On January 27, a tremor rocked Wall Street as widespread concern erupted over the significant investments in artificial intelligence (AI) by American firms in light of a disruptive innovation from the Chinese startup DeepSeekThis day witnessed a catastrophic drop in the market capitalization of U.Stech companies, leading to what some have termed an “epic crash,” with losses exceeding a staggering $1 trillion.
The fallout was immediate and profoundNotably, shares of Nvidia plummeted by 17%, erasing nearly $600 billion in market valueSimilarly, Broadcom faced a decline of over 17%, resulting in a loss of about $200 billionThe Philadelphia Semiconductor Index experienced a decline exceeding 9%, marking its steepest drop since March 2020.
Not stopping there, Taiwan Semiconductor Manufacturing Company (TSMC) saw its stock value decrease by 13%, pushing its total market capitalization below the $1 trillion mark
Advertisements
Discussions began to swirl in Silicon Valley as major AI players scrutinized DeepSeek’s technology to ascertain whether it represented a genuine breakthrough in independent research, or if it was merely building upon existing Western foundational models.
On the heels of these events, DeepSeek launched their latest visual model called Janus-Pro shortly thereafterThis advanced iteration of the previously released Janus model significantly enhanced multi-modal understanding and text-to-image command complianceImpressively, Janus-Pro surpassed stalwarts like DALL-E 3 and Stable Diffusion across multiple benchmarks.
The disruptive wave that emerged following the late 2023 release of OpenAI’s generative AI application, ChatGPT, had already spurred a global frenzy of investment into the AI sectorAs leading American tech giants leaned more heavily into AI, they found their funding appealing to capital markets
Advertisements
This created a conducive environment for Silicon Valley firms to thrive, leading to astronomical valuationsOpenAI, for instance, soared past a valuation of $150 billion, while Elon Musk's xAI recently garnered investments valuing it at $50 billionHowever, these companies continued to burn through cash without a clear pathway toward profitability.
At a private dinner last year in Silicon Valley, a founder of a leading American AI firm was queried about the differentiating factors and potential "moats" his company possessed compared to others building foundational modelsIn a light-hearted manner, the founder from another AI enterprise joked, “He has a moat; no one else has raised billions like he hasThat’s his moat.”
DeepSeek’s emergence has unsettled traditional market perceptions primarily because it is challenging the entrenched notion in Silicon Valley that only immense financial investment can create advantages—or moats—in the competitive landscape
Advertisements
The prevailing view posits that significant capital allocation is necessary to acquire advanced chips from NVIDIA and to erect large-scale data centers to host them.
However, what is shocking is that DeepSeek’s R1 inference model was developed using less than $6 million in initial computing power alongside a generation of AI chips that was becoming increasingly obsolete in the marketTo put this in perspective, the budget of $6 million would scarcely suffice to cover the annual salary of a senior executive at one of America’s tech behemoths.
The dynamics of the “AI power game” appear to be shifting“The power dynamics in AI are rebalancing; the big companies may not necessarily be the guaranteed winners,” remarked one investor to reporters from a leading financial publication.
The collapse in the capital markets also spotlighted a significant oversight within the American AI sector regarding Chinese technological advancements, with few paying attention to DeepSeek
- US Stocks Continue to Rise in January
- Nasdaq Surges Over 2%
- US-Canada Trade Spat: Will Oil Prices Surge?
- Bank of Japan Resumes Interest Rate Hikes
- FOF Asset Allocation: Strategies for Steady Progress
Over the past year, many Silicon Valley titans have become too engrossed in their self-serving “bigger than bigger” race in the development of large models and have not devoted enough attention to converting these innovations into thoughtfully designed products that truly serve businesses and consumersThe rise of DeepSeek could represent a pivotal moment in redefining the competitive landscape within the AI industry.
One investor observed, “It wasn't until the release of R1 that real attention started to focus on DeepSeekIt not only aligned with ChatGPT in terms of ranking capabilities in AI models but also showcased the ability to execute reasoning chains, winning over public favor.”
Currently in Silicon Valley, executives and technical staff from numerous AI firms are hastily analyzing DeepSeek’s published research and technology to decipher the mechanisms behind their model
Their investigations center on whether DeepSeek’s results signal an independent technological breakthrough and to what extent the company relies on the foundational models established by American corporations.
According to industry insiders, the likes of OpenAI and Meta are conducting systematic research into DeepSeekThey are beginning to realize the urgent need to take the progress of Chinese companies seriously, as such advancements could also present opportunities for U.Sfirms to innovate and enhance their existing models.
Demis Hassabis, CEO of Google’s DeepMind, noted at the recent Davos Forum that DeepSeek’s models were somewhat “unexpected,” yet he expressed uncertainty about the model's inner workings and its dependency on the results generated from other American companies’ models.
While the market expressed trepidation, the consensus within the industry largely remains one of recognition for DeepSeek’s innovations
Their success has illuminated new prospects for the development of what is referred to as the “measuring tape law.” Just months prior, this concept faced challenges, with members of the tech industry fearing that artificial intelligence capabilities were nearing a plateauYet, DeepSeek’s experiences suggest that solutions to these limitations can indeed be found.
This points to the conclusion that the future victors in AI will require not only raw computational power but also innovative approaches to enhance efficiency within limited resources—something that the giants of Silicon Valley seem to have overlookedMeanwhile, tech giants in the U.S., including Google and Meta, continue to splurge on acquiring cutting-edge AI chips from NVIDIA.
Aidan Gomez, CEO and co-founder of AI startup Cohere, remarked, “The future of large models belongs to those who focus on efficient technologies rather than merely accumulating more compute power.” He further asserted, “We have always believed in this, and DeepSeek has served as the catalyst that has sparked a significant response across the entire sector.”
NVIDIA countered that DeepSeek’s work signifies “outstanding advancements in AI.” The company added, “DeepSeek's research illustrates how one can leverage widely available models and comply fully with export control regulations to build new models.”
The technology firm emphasized that reasoning processes typically require a substantial number of NVIDIA GPUs and high-performance networking
“Currently, we advocate for three laws of expansion: the persistent applicability of pre-training and post-training laws, along with a new testing-time expansion law,” NVIDIA remarked.
Greg Allen, director of the Wadhwani Center for AI at the Center for Strategic and International Studies, stated that DeepSeek has effectively surmounted the so-called “interconnect speed barriers” within chips, allowing them to construct their models—a consequence of the initial failure of recent export controlsHe cautioned that a new round of stricter chip export regulations implemented in October 2023 will impose greater constraints on DeepSeek’s ability to scale and continue developing their models, which require significant computational strength.
DeepSeek's founder and CEO, Liang Wenfeng, highlighted in a previous interview, “Money has never been our problem