Deep Dive: The DeepSeek Boom — Who Hit the Jackpot? Who Took the Hit? | Yunqi Capital Attent!on Podcast

DeepSeek's open-source push continues to intensify, and the chain reactions it has triggered are still rippling through the industry. As a major focus of tech entrepreneurship and venture capital investment in recent years, how is the AI investment narrative shifting? What changes are entrepreneurs and investors on the ground actually sensing?

In a recent episode of Attent!on, a tech podcast produced by Yunqi Capital, two AI entrepreneurs — Han Xiao (Founder & CEO of Jina AI) and Shaun (Co-founder & CEO of HeyRivia) — joined Emily, Vice President at Yunqi Capital, for a wide-ranging discussion. Below are highlights from their conversation.

Scan the QR code above or follow "Attent!on" on Xiaoyuzhou to listen to this episode

· Guest Introductions:

Han Xiao, Founder & CEO of Jina AI

Founded in 2020, Jina AI is a leading search AI company. Its search infrastructure platform encompasses embedding models, rerankers, and small language models, enabling enterprises to build reliable, high-quality generative AI and multimodal search applications.

Shaun, Co-founder & CEO of HeyRivia

Founded in 2024 and based in the Bay Area, HeyRivia provides AI call assistants for healthcare service providers.

Emily, Vice President at Yunqi Capital

An investor on Yunqi's tech team, Emily focuses on frontier AI technologies and innovative projects. Previously a Silicon Valley entrepreneur.

Linda, Managing Director at Yunqi Capital (Host)

The following is an edited excerpt from the podcast.

Breaking Through: The Story Behind DeepSeek's Rise

Linda, Managing Director at Yunqi Capital

Let's start with DeepSeek's models themselves. Before this wave of mainstream attention, what impression did DeepSeek leave on you? The R1 release had actually been in development for some time — they simply chose that particular moment to launch it. What struck each of you as most impressive?

Han Xiao, Founder of Jina AI

I've spent a long time overseas. From an international perspective, DeepSeek and Alibaba's Qwen had long been considered China's top-tier foundation models. DeepSeek left a very low-key impression — compared to Qwen and others, they did very little PR. Second, their model performance and cost efficiency were excellent.

But they hadn't stood out previously because, at this stage, large models are still climbing exponentially. Though some keep saying we've "hit the wall," there's actually still room for performance gains. So people weren't particularly enthusiastic about models with middling performance. That was the general sentiment before DeepSeek released V3.

The R1 launch was impeccably timed. Of course, from the outside we don't know whether this timing was deliberate or not, but when OpenAI released the o1 preview last September, the signal spreading through AI circles was that training-time scaling had reached its limit and would shift to test-time scaling — meaning the reasoning chain would consume more resources during inference rather than training, with the model continuously reflecting and thinking to produce better results.

After OpenAI spent two to three months getting this message across, people bought in and recognized that test-time was where attention should go. The problem was that OpenAI had been deliberately opaque, never revealing how this was achieved, fearing reverse engineering. Then DeepSeek R1 dropped like a bombshell, reproducing an "Eastern version" that sent massive shockwaves through the entire AI industry.

Shaun, Co-founder & CEO of HeyRivia

I first heard about DeepSeek last year when they released APIs costing one-tenth or even one-hundredth of Llama's. Many friends around me immediately switched to DeepSeek. A lot of people were working on very basic tasks that didn't require a massive, highly intelligent model. So if costs could drop to 1% or even 0.1%, many were willing to try.

The biggest change in DeepSeek's approach from last year to this year was its open-source methodology — it open-sourced its paper. Across Silicon Valley, whether OpenAI or Anthropic, technical discussions in papers are always cagey; they rarely tell you how they actually implemented things. After R1's paper came out, many people began replicating its capabilities with smaller models and achieved excellent results. So the industry was extremely interested in a model whose training process was knowable.

And it was probably the first open-source thinking model. Previous models, whether V3, Llama, or Qwen, weren't thinking models — just standard base models. But this model really made people feel like AI had become more intelligent. You could see how the AI was thinking, which excited many Bay Area developers. With this added thinking process, you could not only prompt the model on the problem itself, but also prompt its thinking process, leading to much better final results.

Emily, Yunqi Capital

I strongly agree with both points Han Xiao and Shaun raised. First, even today, DeepSeek remains the only open-source model in the first tier, and its open-source approach has been genuinely substantive — not just an MIT license, but a paper that relatively thoroughly explains the entire training process and data methodology. Even now, as SOTA models continue to compete, very few have the conviction to go open-source. That's a deeply disruptive move.

Second is the cost of training and using models. This is something Yunqi has long believed in, and why we started investing in the application layer relatively early — we felt that limited compute wasn't a permanent constraint, that innovative methods would emerge to improve compute utilization, whether through more efficient architectures or more effective training methods. As it turns out, costs have fallen even faster than we imagined.

A third point we haven't discussed today: DeepSeek's team structure and organizational approach also offer many lessons. The founding team is notably localized and young, which overturns previous consensus — there was a sense that only people from Silicon Valley or its core labs had a meaningful lead. But it turns out that in large model training or deep learning, conviction matters enormously. I read an early talk by DeepSeek's CEO Liang Wenfeng where he made this point, but I don't think it's been widely enough appreciated. In large model training, because it's fundamentally a black box, having conviction is critical — you have to believe the capability is achievable to keep experimenting, because the process is genuinely grueling.

Industry Ripples: How DeepSeek Is Reshaping the AI Ecosystem

Linda, Managing Director at Yunqi Capital

We just touched on something important — open source. The substantive open-source approach and the dramatic price reductions are significant factors. I'd like to ask both entrepreneurs: what opportunities and challenges have these characteristics created for your respective businesses?

Han Xiao, Founder of Jina AI

From a cost and business model perspective, there's not much to analyze about DeepSeek itself — that's not where its business model focuses, and at least for now, profitability isn't a concern.

For other players in the AI landscape, particularly those at the same model layer — companies like Together AI, Cohere, Mistral — the pressure is substantial. Their models don't outperform DeepSeek yet they charge premium prices, so they'll need to adjust pricing strategies.

For many cloud providers, DeepSeek isn't a threat; quite the opposite. They're seeing substantial enterprise demand for on-premise deployment. So across the ecosystem, it's a mixed bag — some are pleased, others worried. Those who benefit are likely players not on the same plane, either upstream or downstream in the inference platform stack. Those suffering are at the same layer, building similar large AI models — they'll face more intense competition.

For Jina specifically, 2025 may be another "year of search" — every year is the year of search, we always think this year will be different. 2024 was about RAG, 2023 about vector databases, and 2025 is shaping up to be agentic search, search powered by autonomous agents. We're going deeper — we just released a deep search feature based on chain-of-thought reasoning that continuously reasons while searching.

Why is this possible now when it wasn't in 2024 or 2023? We have OpenAI and DeepSeek to thank for shifting mindsets from "search must complete in a few hundred milliseconds" to "it's acceptable if it takes a minute." After Sam Altman and Noam Brown started talking about test-time compute, people gradually accepted that model inference can be slow — as long as the final result is directly usable. This search philosophy is essentially autonomous search, a major trend for 2025.

Now Baidu and WeChat have integrated DeepSeek to enhance search capabilities, while Perplexity, OpenAI, and Google released their own deep search products earlier. From the perspective of market leaders like Google, this is a perilous moment. I've been telling my colleagues: today is the best time to "rebel against Google."

Shaun, Co-founder & CEO of HeyRivia

We're in the AI Agent space, so DeepSeek's breakthrough means we can access higher-level intelligence at lower cost. Traditionally, to cover my costs, I could only afford to replace workers making around $10 per hour. Once compute costs scaled up, the math didn't work. Now with costs dramatically reduced, even replacing workers at $3-4 per hour becomes profitable — that's fundamentally different.

(Q: Has there been significant improvement in multi-step task handling?)

A standard scenario in care management is creating a ticket when communicating with someone who has care management needs. Current AI agents are essentially "competing for work" within these ticket systems. For example, with a patient who needs weight-loss medication, you need to verify whether they've used the same medication before — requiring face-to-face communication, photo proof, and many steps, making the cycle very long. Now this can basically be resolved in 1-2 days.

But healthcare overall is a relatively traditional industry, and some clients have concerns about R1. Some believe there must be "backdoors" in R1, and sometimes it's not easily explainable. So even though R1 costs less, clients may prefer other more established models.

Linda, Managing Director at Yunqi Capital

Emily, could you share some memorable DeepSeek application cases and what insights they offer?

Emily, Yunqi Capital

On capability ceilings — it's truly impressive, and precisely because of this, things we previously thought impossible are now achievable. One clear application is in finance. I've recently been using DeepSeek to analyze investment projects, and I'd say it basically reaches junior analyst level, because its information retrieval, summarization, and analysis are quite strong.

A second personally fascinating point: DeepSeek is actually quite emotionally intelligent. Its capabilities in psychology, personality analysis, and language analysis are remarkably strong. I'm not sure whether the training data contained substantial content in these areas.

So returning to an investment and entrepreneurship perspective, from a product-market fit standpoint, you need some conviction about model capability ceilings, and you should target domains that can leverage higher model ceilings. Previously, when investing in projects or brainstorming with founders, we'd face a difficult question: copilot or fully automated agent? Looking back today, fully automated agent is clearly the answer. And the chosen domain or industry should have relatively high value-add, because such industries face lighter commercialization pressure — people are more willing and able to pay. As a startup, your first year may run more smoothly, because in such an intense competitive environment, there's higher expectations for execution speed.

Escalating Competition: How Grok 3 Shakes Up the Landscape

Linda, Managing Director at Yunqi Capital

Elon Musk has made another big move recently. The newly released Grok 3 shows significant improvements in reasoning, coding, and multimodal capabilities. I'd like to hear both entrepreneurs' perspectives on this model.

Shaun, Co-founder & CEO of HeyRivia

I started trying Grok 3 on day one. The features released so far include thinking, but not yet deep search — at least that was the case last I checked. Personally, I didn't feel it was dramatically better than previous versions, though perhaps I've just gotten accustomed to deep research lately. I may need more experimentation. At minimum, Grok's release made people realize we haven't hit the wall — adding more compute, more machines, running longer, needing more energy — all of this is still viable. They said they're still adding more machines, with more big moves coming. Bottom line: everyone still needs to stockpile GPUs.

I did a rough comparison between its deep research and current Grok thinking — Grok seems to think more, though deep search will be released later. But I have some doubts about its coding capabilities. Some posts I've seen don't match the claimed performance, though again, more testing is needed.

Han Xiao, Founder & CEO of Jina AI

From the model layer, the more intense the competition, the better. For the application layer, it's mainly about which model offers strong performance at reasonable cost. I'd been paying attention to Google Gemini 2.0 Flash — actually a very good model, but perhaps because Google was relatively low-key. They've released several solid models, but people expect more from Google, so they didn't get much attention. We actually use Google's models extensively, primarily for large context windows — they need to be long, because we do search. And now with many agents using reasoning models, adding reasoning chains to agent prompts immediately adds thousands of tokens. With continuous reasoning, this places extremely high demands on a model's "needle in a haystack" capability across long contexts.

As reasoning models rise, many design paradigms will shift. I posted on Twitter asking: how many chat apps render thinking separately? When I checked last week, among several major local chat apps, 50% didn't separately render thinking. From the application layer, there's much to do with thinking; from the model design layer, there's also much to explore — prompt design, long-context retrieval capabilities, many things can be inserted. I believe that in 2025, whether Grok, OpenAI, or Llama, everyone will increasingly move toward thinking/reasoning model capabilities. But perhaps by May or June, people may tire of thinking and want a regression — maybe this thinking process is just gimmicky, can we not show it?

Listen to the full episode by subscribing to the "Attent!on" channel on the Xiaoyuzhou app~

Podcast Timeline

Part 1 Breaking Through: The Story Behind DeepSeek's Rise

00:05:12 Affordable DeepSeek — why did it blow up only now? 00:07:03 DeepSeek R1's launch timing — was there method behind it? 00:10:32 The "accidental" breakthrough: substantive open source, substantive thinking 00:12:37 The DeepSeek shockwave — how did Silicon Valley react? 00:14:57 Updated investor perspectives: from costs to team structure 00:19:34 A Silicon Valley founder's view: has reverse tech export been achieved?

Part 2 Industry Ripples: How DeepSeek Is Reshaping the AI Ecosystem 00:20:56 Upstream and downstream impacts: model providers, cloud providers, hosting platforms — who wins, who loses? 00:24:37 From RAG to Agentic Search — another "year of search" approaching? 00:31:11 Higher-dimensional intelligence vs. client concerns: opportunities and challenges at the application layer 00:36:39 Finance, psychology, emotion — how far has DeepSeek's application potential been explored? 00:41:31 Entrepreneurship insights: "conviction" in model capability ceilings, deep domain expertise 00:44:19 How Elon Musk's Grok 3 shakes up the foundation model "battlefield"

Part 3 On Trends: Key Inflection Points for 2025 00:52:04 Investor perspective: three things to watch in 2025 AI ecosystem (multimodal capability improvements, alignment, AI + hardware) 00:56:31 AI entrepreneur survival rules: mentally prepare for intense competition; stay sharp on AI iteration speed and capability ceilings; stay lean and mean