Identifying and evaluating early stage startups is challenging for leaders in corporate innovation, corporate strategy, venture capital, and more. Download the report to learn who made it on our list of top 100 startups.
Below is an Edge Insight Report diving into the latest news and developments on DeepSeek. Typically this is reserved for our members, but we're sharing this report publicly due to its significant disruptive impact.
DeepSeek has disrupted the AI market landscape with its latest reasoning model, R1, challenging conventional development approaches and rivaling Western AI leaders like OpenAI and Anthropic. The company’s innovations have also sparked significant market and industry reactions, raising questions about the need for high-performance chips in training and inference, a shift that particularly impacted NVIDIA, which has heavily benefited from the GenAI boom.
This report examines how DeepSeek leveraged existing techniques, such as reinforcement learning and mixture-of-experts (MoE) architecture to enhance performance while significantly reducing development costs. It also explores industry reactions, including NVIDIA’s historic stock drop, the immediate responses from Silicon Valley companies, and the broader implications for AI innovation and geopolitical tech competition. Finally, we provide commentary from several subject matter experts, who offer strategic perspectives on the potential impact of this event.
What exactly happened?
Reasoning models represent the next frontier of large language models (LLMs) led by OpenAI’s o1, its upcoming o3, and Google’s Gemini 2.0. In this rapidly evolving space, DeepSeek's latest reasoning model, R1 (launched on 20th January), has disrupted Western GenAI market dynamics. R1 was based on the architecture of DeepSeek’s earlier model, V3, which was developed in just two months for ~USD 5.6 million (based on an estimated rental price of USD 2 per GPU hour for the NVIDIA H800—a nearly two-year-old NVIDIA GPU). In contrast, according to Anthropic CEO Dario Amodei, Claude 3.5 Sonnet cost a “few USD 10 millions” to train. Additionally, R1’s open-source nature enables local deployment, giving companies greater control over data security and mitigating privacy concerns.
DeepSeek was established in 2023 by Liang Wenfeng, co-founder of High-Flyer, a quantitative hedge fund specializing in AI-driven strategies. The AI startup is said to have originated from High-Flyer’s AI research division in April 2023 to focus on developing LLMs and advancing artificial general intelligence (AGI).
Timeline of DeepSeek’s model releases
Price and performance analysis of select models
What makes R1 special?
R1 stands out for its open-source approach and its innovative integration of existing LLM technologies, resulting in improved performance at a significantly lower cost. For instance, OpenAI charges ~USD 60 per million output tokens for o1, whereas DeepSeek offers R1 at just USD 2.19 per million output tokens (as of January 30, 2024). This significant cost difference makes advanced AI capabilities more accessible to a broader audience. Moreover, unlike traditional LLMs that depend on supervised fine-tuning, R1 uses a reinforcement learning-based method (the model learns to perform a task through trial and error and without any instructions from a human user), allowing for autonomous chain-of-thought reasoning, self-verification, and reflection. Additionally, by using cold-start data before applying reinforcement learning, R1 addresses common challenges like repetitive outputs and poor readability, while maintaining high efficiency. It also uses NVIDIA's assembly-like Parallel Thread Execution (PTX) programming instead of NVIDIA's CUDA (which has been the industry norm) for some functions, enabling the company to implement fine-grained optimizations like register allocation and thread/warp-level adjustments.
R1 also benefits from the V3 model’s "mixture-of-experts" (MoE) architecture. Unlike traditional models, like GPT-3.5, which use all parameters for inference, MoE models divide the network into multiple specialized submodels or "experts," activating only the most relevant ones for a given task. As a result, the V3 model activates only 37 billion of the 671 billion parameters during interactions, making DeepSeek's R1 computationally efficient and highly scalable.
Moreover, DeepSeek has further advanced the democratization of R1 by releasing distilled versions of it based on Meta's Llama and Alibaba’s QWEN architectures. The team generated 800,000 training samples using the full R1 as a teacher model, carefully transferring its reasoning patterns and capabilities into more compact architectures. These versions make the model's advanced reasoning capabilities accessible to users with limited computing resources, such as those using standard computers or even smartphones. Additionally, by releasing these distilled versions under the MIT license, DeepSeek ensures that the R1 series supports commercial use (including modifications and derivative works) and helps companies and individual developers who might otherwise be priced out of using such sophisticated AI systems.
What was the reaction to this release?
1. NVIDIA sees the biggest one-day loss in US history
DeepSeek's launch showed that powerful AI models can be built without always needing the newest and most expensive hardware, as they achieved impressive results using two-year-old GPUs. This discovery affected the tech market, particularly NVIDIA, which saw its stock drop 17% on January 27, losing around USD 600 billion in market value—the biggest one-day loss in US history. While the stock bounced back by 7% the following Tuesday as investors saw a buying opportunity, it fell again by 6% on Wednesday with news of possible new US restrictions on sales to China.
2. OpenAI, Meta, and Anthropic react
The response from US companies has been mixed—Meta set up "war rooms" to analyze DeepSeek’s models, aiming to understand their efficiency with minimal resources. OpenAI, meanwhile, alleges that DeepSeek leveraged its models via knowledge distillation. While many dismiss concerns over a threat to US AI leadership as overstated, Anthropic CEO Dario Amodei supports tighter chip export controls to China. He also views DeepSeek’s advancements as part of AI’s evolving learning curve, where continuous innovations and shifts in paradigms, such as moving toward reinforcement learning from supervised learning, are enhancing model efficiency.
What do experts say?
The following are some "expert opinions" on this development from AI industry professionals, sourced from our sister platform Speeda’s Experts Network Service. This service connects over 80,000 professionals globally to our Speeda clients, giving them a platform to use their expertise in shaping strategic decisions and advancing market research across industries.
Key takeaways from expert opinions
DeepSeek's success underscores a shift from hardware reliance to optimized training methods, demonstrating that AI models can deliver high performance with fewer resources. This challenges the conventional belief that AI models require immense compute power to succeed. However, this approach isn't entirely new or proprietary—it’s an optimization that other companies, particularly those with expertise in model efficiency, could readily adopt.
Businesses can also leverage open-source models like LLaMA, Falcon, or Mistral, fine-tuning them to meet specific needs while minimizing reliance on proprietary systems.
China's advancements in AI, driven by DeepSeek's progress, bolster its competitive position and could trigger tighter trade and technology restrictions. While the US may respond with additional tariffs or export controls, China's ability to innovate with existing resources poses a challenge to the effectiveness of such measures.
Expert opinions
Expert
Opinion
Senior Strategic Trade Analyst - US
The shocking developments last week of DeepSeek surpassing the AI computing progress on a few key AI tests have stoked much debate in the high technology realm and among governments. China is now poised to have more leverage regarding export controls and potential tariffs. Why? Because Chinese industry and academia have now shown that they were able to reach never-before-seen AI computational levels without the use of US-designed and Taiwan-produced advanced semiconductors or ASML/European lithography. The immediate effect is global investors are rethinking pouring substantial sums into Western (esp. American) AI firms when Chinese firms have shown they can accomplish the same metrics at far less investment. This is the main reason stock markets in the US and Europe opened on Monday to notable price corrections for several of the blue-chip AI stocks such as NVIDIA, MSFT, and AMD. This breakthrough was made all the more dramatic just a few days after the Stargate venture was announced by President Trump between ChatGPT (MSFT), Oracle, and SoftBank of Japan to the tune of USD 100 billion (and potentially USD 500 billion with additional investors). It would seem the PRC is now in the AI driver's seat.
Policy Analyst - Canada
For the short term, it seems that DeepSeek's allegedly low-cost AI model has given China's AI industry a leg up in the tech competition with the US. If the actual data regarding the training cost of the model is accurate, it means that the company can compete with US counterparts like OpenAI at a significantly lower cost. Also, it has been reported that DeepSeek has used 2,000 of NVIDIA's H800 chips, which complies with the US export control policy. To what extent the export control policy, which forbids US companies from selling advanced AI chips and chip-making equipment to China, will continue to be effective is questionable for now.
DeepSeek has affected Washington's strategic calculations on the AI race with China. So more restrictive policies can be expected from the White House. For instance, Trump has just announced that tariffs will be imposed on imports from Taiwan, including semiconductors and other source materials for chip production. But, it's also time for the US to reassess its sanction tools. Companies like DeepSeek can find innovative ways to sidestep policy constraints, which could potentially contribute to China's self-reliance in the tech space.
Assistant Professor in Public Policy - US
It depends on whether or not you think DeepSeek is hiding anything. They release their models to the public, meaning anyone in any country can use them, so in that sense, there is no net impact on geopolitics. If, however, you believe they have proprietary models they only share with the Chinese Government, then the implication is that US export controls are not as effective as believed. The export controls assume that the best models require the best hardware, which ignores the software side of AI development. DeepSeek has shown that software improvement, making the learning algorithms more efficient, is still a frontier and can circumvent hardware restrictions.
The biggest advantage it confers is that it shows that China can produce world-class intellectual property. DeepSeek's staff all have their PhDs from Chinese institutions. In conjunction with China's dominance of EVs and solar technology, the West is coming to the realization that China is becoming an economy of global leaders, not followers.
AI Research Engineer - US
DeepSeek’s sudden progress undermines the assumption that only deep-pocketed labs with advanced GPUs can maintain an AI edge, allowing China to bypass US hardware restrictions and making American export controls look less effective. In the near term, that shift challenges Western firms that rely on proprietary, high-compute models, and it could encourage more cross-border AI projects to tap DeepSeek’s low-cost, open-source approach. Over time, a maturing Chinese open-source ecosystem might lock more developers worldwide into Beijing-centric standards, thus diminishing the influence of US tech giants and strengthening China’s hand in any tech-related tariff or trade negotiations. As a result, American policymakers may tighten chip export rules further or reshape them, but this could spur additional algorithmic innovation in China.
IT Manager - Greece
China will gain a competitive advantage in AI, which will raise concerns about AI use (or rather misuse) for military purposes, autonomous weapons etc. Also, it will raise competition with the US, which may have side effects—The race for fast development/implementation may lead to overlooking some areas, like exhaustive testing, biases, or ethical concerns.
Advantages for China: the huge population, meaning access to huge amounts of data, also combined with the weakest data privacy regulations. Governmental support, financial aid and other benefits, and also access to the highest talented and skilled executives. This advantage of China can harden international collaboration, especially with US or other Western countries, because of intellectual property concerns, security, export controls.
DeepSeek's emergence will strengthen global competition, which will result in trade imbalances and can also put pressure on data privacy regulations, investments, and also international collaboration in other areas, not just AI and technology.
The US will also continue investing in AI and the race will continue, definitely disregarding serious side considerations, in their effort to gain advantage.
Software Architect - US
Immediately, I see DeepSeek's advancements challenge the dominance of Western/US AI developers. This intensifies competition for talent, resources, and market share in the AI sector. It also suggests significant progress can be made with fewer resources than previously thought. This could empower smaller players to challenge the established giants. I also think this news will lead to a fast pace of innovation in AI. Long term, if DeepSeek and other Chinese AI companies continue to advance rapidly, it could shift global AI leadership from the US to China, impacting tech, defense, and economics. There may be impact on international collaboration and raise concerns for IP theft, data security, privacy, ethics, and military applications.
As for strategic advantages, DeepSeek's ability to develop AI models with low cost could give China an edge in deploying AI solutions quickly. They also have access to large data sets due to China's population and government policies. This is crucial for model training.
At the same time, DeepSeek's rise could result in stricter export controls on AI tech, especially military apps. West could limit the export of AI chips, SW, and expertise. It could also play a role in trade negotiation.
Professor and Consultant - US
It is early to speculate as to the long-term repercussions of DeepSeek, in that it is unknown whether its development was the result of chips that China has access to or whether it resulted from stored chips that were put into use after export controls began.
What is certain is that in the short run, this is a major "coup" for China's technological advancement and verification that its model of high-tech development works pretty efficiently right now. For the US, this means more funding for AI at the state and private sector level, more mergers and consolidations in the sector (ChatGPT will want to appear as the unparalleled rival to DeepSeek soon), and certainly even more export controls to try and curb China's capacity to "translate" imports into productive civilian/AI use, as well as bans on technology that could be used for military purposes.
The link with tariffs, however, is unclear and seems to me improbable. Trump is ready to cut a deal with China, has stated that, and DeepSeek's breakthrough has gained his personal admiration (as a symbol of the country's capacity to do "big" and to do it quickly). Negotiations over tariffs are more affected by TikTok instead.
Entrepreneur, Innovator, Public Speaker - US
What should be noted in DeepSeek's recent rise is not only the fact that an AI model was trained using less capital; it did so in spite of the US' crackdown on chip access. The reason it is so significant is the US believed that it could stifle Chinese innovation by precluding access to the necessary chips; what DeepSeek showed was it could maximize the H800 chips (equivalent to the H100s) using better models. This is the leap-frogging China is good at. Short term: There could be a reduction of VC capital distribution to US AI companies in an effort to force them to create better, more efficient (in cost and energy consumption) AI tools. Long-term: China has an advantage right now; they have a strong government that supports innovation on top of VC access, whereas the US President just shut down all funding. AI is a long-term game and disruptions to funding could create a significant disparity in product development.
There is also conflict starting in the US administration, as seen when Musk was angered when Trump announced an AI plan with Sam Altman. None of this is conducive to facilitating progress in the US AI race. If I were to bet, I would bet on China.
CEO of AI Company - US
The ongoing DeepSeek Saga has been widely misinterpreted in the context of market dynamics and geopolitical competition. From the outset, large language models (LLMs) were destined to become commoditized, shifting the battleground toward specialized, industry-specific applications. The moment open-source models like LLAMA entered the scene, the true competition moved to the vertical application layer. Success in this space will be defined by those who master unique workflows, leverage proprietary data, and refine algorithmic advantages. The US didn’t dominate the dot-com era through sheer computing power but through groundbreaking business models like Uber and Airbnb. Similarly, the AI race will be won by those who create impactful applications using foundational models, not merely by those focused on infrastructure.
CEO of Defense and Space Manufacturing Tech Company - Netherlands
Actually, I ran some experiments with DeepSeek, and I am not unimpressed. Since the start of the AI “hype,” I have been a critical thinker and it appeared to be the only trending thing in business, far overvalued. On this, if it is good or not, I think DeepSeek helps us to give a healthy correction to market values as a wake-up call. There are many other sectors than AI that could deserve investment, and this is a signal for that.
As for strategic advantage, I just don't see it as that, apart from that it could catalyze US aggression even more, isolating North America. This of course could mean bonds between China and Europe would be reinforced, but then there is the Taiwan conflict.
The bottom line is DeepSeek helped to humble the AI market, which was needed, creating chances for markets forgotten. Strategically, it just levels out the field for AI at the moment, easing US domination. This can be either good or create more instability depending on governmental reactions. For now, the US is acting extremely unpredictable. For doing business, the EU is currently the safest harbor.
Technology Consultant, Writer, and Researcher - Netherlands
The work of DeepSeek is impressive, but it's not entirely game-changing (as the recovery in NVIDIA's stock price suggests). Its primary innovation has been narrowing the field of queries to bring massive savings in cost efficiency. Now that approach has been proven to barely diminish (and sometimes even improve) response quality, other developers will be able to easily replicate their results (e.g., OpenAI).
What has changed is the awareness that all nations will be able to afford to develop AI models. Rather than global AI dominance, DeepSeek has shown AI will not be a new nuclear weapon that only a few states can afford and wield.
The US still retains the most advanced chipsets for cutting-edge research. DeepSeek's work was built on that foundation, not wholly original.
The export controls on chipsets will continue to give the US an innovation edge for their models, but the limitations on China pushed their innovation in an ironically democratizing direction. This is unlikely to have a significant impact on tariff negotiations. Trump's approach to tariffs isn't good for anyone, including everyday US people, and DeepSeek's arrival won't make an already bad approach much worse.
Jehan has over six years of experience in capital markets and industry research, currently covering FinTech, Web3, and Machine-Learning at SPEEDA Edge. Previously, he was a Financial Analyst at a leading investment research company. He holds a BSc in Accounting and Finance from the University of London and is an ACCA member. In his free time, Jehan enjoys playing football and immersing himself in video games.
