My Thoughts on DeepSeek

People keep asking, “What’s the deal with DeepSeek? Why is it shaking up the entire AI world?” So here’s the story in my own words, with a bit more punch.

What is DeepSeek?

DeepSeek is a Chinese AI startup that began inside a stock-trading firm called High-Flyer, led by Liang Wenfeng. They’ve been building “large language models” (like ChatGPT), but at a supposed fraction of the cost. Their big claim is that you don’t need a war chest of billions and the world’s fanciest chips to train top-level AI.

Why is DeepSeek Such a Big Deal?

1. Low Training Costs: DeepSeek-V3 allegedly cost under $6 million to train, while many folks believe OpenAI, Google, and Meta spend hundreds of millions per model. Everyone’s asking: “How can they do it that cheap?”
2. Open Source: DeepSeek shares its code and model weights. This is the exact opposite of what big U.S. labs do—where everything’s locked down. That’s a major shift.

How Did They Pull This Off?

Conventional wisdom says you need thousands of premium GPUs to build a massive model. But DeepSeek claims they only used around 2,000 H800 chips (due to U.S. export bans on the top-tier H100s). By selectively “lighting up” parts of the network—what they call Mixture of Experts and other memory optimizations—they got top-tier results. Or so they say.

Which Models Has DeepSeek Made?

• DeepSeek-V2: Debuted “Mixture of Experts” (MoE). This approach only activates specialized sub-networks when needed, cutting down compute costs.
• DeepSeek-V3: Their headline model with the alleged $5.6 million final training tab, said to rival the likes of GPT-4 or Claude. It’s also the basis for their consumer-facing app, which rocketed to the #1 free download in the U.S.
• DeepSeek-R1: Billed as a “reasoning” model for logic, math, coding, etc. R1-Zero trains largely by itself via reinforcement learning, prompting some people to call this a “Sputnik moment” for AI (the January 20 release date is now infamous).

A New ‘AlphaGo’ Moment?

Back in 2016, Google’s AlphaGo beat the world champion Lee Sedol at the game of Go, shocking many, especially in China—where Go is culturally significant. That event was China’s “wake-up call” for AI. Now, some analysts say DeepSeek might be the West’s equivalent “AlphaGo moment” but in reverse: This time it’s a Chinese company catching the U.S. off-guard. The question is whether American tech leaders will see it as a call to action and double down on AI investment.

Moonshot or H-Bomb?

Some are also comparing DeepSeek’s emergence to big historical shocks:

• Sputnik Moment (1957): When the Soviet Union launched the first satellite, it jolted the U.S. into the Space Race.
• Hydrogen Bomb: A technology arms race that no one wanted to lose, fueling rapid scientific advances.
• Man on the Moon (1969): Achieving what once seemed impossible in record time.

In all these cases, one side’s breakthrough triggered a massive response from the other. DeepSeek’s success—despite lower-end hardware and minimal budgets—might push the U.S. to rethink AI strategy, just like Sputnik or AlphaGo did.

Does $6 Million Cover Everything?

Probably not. A lot of analysts doubt they could build a model that powerful for so little. Bernstein’s Stacy Rasgon put it bluntly: “Did DeepSeek build OpenAI for $5 million? Of course not.” Some suspect that $6 million is just the final run, excluding all the hardware purchases, R&D, data pipeline costs, and so on.

The Nvidia Stock Meltdown

After DeepSeek’s R1 model and its accompanying hype dropped, Nvidia’s stock cratered—falling about 16% in a single day, the biggest single-day dollar-value loss in Wall Street history for any chipmaker. That alone wiped out hundreds of billions in market cap. Investors freaked that if DeepSeek can succeed with “nerfed” chips, maybe the GPU demand gravy train isn’t as guaranteed as we thought. TSMC also slid 14%, and Microsoft lost nearly 4%.

Elon Musk’s Skepticism

Elon Musk chimed in on X (formerly Twitter) to say he doubts DeepSeek’s claim of using “only 10,000 A100s.” He speculated it’s more like 50,000 HPC GPUs if they’re truly at the level they say. Alexandr Wang from Scale AI backed him, hinting there’s no way DeepSeek got so far, so fast, with so little hardware. Musk basically calls the $6 million figure “too good to be true.”

Trump’s Take on DeepSeek

President Donald Trump didn’t mince words either, calling DeepSeek a “wake-up call” for America’s technology industry. He warned that if Chinese firms can build serious AI with cheaper, restricted hardware, the U.S. must up its game. Yet he also suggested that “doing it cheaper” might ultimately help American companies if it opens up more avenues for innovation at lower cost.

Sam Altman Weighs In

Sam Altman, who runs OpenAI (maker of ChatGPT), called DeepSeek’s new models “impressive,” but said his team still believes raw computing power matters. “We can obviously deliver much better models,” he posted, “but it’s legit invigorating to have a new competitor!” That’s as close as you get to grudging respect in the AI world. He also teased future releases from OpenAI that might overshadow DeepSeek.

Are They Really That Good?

DeepSeek publicly released the model weights for V3 and R1, allowing outside researchers to poke around. So far, no one’s found major flaws in the performance metrics. But the question remains: Is it truly a sub-$6 million success, or did they sink more money in behind the scenes? Either way, the output is good enough to scare the daylights out of U.S. tech giants.

Is China Overtaking the U.S. in AI?

That’s the trillion-dollar question. Google, OpenAI, Meta, and Microsoft still hold massive resources, brand power, and advanced data centers. But if a small Chinese startup can come this close with arguably weaker hardware, it says a lot about the power of open-source collaboration and forced innovation under sanctions. Some are dubbing January 20, 2025, as a new “Sputnik moment,” pushing the U.S. to respond—and quickly.

Why Does “Open Source” Matter?

• Attracts Talent: Top researchers love open models so they can see exactly how everything works.
• Speeds Up Progress: More eyes means more innovations, bug fixes, and spin-offs.
• Centers AI in China: If the best open models come from Chinese labs, global AI development may anchor there—particularly in advanced reasoning or large-scale chatbots.

So Why Are Investors Freaking Out?

They see that “commodity hardware + clever code” can churn out advanced AI. If that’s the case, the GPU gold rush could slow down, and the multi-billion-dollar bets from U.S. megacorps might not pay off as predicted. Nvidia’s meltdown might be the first sign. It doesn’t help that TSMC also took a hit. Some are calling it a re-pricing of AI expectations in the market.

Is This Heading Toward AGI?

DeepSeek’s R1-Zero can solve logic puzzles, reason about code, and refine its own steps. That’s advanced but not quite human-level thinking. We’re still a ways off from “true” AGI, but the speed of progress is jaw-dropping. A year ago, we thought you needed the biggest supercomputers on the planet. Now, apparently, you can do it with cheaper chips.

What Does It Mean for Big Tech?

• Meta, Apple, Amazon, Microsoft: Might see opportunities in cheaper, more efficient AI. The big question is whether they’ll pivot or keep throwing money at massive, closed models.
• OpenAI, Anthropic: They have to prove they’re worth their massive training budgets if a competitor can do 80-90% of the same work at a fraction of the cost.
• Google: Could see new threats to search if DeepSeek (or others like it) master large-scale indexing and Q&A. Google’s “Gemini” is rumored to be huge, but the public is now more skeptical of the “big-spend = big win” strategy.

Next Moves?

We will see the U.S. tighten chip sanctions even more—but that may just force Chinese companies to be even more clever, like DeepSeek. Or the U.S. might invest heavily in next-gen research, to ensure it stays ahead. One thing’s for sure: The old assumption that AI requires endless GPU horsepower is taking a beating. This is a big shift.

My Prophecy on What Comes Next

So what does the future look like now that DeepSeek has rocked the boat? Here’s my take:

1. Incoming Model Flood: Over the next couple of weeks, we’ll see a swarm of new models drop—from American AI startups, from European labs, from everyone who wants to replicate DeepSeek’s open-source approach. Because the process is right there for anyone to adapt, the field is about to get crowded fast.
2. Stock Market Yo-Yo: Yes, the market will bounce back—these panic sell-offs usually do—but it won’t recover those sky-high valuations immediately. Wall Street will likely wait for a real “wow” moment from one of the big guys (Anthropic, OpenAI, Google) before letting AI hype run wild again.
3. Pressure on Pricing: OpenAI, Anthropic, and others will have to rethink their price structures. With more cheap (or even free) models emerging, they can’t keep charging a premium. By the end of this year, we’ll probably see GPT-level or “o1-level” models running locally on regular machines—or offered on a near-free basis by some cloud providers.
4. Tighter Sanctions, Less Effect: Expect the U.S. to clamp down further on mid-tier GPUs like the H800—because that’s exactly what DeepSeek used. But honestly, it might be too late. Chinese companies will just find another workaround. China has that top-down push to lead in AI by 2030, and it’s not going to slow down.
5. China’s Long Game: In my view, China ends up winning this game because they’re all-in. The U.S. has entrepreneurial energy, but China’s got a national directive—and a track record of mobilizing quickly when they set a target. By the time the U.S. sorts out new sanctions, Chinese labs will have leapfrogged to new approaches.
6. Europe Trails: Meanwhile, Europe is, let’s be honest, playing catch-up. Mistral is not, and probably never will be, state-of-the-art. Europe’s stance on AI regulation may further slow them down. So Europe might end up just supporting whichever major AI ecosystem proves dominant.

Bottom Line: The year ahead is going to be crazy. Free or nearly-free GPT-level models, new releases from the biggest players, frantic hardware sanctions, and a global race to replicate or outdo DeepSeek’s approach. Buckle up—because 2025 is shaping up to be the year AI goes absolutely mainstream, and nobody wants to be left behind.

Key Topics (to Sound Extra Smart):

Moonshot vs. Bombshell: Does this parallel the “Man on the Moon” moment, or a “Hydrogen Bomb” arms race in AI?

Distributed Training & Mixture of Experts: The secret sauce behind how DeepSeek can use fewer GPUs.

Emergent Reasoning: How R1-Zero seems to figure out logic with minimal human guidance.

Sanctions & Innovation: How U.S. hardware bans pushed Chinese AI startups to new efficiencies.

Open-Source Advantage: Why DeepSeek’s open weights challenge the dominant, closed U.S. AI labs.

Investor Overreaction: Nvidia’s stock drop and what it signals about the future of AI hardware demand.

Geo-Political Tension: Trump’s wake-up call for U.S. tech; Musk’s skepticism fueling public debate.

AI Commoditization: If DeepSeek can do more with less, does that mean the end of expensive AI?

Potential AGI Pathways: R1-Zero’s emergent thinking and the race toward advanced general intelligence.

‘AlphaGo’ and ‘Sputnik’ Moments: How DeepSeek’s breakthrough could mirror historic tech shocks.

Footnotes

1. NYT Coverage of DeepSeek’s Cost Efficiency — The New York Times: China’s AI Startup DeepSeek Takes On U.S. Giants – How they trained a huge model for under $6 million. ↩
2. Nvidia’s H800 vs. H100 Chips — PCMag Breakdown – Explains the difference and why H800 is considered “nerfed.” ↩
3. Mixture of Experts (MoE) Approach — Stanford AI Lab Blog – Short explainer on how distributing tasks to specialized “experts” can save compute. ↩
4. DeepSeek-R1 and Reinforcement Learning — MIT Technology Review – Discussion on R1-Zero’s reinforcement approach. ↩
5. Market Reaction to DeepSeek — Financial Express – Coverage on Nvidia and TSMC stock drops. ↩
6. China’s Growing AI Momentum — South China Morning Post – How open-source collaboration is helping narrow the gap. ↩
7. Nvidia Stock and AI Hype — Bloomberg – Reports on the meltdown and future GPU demand concerns. ↩