DeepSeek, Stargate and the Global AI Arms Race: China’s Pressure vs. U.S. Response
The U.S. and China are locked in a silent AI battle. Learn how DeepSeek, Moonshot AI, and the ambitious Project Stargate are driving innovation, shaping the next decade of Artificial Intelligence
Hello and welcome back to my cosy space of Tech Trendsetters where we draw parallels between science, tech, business, and well, pretty much everything that catches my attention in this space. Today, I want to talk about something that's been on my mind lately – DeepSeek's emergence and what it tells us about the pressure China is putting on the AI development scene, and how the US is fighting back.
I often see people getting overly excited with the AI tools that emerge in a market, but almost nobody is aware that it's not only the West – or I dare to say United States – who has ambitious plans for AGI/ASI (Artificial General/Super Intelligence). But also the East. And let me be realistic: they are very much on par with US in this race. What's more interesting is that they might have a more pragmatic and potentially more successful approach to AI development than many realize.
The Chinese perspective on AI development
To understand China's AI strategy, we need to start with one of its most influential voices: Kai-Fu Lee.
For those unfamiliar (me included until I fell down this particular rabbit hole), Lee is a unique bridge between Eastern and Western tech worlds. He's held executive positions at Apple, Microsoft, and Google, founded China's prominent tech investment firm Sinovation Ventures, and now leads Zero One, one of China's leading AI companies. His perspective is particularly valuable because he understands both Silicon Valley's approach and China's technological ambitions from the inside.
In a recent revealing interview, Lee outlined what I believe is China's dual-track approach to AI dominance, and it's far more sophisticated than most Western observers realize. Let me break it down:
Plan A: The Direct Competition
China isn't shying away from the AGI race, but they're realistic about the challenges. Kai-Fu Lee points out that while AGI development by 2030 is likely, the path there requires:
Astronomical investments in hardware and model development;
Massive scaling of computational resources;
Breakthrough advances in fundamental AI models;
The candid assessment from Lee is that the U.S., particularly through companies like OpenAI, might win this direct race. As he puts it:
“The first company to create AGI and crush its rivals will inevitably become a commercial hegemony monopolist”
He specifically points to OpenAI as potentially becoming "the biggest monopolist in history." And he is not mistaken (we will prove it later in the episode).
Plan B: The Ecosystem Strategy
But here's where China's approach gets interesting. Rather than putting all their eggs in the AGI basket, they're building what Lee calls a "moat" through three key strategies:
Cost-Effective Innovation
Unlike U.S. companies that "use resources to attract very smart PhDs" and give them unlimited resources to experiment, Chinese companies focus on making world-class models with cheap inference costs. As Lee explains, "We thought from the beginning that our goal was not to spend money on the world's most expensive AGI, but to make a world-class model that must have cheap reasoning."Application Ecosystem Development
China is betting heavily on building a robust ecosystem of AI applications. Lee points out that "When it comes to implementation, China can catch up with or even surpass the United States." This isn't just boasting – he cites how China's mobile internet applications eventually surpassed U.S. alternatives.Accessibility Focus
Perhaps most importantly, China's vision includes making AI truly accessible to everyone. As Lee notes, "What does 'available to everyone' mean? If you create OpenAI in the United States and don't let Chinese people use it, then it's not available to everyone." I must admit, he has a point.
The Pragmatic Advantage
What makes this Plan A/B approach particularly clever is its pragmatism. And if you know me, common sense is what I really value. While U.S. companies hire the brightest minds and burn through GPU resources creating impressive but hard-to-commercialize results (Lee compares it to trying to fit a luxury kitchen in a tiny apartment). Chinese companies are building scalable, commercially viable solutions. You can read more about these pragmatic solutions in one of our previous episodes:
This strategy acknowledges a crucial reality: even if China doesn't win the race to AGI, they'll have built something equally valuable – an AI ecosystem that could maintain technological independence even under U.S. AI hegemony. As Lee puts it, "If AGI is bound to happen, let's say seven years from now, when OpenAI 'rules the world', and it wants to crush us, we at least have room to resist."
In my view, this dual-track approach demonstrates a level of strategic thinking that many in the West might be underestimating. While Silicon Valley focuses on the moonshot of AGI, China is building both the infrastructure and ecosystem that could define the practical implementation of AI technology for years to come.
Aren’t They Just Afraid?
You might be wondering – is this just China's way of admitting they can't win the AGI race? Let me break down what this actually means in practical terms. China's goal isn't to build the world's most expensive AGI – instead, they're pursuing two clear objectives:
Creating fundamental models optimized for cost-effective inference;
Focus on reducing operational costs rather than just pushing model capabilities;
Prioritizing efficient resource utilization over raw performance;
Building models that can be practically deployed at scale within the real, existing industrial infrastructure;
Making AI inference affordable enough for widespread commercial adoption;
Rapidly developing commercial applications with advanced interfaces;
Building conversational AI systems that can handle real business tasks;
Developing delegation-based interfaces where AI can autonomously execute complex operations;
Creating practical applications that solve immediate business problems;
Scaling these solutions across different real industries and real use cases;
This approach might lack the glamour of pursuing AGI, but it's potentially more transformative for the immediate future of AI adoption. Overall while the West chases theoretical breakthroughs, China is building the practical foundation for widespread AI implementation. I can’t say for sure which approach is better, but it’s clear that both strategies reflect differing priorities and philosophies toward innovation.
DeepSeek and Moonshot: China's AI Giants Showcases
If you thought our above discussion about China's pragmatic AI approach was just theoretical, let me share something that literally happened two days ago. The Chinese AI scene erupted with two major releases that perfectly validate everything we just discussed about their strategic approach to AI development.
DeepSeek R1 on par with OpenAI o1
DeepSeek's release of their R1 model is a perfect case study of what we've been discussing about China's AI strategy. While everyone's been focused on OpenAI and Anthropic, DeepSeek has quietly developed a model that matches – and in some cases exceeds – their Western counterparts' capabilities.
Let's look at the numbers:
R1 scored 79.8% on the AIME 2024 math benchmark, slightly edging out OpenAI's o1 at 79.2%.
In programming tasks, it achieved 96.3% on Codeforces.
These aren't just impressive numbers; they're indicators of China's growing technical capabilities in AI development. It’s also worth to mention that DeepSeek:
has no external funding (fully self-funded by their hedge fund's resources);
has a really small team, only 140 members, compared to OpenAI's 1,200 and Anthropic's 500);
has entirely localized operations, showcasing their real domestic expertise.
But what really validates our earlier discussion about China's practical approach is how DeepSeek is implementing this technology:
First, they're making it accessible. Unlike many Western models hidden behind strict access controls and paywalls, R1 is fully available through their official website with 50 free messages daily. They've also released to open-source the model's weights, code, and technical documentation – a level of transparency that's quite rare in the AI world.
Second, and perhaps more importantly, let me tell you about R1-Zero – this is the most technically fascinating model in their lineup. What makes it revolutionary is that DeepSeek managed to train this AI without using any human-labeled data at all. This is a big deal in AI development. Typically, AI models require massive datasets where humans have manually labeled and verified the information – an incredibly expensive and time-consuming process.
The most telling aspect of DeepSeek's strategy is their focus on scalability. They've created eight versions of the model, including six distilled versions ranging from 1.5B to 70B parameters (even a 1.5B model performs better than Sonnet and the original 4o on mathematical benchmarks). This again, is about making AI practically deployable across different scales of infrastructure.
I can clearly see that overall approach perfectly aligns with what Kai-Fu Lee described in our earlier discussion – focusing on cost-effective innovation while building a robust ecosystem of AI applications.
Moonshot AI – Another Promising Challenger to OpenAI's o1
Just when I thought DeepSeek's announcement was exciting enough, Moonshot AI stepped into the ring – literally within an hour of DeepSeek's release – with their Kimi K1.5 model. That means it's not just one model rivaling OpenAI's o1 (currently the most advanced model), but two – and both coming from the same source we all know.
Founded in 2023, Moonshot AI has grown rapidly, receiving over $1 billion in funding from Alibaba by February 2024 and a valuation of $2.5 billion. By August, the valuation had reached $3.3 billion. Although Kimi k1.5 is planned as a competitor to ChatGPT, the company has not yet made the model available to the public.
What fascinates me about Moonshot's release isn't just its timing, but how it demonstrates a different angle of China's AI capabilities. While DeepSeek focused on pure reasoning and mathematics, Moonshot went multimodal – meaning their AI can work with both text and images. Smaller, but more focused.
And here's where the practical approach we discussed earlier really makes sense – they're doing all of it at prices that almost make me do a double-take. We're talking about API costs that are ~95% lower than OpenAI's rates ($0.10 vs $15 per million tokens for input, and $2.20 vs $60 for output).
United States Responds with Project Stargate
Since Trump's inauguration (ya it was just a couple of days ago), we've witnessed a dramatic shift in how the U.S. approaches the AI arms race. It's no longer just about technological superiority – it's become a matter of U.S. national security and economic dominance.
If you've read my last episode on “The Manhattan Project of Our Time” you'll understand why this isn't just another tech race – it's potentially the most crucial technological competition of our century.
And now, we're seeing the U.S. make its most decisive move yet with Project Stargate.
The Money Talk
Let me tell you, when I first saw these numbers, I had to double-check all my sources. We're talking about a $500 BILLION investment over four years – that's nearly twice what the entire Apollo Program cost in today's dollars. The first $100 billion is already lined up, and this isn't even government money we're talking about.
To put this in perspective, Google – one of the world's biggest tech spenders – plans to spend about $50 billion on capital expenditures in 2024. Project Stargate is planning to spend ten times that amount over just four years. This isn't just an investment; it's a statement of intent.
The Power Players
What makes this particularly fascinating is the coalition behind it. We're seeing companies like OpenAI, Softbank, Nvidia, Oracle, and Microsoft coming together, along with international investors like MGX (Abu Dhabi sovereign wealth fund) from the UAE. It's like watching a tech version of the Avengers assembling, but instead of fighting Thanos, they're building what could become the world's most powerful AI infrastructure.
The dynamics are shifting too – Microsoft will no longer be OpenAI's exclusive technology partner, allowing OpenAI to build and manage its own infrastructure. I bet the plan is to potentially positioning itself as a dominant monopolist in this field.
Infrastructure Scale
The scale is mind-boggling: ten interconnected data centers initially, expanding to twenty. They're starting in Texas, but this will spread across the country. The sheer scale of this infrastructure is staggering – we're looking at facilities that could consume up to 5 gigawatts of power, equivalent to 2-3 nuclear power plants, spread across "several hundred acres." To put this in perspective, Goldman Sachs predicts that by 2028, AI will account for about 19% of all data center energy consumption. They are creating an entirely new level of computational infrastructure.
Healthcare Revolution
The project's healthcare applications are particularly exciting. Larry Ellison from Oracle described a future where we could develop personalized cancer vaccines in just 48 hours by analyzing tumor fragments in blood and rapid sequencing.
But it goes even further. OpenAI has partnered with Retro, a startup aimed at increasing average life expectancy by ten years. They've developed a specialized model called GPT-4b micro, trained on protein sequences and biological interactions. With GPT-4b micro, they've created versions of necessary proteins that are 50 times more effective. Sam Altman himself has invested $180 million in this venture, showing just how serious they are about using AI for life extension.
By the way, the initial general goal voiced here is to extend human lifespans by a 10 years.
Government Approach
While the U.S. government isn't directly investing federal funds, they're showing unprecedented support through emergency orders to streamline construction and cut through bureaucratic red tape. This level of support shows just how seriously the U.S. is taking the AI race.
Trump also rescinded former President Biden's executive order on the safe development of artificial intelligence. Yes, now it’s officially not safe. This obviously speeds up AI development, but raises many questions about the risks we've discussed earlier in our AI alignment episodes.
Historical Context
To put this investment in historical context: while the Manhattan Project represented about 1.5% of U.S. GDP in the 1940s, Project Stargate is estimated at 1.7% of 2023 GDP. When you're matching or exceeding the relative scale of the project that ushered in the atomic age, you must know this isn't just another tech initiative – it's a national priority.
What makes me think: this game spreads even further than we might imagine. While public news focuses on building bigger computers or faster AI models, in fact, this is about establishing dominance in what could be the most transformative technology of our time.
When OpenAI researcher Noam Brown says, "Investments like this are only possible when the science is rigorously validated and people believe it will be transformative," he's also hinting at something bigger – they believe they're on the verge of something revolutionary.
As you can see, the AI race between the U.S. and China is no longer just about technology – it’s about shaping the future of global power. Both nations are carving distinct paths, each with own transformative implications.
What’s clear is that this silent battle is reshaping how AI will impact industries, nations, and lives. Including our own. Until that happens, stay sharp, stay healthy, and hang in there until any healthcare breakthroughs to come! See you next time!
🔎 Explore more: