As China heads into the Lunar New Year, DeepSeek has “gone quiet” into “holiday mode.” With its Hangzhou headquarters deserted, the rest of the world ponders the shock and reverberation after the company’s release of its artificial intelligence (AI) reasoning model R1 and non-reasoning model V3. These models perform on par with OpenAI’s o1 reasoning model and GPT-4o, respectively, at a minor fraction of the price. On January 27, the U.S. stock market and tech stocks took one of the biggest tumbles in history, with chipmaker Nvidia falling 18%, losing $589 billion in market value. However, the company’s shares recovered the day by about 9%.

Beyond the upheaval caused to the stock market, the implications for the ongoing AI competition between the U.S. and China continue to unfold. To unpack how DeepSeek will impact the global AI ecosystem, let us consider the following five questions, with one final bonus question.

How did DeepSeek get to where it is today?

DeepSeek began in 2023 as a side project for founder Liang Wenfeng, whose quantitative trading hedge fund firm, High-Flyer, was using AI to make trading decisions. But Liang started accumulating thousands of Nvidia chips as early as 2021. Although Liang, as well as DeepSeek, has been relatively low-profiled and did not give a lot of interviews, in a Chinese-language feature in July 2024, he discussed his technology vision, strategy and philosophy in detail.

Liang was a disruptor, not only for the rest of the world, but also for China. His fundamental belief is that most Chinese companies were simply used to following not innovating, and it was his vision to change that. To him, what China and Chinese companies lack is not capital, but rather confidence and the ability to organize and manage talents to realize true innovations.

While most other Chinese AI companies are satisfied with “copying” existing open source models, such as Meta’s Llama, to develop their applications, Liang went further. His ultimate goal is to develop true artificial general intelligence (AGI), the machine intelligence able to understand or learn tasks like a human being. He decided to focus on developing new model structures based on the reality in China with limited access to and availability of advanced AI processing chips.

The talent hired by DeepSeek were new or recent graduates and doctoral students from top domestic Chinese universities. The company’s organization was flat, and tasks were distributed among staff “naturally,” shaped in large part by what the employees themselves wanted to do. The bottom-up organization of DeepSeek as a startup looked as “Silicon Valley” as it could be, and they appeared to have beaten its real Silicon Valley rivals in the U.S. at their own game. According to benchmarks, DeepSeek’s R1 not only matches OpenAI o1’s quality at 90% cheaper price, it is also nearly twice as fast, although OpenAI’s o1 Pro still provides better responses.

Already, DeepSeek’s success may signal another new wave of Chinese technology development under a joint “private-public” banner of indigenous innovation. In an interview by Liang with Chinese technology news portal 36Kr in July 2024, he said: “We believe China’s AI technology won’t keep following in the footsteps of its predecessors forever. U.S. semiconductor giant Nvidia managed to establish its current position not simply through the efforts of a single company but through the efforts of Western technology communities and industries. The Chinese AI industry needs to create such an ecosystem. Development of domestically-made chips has stalled in China because it lacks support from technology communities and thus cannot access the latest information. That is why China needs people at the forefront of technology.”

Did DeepSeek really only spend less than $6 million to develop its current models?

According to the DeepSeek-V3 Technical Report published by the company in December 2024, the “economical training costs of DeepSeek-V3” was achieved through its “optimized co-design of algorithms, frameworks, and hardware,” utilizing a cluster of 2,048 Nvidia H800 GPUs for a total of 2.788 million GPU-hours to complete the training stages from pre-training, context extension and post-training for 671 billion parameters. The total training cost of $5.576M assumes a rental price of $2 per GPU-hour. The technical report noted that this cost figure excluded “the costs associated with prior research and ablation experiments on architectures, algorithms, or data.”

It should be noted that such parameters on the amount and the specific type of chips used were designed to comply with U.S. export controls released in 2022. According to Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies (CSIS), the total training cost could be “much higher,” as the disclosed amount only covered the cost of the final and successful training run, but not the prior research and experimentation. Also, unnamed AI experts also told Reuters that they “expected earlier stages of development to have relied on a much larger quantity of chips,” and such an investment “could have cost north of $1 billion.” Another unnamed source from an AI company familiar with training of large AI models estimated to Wired that “around 50,000 Nvidia chips” were likely to have been used.

Understandably, with the scant information disclosed by DeepSeek, it is difficult to jump to any conclusion and accuse the company of understating the cost of its training and development of the V3, or other models whose costs have not been disclosed. DeepSeek chose to account for the cost of the training based on the rental price of the total GPU-hours purely on a usage basis. It did not take into account the investment it made to purchase thousands of varying models of Nvidia chips, and other infrastructure costs.

Based on reports from the company’s disclosure, DeepSeek purchased 10,000 Nvidia A100 chips, which was first released in 2020, and two generations prior to the current Blackwell chip from Nvidia, before the A100s were restricted in late 2023 for sale to China. The company also acquired and maintained a cluster of 50,000 Nvidia H800s, which is a slowed version of the H100 chip (one generation prior to the Blackwell) for the Chinese market. DeepSeek likely also had access to additional unlimited access to Chinese and foreign cloud service providers, at least before the latter came under U.S. export controls. Even if the company did not under-disclose its holding of any more Nvidia chips, just the 10,000 Nvidia A100 chips alone would cost close to $80 million, and 50,000 H800s would cost an additional $50 million.

In other words, comparing a narrow portion of the usage time cost for DeepSeek’s self-reported AI training with the total infrastructure investment to acquire GPU chips or to construct data-centers by large U.S. AI companies is neither a fair or a direct comparison. Moreover, such infrastructure is not only used for the initial training of the models — it is also used for inference, where a trained machine learning model draws conclusions from new data, typically when the AI model is put to use in a user situation to answer queries.

While there is no current substantive evidence to dispute DeepSeek’s cost claims, it is nonetheless a unilateral assertion that the company has chosen to report its cost in such a way to maximize an impression for being “most economical.” Notwithstanding that DeepSeek did not account for its actual total investment, it is undoubtedly still a significant achievement that it was able to train its models to be on a par with the some of the most advanced models in existence. Its innovative optimization and engineering worked around limited hardware resources, even with imprecise cost saving reporting.

What will dictate the future of AI development, scaling or more innovative optimization?

The saying, “necessity is the mother of all invention,” has been around for centuries. Facing ongoing U.S. export restrictions to China over technology products and services, China has taken up the urgency resulting from scarcity to escalate its focus and expedite its development efforts.

Contrast the Chinese scenario with the U.S. AI industry, which is already dominated by Big Tech and well-funded “hectocorns,” such as OpenAI. With a valuation already exceeding $100 billion, AI innovation has focused on building bigger infrastructure utilizing the latest and fastest GPU chips, to achieve ever larger scaling in a brute force manner, instead of optimizing the training and inference algorithms to conserve the use of these expensive compute resources. Big Tech and its investors subscribe to the same “big and bigger” mentality, in pursuit of ever-rising valuations and a self-fulfilling loop of perceived competitive advantages and financial returns. Culturally, some argue that even recent “scrappy, disruptive startups,” such as OpenAI a couple of years ago, have already “matured into the kind of big, connected firms that get caught on their back foot by faster-moving rivals.”

What makes DeepSeek particularly interesting and truly disruptive is that it has not only upended the economics of AI development for the U.S. AI industry and its investors, but it has also already done the same to its Chinese AI counterparts. When DeepSeek-V2 was released in June 2024, according to founder Liang Wenfeng, it touched off a price war with other Chinese Big Tech, such as ByteDance, Alibaba, Baidu, Tencent, as well as larger, more well-funded AI startups, like Zhipu AI. On the other hand, compared to Huawei’s foray into developing semiconductor products and technologies, which is often considered to be state-backed, it seems unlikely that DeepSeek’s rise has been similarly state-planned. 

Back in the U.S., contrary to the strong reaction from the stock market, the political response to DeepSeek was rather subdued. President Donald Trump called it “a wake-up call for our industries that we need to be laser focused on competing.” He also said he thought DeepSeek to be “very much a positive development,” because “instead of spending billions and billions, you’ll spend less, and you’ll come up with, hopefully, the same solution.”

There is good reason for the President to be prudent in his response. It was only days after he revoked the previous administration’s Executive Order 14110 of October 30, 2023 (Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence), that the White House announced the $500 billion Stargate AI infrastructure project with OpenAI, Oracle and SoftBank.  The U.S. industry could not, and should not, suddenly reverse course from building this infrastructure, but more attention should be given to verify the long-term validity of the different development approaches.

In the long run, once widespread AI application deployment and adoption are reached, clearly the U.S., and the world, will still need more infrastructure. Some market analysts have pointed to the Jevons Paradox, an economic theory stating that “increased efficiency in the use of a resource often leads to a higher overall consumption of that resource.” That does not mean the industry should not at the same time develop more innovative measures to optimize its use of costly resources, from hardware to energy.

What will be the policy impact on the U.S.’s advanced chip export restrictions to China?

Compared to the swift revocation of former President Joe Biden’s executive order on AI, President Trump has not addressed the issue of the ongoing export restrictions to China for advanced semiconductor chips and other advanced equipment for manufacturing. It is likely that the new administration is still working out its narrative for a “new policy,” to set itself apart from the Biden administration, while continuing these restrictions. Of course, there is also the possibility that President Trump may be re-evaluating these export restrictions in the wider context of the entire relationship with China, including trade and tariffs.

DeepSeek has now put new urgency on the administration to make up its mind on export controls. If Chinese companies can still access GPU resources to train its models, to the extent that any one of them can successfully train and release a highly competitive AI model, should the U.S. redouble these export restrictions? As export restrictions tend to encourage Chinese innovation due to necessity, should the U.S. turn around and remove these controls, and let U.S. companies such as Nvidia profit from selling to China?

First, the fact that DeepSeek was able to access AI chips does not indicate a failure of the export restrictions, but it does indicate the time-lag effect in achieving these policies, and the cat-and-mouse nature of export controls. DeepSeek acquired Nvidia’s H800 chips to train on, and these chips were designed to circumvent the original October 2022 controls. Further restrictions a year later closed this loophole, so the now available H20 chips that Nvidia can now export to China do not function as well for training purpose. However, according to industry watchers, these H20s are still capable for frontier AI deployment including inference, and its availability to China is still an issue to be addressed.

Despite these shortcomings, the compute gap between the U.S. and China would continue to widen due to export controls, a fact cited by DeepSeek as its own primary constraint. The company acknowledged a 4x compute disadvantage, despite their efficiency gains, as reported by ChinaTalk. For the U.S. to maintain this lead, clearly export controls are still an indispensable tool that should be continued and strengthened, not removed or weakened.

Fortunately, early indications are that the Trump administration is considering additional curbs on exports of Nvidia chips to China, according to a Bloomberg report, with a focus on a potential ban on the H20s chips, a scaled down version for the China market. However, the source also added that a quick decision is unlikely, as Trump’s Commerce Secretary nominee Howard Lutnick is yet to be confirmed by the Senate, and the Department of Commerce is only beginning to be staffed. The downside of this delay is that, just as before, China can stock up as many H20s as they can, and one can be pretty sure that they will.

However, on the opposite side of the debate on export restrictions to China, there is also the growing concerns about Trump tariffs to be imposed on chip imports from Taiwan. Last year, Taiwan’s exports to the U.S. rose 46% to $111.3 billion, with the exports of information and communications equipment — including AI servers and components such as chips — totaling for $67.9 billion, an increase of 81%. This increase can be partially explained by what used to be Taiwan’s exports to China, which are now fabricated and re-exported directly from Taiwan.

In a speech on the same day as the stock market crash because of the DeepSeek news, Trump addressed the House Republican Issues Convention and complained about companies “leaving us and [going] to Taiwan,” and said that he would place tariffs of up to 100% “on foreign production of computer chips, semiconductors and pharmaceuticals to return production of these essential goods to the United States.” If this really happens, it would severely harm U.S. companies such as AMD, Apple, Nvidia and Qualcomm that procure from Taiwan’s TSMC and others. These companies will undoubtedly transfer the cost to its downstream buyers and consumers. For the U.S. AI industry, this could not come at a worse moment and may deal yet another blow to its competitiveness.

While bringing back manufacturing to the U.S. will take years to realize, even with the right policies, levying excessive tariffs can hurt the industry and cause inflation immediately. One would hope that the Trump rhetoric is simply part of his usual antic to derive concessions from the other side. Indeed, Taiwan’s Premier Cho Jung-tai has responded to Trump’s comments, saying that the government would urgently consider making more cooperative plans and future assistance programs for the industrial sector.

Are there concerns about DeepSeek’s data transfer, security and disinformation?

Isaac Stone Fish, CEO of data and research firm Strategy Risks, said on his X post that “the censorship and propaganda in DeepSeek is so pervasive and so pro-Communist Party that it makes TikTok look like a Pentagon press conference.” Indeed, with the DeepSeek hype propelling its app to the top spot on Apple’s App Store for free apps in the U.S. and 51 other countries, DeepSeek has quickly become the TikTok for those who think of themselves as technically savvy and “know what they are doing.”

DeepSeek is not hiding that it is sending U.S. and other countries’ data to China. Its Privacy Policy explicitly states: “The personal information we collect from you may be stored on a server located outside of the country where you live. We store the information we collect in secure servers located in the People's Republic of China.” In its terms of use, it also clearly says: “The establishment, execution, interpretation, and resolution of disputes under these Terms shall be governed by the laws of the People's Republic of China in the mainland.”

What type of data may be at risk? In addition to all the conversations and questions a user sends to DeepSeek, as well the answers generated, the magazine Wired summarized three categories of information DeepSeek could collect about users: information that users share with DeepSeek, information that it automatically collects, and information that it can get from other sources. This information will include personal information provided by the users during registration, the users’ text or audio inputs and prompts, all uploaded files, chat history, and keystrokes tracking, etc.

As is often the case, collection and storage of too much data will result in a leakage. Cloud security firm Wiz Research recently discovered an “exposed database leaking sensitive information, including chat history” from DeepSeek, with over a million lines of log streams with “highly sensitive information.” The company informed DeepSeek, which “promptly secured the exposure.”

Another area of concerns, similar to the TikTok situation, is censorship. Numerous reports have indicated DeepSeek avoid discussing sensitive Chinese political topics, with responses such as “Sorry, that’s beyond my current scope. Let’s talk about something else.” This shouldn’t be a surprise, as DeepSeek, a Chinese company, must adhere to numerous Chinese regulations that maintain all platforms must not violate the country’s “core socialist values,” including the “Basic security requirements for generative artificial intelligence service” document. Companies are required to conduct security reviews and obtain approvals before their products may be launched.

AI security tool builder Promptfoo tested and published a dataset of prompts covering sensitive topics that were likely to be censored by China, and reported that DeepSeek’s censorship appeared to be “applied by brute force,” and so is “easy to test and detect.” It also expressed concern for DeepSeek’s use of user data for future training.

Besides concerns for users directly using DeepSeek’s AI models running on its own servers presumably in China, and governed by Chinese laws, what about the growing list of AI developers outside of China, including in the U.S., that have either directly taken on DeepSeek’s service, or hosted their own versions of the company’s open source models? AI search company Perplexity, for example, has announced its addition of DeepSeek’s models to its platform, and told its users that their DeepSeek open source models are “completely independent of China” and they are hosted in servers in data-centers in the U.S. and EU countries.

According to cybersecurity company Ironscales, even local deployment of DeepSeek may still not completely be safe. First, without a thorough code audit, it cannot be guaranteed that hidden telemetry, data being sent back to the developer, is completely disabled. The protection of sensitive data also depends on the system being configured properly and continuously being secured and monitored effectively. In other words, it is difficult to ascertain the absence of any “backdoors” without more thorough examination, which takes time. Moreover, there is also the question of whether DeepSeek’s censorship may persist in a walled version of its model.

The present hype for not only casual users, but AI firms across the world to rush to integrate DeepSeek may cause hidden risks for many users using various services without being even aware that they are using DeepSeek. For developers to “securely experiment,” DeepSeek-R1 is now available as an NVIDIA NIM micro-service preview. Similarly, it is also now on the model catalog on Microsoft’s Azure AI Foundry and GitHub, and Microsoft claims they have put DeepSeek-R1 to “rigorous red teaming and safety evaluations, including automated assessments of model behavior and extensive security reviews to mitigate potential risks.”

But for casual users, such as those downloading the DeepSeek app from app stores, the potential risks and harms remain high. Tests have shown that, compared to other U.S. AI models, it is relatively easy to bypass DeepSeek’s guardrails to write code to help hackers exfiltrate data, send phishing emails and optimize social engineering attacks, according to cybersecurity firm Palo Alto Networks. Another security firm, Enkrypt AI, reported that DeepSeek-R1 is four times more likely to “write malware and other insecure code than OpenAI's o1.” A senior AI researcher from Cisco commented that DeepSeek’s low-cost development may have overlooked its safety and security during the process.

Also, according to information reliability firm NewsGuard, DeepSeek’s chatbot “responded to prompts by advancing foreign disinformation 35% of the time,” and “60% of responses, including those that did not repeat the false claim, were framed from the perspective of the Chinese government, even in response to prompts that made no mention of China.” Already, according reports, the Chief Administrative Officer of the U.S. House of Representatives issued a notice to congressional offices that “DeepSeek is under review by the CAO and is currently unauthorized for official House use.”

What is President Trump’s attitude, regarding the importance of the data being collected and transferred to China by DeepSeek? Recently, commenting on TikTok, Trump downplayed its potential threats posed to U.S. national security by doubting whether it is “that important for China to be spying on young people, on young kids watching crazy videos.” Will he be as lenient to DeepSeek as he is to TikTok, or will he see higher levels of personal risks and national security that an AI model may present?

On the other hand, European regulators are already acting because, unlike the U.S., they do have personal data and privacy protection laws. The Italian privacy regulator has just launched an investigation into DeepSeek, to see if the European Union’s General Data Protection Regulation (GDPR) is respected. Given that DeepSeek openly admits user data is transferred and stored in China, it is very possible that it will be found to be in violation of GDPR principles. In fact, the DeepSeek app was promptly removed from the Apple and Google app stores in Italy one day later, although the country’s regulator did not confirm whether the office ordered the removal. Separately, the Irish data protection agency also launched its own investigation into DeepSeek’s data processing.

Did DeepSeek cheat? 

In the days following DeepSeek’s release of its R1 model, there has been suspicions held by AI experts that “distillation” was undertaken by DeepSeek. Finally, on January 29, the Financial Times reported that OpenAI confirmed that it had seen “some evidence” of distillation, which “it suspected to be from DeepSeek.”

Distillation, or “knowledge distillation,” is a machine learning technique where knowledge from a large, pre-trained model, the “teacher," is transferred to a smaller, more compact model, the “student.” The goal is to enable the student model to perform like the teacher but with reduced or limited computational resources. While the technique is well-known and common, OpenAI forbids any of its users from using distillation to build a rival model, according to its terms of use, as in using “output to develop models that competes with OpenAI.”

According to Bloomberg, Microsoft’s security researchers observed activities of exfiltration of large amounts of data using OpenAI’s application programming interface (API), which were only available to OpenAI users under paid licenses, in the fall of last year. Microsoft, one of OpenAI’s major partners and investors, notified the company, with the information that the activities were suspected to originate from DeepSeek.

Some users have also reported that DeepSeek’s AI models have often responded to queries saying it is “AI developed by Microsoft,” or “built on OpenAI’s GPT-4 architecture.” The model’s “thoroughness and insistence” about “its own identity as Microsoft Copilot” may point to the kind of data the DeepSeek models have absorbed from OpenAI during training.

On January 28, David Sacks, the White House AI and crypto czar, said in a Fox interview that there was “substantial evidence" that DeepSeek “distilled the knowledge out of OpenAI’s models.” He went on to also say that he expected in the coming months, leading U.S. AI companies will take steps to “try and prevent distillation” to slow down “some of these copycat models.”

OpenAI confirmed to Axios that it had gathered “some evidence” of "distillation" from China-based groups and is “aware of and reviewing indications that DeepSeek may have inappropriately distilled” AI models. OpenAI will work closely with the U.S. government to protect its technology and its “most capable models.”

Will such allegations, if proven, contradict what DeepSeek’s founder, Liang Wenfeng, said about his mission to prove that Chinese companies can innovate, rather than just follow? In recent days, the Chinese government, specifically the Zhejiang Provincial Committee Publicity Department, also jumped on the DeepSeek bandwagon and published an article touting the company’s innovation, confidence, composure, and the trust in its young talent. The allegation of “distillation” will very likely spark a new debate within the Chinese community about how the western countries have been using intellectual property protection as an excuse to suppress the emergence of Chinese tech power. The Chinese technological community may contrast the “selfless” open source approach of DeepSeek with the western AI models, designed to only “maximize profits and stock values.” After all, OpenAI is mired in debates about its use of copyrighted materials to train its models and faces a number of lawsuits from authors and news organizations. OpenAI said last year that it was “impossible to train today’s leading AI models without using copyrighted materials.” The debate will continue.

Since AI models can be set up and trained rather easily, security remains critical. Despite recent advances by Chinese semiconductor firms on the hardware side, export controls on advanced AI chips and related manufacturing technologies have proven to be an effective deterrent.

Finally, what inferences can we draw from the DeepSeek shock?

First, the U.S. is still ahead in AI but China is hot on its heels. David Sachs, the U.S.’s AI and crypto czar, has acknowledged that Chinese companies are “catching up very fast," as “the DeepSeek-R1 model is basically comparable in capabilities to the OpenAI o1 model,” which came out about four months ago. He positions the U.S. in a mere three to six month lead on the Chinese. Although DeepSeek has drawn a lot of attention from recent and upcoming advances in U.S. AI models in the past weeks, OpenAI’s announced but yet to-be-released o3 model is expected to deploy enhanced problem-solving capabilities and improved logical reasoning. The DeepSeek episode may well turn out to be a much-needed alarm and reminder for the U.S. AI industry to invest to improve the safeguarding of its intellectual property and enforcing the rules of the game as related to data training for models.

Second, the genie is out of the bottle.  No longer can U.S. or any AI companies overly rely on brute-force scaling. Even if it can be proven that DeepSeek did engage in distillation that violated OpenAI’s terms of use, there are still ample undeniably and remarkable technical innovations made. From improved chain-of-thought (CoT) reasoning capabilities to the novel reinforcement learning (RL) approach adopted, DeepSeek revealed various cost-effective training measures. Thanks to DeepSeek, similar innovations may finally catch the attention of more Silicon Valley investors. Marc Andreessen, U.S. investor, posted in X that “DeepSeek R1 is AI's Sputnik moment.” Such a realization should be shared by government leaders, as well as investors, and developers. If the U.S. can see this as its “Sputnik moment,” China has the same right and much of what it takes to seize their “Sputnik moment,” too. The race for achieving better and more advanced reasoning and optimization techniques for AI training has arrived with increased urgency.

Third, the U.S. government and industry must realize that China has a thick playbook to disrupt the U.S.-led AI ecosystem, and it is flexing its muscles. Not only does it possess the talent pool to train and develop extremely competitive and advanced AI models domestically, it even manages to wipe out $1 trillion of U.S. market value in one day, including close to $600 billion of Nvidia’s value. DeepSeek-R1 was “coincidentally” released on the same day as President Trump’s inauguration. For China’s intent and purpose, it is probably not just about disrupting the technology and the industry but also disrupting the market and the economy of its adversary.

Fourth, the U.S. must redouble its effort in supporting fundamental research and talent development. Ritwik Gupta, AI policy researcher at the University of California, Berkeley, pointed out that “China had a much larger pool of systems engineers than the U.S. who understand how to get the best use of computing resources and run models more cheaply.” This is not just about achieving results with limited resources. Over the last few years, China has focused its manpower development effort with domestic universities and benefited from more technology talents staying in China rather than going abroad, during and after COVID. Part of that is a consequence of more hostile U.S. education and immigration policies toward foreign students and talents. In terms of support for academic and non-commercial AI research, the AI Executive Order of the Biden administration, which included support for the pilot program to implement the National AI Research Resource (NAIRR), has been revoked by the Trump administration. Furthermore, Trump’s attempt to freeze federal grants covered “everything from AI research to semiconductor manufacturing,” including NAIRR under the National Science Foundation, as well as many university research projects. Even the $3.87 billion investment by Korean semiconductor company SK Hynix in West Lafayette, Indiana, was put on pause due to federal funding freeze ordered by the President, because it’s tied to a smaller $2.1 million federal grant to upgrade the local municipal infrastructure to support the new plant. Fortunately, the President rescinded the freeze after just one day, but the constant risk of executive oscillation — from immigration to tariffs to federal assistance — will continue to cast huge uncertainties for the U.S. government’s support for technology leadership.

Fifth, DeepSeek is also disrupting its Chinese AI competitors and may contribute to restructure the future AI ecosystem of China and the world. The fact that a small subsidiary of a local financial services startup was able to make such breakthrough in engineering has given the country’s government, industry and people a huge boost of confidence. Many more government support initiatives are still in the pipeline, with coordination from other critical sectors, including financing from state-owned banks. The Chinese government appears to encourage such fierce competition in AI domestically. DeepSeek has already touched off a price war among AI models in China in the past year. Alibaba has also just released a new version of its AI model and claims that it surpasses DeepSeek’s benchmark performances. On a different note, AI analyst Alexander Doria reported that while DeepSeek’s models were said to have been trained on Nvidia chips, it relied on Huawei’s Ascend 910C chips for inference to generate responses. If this is true, DeepSeek users all over the world, including U.S. users, may be generating huge demands and utilization for Huawei’s chips, making contributions to the Chinese AI and semiconductor ecosystem. This should be yet another policy concern to U.S. policymakers.

Finally, the U.S. will need to develop a response to China’s open source strategy on AI and other technology areas. DeepSeek and China, pride themselves on the open source philosophy, and, partly because it is not even a tech startup but a “side project” of a quant trader, it claims that its mission is not profit. The company, to an extent, picked up the non-profit mantle right where OpenAI abandoned it. China will contrast its “openness” with the U.S.’s proprietary models. This is a welcome narrative, fitting for a new iteration of China's AI Digital Silk Road strategy, which should be music to the ears of developers and governments in the Global South, hungry for affordable AI technologies that they can leverage to utilize their homegrown talents to develop domestic tech sectors and local applications. China’s concept of open source technology is not limited to AI; however, it remains tightly controlled by its government, making their “open source” platforms and technologies susceptible to vulnerabilities and manipulation. If the U.S.’s strategy is only focused on supporting Big Tech investment, it will risk losing the world market further to China. A repeat of the unintended incubation of Huawei looms, only potentially at an even bigger scale. There must have been a good reason for OpenAI to have started out as a non-profit initiative. Now is a good time to revisit that root.