold-Constrained Hyper-Connections (mHC), a stabilization technique that prevents training from diverging at the trillion-parameter scale by constraining signal amplification. DeepSeek published the underlying research paper describing how mHC makes training trillion-parameter MoE models feasible without the catastrophic instabilities that typically derail such efforts.

The second is Engram Conditional Memory, the same system powering the million-token context window. By using hash-based O(1) lookups for static knowledge retrieval, Engram eliminates the quadratic scaling problem that makes standard attention prohibitively expensive at long context lengths. The result: 97% needle-in-a-haystack accuracy versus standard attention’s substantially lower performance.

The third innovation is Dynamic Sparse Attention with a Lightning Indexer, which reduces compute overhead by an estimated 50%. Together, these three breakthroughs allowed DeepSeek to train a trillion-parameter model on the kind of budget that a well-funded Series B startup could theoretically afford.

In simpler terms: Running all trillion neurons simultaneously would be ruinously expensive. DeepSeek’s approach is to selectively activate only the 37 billion most relevant ones per token, and to use a dictionary-like instant lookup system for long documents instead of reading every word sequentially. The result is frontier-class capability at roughly 1/20th the cost.

No Nvidia Required: The Huawei Ascend Chip Strategy

DeepSeek V4’s most provocative decision was not architectural — it was strategic. The company optimized the model for Huawei’s Ascend 910C and Cambricon chips, while simultaneously denying early access to Nvidia and AMD. In an industry where every major model launch is built on Nvidia’s CUDA ecosystem, this is an open declaration of technological independence.

Huawei’s Ascend 910C delivers approximately 60% of the Nvidia H100’s inference performance, according to DeepSeek’s own researchers — a meaningful gap. But Huawei’s CloudMatrix 384 architecture narrows it considerably, reporting 6,688 tokens per second on prefill and 1,943 tokens per second on decode per NPU. These figures suggest competitive inference economics even if raw silicon performance lags behind Nvidia’s best.

The demand signal is already visible. Alibaba, ByteDance, and Tencent have placed bulk orders for hundreds of thousands of Huawei AI chips. This is not a pilot program or a research experiment — it is a coordinated bet by China’s largest technology companies that Huawei’s silicon can sustain production-grade AI workloads.

Legal scholar Angela Zhang observed that “export restrictions have paradoxically accelerated China’s efficiency innovation,” a dynamic that the Ascend chip strategy embodies. The harder the U.S. squeezes the supply of advanced chips, the more aggressively China invests in building its own.

Data center server racks used for AI inference workloads
AI inference infrastructure built on Huawei Ascend chips is challenging Nvidia’s dominance in the data center — Photo by Elimende Inagella on Unsplash

Benchmarks Under Scrutiny: How DeepSeek V4 Stacks Up Against GPT-5 and Claude

[Post-launch note] The official April 2026 release materials from DeepSeek do not confirm the leaked benchmark figures below.

Leaked internal benchmarks — and the emphasis here belongs squarely on “leaked” and “internal” — paint a compelling picture. DeepSeek V4 reportedly scores 90% on HumanEval and above 80% on SWE-bench Verified, the latter putting it in the same tier as Claude Opus 4.5’s 80.9%. These are serious numbers, if they hold up under independent scrutiny.

The pricing comparison is where the disruption becomes tangible. DeepSeek V4’s expected API cost of $0.10 to $0.30 per million tokens is roughly 6 to 17 times cheaper than GPT-5.2 at $1.75 per million tokens, and 17 to 50 times cheaper than Claude Opus at $5.00 per million tokens. If the performance benchmarks are even approximately accurate, the price-performance ratio represents a structural challenge to the business models of U.S. frontier labs.

The takeaway: verification is everything. DeepSeek V4 Lite, a smaller variant with approximately 200 billion parameters, was previewed on March 9, 2026.

The full model’s official release is imminent. Whether the independent benchmark results replicate the internal claims will determine whether DeepSeek V4 is a genuine inflection point or a well-marketed incremental improvement.

💡 Related Posts
Nvidia Bets $2B on Silicon Photonics to Break AI’s Data Wall

The Distillation Controversy: U.S. AI Labs Strike Back

Performance claims are shadowed by a bitter intellectual property dispute. In February 2026, Anthropic publicly accused DeepSeek and two other Chinese AI companies — Moonshot AI and MiniMax — of orchestrating “industrial-scale distillation attacks” using 24,000 fraudulent accounts that generated over 16 million exchanges with Claude. The targets, according to Anthropic, were Claude’s most differentiated capabilities: agentic reasoning, tool use, and coding.

OpenAI sent a parallel open letter to Congress, warning that DeepSeek was conducting “ongoing attempts to distill frontier models” using new obfuscation techniques. Both companies frame the dispute as a matter of national security. Anthropic explicitly warned about “authoritarian governments deploying frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance.”

The counterargument is equally pointed. Industry experts note that “the boundary between legitimate use and adversarial exploitation is often blurry” when it comes to distillation. Dan Hendrycks, director of the Center for AI Safety, offered a more structural observation: “If you know which research dead ends to avoid, replication is much easier.” Why this matters: the implication is that algorithmic insight leakage through informal channels may be as significant a threat as hardware export controls are designed to address.

💡 Related Posts
78 Laws in 27 States Turn AI Chatbot Therapy Into a Felony

Apache 2.0 at Trillion Scale: What Open Source Means for the Developer Ecosystem

A trillion-parameter model released under the Apache 2.0 license is unprecedented. No MAU caps, no acceptable-use restrictions, no commercial limitations — full freedom to modify, deploy, and redistribute. For startups, research labs, and enterprises that need on-premises AI for data sovereignty reasons, this changes the calculus entirely.

The hardware requirements for running the model locally are surprisingly accessible. INT8 quantization enables inference on two RTX 4090 GPUs (48GB VRAM total), while INT4 quantization brings it within reach of a single RTX 5090 (32GB VRAM). Full-precision inference still demands a multi-node GPU cluster, but quantized versions should handle most enterprise workloads adequately.

The strategic implication cuts deep. GPT-5 and Claude operate exclusively through API subscriptions, meaning every query routes through U.S.-controlled infrastructure. DeepSeek V4’s open-source release allows any organization to run a frontier-class model on its own servers, fully air-gapped from external dependencies.

For governments, defense contractors, and enterprises with strict data residency requirements, the appeal is obvious. Data sovereignty is no longer a theoretical concern — it is an operational constraint that open-source frontier models can directly address.

Close-up of semiconductor circuit board components
At the heart of the U.S.-China AI supply chain split lies a battle for semiconductor self-sufficiency — Photo by Anne Nygård on Unsplash

The Global AI Supply Chain Is Splitting in Two

DeepSeek V4 is not just a model — it is evidence that the global AI supply chain is undergoing a structural bifurcation. On one side: the U.S.-centric stack built on Nvidia GPUs and CUDA; on the other, a China-centric stack anchored by Huawei’s Ascend chips and CANN software framework.

DeepSeek’s decision to deny Nvidia and AMD early access was not an oversight — it was a deliberate acceleration of this split. The two parallel ecosystems are now developing independent toolchains, hardware pipelines, and developer communities.

The enforcement apparatus behind U.S. export controls is struggling to keep pace. According to CSIS, TSMC produced over 2 million Ascend 910B logic dies for Huawei through shell companies, circumventing restrictions that were designed to prevent exactly this outcome. The existence of multiple viable workaround channels raises fundamental questions about the long-term effectiveness of the current regulatory framework.

The resource constraints are structural. The Bureau of Industry and Security (BIS) operates with fewer than 600 employees and a budget of approximately $200 million — tasked with overseeing all dual-use technology exports, not just semiconductors. In other words, the entire enforcement mechanism for the most consequential technology competition of the decade is running on a shoestring.

If DeepSeek V4’s Huawei chip optimization proves viable in production workloads, the implications extend far beyond one model. A successful demonstration would undermine the core premise of U.S. export controls — that restricting access to advanced chips can meaningfully slow Chinese AI development. The AI industry is approaching a threshold: from a single global supply chain optimized for performance, to two competing supply chains optimized for independence.

Key point: Over 2 million Ascend dies produced through TSMC shell companies and a BIS enforcement team of fewer than 600 employees underscore the structural limits of export controls.

Two data points should be on every watchlist. The first is DeepSeek V4’s independent benchmark performance once the full model ships — if third-party tests replicate the internal claims, the pricing premium commanded by U.S. frontier models loses its justification.

The second is the real-world inference throughput of Huawei Ascend chips running production workloads at scale — if the economics hold, Nvidia’s monopoly on AI compute faces its first structural crack. The global AI industry is not heading toward fragmentation as a distant possibility; it is fragmenting now.

References

Stay Ahead of the Curve

Get the latest insights on AI, tech, and markets — delivered in both Korean and English.

Explore More Posts