Over the years, OpenAI has had its fair share of competition and panic from Anthropic, xAI, Google, and others. However, nothing had ever brought the giant and the entire American AI ecosystem to their knees, as DeepSeek has.
The following should serve as a primer to everything you need to know about the DeepSeek drama – all at once.
DeepSeek Just Beat OpenAI’s Best Model
DeepSeek’s latest reasoning model, R1, has outperformed OpenAI’s o1, the company’s most powerful model available for public use. On multiple benchmarks, DeepSeek R1 scored higher than o1 and is on par with the rest.
Source: DeepSeek R1 Report
DeepSeek R1 is available for open-source use under an MIT license. Unlike OpenAI’s o1, it displays all the steps it uses to reason. It is also available for free on the DeepSeek app for the web, iPhone/iPad, and Android.
DeepSeek Beat ChatGPT to be the #1 App
DeepSeek’s official app has dethroned OpenAI’s ChatGPT and other competing AI apps in the US App Store’s Top Charts for iPhone and iPad. According to an AppFigures report four days ago, DeepSeek’s free app was downloaded over a million times from Apple’s App Store and Google Play Store.
Over $500 Bn Wiped Out from NVIDIA
Recently, NVIDIA’s market cap dropped by around $589 billion in a single day. Is DeepSeek to blame? The model achieved superior performance with as few resources as possible, raising concerns about the amount of capital and computing resources needed to build powerful AI models.
DeepSeek prioritises using efficient techniques in the model’s architecture for improved performance rather than relying on high levels of computing power.
One of DeepSeek’s previous models, V3, used 2048 NVIDIA H800 GPUs to achieve performance better than most open-source models. Andrej Karpathy, former OpenAI researcher, said the DeepSeek V3’s level of capability is ‘supposed to require clusters of closer to 16,000 GPUs’.
DeepSeek trained the model for a mere $5.5 million. Last year, a technical paper revealed that the most expensive publicly announced training runs to date are OpenAI’s GPT-4 at $40 million and Google’s Gemini Ultra at $30 million.
Moreover, both Meta and xAI have revealed that they’re using more than 100,000 GPUs or more to train their upcoming models.
While DeepSeek officially revealed that it used NVIDIA H800 GPUs for DeepSeek V3, the company did not reveal the GPUs used for R1.
The NVIDIA H800 was a GPU designed for the Chinese market, bypassing the then-US export controls. Its data transfer rate was 50% lower than that of the NVIDIA H100.
Soon after, the export of H800 was banned as well.
Alexandr Wang, CEO of Scale AI, said that DeepSeek has about 50,000 NVIDIA H100 GPUs – but they cannot disclose the same due to US export controls.
DeepSeek is Nothing But a Side Project
High Flyer, the Chinese hedge fund company that developed the AI model, was founded in 2015. It started work on the DeepSeek project in 2023.
Meanwhile, Han Xiao, CEO of JinaAI, said the company owns “a lot of GPUs” for trading purposes and that “DeepSeek is their side project for squeezing those GPUs”.
“I heard people say they were running DeepSeek LLM as a side project bc [because] of leftover GPU,” he added.
He also credits the CEO with the model’s success, describing him as a ‘low-key guy’ who is smart, has no ego hassles and is always engaged in learning without wasting time on public exposure.
“They [the founders] spent years in quant [quantitative analysis] – where the community values leverage and efficiency much more than headcount. And one person can and should manage 7-digit dollar portfolio without panic. So ‘lean and mean’ is deeply rooted in their culture,” he added.
An Inspiration for the United States?
US President Donald Trump said, “The release of DeepSeeek AI from a Chinese company should be a wake-up call for our industries,” adding that he sees DeepSeek’s ability to produce an AI model using cheaper methods as a positive.
OpenAI CEO Sam Altman was impressed too. “DeepSeek’s R1 is an impressive model, particularly around what they’re able to deliver for the price,” he said.
Meanwhile, NVIDIA issued a statement after the market bloodbath, saying, “DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely available models and compute that is fully export control compliant.”
Some believe that US restrictions on China backfired, leading them to create DeepSeek. Amjad Masad, CEO of AI-enabled coding platform Replit, said on X, “The Chinese [have] innovated a way to train large models for cheap. Regulators never consider second-order effects.”
This also serves as an example for leading companies to not be complacent. Many also speculate if Trump was right in removing previous president Joe Biden’s executive order (EO) on AI that sought stringent measures on AI safety, which was deemed to slow down progress.
David Sacks, the White House AI and crypto czar, pointed out the same, “DeepSeek R1 shows that the AI race will be very competitive and that President Trump was right to rescind the Biden EO, which hamstrung American AI companies without asking whether China would do the same.”
DeepSeek Isn’t Flawless, After All
However, beneath the shine, flaws remain. DeepSeek’s AI models are subject to censorship, which prevents them from answering controversial questions related to China. We also tried asking a question about Arunachal Pradesh, the conflicted Indian state, but the model did not give an answer.
However, this is not a problem with using DeepSeek on first-party mobile and web apps, not if you plan to deploy the model locally. The model’s open-source nature allows developers to remove restrictions and modify it according to their needs.
Moreover, data collection concerns have also come to light. Luke de Pulford, a human rights activist, observed in DeepSeek’s privacy policy that it “collects your IP, keystroke patterns, device info…and stores it in China”.
But again, the model can be downloaded and used locally, eliminating the concerns of data sharing via the internet.
Another problem DeepSeek faced was a large-scale cyber attack. This led the company to temporarily stop onboarding new users. Moreover, it was also affected by outages on its website, but DeepSeek did resolve those issues soon after.
Not Done Yet
That’s about all the major developments DeepSeek has made in the past few days, but the company isn’t done yet. It has announced Janus Pro, an AI image generation model that is claimed to offer better results than OpenAI’s DALL-E 3.
It will only get interesting if DeepSeek enters the video generation competition – so Sora and Google Veo, better watch out!
