Good Morning,

We are on the cusp of Chinese New Year and something is happening that’s a bit dark on the U.S. frontier.

Silicon Valley has adopted a 996 lifestyle for AI research. As models get more fine tuned to agentic capabilities, something remarkable is happening. A few Chinese AI startups are now showing signs of being more innovative at the frontier of open-weight models (than Silicon Valley and the rest of the world), this - with a fraction of the capital and with significantly less access to the best AI chips. How could this be?

Share

After the DeepSeek moment of last year (January, 2025), China has surged ahead in Open-source AI models. A flurry of Chinese IPOs means the following AI startups and open-source incumbents are going to be a big deal:

• Qwen (Alibaba Cloud)

• Zhipu AI (Z.AI)

• Moonshot AI

• DeepSeek

• Minimax

It’s not just the quality of their models, it’s the pace of iteration. 🚀 The DeepSeek moment of last year snowballed into a very different open-source AI reality.

One year after Chinese startup DeepSeek rattled the global tech industry with the release of a low-cost artificial intelligence model, its domestic rivals are better prepared, vying with it to launch new models, some designed with more consumer appeal. - Reuters

Read more DeepSeek articles

There’s a buzz for what DeepSeek is ready to announce, and high expectations. There’s software, reinforcement learning, LLM training and hardware innovations that DeepSeek is making that has rubbed off on China’s vibrant open-weight LLM ecosystem. (The U.S. has no equivalent).

The bifurcation of closed models and open-models has never been this intense. While American companies get the hype, Chinese models are helping developers build real products.

DeepSeek as a Research Lab is taking a different approach

There is a sense of anticipation because DeepSeek is currently on the verge of its next major release cycle. As of February 2026, the primary focus is on DeepSeek V4, which is expected to launch around the Lunar New Year (Tuesday, February 17th begins).

Rumored to be codenamed "MODEL1," DeepSeek V4 is expected to be a total architectural overhaul rather than just a minor update. Meanwhile its next flagship model DeepSeek-R2 has been significantly delayed and is still likely perhaps weeks away.

What will be DeepSeek’s next play?

• Engram: A breakthrough “conditional memory” system that separates factual recall from reasoning. This allows the model to access vast amounts of information (over 1 million tokens) without the usual performance “amnesia.”

• mHC (Manifold-Constrained Hyper-Connections): A technique designed to stabilize training and improve scalability while reducing energy consumption.

• DeepSeek Sparse Attention (DSA): This enables an massive 1 million+ token context window. It allows the model to “read” an entire codebase—thousands of files—in one go, enabling true multi-file debugging and refactoring.

• MODEL1 Architecture: A tiered storage system for the KV cache that reduces GPU memory consumption by 40%, making it much cheaper to run than previous frontier models.

I asked China AI blogger to take a deeper look into DeepSeek’s expected moves in the days and weeks ahead. DeepSeek has been busy behind the scenes the past year and we don’t know its full capabilities as it comes out of stealth so to speak with major releases.

Join Recode China AI for China AI spotlights and deeper resources, trends and stories around the ecosystem.

Recode China AI

China AI Spotlight: Your weekly guide to China's AI breakthroughs, trends, and stories.

By Tony Peng

1. Zhipu AI and MiniMax Just Went Public, But They’re Not China’s OpenAI

2. DeepSeek-R1 and Kimi k1.5: How Chinese AI Labs Are Closing the Gap with OpenAI’s o1

3. The Most Important AI Panel of 2026: Can China Lead the Next Paradigm?

While Western closed-source models double down on AI coding capabilities, Chinese open-source models are showing more agentic, browser and fundamental capabilities. Major IPOs and consolidation is occurring and China’s AI engineering prowess is starting to show ultra competitive results. The convergence of China’s AI ecosystem and AI chip makers hurrying to catch up with the rest of the world in semiconductors and HBM, China’s place in the world of AI is likely on the ascent.

I try my best to give global coverage around AI and also to call upon people with “boosts on the ground” in East Asia. That’s why I respect people like , and so much, they are the ones immersed and connected in the latest AI news in China. You can also get full access to this post soon on Tony’s blog.

But how good will DeepSeek’s latest models be? 🤔

“I believe V4 and R2 will remain among the best open-source LLMs available, potentially even narrowing the gap with leading proprietary models. “ - Tony Peng (former AI reporter for Synced)

That being said let’s try to dig into that anticipation and those key details of what we know before the unveiling event:

👀DeepSeek’s Next Move: What V4 Will Look Like

By Tony Peng

Sparsity is DeepSeek’s sauce to scale intelligence under hard constraints

The widely anticipated DeepSeek V4 large language model (LLM) is expected to meet the public in mid-February, right before the Chinese New Year. V4, DeepSeek’s upcoming generation base model, is supposed to be the largest challenge of open-source models to the proprietary models, and the most anticipated releases in the AI space for early 2026.

According to Reuters and The Information, DeepSeek V4 is optimized primarily for coding and long-context software engineering tasks. Internal tests (per reporting) suggest V4 could outperform Claude and ChatGPT on long-context coding tasks.

2025 was a watershed moment for DeepSeek and the following open-source movement fueled by Chinese AI labs. The release of V3 and R1 upended multiple AI narratives: that only spending hundreds of millions could produce a frontier LLM, that only Silicon Valley companies had talents to train competitive models, that the U.S.-China gap in AI was widening due to chip shortages.

Throughout the rest of 2025, DeepSeek continued churning out notable models, including DeepSeek-V3.2-Thinking and DeepSeek-Math-V2, which won the International Olympiad in Informatics. Yet the widely anticipated DeepSeek-V4 and DeepSeek-R2, which were reportedly slated for release in the first half of 2025, were delayed. According to reports, DeepSeek CEO Liang Wenfeng was dissatisfied with the results and chose to delay the launch.

The Financial Times offered an alternative explanation: DeepSeek initially attempted to train R2 using Huawei’s Ascend AI chips rather than Western silicon like Nvidia’s GPUs, partly due to pressure from the Chinese government to reduce reliance on U.S.-made hardware. The training runs encountered repeated failures and performance issues stemming from stability problems, slow chip-to-chip interconnect speeds, and immature software tooling for Huawei’s chips. Ultimately, DeepSeek had to revert to Nvidia hardware for training while relegating Huawei chips to inference tasks only. This back-and-forth process and subsequent re-engineering significantly delayed the timeline.

Sparsity Through Iteration

DeepSeek’s architectural evolution has been driven by a consistent principle: Sparsity.

While computation can be scaled relatively easily by adding more data, more parameters of the model, and more chips, sparsity is the most straightforward way DeepSeek can scale intelligence under constraints, including compute, memory bandwidth, and chips. These constraints shaped every iteration of their model architecture, from DeepSeekMoE’s initial Mixture of Experts (MoE) through the attention optimizations in V2, V3, and V3.2.

In a Transformer, you have two main computational blocks: Attention and Network (specifically Feed-Forward Network). DeepSeek started by making the network layers sparse using MoE. Instead of every token going through the entire network, they route each token to only a small, relevant subset of parameters as experts.

• DeepSeekMoE laid the groundwork with a MoE architecture containing 64 specialized experts and 2 shared experts. For each input, the model would route it to the 6 most relevant experts (topk=6).

• DeepSeek-V2 expanded the expert pool to 160 specialized experts while keeping 2 shared experts and the same topk=6 routing, with improvements focused on distribution and efficiency.

• DeepSeek-V3 scaled further to 256 experts with 1 shared expert, and increased routing to topk=8. The architecture became more sophisticated in how it managed experts, and introduced a new communication system called DeepEP that makes expert coordination more efficient.

Once the network sparsity was well-optimized, attention became the next target for improvement. As the core module of the Transformer architecture, attention is where each token in a sequence analyzes and weights the importance of every other token to understand context and relationships. Attention presents a different challenge as computational complexity exponentially grows along with the increasing context window.

• Multi-head Latent Attention (MLA) was DeepSeek’s first major attention innovation introduced in V2. Instead of storing complete key-value information for every token (as standard Multi-Head Attention does), MLA compresses this information into a smaller representation. This compression reduces how much data needs to be moved in and out of memory.
  

• Then in a February 2025 paper, DeepSeek introduced Native Sparse Attention (NSA) with optimized design for modern hardware. Tokens are processed through three attention paths: compressed coarse-grained tokens, selectively retained fine-grained tokens, and sliding windows for local contextual information.
  

• DeepSeek Sparse Attention (DSA), introduced in V3.1 and V3.2, streamlined the NSA design. Instead of selecting blocks of tokens, DSA selects individual tokens. It uses a lightweight indexer model to identify the 2,048 most relevant tokens from the full context. This indexer is trained through a process where it learns to mimic the full attention pattern—the model first trains with full attention, then the indexer learns to predict which tokens the full attention would focus on. DSA can work with MLA.
  

DSA+mHC+Engram

With the foundation of DSA established in V3.2, DeepSeek appears ready to push optimization further in V4. The evidence points to two other papers that DeepSeek released over the past few months.

Manifold-Constrained Hyper-Connections (mHC)

mHC represents a rethinking of how information flows through deep neural networks. While traditional networks pass information sequentially from one layer to the next, mHC introduces richer connectivity patterns between layers—essentially creating multiple pathways for information to flow across the model’s depth.

To understand mHC’s, we must first revisit residual connections, the backbone of modern deep neural networks regardless of architecture (CNN or Transformer). Proposed in the landmark 2015 paper Deep Residual Learning for Image Recognition, residual connections add the input (x) of a block directly to its output F(x), typically formatted as y = F(x) + x.

This simple shortcut proved revolutionary. By allowing gradients to bypass layers, residual connections solved the vanishing gradient problem that had plagued deep networks, enabling effective training of architectures with hundreds or even thousands of layers.

Then in 2024, ByteDance researchers proposed hyper-connections (HC) as an alternative approach. Rather than simple additive shortcuts, HC creates richer connectivity patterns that allow information to flow more freely between non-adjacent layers. This architectural flexibility offers advantages over standard residual connections, particularly in avoiding the gradient-representation tradeoffs.

But HC introduced its own challenge. As training scales increase and connectivity grows richer, the risk of training instability rises. More pathways for information flow means more opportunities for gradients to either explode (grow uncontrollably large) or vanish (shrink to near-zero) during training.

This is where DeepSeek mHC’s innovation emerges. The approach treats the model’s parameter space as existing on a high-dimensional geometric structure, a manifold. By imposing mathematical constraints based on this manifold geometry, mHC creates a guardrail system that maintains stable information flow even as connectivity increases.

Think of it as a multi-lane highway with intersections. Vehicles can change lanes, merge, or split—increasing routing flexibility. But the total traffic flow must remain constant and balanced. Information can take different paths through the network, but the overall flow is constrained to prevent either congestion (gradient explosion) or emptiness (gradient vanishing).

The mathematical formulation ensures that information transformations remain well-behaved across the manifold. This prevents the instabilities that would otherwise emerge from unconstrained hyper-connections, enabling deeper and more flexible architectures while maintaining training stability.

For V4, mHC could appear strategic. As DeepSeek pushes toward increasingly sparse and selective processing through DSA, Engram, and other mechanisms, having stable hyper-connections allows these components to interact more effectively across layers.

Engram

Engram, detailed in DeepSeek’s January 2026 paper Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models, introduces a new dimension to model architecture by adding conditional memory to the Transformer through efficient lookup mechanisms.

Traditional Transformer-based LLMs compress all learned knowledge into neural network weights. Whether answering a simple factual query like “Barack Obama was a U.S. president” or solving a complex mathematical proof, the model must route every computation through the same expensive neural processing.

Consider the phrase “New York City.” A standard Transformer must learn that “New,” “York,” and “City” together form a specific entity, then rebuild that relationship through attention computation every single time. This is static knowledge that never changes, yet the model treats it like novel information requiring full neural processing each time.

Engram’s premise is elegantly simple: not all knowledge requires neural computation. Static facts and established patterns can be stored in a complementary memory system and retrieved efficiently when needed. This mirrors biological memory: You don’t re-calculate that 2+2=4 each time; you simply recall it.

Engram implements this through a modernized N-gram lookup module operating in constant time (O(1))—retrieval speed remains constant regardless of how much information is stored. This creates a fundamental architectural separation:

• Neural computation (attention and MoE): Complex reasoning, novel synthesis, context-dependent processing

• Memory lookup (Engram): Static knowledge, established patterns, factual recall

The architecture optimizes the balance using a U-shaped scaling law—a mathematical framework determining ideal parameter allocation between neural computation and memory lookup at different model scales. This is crucial because the tradeoff isn’t straightforward: too much reliance on lookup tables risks brittleness; too much neural computation wastes resources on static knowledge. The U-shaped law identifies the sweet spot where both systems work synergistically.

Memory efficiency improves dramatically because Engram offloads static knowledge from expensive GPU memory to host CPU memory. Empirical results demonstrate that offloading a 100B-parameter lookup table to host memory incurs negligible overhead (less than 3%). This enables massive knowledge bases without proportional GPU memory costs, while supporting much longer effective context windows.

For reasoning and knowledge-intensive tasks, Engram delivers measurable improvements over comparable MoE-only architectures. Benchmarks show particular gains in multi-hop reasoning and long-context understanding—tasks that benefit from quick access to established knowledge while applying neural computation to novel reasoning steps.

Model1: New Clues to DeepSeek V4?

Just over a year after DeepSeek-R1 became the most-liked model on Hugging Face, sharp-eyed developers have spotted a mysterious “Model1” in recent code updates to the FlashMLA library. The timing is suggestive—could Model1 be the codename for DeepSeek V4?

Analysis of recent commits reveals several architectural signatures suggesting Model1 is an entirely new flagship model:

• The 512-Dimensional Shift: Model1 switches from V3.2’s 576-dimensional configuration to 512 dimensions, likely optimizing for NVIDIA’s Blackwell (SM100) architecture where power-of-2 dimensions align better with hardware.

• Blackwell GPU Optimization: New SM100-specific interfaces, CUDA 12.9 requirement, and performance benchmarks showing 350 TFlops on B200 for sparse MLA operations represent deep integration with next-generation hardware.

• Token-Level Sparse MLA: Separate test scripts for sparse and dense decoding indicate parallel processing pathways. The implementation uses FP8 for storing KV cache and bfloat16 for matrix multiplication, suggesting design for extreme long-context scenarios.

• Value Vector Position Awareness (VVPA): This new VVPA mechanism likely addresses a known weakness in traditional MLA—positional information decay over long contexts. As sequences extend into hundreds of thousands of tokens, compressed representations can lose fine-grained positional details. VVPA appears designed to preserve this spatial information even under aggressive compression.

• Engram Integration: References to Engram throughout the codebase suggest deep integration into Model1’s architecture.

Concluding Remarks

I believe V4 and R2 will remain among the best open-source LLMs available, potentially even narrowing the gap with leading proprietary models. However, since DeepSeek R1’s release over a year ago, the competitive race to push LLM capabilities forward has only intensified. The “DeepSeek effect” has motivated several other Chinese AI labs—including Moonshot AI, MiniMax, and Zhipu—to redouble their efforts in releasing top-tier LLMs. This doesn’t even account for tech giants Alibaba and ByteDance, both of which are producing frontier models while simultaneously expanding into chatbots, AI cloud services, chips, and hardware.

Given this rapidly evolving landscape, expecting another watershed “DeepSeek moment” like the one in early 2025 seems nearly impossible.

Where to start? I’m interested in giving global macro coverage of AI trends and that wouldn’t be possible without talking about China’s progress in 2025. is one of my favorite sources: he’s ex-Baidu Global Head of Comms and a former AI reporter; a longtime AI observer with a keen focus on China’s AI development.

If you aren’t premium and want to read his entire 2025 rundown, please subscribe to his Newsletter below. He’ll be publishing it (this deep dive) at end of 2025 in a few days.

Recode China AI

China AI Spotlight: A weekly guide to China’s AI Breakthroughs, trends and stories. By .

Recode China AI

China AI Spotlight: Your weekly guide to China's AI breakthroughs, trends, and stories.

By Tony Peng

Who is ready for another DeepSeek-R2 flagship model launch? I suspect we might have another interesting time of it soon (within weeks).

Subscribe to Recode China AI

Tony’s writing is clear and his range of topics is vast. His coverage of China BigTech advances in AI is itself notable:

1. Meet Tesla FSD’s Fiercest Competitors in China

2. 💪🏻AI Makes Alibaba Great Again

3. 👀TikTok’s Parent ByteDance Quietly Builds Its AI Empire

Highlights of China in AI in the year 2025

Here are some of my notes: open-weight models, agentic AI and reasoning model upgrades takes the prize. China also is early on the AI IPO front with many significant physical AI and humanoid robotics companies.

If you know anyone curious about what China is up to in AI, this could prove helpful:

Share

China managed to do this with little access to leading AI chips and with a fraction of the capital allocation (capex) of the U.S. We do know that the likes of Alibaba and ByteDance are starting to enter the capex territory of the global leaders in AI. China’s progress in artificial intelligence in 2025 is like a global undercurrent of the future.

Some Themes I noticed

⚫ Open-Source Dominance: For the first time, Chinese open-source AI models accounted for 17.1% of global downloads, surpassing the U.S. to rank first in the world.

🟠 Ecosystem Adoption: Massive User Adoption: By February 2025, China’s generative AI user base hit an estimated 570 million, driven by "super-apps" like and new winners like Doubao and DeepSeek. To illustrate the momentum of DeepSeek’s moment - China’s user base of generative AI doubled in the first six months of 2025, a growth rate of 106.6%.

🟢 China’s Six Little Dragons of AI grew up a bit: Emergence of Advanced Reasoning Models that are open-weight: The releases like Kimi K2 (Moonshot AI), GLM-4 (Zhipu AI. now Z dot AI open-sourced GLM-4.7), and MiniMax models closed the gap with US frontiers in complex reasoning, with steady streams of free, high-performance updates attracting developers and builders.

🔵 China showed with startups like Manus and Genspark, they will develop AI agentic product companies. The "AI Agent" Year: 2025 is widely cited as the year AI agents moved from chat interfaces to autonomous task execution (e.g., the "Manus" agent achieving 57.7% accuracy on high-level GAIA benchmarks). China integrated reasoning models into products very quickly, including many catering to Western and global audiences.

🟣 AI chip startups and IPOs: While MiniMax and Zhipu dropped flagship models in December ahead of Hong Kong IPOs. China’s AI chip startups went or are about to go IPO too: Moore Threads Technology Co. saw its stock jump 425% on the first day of trading in Shanghai, followed by the 693% gain for MetaX Integrated Circuits Shanghai Co.’s listing. Cambricon Technologies also went public meanwhile Enflame, Biren and Kunlunxin will be early 2026 most likely as well. Huawei Ascend had about 35-40% of the Chinese market in 2025 as far as I can tell. Bernstein Research thinks this could look like 50% in 2026.

🟡 Sovereign Compute & Chips: Huawei’s Ascend 950/960 series and the return of Kirin chips (Kirin 9020) have begun powering over 50% of domestic data centers, reducing reliance on Nvidia. These are Huawei’s in-house AI chip efforts. Apparently AMD has a new order from Alibaba, so things are getting interesting as Nvidia’s efforts and GPU sales in China have come under many revisions.

⚪ China’s Semiconductor push: China is reportedly to have recently successfully built and is testing a prototype Extreme Ultraviolet (EUV) lithography machine capable of generating EUV light, a major milestone in its push for semiconductor self-sufficiency. The Chinese team, which includes former ASML (Netherlands) engineers recruited with large bonuses and working under aliases, used reverse-engineering on older ASML machines and components salvaged from secondhand markets. The prototype was developed in a high-security laboratory in Shenzhen as part of a six-year, government-led "Manhattan Project" initiative to bypass Western export controls (coordinated in large part by Huawei). While generating light, it hasn't produced working chips yet, with targets set for 2028 (government goal) or 2030 (more realistic estimate) for actual chip output, significantly faster than anticipated. China also has a hoard of humanoid robotics startups and robot prototypes that showed some rapid progress in Physical AI capabilities in 2025 as well. China’s native semiconductor ecosystem is still likely 6-12 years behind TSMC, Taiwan or the likes of Nvidia.

China’s Innovation Market is Dynamic

China’s ecosystem of AI startups, incumbents, academic AI researchers, and AI chip startups is likely more vibrant, manifold and dynamic than in the West where just a few BigAI companies tend to dominate. The physical AI spectrum will become more clear in 2026 as well. With China’s abundant energy grid, if they wanted to scale datacenters they could do so fairly cheaply as compared to the United States. Meanwhile Huawei is estimated to produce roughly 800,000 to 1,000,000 AI chip dies in 2025, with plans to double this output for the Ascend 910 series in 2026. From DeepSeek to Huawei to the incredible progress by Alibaba Cloud’s Qwen team, China suddenly looks like they are on an upwards trajectory worthy of their AI engineering prowess.

🔴 AI Smartphone Surge: AI-powered smartphones (for instance capable of real-time voice translation and on-device editing) now make up over 30% of all shipments in China. China’s AI innovation in mobile is often overlooked and how their super-apps have integrated it is fairly rapid and advanced. In autonomous vehicles, Baidu’s Apollo went global in 2025.

Do you enjoy podcasts?

Was DeepSeek-R1 China’s Sputnik moment?

2025 was the year watching China in AI, robotics and technology became a lot more interesting. Qwen (Alibaba) models basically took over the spot Meta (Llama) had that new foundational open-source models startups use.

I believe a lot of approaches by the United States and China are ultimately complementary for global consumers, software developers and innovation for human civilization.

While many leading U.S. labs like OpenAI and Google focus on proprietary, "closed-source" frontier models, Chinese firms have embraced a more "open" approach, making their models more accessible and customizable for developers. Chinese open-source models power nearly 30% of global AI usage, a significant jump from 1.2% in late 2024. How high can it go in 2026?

Now let’s turn to Tony’s deep dive:

Top 10 China AI Stories in 2025: A Year-End Review

By Tony Peng

China’s open-weight AI strategy took a huge leap ahead this past year.

Go Premium

Listen 🔊

Read more

On August 13th, 2025 we learned that the much anticipated DeepSeek-R2 model would be delayed reportedly due to unsuccessful attempts to train it on Huawei’s AI chips instead of those of Nvidia, according to the Financial Times.

If you enjoy the topic of this article, why not share it with a friend.

Share

Previous articles on DeepSeek

• 🌊 China’s DeepSeek AI Shakes Up the Game (January, 2025)

• Is DeepSeek the new DeepMind? (January, 2025)

• In the Wake of DeepSeek: What Does it Mean For China’s ByteDance, AI Four Dragons and Six Tigers? (March, 2025)

Read More

Denis Kalinin (https://substack.com/@kalinindenis) of DeepTech Asia (www.deeptech.asia) Newsletter explores the latest trends in AI, robotics, new energy, computing and space. His writing focuses on emerging innovation and investment trends across Asia, with a particular emphasis on China. Denis also works as the Head of APAC at a deep tech VC firm Runa Capital. With a M.S. from University for International Business and Economics in Beijing and exchange program experience at Free University of Berlin, Denis was based in Singapore for a time.

Why is DeepSeek generally considered important?

DeepSeek is a significant player in the AI landscape due to its innovative approach to developing cost-efficient, high-performance, open-source AI models, challenging the dominance of Western tech giants and reshaping the global AI race.

Honorable mention (recommended reads related to our topic): , , , , , , , , , , , , , , , , , , , , , and so many others.

DeepTech Asia 🐉

DeepTech Asia

Latest trends in Asia, the world's biggest and most dynamic region that is becoming the powerhouse of the future technologies.

By Denis Kalinin

Pledge your Support to DK

Go Deeper into China & Asia Tech

• Chinese Tech Going Global. Playbook

• China: Private Space Ecosystem of the Rising Superpower

• Demystifying VC Exits in Asia. Data vs Perception

• Would you invest in DeepSeek if you had a chance?

• Edition #1. No Winter Break in the US-China Tech Race

This article was originally timed for a release with DeepSeek-R2, however since it has been delayed at least twice, we decided to share it early.

Was DeepSeek Such a Big Deal?

This article cuts through the media noise around DeepSeek and offers a clear, fact-based analysis of its long-term impact on global AI, open-source, China’s AI, NVIDIA, and beyond.

In January 2025, an unknown Chinese startup, DeepSeek, stunned the global tech community with the release of its R1 reasoning model—delivering breakthrough performance at a fraction of the typical training cost. This led to a record NVIDIA stock sell-off. Leading media outlets erupted with apocalyptic headlines predicting the downfall of OpenAI and the collapse of America's tech dominance.

Fast forward to early September 2025:

• NVIDIA’s stock is up 22% from its “pre-DeepSeek: levels.

• OpenAI not only survived – it raised another $40B in April 2025, marking the largest private tech funding round in history.

• Other American AI players like xAI and Anthropic continue to attract billions in fresh capital too.

• The launch of DeepSeek’s new R2 model has been repeatedly delayed — originally planned for May, it still has no confirmed release date.

So, was DeepSeek truly as revolutionary as the media portrayed it seven months ago — or was it simply a hype-fueled moment of panic?

In this article, we examine the real long-term consequences of DeepSeek for AI infrastructure and applications, both in China and around the world—separating the media frenzy from DeepSeek’s actual impact.

1. DeepSeek vs Global AI

When DeepSeek R1 was released, there was an expectation that major LLM companies like OpenAI, xAI, Anthropic and others would re-consider their existing strategies, considering that companies like DeepSeek can train AI models as good as OpenAI, but at a fraction of the cost.

Moreover, DeepSeek’s models are fully open-source, so now anyone can adjust their architectural innovation and training techniques to make their models as efficient and powerful as DeepSeek.

However, Anthropic expects to burn ~$3B in 2025, xAI – $13B and OpenAI more than $14B.

It was widely expected that DeepSeek would trigger a price war in the global LLM market—after all, its models launched at 50–100x lower prices than those of OpenAI, xAI, and Anthropic. But are these companies burning billions just because they were forced to slash prices?

As the chart below shows, LLM prices have indeed declined since January 2025, but this trend began before DeepSeek at the beginning 2024. DeepSeek just accelerated it.

Even after these reductions, the leading global LLM providers still charge 14–15x more than DeepSeek for its APIs (see table below).

Why hasn’t DeepSeek disrupted the business model of global LLM providers?

1. Product development cost. DeepSeek remains laser-focused on research, raw model performance, and the long-term pursuit of AGI—rather than building tailored products for specific industries or use cases. In contrast, companies like OpenAI and xAI are pursuing a dual strategy: pushing the frontier of model development while also investing heavily in security features, multimodal capabilities, user experience, ecosystem integrations, and industry-specific applications in fields like healthcare, finance, and law.

2. Innovators vs followers. DeepSeek and other Chinese LLMs are still in catch-up mode when it comes to matching OpenAI’s performance. However, the path to achieving comparable results more efficiently—particularly through techniques like model distillation—demands significantly fewer resources than pushing the boundaries of frontier research.

3. Extra deployment costs. Most global companies are expected to adopt the open-source version of DeepSeek, rather than its cloud-hosted offering on Chinese infrastructure—primarily due to concerns around data privacy and regulatory compliance.This raises a key question: Is the total cost of deploying and running an open-source model truly lower than that of using closed-source alternatives—especially at scale, with growing inference volumes? Several recent analyses (e.g. here) explore this trade-off in greater detail.

4. Geopolitics and privacy. DeepSeek’s global market share remains small—currently under 2%. As noted earlier, its cloud-hosted version is generally not considered a viable option for international customers due to privacy and data sovereignty concerns. Even when self-hosting the open-source version is technically and economically feasible, many Western companies remain hesitant, largely due to the perception risks of using a Chinese-developed AI model.

When I surveyed European startups for a comment to Sifted.eu, several asked to remain anonymous—not because of technical limitations, but because they feared potential backlash from customers, despite running the model entirely on local infrastructure.

Impact on open source

In my view, DeepSeek’s biggest long-term impact will be on the global open‑source LLM ecosystem.

On one hand, its rise has pushed previously closed players toward open source. We’ve already seen this in China, where giants like Alibaba and Baidu opened parts of their models, and globally with OpenAI releasing its first open-source model, gpt‑oss.

On the other hand, DeepSeek has raised the stakes for existing open-source developers like Meta’s LLaMA and Mistral. Competition is heating up fast. As Kevin Xu suggested in a recent Substack post, Meta’s next AI model may move away from full open-sourcing, and I tend to agree.

Looking purely at performance, pricing, and speed to market over the past year, it’s hard to see why customers would opt for self-hosted versions of LLaMA or Mistral over DeepSeek or Qwen—unless there’s a strong institutional preference to avoid Chinese technology, as is often the case in sectors like government, defense, or finance.

The growing adoption of DeepSeek among B2B users is also reflected in the rising token consumption on third-party hosting platforms.

II. DeepSeek vs Chinese AI

So far, DeepSeek’s impact has been far more profound within China than on the global AI ecosystem.

China is home to over 200 companies developing their own LLMs, including major tech giants such as Alibaba, Tencent, Baidu, and Bytedance. Alongside them, a new wave of next-generation AI startups—often referred to as the “Six Tigers”—has emerged with a singular focus on advancing frontier LLMs. These include Zhipu AI, Moonshot AI, Baichuan, Stepfun, Minimax, and 01.ai.

The first major shock came on May 6, 2024, with the release of DeepSeek-V2. Its API launched at just $0.14 per 1 million input tokens—a dramatic 5–10x lower than the prevailing market rates at the time.

This immediately triggered a price war. Alibaba quickly slashed the price of its Qwen model from $1.10 to $0.07 per million tokens, and just a week later, Bytedance entered the fray, launching its new model, Doubao, with an aggressively low entry price of $0.04 for the lite version.

By the time DeepSeek’s R1 and V3 models gained global attention in December 2024 and January 2025, the average price of LLMs in China had already fallen by 92% compared to May 2024. Following the R1 release, Chinese providers were forced to cut prices further—though the drop in early 2025 was less dramatic, averaging around 30%.

Source: DeepTech Asia. Based on [1], [2], [3], [4]

The response to DeepSeek’s aggressive pricing was far more intense in China than in the rest of the world — largely because, unlike OpenAI, xAI, or Anthropic, domestic AI companies viewed DeepSeek as a direct competitor. As I explored in my recent piece on China’s strategy for nurturing global champions, price wars are a familiar pattern across most Chinese industries — from solar panels and EVs to consumer electronics.

By January 2025, DeepSeek R1 had become the top-performing LLM in China, but rivals like Moonshot AI, Stepfun, and Baichuan were not far behind, creating a fiercely competitive domestic landscape.

Since January 2025, only Alibaba and Minimax have released new models that rival DeepSeek R1 in performance.

If 2024 was the year of AI price wars in China, then 2025 became the year of strategic pivots. Beyond DeepSeek, Alibaba, and Minimax, only a handful of players — such as Bytedance, StepFun, and Moonshot AI — chose to stay in the game of developing large, high-performance models, while many others began to scale back or refocus their efforts.

Source: DeepTech Asia. Based on [1], [2], [3], [4], [5], [6].

Zhipu AI is focusing on commercial traction through developing enterprise and consumer grade applications ahead of its upcoming IPO. Baidu and Tencent doubled down on integrations with DeepSeek and leveraging its extensive ecosystem. 01.ai started focusing on enterprise-grade solutions powered by DeepSeek, while Baichuan prioritized healthcare applications. Minimax, despite still heavily investing in its foundational model, leverages its video and image generating consumer applications.

I think all these pivots make a lot of sense. DeepSeek positions itself as a research-heavy infrastructure company, rather than a B2C customer facing app. And it clearly can’t do both, considering that the resources of its parent company are not as vast as OpenAI or Bytedance. That is why despite DeepSeek’s growing B2B adoption, its consumer web traffic is decreasing after riding the hype wave.

This level of specialization among Chinese LLM players stands in stark contrast to the global landscape, where leading providers like OpenAI, Anthropic, and xAI typically pursue end-to-end strategies—building both core infrastructure and industry-specific applications.

Another consequence of DeepSeek’s rise that I mentioned above is that it pushed many previously closed-source LLMs in China to adopt partially or fully open-source strategies.

On one hand, DeepSeek has set a new industry benchmark by offering basic LLM infrastructure for free. On the other, intensifying competition in China is likely to push domestic LLM providers to expand globally. To succeed internationally, many of them will need to embrace open-source strategies, as discussed earlier.

In parallel, DeepSeek appears to have improved the overall perception of Chinese AI models—even outside China. For instance, I've observed a growing number of European startups adopting Alibaba’s Qwen, signaling a shift in sentiment and trust.

III. What Comes After DeepSeek

DeepSeek’simpact on Chinese companies building AI applications has also been profound — both directly and indirectly.

Lower infrastructure costs

The most immediate effect has been the LLM price war triggered by DeepSeek’s ultra-low pricing. Unlike their global counterparts, Chinese developers are not limited to open-source versions; they can leverage cloud-based LLMs at a fraction of the cost—a major advantage, especially given the complexities of self-hosting under current U.S. export restrictions.

This creates a unique cost advantage for Chinese AI developers, allowing them to build and scale more efficiently. However, it may also discourage international expansion, since low-cost LLM access is largely restricted to domestic users—or more precisely, to customers comfortable with Chinese cloud infrastructure handling their data.

Better funding environment

Over the past four years, Chinese AI startups have faced a tough funding environment, with venture and private equity investment in sharp decline.

Source: DeepTech Asia, based on ITjuzi

AI funding in China began to show signs of recovery in 2024—likely aided by falling infrastructure costs. However, the real momentum shift came in the first half of 2025, when the number of funding rounds by Chinese AI startups rebounded to levels not seen since early 2022. While the total capital raised hasn't grown significantly, this trend suggests a reallocation toward early-stage startups rather than continued concentration in already well-funded players.

In my view, the "DeepSeek moment" in January 2025 was a key catalyst. Its international breakthrough not only shifted the global narrative on Chinese AI, but also revived investor confidence in the broader ecosystem. Given how closely investor sentiment tracks media perception, DeepSeek’s global spotlight likely encouraged both domestic and international funds—including Temasek, Prosperity7, EQT, and even Benchmark—to re-engage with Chinese AI startups.

Flexible startup culture

DeepSeek is truly unique in its culture that is drastically different from many other tech companies in China like Bytedance or Alibaba. This topic was covered in much detail by JS Tan (here and here) and Jon from Asianometry (here), so I’ve just summarized their points below.

Source: DeepTech Asia, based on Value Added and Asianometry

The key takeaway is that DeepSeek’s founder, Liang Wenfeng, is challenging the conventional belief that Chinese companies need high-profile executives, massive teams, and grueling work culture to compete globally.

In many ways, DeepSeek operates more like a U.S. startup than a traditional Chinese one—emphasizing agility, creative problem-solving, and flexible execution over brute force and hierarchy.

Inspiration to go global

Bytedance was arguably the first “truly global” Chinese tech company, with TikTok reaching around 760 million monthly active users (MAU) in China and 1.8 billion MAU internationally. However, Bytedance was founded in 2012, and its global expansion only began in 2017 with the acquisition of Musical.ly. Its first product, the news app Toutiao, was entirely China-focused.

In contrast, DeepSeek has demonstrated that a Chinese company can be global from day one—not just in market reach, but also in technical competitiveness. It has contributed to a more positive global image of China’s tech ecosystem.

DeepSeek’s success with international users—amplified by strong global media coverage—encouraged other leading players like Manus and Minimax to target global markets early. While DeepSeek wasn’t directly responsible for their success, it undoubtedly helped legitimize their global ambitions and accelerated their credibility abroad.

In contrast to the US or Europe, most AI application startups in China that raised money in the first half of 2025 were hardware-related, developing either robots (Flexiv, Noematrix), drones (i-Kingtec, SkySys), autonomous vehicles (Zelos, Boonray, Yihang) or industrial solutions (Ensonic).

A wave of early-stage AI agent startups has emerged in China following DeepSeek’s breakthrough—many of them raising their first institutional rounds in 2025 and targeting global markets from day one. Companies like Shulex, WaveSpeed, and LynxAI are among this new cohort. I’m sure almost nobody outside of China has ever heard about them yet. Some of these companies might fail quickly, but the others may become the “next Manus”.

DeepSeek’s success has redefined the global ambition of Chinese startups, proving that it’s possible to compete internationally without raising billions or relying on government backing. In interviews, Liang Wenfeng has hinted that this very aspiration—to build a lean, globally relevant AI company from China—was one of his core motivations for starting DeepSeek.

Challenges remain

Read more

April 2nd, 2025. —

• Welcome to Liberation day, where tariffs super-charge the world, trade, macro economics and geopolitics with grave uncertainty!

• However on the AI News spectrum, there’s also a certain measure of uncertainty with regards to all of this AI capex spending on AI chips, infrastructure and pesky closed source models. Is it all worth it in the end?

Does closed-source even win against open-source in the end? Open-source software often benefits from a broader range of innovations due to community collaboration and transparency and those that win on cost, democratization and go-to-market virality may win and end up disrupting incumbents.

Wrong Side of History

“Shortly after DeepSeek’s model was released in January, Altman said that OpenAI was “on the wrong side of history” regarding open models, signaling a likely shift in direction.” - Wired

Outside of scale, generational funding and tons of features in ChatGPT, what is OpenAI’s moat exactly?

Today we have a guest post on DeepSeek’s Open-Source impact by esteemed of the Newsletters The Ontologist, Generation AI, and the Cagle Report, among others. I’ve been reading and talking to Kurt for years.

Whether it was an April’s fools joke or not, the Studio Ghibli stunt was bad enough but yesterday OpenAI says it too is building an “open-source” model. DeepSeek’s downloads really surged in the last quarter. But OpenAI’s, well ChatGPT’s revenue, has surged 30% in just three months.

With the usual vague bravado of a high sales entity and the freakish charm of manipulative Sam Altman, OpenAI says that it intends to release its first “open” language model since GPT‑2 “in the coming months.”

The DeepSeek Threat 🐲

DeepSeek gets more mobile and web traffic than most of OpenAI’s other other rivals, so OpenAI is going after DeepSeek with yet another PR stunt: just don’t call anything OpenAI does, real open-source innovation.

According to the Top 100 Gen AI report by a16z and indeed traffic websites, the radically open-source DeepSeek is the biggest competitor now of OpenAI.

Let's take a closer look at a quick snapshot of web traffic through March, 2025: According to data from analytics firms Similarweb and Sensor Tower let’s take a look at some of the traffic trends around Gen AI chatbots and apps:

• DeepSeek’s chatbot eclipsed 16.5 million visits in March, 2025.

• Gemini’s web traffic grew to 10.9 million average daily visits worldwide in March, up 7.4% month-over-month.

• Anthropic’s Claude reached 3.3 million average daily visits in March.

• Microsoft’s lame Copilot increased to 2.4 million — up 2.1% from February.

Is “AI Eating the World” ?

• In March 2025, openai.com saw 499 million visits, a significant increase of 251 million visits compared to the previous month. In early 2025, AI isn’t eating the world, but OpenAI is.

When such a deceptive company as OpenAI can raise $40 Billion with funds Softbank needs to take out a gigantic loan for, we have a problem on the internet. An AI Infrastructure filled world funded by third party debt financing and Middle East oil money that Softbank is known for isn’t exactly healthy or sustainable.

• Hard to know if OpenAI has capitulated or if this is just another crazy PR stunt.

• What does 500 million weekly active users get OpenAI besides perhaps $12 Billion in revenue in 2025?

• With a $300 Billion valuation, what if competitors in China are building models, apps and utility at a fraction of the price? OpenAI sound a bit confused about how to respond to DeepSeek:

• OpenAI yesterday also posted a webpage inviting developers to apply for early access to the forthcoming model. Altman said in his post that the company would host events for developers with early prototypes of the new model in the coming weeks. OpenAI needs to change their company name when they go for-profit and much of their latest funding round is contingent on their new business structure.

• I believe OpenAI’s ChatGPT is likely to increase its Plus plans prices and introduce Ads soon into ChatGPT.

• If “AI is eating the world”, what OpenAI becomes has relevance to us all and civilization at large.

“With DeepSeek, everyone’s realizing the power of open weights.” - Clem of Hugging Face.

The Ontologist

Principles of Metadata Architecture, Graph Theory and AI

By Kurt Cagle

Top Guest Works

Kurt Cagle is an extremely prolific writer, author, editor and history of tech nerd. He also doesn’t get the credit he deserves imho. Kurt’s a true historian of data science, tech and AI.

• The Taxonomist, the Ontologist and the Linguist

• All You Need is Context

• Why AI Agents Aren't Agents

• What to Study in 2025 If You Want A Job in 2030

And also what might be DeepSeek’s long-term impact on the Open-source vs. Closed-source debate? DeepSeek’s impact on China AI vs. U.S. BigTech and AI? Read our other articles on DeepSeek for more context.

For less than $2 a week get all access.

Upgrade your Subscription

DeepSeeking Missile: Open Source AI

Written in February, 2025 by Kurt Cagle

Kurt Cagle is the editor of The Cagle Report, a newsletter focusing on knowledge systems, AI, cognitive computing, and how it impacts with the rest of the world.

As I write this, the US Tech Market (especially the Nasdaq) is in freefall, as the release over the last week of both the Chinese DeepSeek LLM and a new multimodal version called Janus Pro has sent shock waves through both the AI tech and investment communities. It’s about time.

DeepSeek has a lot of people in a snit right now. It cost $6 million dollars to train the large language model, and both its code base and its weights have been released as open-source. It is available for free. It can run on older nVidia h100 GPUs. To put this into perspective, it costs OpenAI roughly $100 million per model to train a new model, to get to the latest models is now costing users $200 a month, and, neither its source code nor the training weights (which determine the optimizations made to the model based upon its training sets) are available to anyone outside the organization.

Or, looked at another way, OpenAI spends in one week what cost the DeepSeek team a year in training, and it requires far fewer GPUs to run once it is compiled. Oh, and it reasons better than anything that OpenAI has out now or potentially will have in the next six months.

It does have a few limitations. The service, though not the raw model, will not let you ask questions about Tiananmen Square. It is very likely that the service model prompt feeds are regularly reviewed by the Chinese Communist party. OpenAI won’t talk about David Mayer, Alexander Hanff, Jonathan Turley, Brian Hood, Jonathan Zittrain, David Faber and Guido Scorza. Who are these people? No clue, but they have ceased to exist in the ChatGPT world (though David Mayer, at least, may have been jailbroken). OpenAI prompts may be monitored by specific US agencies. No one will confirm or deny this, of course. In other words, if you want to get your knickers twisted about security or bias, don’t use LLMs in the first place.

The Fallout

Read more

“Computing and new media depend upon flesh-and-bone metabolism. Our "virtual worlds" are made possible by battalions of human beings.” 1

— Thomas S. Mullaney

Good Morning,

In today’s world that which integrates to the Cloud, goes open-source. In 2025 we are finding out that China’s application layer nativity might prove a super-power in Generative AI’s go-to-market’s (GTM) near term future.

When Afra Wang (王曌) wrote her debut ‘China in Flux’ essay recently on her new Newsletter about her latest travels in China, I was fairly mesmerized, but she’s been thinking a lot about history, China and digging into AI, and this means DeepSeek as well. She’s a budding sinologist, but a very global citizen.

🧮 Concurrent, explores the parallel and divergent tech/cultural currents flowing through China and the world.

Share

• She is known for her insightful pieces focusing on culture, technology, and the Chinese diaspora. But her writing cadence is very special to my ear and her research on this piece is very multi-layered.

Concurrent

Concurrent explores the parallel and divergent tech/cultural currents flowing through China and the world.

By afra

Before we get into this next Deep Dive in our DeepSeek series, some China AI News to report on. Afra is by the way the host of the podcast "CyberPink," where she discusses topics related to the intersection of culture and technology in contemporary society.

Baidu’s Surprising new Models release (Sunday)

• Chinese search engine Baidu has launched two new AI models — Ernie 4.5 and Ernie X1.

• In the many 🐲 dragons and many 🐯 tigers war of 100 models in China, 🐼 Baidu claims that Ernie X1’s performance is “on par with DeepSeek R1 at only half the price,”, and it touts Ernie 4.5’s “high EQ,” allowing the model to understand memes and satire.

• Baidu plans to release its next generation model, Ernie 5, later this year.

Baidu’s March 16, 2025 surprise announcement on Sunday were pretty vague and I’m fairly hesitant of them comparing themselves to either DeepSeek or Alibaba Qwen.

Other works by Afra:

• DeepSeek and Destiny: A National Vibe Shift

• On Typing Chinese

China follows the West in AI Infrastructure Capex

Since DeepSeek-R1 was released, Tencent, ByteDance and Alibaba Qwen have released a flurry of models. A day before Nvidia’s GTC, Chinese tech giant Tencent has reportedly purchased a batch of new H20s AI chips, according to the Chinese media outlet Calian News. A report from GuruFocus highlights that Tencent and ByteDance are among NVIDIA’s largest clients in China, with both companies acquiring approximately 230,000 NVIDIA chips in 2024. This is quite a big capex by China’s own version of BigTech we are seeing.

As you might know, Alibaba is set to significantly enhance its AI and cloud computing capabilities through a major capital expenditure of over $52 billion by 2025.

Subscribe to Concurrent

Explore the parallel and divergent tech future-history of China’s impact.

Because who is better to help us understand the future, than a historian? From super-app envy to super-app ban? DeepSeek’s impact is getting tangible and contagious. I believe we have indeed witnessed China’s own ‘ChatGPT’ moment. Deconstructing it is easier said than done.

Share

Audio tl;dr version:

DeepSeek and Tencent Bundle: Explained

By (Afra Wang, March, 2025).

As DeepSeek and Manus2 have seemingly emerged from nowhere to threaten Silicon Valley's carefully constructed moats, the most revealing development isn't in benchmark scores but in how these technologies are being integrated into existing ecosystems. OpenAI's formal proposal to ban DeepSeek—labeled as "state-subsidized" and "state-controlled"—reveals more about American anxieties than Chinese realities.

Two months after DeepSeek's global debut, the rapid integration between DeepSeek and Tencent offers a window into how the Chinese technological ecosystem actually operates: not through monolithic state control as OpenAI suggests, but through state-mediated acceleration that creates powerful connections between intelligence capabilities and distribution channels.

The orchestrated dance: DeepSeek, Tencent, and the state

Some online commentary suggests that Chinese internet giant Tencent has emerged as the most nimble responder—and potentially the greatest beneficiary—of DeepSeek's success.3

The lightning-fast integration of DeepSeek and Tencent exemplifies Chinese technological efficiency at its finest. This state-facilitated acceleration promises to create a powerful multiplier effect between AI capabilities and established digital platforms. However, the state's role here differs from Sam Altman's fear-mongering portrayal of "controlling" oversight. Rather than direct control, what we observe is strategic coordination and implicit endorsement—a distinctively Chinese approach to technological development that Western analyses often mischaracterize.

Last month, on February 18th, we know that DeepSeek founder Liang Wenfeng appeared on CCTV sitting beside Tencent's Ma Huateng at Xi Jinping's symposium for top business leaders. In China's political visual language, such seating arrangements carry potent significance—akin to state certification or even endorsement.

The timing was hardly coincidental. Just one day before this high-profile symposium, Tencent's WeChat announced its partnership with DeepSeek. What followed was a cascade of integration: Tencent Cloud announced that more than ten of its products would fully incorporate DeepSeek's LLMs, spanning development tools, enterprise services, and consumer applications. Simultaneously, WeChat began gray-testing DeepSeek-R1 within its search function, incorporating content from official accounts and supporting social sharing.

Several analyses have already illuminated the significance of this partnership for China's tech landscape. But the implications run deeper. Through this symposium appearance, DeepSeek secured its place on the AI throne in national-level forums, earning the tacit blessing of the state apparatus. This blessing, however, comes with implicit expectations rather than explicit directives—a subtle distinction that OpenAI's proposal to the U.S. government conspicuously fails to acknowledge.

This integration effectively transformed DeepSeek into a nationally endorsed product from both technical and strategic perspectives. Overnight, it gained entry into a digital ecosystem with 1.3 billion user base—becoming what Chinese people call a "national-level" app. WeChat’s scale and its super-app model once inspired Mark Zuckerberg and Elon Musk’s "WeChat envy."4 Now, most users don't access DeepSeek through its official app (which constantly crashes) but rather through Tencent's multi-agent platform Yuanbao.

While facilitating DeepSeek's integration, Tencent advances quietly on multiple fronts. Its recent Hunyuan video generation model has garnered international praise specifically for its practical strengths. As one user noted on Reddit: "I was genuinely impressed by its remarkable speed... Hunyuan's standout advantage lies in its robust community support."

This alliance illustrates the unique complexity of relationships between Chinese private enterprises and government—a dynamic that shifts between leveraging government power and maintaining calculated distance. Huawei offers an instructive parallel: as Eva Dou mentioned in a talk about her book House of Huawei, the company's "success stems from a hybrid model that combines strong state support with entrepreneurial drive." Yet historical evidence shows that early Huawei attempted to preserve independence, with founder Ren Zhengfei reportedly declining loan offers from China's then-Premier. Only after U.S. restrictions in the Trump administration did Huawei's relationship with the government grow increasingly close.

Under current geopolitical tensions, Chinese technology companies find themselves with precious little opportunity to shape their own narratives internationally. Facing intense scrutiny and persistent hostility, many are forced to identify as Chinese firms first, tech firms second. This external pressure has pushed entrepreneurs closer to the state out of necessity. Countries such as Australia, South Korea, and Italy banned the access to DeepSeek while it reveals some security vulnerabilities. As one tech consultancy told RestofWorld regarding TikTok's situation:

"If you are a Chinese company, you are often guilty before proven innocent."

As Ivy Yang noted, DeepSeek's explosive reception in the Western world has paradoxically robbed it of the chance to define itself. Founder Liang Wenfeng's studied reticence and media avoidance have created a gray space where observers project their own understanding. This narrative vacancy functions as a mirror reflecting American anxieties: tech hawks see unstoppable Chinese AI ascendance, tech pundits see it threatens free speech, while capital markets see DeepSeek dismantling OpenAI's moats.

But for someone like me—straddling Chinese and American tech landscapes with one foot in each world—the DeepSeek-Tencent integration reveals a distinctly Chinese technological cartography. Here, the state doesn't merely regulate technology; it orchestrates development, deployment, and adoption in ways Western observers often misinterpret as either ominously authoritarian or admirably efficient. Neither characterization captures the complex reality.

DeepSeek in a Deeply Pragmatic Society

Beyond these high-level relationships, what's happening with DeepSeek on the ground tells an equally fascinating story.

In a previous article DeepSeek and Destiny: A National Vibe Shift, I mentioned how DeepSeek's emergence created a sensation among many people, fostering a feeling that "the world order is rapidly changing, and China's moment of destiny is approaching"—partially self-consolation, partially infused with genuine optimism about China’s position in the world. However, delving deeper into China, you'll find that DeepSeek is hardly a topic dominating everyday conversations. The observations from the ground shows fervent discussions and obsession with AI coexist alongside complete unawareness of it. Both realities are valid, equally true.

During the Chinese New Year holiday when DeepSeek made its splash, I crisscrossed the country for nearly a month. My journey took me from ultra-first-tier cities like Beijing and Shanghai to Hangzhou where DeepSeek was born, then inland to the third and fourth-tier cities in Shanxi province. Along the way, I spoke with my cousin (teaching urban planning in Xi'an), a friend leading business development at a state-owned steel company, a high school classmate working in Beijing's energy sector, and family members scattered across the landscape. These ordinary young professionals and middle-class citizens—China's supposed technological vanguard—hardly mentioned DeepSeek. Game-changer? Not on their everyday radar. This is just my snapshot, of course—one American-educated Chinese person's social circle. But it hints at how far AI must still travel before weaving itself into the daily rhythms of Chinese life.

Most telling was my conversation with a friend at a state-owned steel company who dismissed AI with a wave of his hand. "Manufacturing—that's the real game," he insisted, his voice carrying what I would call "Chinese bronze age manufacturing masculinity"—a value system where tangible production still outranks digital innovation. His attitude more or less echoed the seating arrangement at Xi's symposium for tech and business leaders—that carefully choreographed tableau of power and priority so quintessentially Confucian in its hierarchical design. The chart spoke volumes: Huawei, BYD, Xiaomi—companies forging tangible things—commanded the prime positions. Electric vehicles, consumer gadgets, renewable energy, 5G networks—these are the industries China's leadership truly values. This preference is reflected in the "Made in China 2025" plan, which emphasizes manufacturing as the core of China’s strategic plan. Meanwhile on the same table, Tencent and DeepSeek could only occupy positions at the periphery of this seating chart.

To offer more of my on-the-ground observations, Chinese patterns of DeepSeek usage reveal a different relationship with AI than what I see among tech-savvy Americans. Rather than philosophical exploration or technical fascination, the grassroots Chinese approach is marked by utilitarian opportunism—a "catch the wind and profit" mindset focused on immediate practical applications. DeepSeek's most enthusiastic adopters are not just tech elites but ordinary people seeking quick income. This isn't merely pragmatism but entrepreneurial urgency—a recognition that AI represents a fleeting market opportunity to be seized. Content creation represents another DeepSeek stronghold. Scroll through Chinese social media and you'll encounter a flood of AI-generated material—the carefully constructed captions, the too-perfect product reviews, the suspiciously well-structured travel guides. Influencers have embraced DeepSeek's language processing to churn out scripts and posts at unprecedented speeds, leaving my timeline awash with content bearing strong synthetic flavors. Meanwhile, a parallel economy has emerged around "DeepSeek 101" tutorials teaching ordinary users how to transform simple prompts into cash flow.

On Xiaohongshu, people discuss how to use DeepSeek as a social intelligence amplifier, helping users navigate China's high-context, face-conscious society. Young professionals tap its capabilities to craft diplomatic work messages, perfect WeChat replies to their boss's cryptic messages, or formulate graceful rejections to unwanted social obligations. The platform has become an emotional intelligence coach for a generation facing relentless social pressure.

Perhaps most fascinating is DeepSeek's role as digital fortune teller. Users feed it birth dates and personal details, seeking insights about fate, compatibility, and future prospects—a technological revival of ancient divination practices. Caiwei Chen documented this phenomenon in a fascinating report on China's AI oracles.

The professional realm shows differences, too. While American developers publicly celebrate "vibe coding" with AI tools, their Chinese counterparts operate with more discretion and reservation. They prefer shadow arrangements—"DeepSeek+Cursor combinations" deployed quietly in the background. I recently read about China's "shadow AI" confirmed what I've long suspected: in China’s big tech companies, AI coding assistance has become an open secret. Unlike the American tech scene with its abundance of "super devs" doubling as influencers and thought leaders, China's programming culture fosters a heads-down approach. To be clear, China's tech professionals also celebrate AI's potential to automate coding and improve work efficiency, but the environment lacks vibrant discussions and debates around trendy concepts like "vibe coding," which is seen as a pivotal shift in the nature of work. For many Chinese coders, AI represents tools to help them meet intense corporate output pressures.

Municipal Government Adoption

Local governments across China have embraced DeepSeek with characteristic pragmatism—approaching AI as a tool rather than a revolution. The adoption pattern reveals a fascinating duality: on the one hand, genuine administrative efficiency; on the other, technological theater playing out across provinces. As municipalities compete to showcase their AI credentials, adoption metrics become badges of progressive governance in a distinctly Chinese blend of innovation and performance.

Shenzhen's Futian district exemplifies this approach, deploying 70 "AI civil servants" powered by DeepSeek to handle everything from document processing to emergency response and investment promotion. This model has triggered a cascade effect throughout the country. According to the report, in the Yangtze River Delta region alone, cities like Nanjing, Wuxi, Hangzhou, and Shanghai's Pudong district have all announced localized DeepSeek deployments, primarily focusing on document processing applications.

The landscape is now dotted with "AI virtual humans" and "digital government assistants"—some substantive, others resembling the showy joint ventures between tech companies and bureaucracies that have characterized previous innovation waves. What matters isn't always functionality but participation in the race itself. Hangzhou, as birthplace to both DeepSeek and robotics company Unitree, has emerged as the model other municipalities aspire to emulate, further accelerating resource allocation toward AI initiatives.

The contrast with American adoption is interesting: according to this Pew research, US state governments show dramatically varied approaches—some aggressively pursuing AI implementation while others proceed with elaborate caution and ethical frameworks. Approximately 8% remain in the earliest exploratory stages, still developing basic AI literacy within government ranks. This disparity doesn't simply reflect different technological capabilities but fundamentally divergent approaches to innovation governance.

While American discourse about AI often oscillates between utopian technological determinism and dystopian fears of job displacement, the Chinese approach I've observed is grounded—focused on utility rather than grand philosophical implications.

Conclusion

As AI technology splits along geopolitical lines, we need to actually understand what's happening on the ground, not just repeat the binary talk such as ”American democratic AI good and Chinese authoritarian AI bad.” The real danger isn't Chinese tech getting better—it's that misunderstandings like those in OpenAI's proposal might shut down chances for different approaches.

About the Guest contributor

• Afra is a freelance writer and podcaster with working experience in AI and crypto. Personal site here. She just started a newsletter Concurrent, explores the parallel and divergent tech/cultural currents flowing through China and the world.

• She has 70 readers this morning and the emerging writers around here keep lifting my eyes to the world that came before and to the speed of what the world is becoming.

• It’s getting hard to keep up everyone! You know we’re witnessing history, right?

Learn about the Author

Your support makes it possible for me to discover authors like .

1

From afra’s list of quotes. via . Thomas Shawn Mullaney is an American sinologist. He is a Guggenheim fellow. He is professor of history at Stanford University, working on technology, race, and ethnicity in China. He is 47 years old and we admire him. Interview with of The Gradient.

He’s Professor of History and Professor of East Asian Languages and Cultures, by courtesy. He is also the Kluge Chair in Technology and Society at the Library of Congress, and a Guggenheim Fellow. He is the author or lead editor of 7 book. Get alerts if he shows up around here:

Precursor Magazine

The "Hidden Curriculum" of Academia

By Tom Mullaney

Special thanks to Xinrou Shu for her help in researching, co-writing and editing on this piece.

Good…

Read more

Good Morning,

As we approach information overload 🐋 on DeepSeek’s marvel, amid clear explanations of what MoE (Mixture of Experts) is, while others questioning if DeepSeek actually caused Nvidia’s s…

Read more

🎧 Tl;dr Audio Version: 26:02

This week I am collecting thoughts on DeepSeek, the AI story of the year in 2025 so far. Today we feature a great piece by Judy Lin. She’s able to tease out details and connections about DeepSeek not available to most western audiences.

The emergence of DeepSeek, a Chinese AI startup, has significantly disrupted the market dynamics for Nvidia, a leading player in the AI chip industry, including power companies that are most exposed to the tech sector’s data center boom including the likes of Constellation, Vistra and GE Vernova. For the investors among you, take note.

🔥 AI Chips and Datacenter Energy Stocks Plunge

• Nvidia: down 16.86% Monday

• Constellation: down 20.85% Monday

• Vista: down 28.27% Monday

• GE Vernova: down 21.52% Monday

The NASDAQ was hit hard as a whole with other notable names as well down a lot:

• Broadcom: down 17.40% Monday

• Nebius Group: down 37.44% Monday

🌟 DeepSeek Mania and More Releases

Substack is full this week of articles and notes about DeepSeek, and how it might have been inevitable that something like this would have occured. The attention on them has even resulted in a major cybersecurity incident, as their app was trending in most countries.

It’s a shock, $1.2 trillion has been wiped off the US stock markets just yesterday, which felt like a major blow. DeepSeek also released Janus-Pro-7B (image generation)—that was announced in a technical paper shared on DeepSeek’s GitHub page on Monday.

💡 How Good is Qwen2.5-VL

Meanwhile, Alibaba’s Qwen team on Monday released a new family of AI models, Qwen2.5-VL, that can perform a number of text and image analysis tasks. The models can parse files, understand videos, and count objects in images, as well as control a PC — similar to the model powering OpenAI’s recently launched Operator.

For me in the last few months, DeepSeek and (Alibaba Cloud i.e. Tongyi Qianwen) Qwen have been on fairly equal footing in how innovative they are.

Per the Qwen team’s benchmarking, the best Qwen2.5-VL model beats OpenAI’s GPT-4o, Anthropic’s Claude 3.5 Sonnet, and Google’s Gemini 2.0 Flash on a range of video understanding, math, document analysis, and question-answering evaluations.

Read More about Qwen2.5-VL

DeepSeek’s AI models reportedly rival OpenAI’s for a fraction of the cost and compute.

🚨 DeepSeek Impact on Nvidia

The market sell-off on Monday meant Nvidia lost nearly $600 Billion in value.

DeepSeek's recent advancements in AI technology have raised concerns about Nvidia's dominance and market position, leading to a notable selloff in Nvidia's stock and a broader decline in tech stocks. This is has led to a lot of discussions, conspiracy theories and debate on social media. Especially since the U.S.’s surge in AI Infrastructure investments notably beginning in 2024 and increasingly in the area of 40-60% in 2025 by the major players (at least a quarter of a trillion dollars by BigTech alone) will not see significant ROI anytime soon.

AI has fueled Nvidia's extraordinary rise to a $3 trillion market valuation, so to see it lose nearly 17% of its value in just one day was almost unprecedented.

DeepSeek’s impact on AI Infrastructure related companies and capex remains top of mind as BigTech start to release earnings this week. Downloads for the DeepSeek app exploded shortly after DeepSeek released its new R1 reasoning model on January 20th, and it’s full impacts are still unknown. It’s a major developing story and might help define the U.S. vs. China AI race for years to come.

I’ve been in many discussions with other researchers and writers about DeepSeek. of Tech Soda approached me with a well researched piece that I will attach below. It includes many details previously unavailable to most english readers and very deeply researched holistic coverage.

Explore TechSoda

What is SODA?

• “Semiconductor”

• “Optics

• “Digital”

• “AI”

TechSoda

Stimulating, Original, Dynamic, and Accessible Tech News and Insights covering Semiconductors, Optical, Digital, and AI Industries! © 2024 Judy Lin 林昭儀, All Rights Reserved

By Judy Lin 林昭儀

Judy Lin is based in Taiwan so it’s good for us to get yet another perspective from another place in the world and its impact on TSMC. If you are looking for great information on the semiconductor industry as it intersects with technology, TSMC business and geopolitics, she is a great source as one of Taiwan’s top journalists in the area.

🎓 TechSoda - Selected Articles

• NVIDIA’s 2025 CES Keynote: Are We Witnessing the Dawn of a New Era?

• TSMC 4Q24 Earnings Conference Highlights

• TSMC Gains Freedom to Make Most Advanced Chips Overseas as Taiwan Foresees Trade Issue

• Chip Wars Escalate: Can US Counter China's Legacy Semiconductor Surge with Section 301 Investigation?

Let’s jump into her deep dive.

China’s DeepSeek AI Shakes Up the Game

Implication on US Dominance, Nvidia and TSMC

By Judy Lin, January 27th, 2025.

What’s all the Fuss About?

Chinese AI startup Deepseek ignited a round of self-doubt on the US stock markets, sending AI-related stocks to a tailspin. While this pessimism may be overblown, it is a good time to take a deep dive on Deepseek.

Deepseek unviled V3 in December and R1 in January. Now at the World Economic Forum (WEF) and all over the world, it is the hottest topic people are talking about.

Its app has skyrocketed to the top of the U.S. free app charts just a week after its launch. President Donald Trump called the Chinese company’s rapid rise "a wake-up call" for the U.S. tech industry, as its AI breakthrough sent shockwaves through Wall Street.

To train v3, DeepSeek managed with just 2,048 GPUs running for 57 days. The model's training consumed 2.78 million GPU hours on Nvidia H800 chips – remarkably modest for a 671-billion-parameter model, employing a mixture-of-experts approach but it only activates 37 billion for each token.

Editor’s Note:

DeepSeek also claims to have needed only about 2,000 specialized chips from Nvidia to train V3, compared to the 16,000 or more required to train leading models, according to the New York Times. However this claim is disputed.

The Breakthrough Thesis

In comparison, Meta needed approximately 30.8 million GPU hours – roughly 11 times more computing power – to train its Llama 3 model, which actually has fewer parameters at 405 billion.

Some other calculation shows OpenAI o1 costs $15 per million input tokens and $60 per million output tokens, DeepSeek Reasoner, which is based on the R1 model, costs $0.55 per million input and $2.19 per million output tokens.

Deepseek’s open-source reasoning model R1 is on par with the performance of OpenAI’s O1 in several tests (see chart). They built their model at the cost of US$5.6 million, which is only a fraction of the cost of OpenAI’s O1.

Some said DeepSeek-R1’s reasoning performance marks a big win for China, especially because the entire work is open-source, including how the company trained the model. Nevertheless, Deepseek’s extremely low cost and efficiency for training AI models are inviting investigations on how it is possible to spend only US$5.6 million to accomplish what others invested at least 10 times more and still outperform.

Deepseek shattered the impression that the impression that the US was way ahead of China, as it relates to AI, in large part because China does not have access to the most advanced NVIDIA GPUs. ScaleAI CEO Alexandr Wang told CNBC at the sideline of World Economic Forum (WEF) that Deepseek at least have 50,000 Nvidia H100 chips (though it has not been confirmed), also have many people questioning the effectiveness of the export control.

Now, who is behind Deepseek? Why Deepseek is able to achieve such a great result? What are the possible factors that contributed to the success of Deepseek? How should we correctly interpret the implication of Deepseek’s success on the AI competition between the US and China?

Introduction to Journalist

Hi, I am Judy Lin, founder of TechSoda, a news platform that provides refreshing insights to the curious mind. Why soda? It is the acronym for “semiconductor”, “optics”, “digital”, and “AI”.

I am a senior journalist who covers the macroeconomic and foreign exchange market, banking/insurance/fintech, and technology business news in Taiwan for decades. My research interests in international business strategies and geopolitics led me to cover how industrial and trade policies impact the business of companies and how they should respond or take preemptive measures to navigate the uncertainty.

My studies in international business strategies and risk communications and network in the semiconductor and AI community here in Asia Pacific have been useful for analyzing technological trends and policy twists.

Seeing semiconductors become a strategic industry that many countries hold dear in their national security, I try to make my tech articles accessible to people who are not scientists or engineers but also would like to know more about the semiconductor supply chain. If you find my articles interesting and helpful to you, please subscribe to my newsletters and get the latest updates.

In this article, I will discuss the following topics:

1. What and Who: Background of Deepseek and its founder and lead developer.

2. How Deepseek was able to achieve low cost by getting similar performance

3. Does Deepseek’s success mean we don’t need that many Nvidia’s GPUs? What is the impact of decreased demand for GPUs on the foundry?

4. Why is the Deepseek phenomenon significant: US-China AI War going forward?

• Editor’s Note: the background of the CEO and founder seems particularly relevant to DeepSeek’s emergence.

Background of Deepseek, its founder and principal researcher

Founder

Deepseek was founded in July 2023 by Liang Wenfeng, a graduate of Zhejiang University's Department of Electrical Engineering and a Master of Science in Communication Engineering, who founded the hedge fund “High-Flyer” with his business partners in 2015 and has quickly risen to become the first quantitative hedge fund in China to raise more than CNY100 billion.

He grew up in the 1980s in a fifth-tier municipality in Guangdong.

While most Chinese entrepreneurs like Liang, who have achieved financial freedom before reaching their forties, would have stayed in the comfort zone even if they hadn't retired, Liang made a decision in 2023 to change his career from finance to research: he invested his fund's resources in researching general artificial intelligence to build cutting-edge models for his own brand.

“High-Flyer does big models that are not directly related to quantization and finance, and we have established a new company called DeepSeeker to do this. What we want to do is general artificial intelligence, or AGI, and large language models may be a necessary path to AGI, and initially we have the characteristics of AGI, so we will start with large language models (LLM),” Liang said in an interview.

After DeepSeek launched its V2 model, it unintentionally triggered a price war in China’s AI industry. Founder Liang Wenfeng stated that their pricing was based on cost efficiency rather than a market disruption strategy. However, major players like ByteDance, Alibaba, and Tencent were forced to follow suit, leading to a pricing shift reminiscent of the internet subsidy era.

DeepSeek distinguishes itself by prioritizing AI research over immediate commercialization, focusing on foundational advancements rather than application development. Liang emphasizes that China must shift from imitating Western technology to original innovation, aiming to close gaps in model efficiency and capabilities. He believes open-sourcing and ecosystem-building are more sustainable than proprietary models.

Despite financial and resource challenges, DeepSeek remains committed to AGI research, with a long-term strategy centered on mathematical reasoning, multimodality, and language understanding.

Liang believes hardcore innovation will only increase in the future. It’s not widely understood now because society as a whole needs to learn from reality. “When this society starts celebrating the success of deep-tech innovators, collective perceptions will change. We just need more real-world examples and time to allow that process to unfold,” Liang said in an interview in July 2024.

Interestingly, when a reporter asked that many other AI startups insist on balancing both model development and applications, since technical leads aren’t permanent; why is DeepSeek confident in focusing solely on research?

Liang Wenfeng said, “All strategies are products of the past generation and may not hold true in the future. Discussing AI’s future profitability using the commercial logic of the internet era is like comparing Tencent’s early days to General Electric or Coca-Cola—it’s essentially carving a boat to mark a sword’s position, an outdated approach.”

The people they hire don't necessarily come from computer science departments either. Besides STEM talent, DeepSeek has also recruited liberal arts professionals, called “Data Numero Uno”, to provide historical, cultural, scientific, and other relevant sources of knowledge to assist technicians in expanding the capabilities of AGI models with high-quality textual data.

Since its inception, DeepSeek has maintained an organizational culture that is “rank-less and extremely flat”. Members of DeepSeek are divided into different research groups according to specific goals. Instead of a hierarchical relationship, there is a “natural division of labor,” with each member being responsible for the part of the project that he or she is best at and then discussing the difficulties together.

According to Liang, one of the results of this natural division of labor is the birth of MLA (Multiple Latent Attention), which is a key framework that greatly reduces the cost of model training. “MLA was initially a personal interest of a young researcher, but when we realized that it had potential, we mobilized our resources to develop it, and the result was a miraculous achievement,” said Liang.

Principal Researcher - Talent and Recruitment

Liang’s idealism or curiosity alone cannot make it a success; his recruitment standards and management methods are the key, said Feng Xiqian, a Hong Kong commentator. “Liang’s hiring principle is based on ability, not experience, and core positions are filled by fresh graduates and young people who have graduated for one or two years. As for measuring the ability of newcomers, apart from institutional background (mainly Tsinghua and Peking University students), he also looks at competition results and does not use anything below the gold medal – he only recruits 1% of the top geniuses to do what 99% of Chinese companies can't do.”

Luo Fuli, Principal Researcher of Deepseek, is one of the 139 employees that have demonstrated their exceptional talent at a very young age. A Beijing citizen, she has not turn 30 yet but has already published 41 papers since 2018, and was a recipient of China’s National Scholarship in 2016.

Luo got her bachelor’s degree in computer science from Beijing Normal University and a Master of Science degree in Computational Linguistics from Peking University. She got her first job right after graduating from Peking University at Alibaba DAMO Academy for Discovery, Adventure, Momentum and Outlook, where she did pre-training work of open-source language models such as AliceMind and multi-modal model VECO. She joined High-Flyer in 2022 to do deep-learning research on strategy model and algorithm building and later joined Deepseek to develop MoE LLM V2.

She is said to have accepted the CNY 10 million package offered by Xiaomi’s founder Lei Jun just days before Deepseek-V3 was launched.

How Deepseek was able to achieve low cost by getting similar performance

About getting on par performance

I tend to hold a critical thinking position on this. Deepseek was established in July 2023, and OpenAI was founded in 2015. While OpenAI and many western AI companies had to build their generative AI from ground up, Deepseek as a late comer was able to avoid many pitfalls experienced by those predecessors and build on the foundations of open-source contributors.

Without a doubt, Deepseek must have built on the foundation of some open-source database or pre-trained models. This is because inferencing has to rely on pretrained knowledge. As a company which was established only 1.5 years ago, Deepseek was unlikely to build everything from scratch by itself.

Ethan Tu, founder of Taiwan AI Labs, pointed out that open-source models have results that benefit from the results of many open sources, including datasets, algorithms, platforms. And the U.S. is still a major contributor in open source.

“The release of Deepseek on Jan 27 only tells us that the hegemony of AI is not only in the mastery of computing power, but also in the basic skills of investing in software and applications,” wrote Tu. “The technology part (of Deepseek) is worth learning and admiring, but let’s just consider the ‘China surpassing the U.S. or whatever’ rhetorics as marketing language. The marketing is so successful that related stocks plunged today.”

About low cost

1. More efficient in the way it operates

Of course, necessity is the mother of innovation. Not having access to advanced GPUs also serve as a driver for Deepseek and other Chinese AI companies to innovate on more efficient use of computing power.

Due to the US export control, Deepseek has to come up with a more effective way to train the model. So they combined a series of engineering techniques to improve the model architecture, and finally succeeded in breaking through the technological bottleneck under the export ban. Using fewer computing resources to perform complex logical reasoning tasks not only saves costs but also eliminates the need to use the most advanced chips.

DeepSeek V3 introduces Multi-Token Prediction (MTP), enabling the model to predict multiple tokens at once with an 85-90% acceptance rate, boosting processing speed by 1.8x. It also uses a Mixture-of-Experts (MoE) architecture with 671 billion total parameters, but only 37 billion are activated per token, optimizing efficiency while leveraging the power of a massive model.

However, it is noteworthy that Deepseek is an inferencing LLM, not an instructive LLM. Crane Bamboo (鶴竹子), a Wechat blogger, gave a vivid example. ChatGPT needs detailed instructions from a user to accomplish a task. For example, you want it to analyze the energy industry.

More often than not, ChatGPT or any other instruction-based generative AI models would spill out very stiff and superficial information that people will easily recognize it was written by AI.

Crane Bamboo said the right way to use Deepseek is clearly specify how you want the information to be used. For example:

What Deepseek needs is a scenario and specific requests. Therefore, Deepseek is able to come up with more precise answers catered to the needs of the prompter and cut down on the time for scrambling for a vast scope of knowledge. Therefore, having a more targeted scenario and purpose for the information would significantly decrease the computing power required for each task.

When the users are not satisfied with the answers given by Deepseek, which are too abstract, they can prompt “說人話 (speak to me like a person)”. They would immediately rephrase and make the content more easy for people to understand.

From the examples above it is also fair to say that if users have specific scenarios and purposes in mind right at the onset of prompting, that will also boost the speed of generating the content. If it takes less time to process, it would consume less energy, and thus bring down the costs.

Meanwhile, since it is an inference-based system, it is likely to depend on neural networks, which consumes less energy than merely depend on GPUs and CPUs.

Taiwan-based AI startup Kneron claim to offer much cheaper edge AI inferencing solutions based on neural networks. Their applications are focused on smart mobility, smart security, and smart building.

Ethan Tu, a Taiwanese AI expert commented on Deepseek’s performance on using less chips that computing power is important but so is data and algorithms. Hardware is at the front and software is at the back. MoE is not a new idea, it is a trend, and small models will be the future.

2. Energy costs

Some argue that China burns cheaper coals, so the energy consumed by Deepseek is cheaper. Well, not quite. The increased use of renewable energy and the innovations in energy efficiency are key.

According to China’s Energy Transition Whitepaper released by China’s State Council in August 2024, as of the end of 2023, the installed scale of wind power and photovoltaic power generation had increased 10 times compared with a decade ago, with installed clean energy power generation accounting for 58.2% of the total, and new clean energy power generation accounting for more than half of the incremental electricity consumption of the whole society.

The proportion of clean energy consumption in total energy consumption increased from 15.5% to 26.4%, and the proportion of coal consumption decreased by 12.1 percentage points.

Over the past decade, China has eliminated more than 100 million kilowatts of outdated coal power production capacity, and reduced pollutant emissions from the power industry by more than 90%. The rate of electrification of end-use energy in society as a whole has reached 28%. Compared with 2012, energy consumption per unit of GDP has dropped by more than 26%.

“Green energy technology has realized new breakthroughs. It has built a wind power and photovoltaic industry chain R&D, design and manufacturing system, fully mastered large-scale third-generation pressurized water reactor and high-temperature gas-cooled reactor fourth-generation nuclear power technology and is a global leader in hydropower industry chain system.”

3. Government subsidies

Besides the subsidy provided by the central government, the local municipal and provincial governments also have incentives to support AI companies in China. Those incentives include tax breaks, investments, cheap rents for offices located in AI clusters operated by the local governments and talent training programs.

The average salary of AI-related talent freshly out of schools or graduate schools are around CNY15k-25k, which is already considered very well paid in China.

Does Deepseek’s success mean we don’t need that many Nvidia’s GPUs?

Will Nvidia be affected in the short term by the drastic reduction in the cost of AI training? “I don't think so, because when AI can be so popularized and generalized at a low cost, it will only increase the world's demand for it,” wrote Sega Cheng, CEO and co-founder of iKala, a Taiwanese AI company.

Even if the demand for Nvidia’s GPUs decline, Nvidia accounts for less than 15% of TSMC's revenue and less than 10% of global semiconductor revenue. “As far as Nvidia's major customers such as Open AI, Microsoft, Amazon, Google, Meta are concerned, it is unlikely that the GB200/300/Rubin orders that were previously placed will be drastically reduced in the short term, and it will take time to change the training methodology, so it is very likely that the order adjustments will happen in 2026 and beyond,” opined Andrew Lu, a retired investment bank semiconductor analyst based in Taiwan.

The demands for GPUs as a whole may not decrease, but certainly there will be competition among GPU users for the most energy efficient solutions. We will continue to see cloud service providers and generative AI service providers develop their Application Specific ICs (ASICs) to work with their software and algorithms to optimize the performance.

Those chips will continue to be produced by foundries that are most trusted by the customers.

Many research institutions including Gartner and IDC predict that the global demand for semiconductors will grow by 14%-over 15% in 2025, thanks to the robust growth in AI and high-performance computing (HPC). However, TSMC’s chairman and CEO C.C. Wei said recently that TSMC is confident to see CAGR close to 20% in the next five years. The foundry giant is also increasing capital spending again in 2025.

Since TSMC manufactures some 90% of the chips manufactured by 7nm and more advanced processes, which are the chips needed for HPC and AI computing, hence TSMC is likely to continue enjoying higher-than-average growth in the coming years.

The implication of US export control on Nvidia and TSMC in the short run is still likely to influence the location distribution of AI chips made by the two companies. The more important question is, if the trend is moving towards a more software-defined AI computing future, how would it affect the demand for high-bandwidth memory (HBM) and heat dissipation solutions for AI servers?

Why is the Deepseek phenomenon significant: US-China AI War changing directions?

Export control slowed down the catch-up speed of China’s AI technology development, but the win of Deepseek this round is unlikely to make the US give up the policy. The long game for AI supremacy competition is becoming more complex.

Chris Miller, author of Chip War, revealed at the CommonWealth Economic Forum in early January 2025 how AI is transforming the US-China Chip War into a broader "Cloud War." He was right seeing scaling laws falter and efficiency overtakes raw scale. Whether through breakthroughs in inference compute, efficient algorithms, or geopolitical maneuvering, the Chip War is evolving into a broader contest for technological and economic supremacy in the age of AI, said Miller, who also believes tech decoupling is already in place.

Feng thinks Deepseek completely challenged the conventional thinking in Silicon Valley. “From an objective point of view, it is ironic that the U.S. ban has triggered the small universe of these Chinese geniuses, forcing them to innovate with no other choice.”

The success of Deepseek may attract investment capital and talent away from other Chinese AI startups. Meanwhile, who knows if the Deepseek engineers won’t be lured away by other companies? Inner competition among Chinese AI firms have been fierce, and people have no loyalty for employers.

Ethan Tu already noticed that Deepseek is already censoring prompts to make sure their answers are “politically correct”.

“Performance tests for generative AI platforms are like the entrance exams, I am more concerned about the applications and how they are to make a difference in the society and the wellbeing of humanity as a whole,” wrote Tu, who is an AI expert who has been an advocate for the value of democracy.

The rise of DeepSeek AI marks a pivotal moment in the global AI race, proving that innovation can thrive under constraints. While U.S. export controls aimed to slow China’s progress, they may have inadvertently fueled a wave of ingenuity, forcing Chinese engineers to think differently and push efficiency over sheer scale. Yet, this breakthrough is unlikely to prompt Washington to reconsider its policies. If anything, it reinforces the view that the AI rivalry is evolving into a broader "Cloud War," where technological and economic supremacy will be defined not just by hardware, but by intelligence, adaptability, and strategic maneuvering.

However, the road ahead remains uncertain. DeepSeek’s success could spark a surge of investment in China’s AI ecosystem, but internal competition, talent poaching, and the ever-present challenge of censorship cast shadows over its future. As Ethan Tu warns, true AI impact lies beyond mere performance tests—it’s about how these technologies shape society. For now, as the famous Chinese saying goes, "Let the bullets fly a little while longer." The AI race is far from over, and the next chapter is yet to be written.

References:

1. https://www.patreon.com/c/sefirot/posts/

2. https://www.unite.ai/how-deepseek-cracked-the-cost-barrier-with-5-6m/

3. https://www.aminer.cn/profile/fuli-luo/5d50c3c27390bff0db2a8d52

4. https://abg.baidu.com/view/1264d66d25c52cc58bd6be1b?fr=search-income-top3page

5. https://www.facebook.com/andrew.semis.techs?locale=zh_TW

6. https://www.facebook.com/segacheng?locale=zh_TW

7. The implication of US export control on Nvidia and TSMC

8. https://english.cw.com.tw/article/article.action?id=3915&from=search&fbclid=IwY2xjawIFFRBleHRuA2FlbQIxMQABHfybq3DHPD0zTfcdprloTXIcyLa0pifdJFS4hMrrfZ_i9BKFmhhqtYaoBg_aem_BJIP7m8BrbI-JpdH6XBsQg

Happy Lunar 🐍 New Year from Taichung, Taiwan. I may be posting less in the next week or two due to spending more time with family and loved ones.

DeepSeek is trending

Over the last week especially but over the last month generally, the AI Zeitgeist is flooding with what DeepSeek’s R1 means for the larger ecosystem and the future of AI as a whole. See some articles I’m reading on DeepSeek here (Google Doc).

Try DeepSeek

It’s an important moment in so far as everything from export controls to AI Infrastructure, to capex spend or AI talent moats are being put into question. In China, even how they work is facing a stiff test:

“Chinese tech firms are known for their grueling work schedules, rigid hierarchies, and relentless internal competition. By contrast, DeepSeek's flat management structure focuses on empowering its workers with autonomy and creating a collaborative environment.” - J.S. Tan, Source.

Recent articles on AI Supremacy

1. The AI Semiconductor Landscape

2. How to Use Google's Deep Research, Learn About and NotebookLM Together

3. How to use NotebookLM for personalized knowledge synthesis

4. Top AI Tools of 2024

5. Note to Our Energy Sucking Overlords

• In this article (with over 29+ citations) we try to uncover and understand better what is going on with DeepSeek’s models and its story going viral.

• We unpack what it means for the AI industry and in tandem with the AI Infrastructure explosion in capex spending.

• The article curates quotes, infographics and points of view from a variety of sources, meanwhile I also share my own opinions. I may add more posts on DeepSeek from guest writers this week due to the relevance of their work for the entire AI global community.

Subscribe now

Read more