Recently, the top global startup incubator Y Combinator (YC) provided an end-of-year summary on the AI industry for 2025 in their latest podcast episode. As a bellwether for startups, YC incubates numerous leading AI companies each year, and their internal observations often signal shifts in technology and business. In this discussion, four partners delved into key topics such as the competitive landscape of models, the AI infrastructure bubble, and trends in entrepreneurship and talent.

Model Competitive Landscape

1. Anthropic Surpasses OpenAI:
   Anthropic’s model share has surpassed 52%, officially overtaking the long-time leader OpenAI. From early 2024 to early 2025, Anthropic’s share remained around 25%, achieving a steep growth curve in the past 3 to 6 months. The core driver of this shift is Anthropic’s exceptional coding capabilities, making it the preferred tool for many developers and penetrating various use cases.

2. Gemini’s Explosive Growth:
   Gemini’s market share skyrocketed from single digits to 23% within a year, demonstrating excellent reasoning capabilities even before the release of Gemini 2.5 Pro. In terms of accuracy in information retrieval, Gemini outperforms Perplexity, making it a trusted choice for handling real-time information that requires high reliability, despite being slightly slower.

3. Memory as OpenAI’s Competitive Moat:
   ChatGPT’s ability to understand user personalities, thought processes, and historical contexts creates a highly customized experience with significant switching costs. While Perplexity excels in rapid web research, ChatGPT’s memory feature is becoming a consumer-facing competitive barrier, making it difficult for long-time users to switch to models lacking personalized accumulation.

4. Orchestration Layers as Standard:
   Current AI startups no longer bet on a single model but instead build orchestration layers to achieve technological abstraction. This allows founders to flexibly call the most proficient models for specific tasks. For instance, they might use Gemini 3 for context engineering and then input the results into OpenAI for execution, dynamically replacing models as new ones are released. This approach is supported by startups deeply entrenched in specific verticals, possessing proprietary evaluation standards and unique datasets, making models interchangeable commoditized components.

5. Lack of Deep Consumer Experiences:
   Despite powerful models, there is still a lack of fully automated applications capable of handling complex high-value transactions. This has led users and startups to engage in manual arbitrage: opening multiple model tabs, assigning them the same tasks, and comparing outputs, even having different models audit each other. Until fully automated applications mature, users must rely on extensive prompt engineering and cross-model validation to ensure accuracy.

AI Infrastructure Bubble

6. Overabundance of Infrastructure as a Boon for Entrepreneurs:
   The surplus of computing resources signals opportunity. YC reflected on the late 1990s telecom bubble, where hundreds of billions in excess bandwidth ultimately birthed great products like YouTube. In the AI era, fierce competition among Nvidia, Google, and AMD means computing resources will become increasingly cheap and abundant. For startups, this not only lowers costs but also provides a fertile ground for creative expression.

7. Transition from Installation to Deployment:
   Economist Carlota Perez theorizes that technological revolutions occur in two phases: installation and deployment. The frenzy of purchasing GPUs and building data centers in 2023 belongs to the heavy asset installation phase, while we are now transitioning to the deployment phase. This is highly beneficial for startups, as they do not need to engage in costly infrastructure competition but can build next-generation applications on existing AI infrastructure, often leading to the emergence of future giants like Facebook or Google during this phase.

8. Energy Bottlenecks and Space Data Centers:
   Power generation capabilities are a critical limiting factor in AI development. With land and power infrastructure failing to keep pace with demand, the seemingly absurd idea of space data centers has become a direction followed by giants like Google and Elon Musk. YC’s internal company Zephyr Fusion is researching how to achieve gigawatt-level energy supply through space fusion reactors, which may be the only reasonable solution to computing energy shortages in the next decade.

9. Stabilization of the AI Economic System:
   Compared to the turmoil and confusion at the end of 2024, the AI economy in 2025 has entered a relatively stable system. The division of labor among model, application, and infrastructure layers has become very clear, with profit opportunities for all parties. There is now a mature operational manual for building an AI-native company based on large models, eliminating the need for trial and error, thus laying a foundation for sustainable growth in the industry.

10. Scaling Laws and Human Resistance as Buffers:
    Despite rapid technological advancements, according to scaling laws, technological growth is logarithmic and will gradually encounter limits. More importantly, humans inherently resist change, creating a buffer that allows society, culture, and government ample time to digest AI technology.

Trends in Entrepreneurship and Talent

11. Vibe Coding Matures:
    By 2025, Vibe Coding has officially evolved from a founder’s behavioral pattern into a mature industry category. This development method allows developers to focus on high-level logic and “vibe,” rapidly generating code and iterating prototypes through large models. While it cannot yet deliver 100% stable production code, it significantly accelerates the speed of validating ideas, transforming early workflows in startups.

12. Vertical Models Outperform General Models:
    The knowledge of building models is becoming democratized, no longer the patent of a few geniuses. Many startups utilize open-source models, combining proprietary datasets from vertical domains for reinforcement learning (RL) fine-tuning. Data shows that these small models (e.g., with only 8B parameters) can completely outperform OpenAI’s general large models in specialized benchmark tests, proving that “proprietary data + post-training infrastructure” is the hard power of startups.

13. Anti-Signaling Trend:
    The AI era has birthed a new standard of success: no longer flaunting how much money was raised or how many people were hired, but instead showcasing the team’s efficiency and revenue. Gamma achieved $100 million in annual recurring revenue (ARR) with just 50 employees. While a “one-person trillion-dollar company” has yet to emerge, stories of small teams generating massive revenue are becoming the new norm for AI-native companies, where high productivity has replaced scale expansion.

14. Talent Combination Becomes Commonplace:
    A decade ago, the combination of top research thinking, strong engineering skills, and sharp business acumen was extremely rare, but by 2025, such talent has emerged in large numbers. The decentralization of relevant building knowledge will lead to an industry explosion as scarce skills become common skills. More application-oriented AI companies will rise in the future.

15. Efficiency Gains vs. Customer Expectations:
    Despite AI significantly enhancing individual efficiency, leading AI startups are still actively hiring as before. This is because while AI lowers production costs, it also raises customer expectations for product functionalities, necessitating the recruitment of more high-end talent to meet the growing market demands and faster delivery cycles. The current bottleneck lies not in creativity but in recruiting outstanding personnel capable of executing high-quality work using AI technology; the competition for talent remains fierce.

Podcast Transcript Highlights

1. Model Throne Transition: The Rise of Anthropic and Gemini
   Garry: Welcome back to the latest episode of “The Light Cone.” Today, we will discuss the most surprising phenomena we’ve seen in 2025. Diana, you discovered a very crazy example that can almost be seen as a transition of the “flag bearer” in the AI field: there has been a significant shift in entrepreneurs’ preferences for large language models (LLMs) in this YC batch.

Diana: Yes. We just wrapped up the winter batch selection for 2026. During the application process, we ask all founders about their tech stack and preferred models. The results were shocking. For a long time, OpenAI was the undisputed winner in YC batches. I remember when we started this podcast series, OpenAI’s market share was over 90%, although it has been declining. But in this batch, Anthropic has surprisingly become the top choice, slightly surpassing OpenAI. Who would have thought?

Garry: Anthropic maintained around 25% for most of 2024 to early 2025, but the situation reversed in the past 3 to 6 months.

Diana: Anthropic’s growth curve has steeply risen like a “hockey stick,” finally breaking through 52%.

Garry: What do you think is the reason?

Diana: I believe it mainly lies in the logic of model selection. As we have seen this year, many projects based on “Vibe Coding” tools and coding agents (AI Agents) have succeeded, creating significant value. It turns out that the models performing best in this area are from Anthropic.

This is not a coincidence. From our previous conversations with Tom Brown (co-founder of Anthropic), coding ability is one of their internal evaluation metrics, and they consciously prioritize it as a guiding star metric, which directly reflects in the quality of model outputs. Thus, for many entrepreneurs building products, Anthropic has become their first choice.

Jared: In fact, many people use it in non-coding scenarios as well; I think there is a kind of penetration effect. Once people get accustomed to Claude’s personalization in personal coding, they tend to choose it for other tasks as well, even if their applications do not involve any code development.

Garry: How is Gemini performing? Where does it rank?

Diana: Gemini’s growth is also quite rapid. Last year, its share was probably in the single digits (around 2% to 3%), but this winter batch has risen to about 23%. Our team has also extensively used Gemini 3.0, and its quality and efficiency are indeed impressive.

These models seem to have their own personalities. OpenAI gives a cool “black cat” vibe; Anthropic feels like a relaxed, friendly, and very helpful “golden retriever.” That’s my intuitive feeling when interacting with them.

Harj: Exactly, I’ve been using Gemini as my go-to model this year. Even before the release of 2.5 Pro, I felt its reasoning capabilities were better. I now use Gemini more often instead of Google Search. I trust Google’s Grounding API, which can provide real-time and accurate information using Google’s search index.

In terms of accuracy in information retrieval, I believe it is better than all the tools currently available, even outperforming Perplexity. Although Perplexity is faster, it is not always accurate; Gemini, while slightly slower, often provides very reliable answers if you ask it about today’s events.

Garry: You all have switched tools, but I haven’t moved away from ChatGPT. I find its “memory” feature very sticky; it understands my personality and knows how I think.

I use Perplexity for some quick web research tasks because I feel ChatGPT is somewhat lagging in web searches. I believe “memory” is becoming a consumer-facing moat. I don’t expect Gemini to have the same personality as ChatGPT; they feel like completely different entities.

Harj: What surprises me is that there still haven’t been more deep consumer applications emerging in the market. Looking back, my biggest change this year has been the dramatic increase in prompt engineering and context engineering workload. For instance, I recently bought a house, and throughout the process, I was deeply communicating with ChatGPT. I fed it every inspection report and all the PDF documents, asking it to summarize key points to establish information parity in negotiations with the agent.

Garry: While you can throw PDFs into a model for it to summarize, in high-value transactions, if you don’t do a lot of prompting and guidance, it’s hard to fully trust the model’s accuracy. You still need to put in the effort to verify.

Harj: Indeed, I haven’t seen any application that can fully automate these tasks.

2. New Normal in the Industry: Multi-Model Orchestration
   Garry: Have you seen Karpathy’s “LLM Arena”? I am currently doing this in a more primitive way: I open tabs for Claude, Gemini, and ChatGPT simultaneously, give them the same tasks, and compare outputs. I usually throw suggestions from other models to Claude for mutual auditing.

Diana: This consumer-side behavior, akin to “manual arbitrage,” is quite interesting; startups are doing the same thing. I’ve interacted with several founders who were once die-hard OpenAI fans but are now trying cross-model strategies.

I recently spoke with some founders of AI companies, most of whom have reached Series B and are of considerable scale. They are now doing something interesting: building an orchestration layer to achieve technological abstraction. This means that when new models are released, they can flexibly replace them and even call the most proficient models based on specific task requirements. For example, I heard one startup say they would first use Gemini 3 for context engineering and then input the results into OpenAI’s model for execution. As new models continue to emerge, they are iterating on the strongest AI agents across various categories.

This approach is feasible because they possess proprietary evaluation standards and datasets. As AI Agent companies in highly regulated industries, they have mastered the most suitable data for their scenarios. I believe this has become the new normal in the industry. While model companies have invested heavily to make underlying intelligence faster and stronger, application-layer companies will also do their best to leverage these results. This is reminiscent of the Intel and AMD era, where new architectures continuously emerged, allowing users the freedom to choose and switch.

Harj: Indeed. In high-level discussions, the most anxious question often revolves around where the value will ultimately flow. Will it flow to the model layer or the application layer? This seems to change every so often.

Sometimes when model companies release very impressive products, like Claude Code, people feel that model companies will dominate the application layer. However, at least from the past few months’ experiences, such as Gemini’s explosive growth, we have returned to a world where models are “commoditized” to each other. If this trend continues, startups in the application layer will welcome another golden year.

3. Beneath the Bubble: Infrastructure Frenzy and Energy Bottlenecks
   Diana: I would love to hear Jared’s thoughts, especially since there has been a lot of negative discussion on Twitter about the “AI bubble.”
   Jared: Yes, this is also a question undergraduates often ask me. They see the hundreds of billions and even trillions of dollars being invested between Nvidia and OpenAI and worry whether this is a bubble or some form of false prosperity.

Garry: In fact, this phenomenon is quite remarkable. We can reflect on the telecom bubble of the late 1990s. Hundreds of billions and even trillions of dollars were piled into telecom infrastructure, leading to severe bandwidth oversupply. But it was precisely this cheap and abundant bandwidth that gave birth to products like YouTube. Without that excess, YouTube might have appeared many years later. We are in the intelligent era, where computing resources are like thinking, working rocks; as long as they are powered, they become smarter. For today’s undergraduates, the “overabundance” of infrastructure is actually an opportunity. If resources were scarce, competition would be smaller, but prices would be higher, and underlying profits would be thicker.

Looking at this year’s news, although Nvidia’s stock price has fluctuated, I remain optimistic in the long run. The current situation is that people are no longer betting solely on Nvidia; AMD and Google’s TPUs are also making strides. Competition means more computing resources, not fewer. This may put pressure on large AI labs as they compete with each other, but for students starting businesses in the application layer, this is undoubtedly a significant advantage.

Harj: Indeed. Regarding the bubble, if you are a large company like Comcast or Nvidia, you do need to worry about whether people are overbuilding GPU capacity. But students are not infrastructure builders; they are more like the YouTube of their time. Even if Nvidia’s stock price drops next year, it doesn’t mean it’s not a good time to start an AI business now.

Jared: It’s like Zuckerberg mentioned in a podcast; large companies like Meta have to invest heavily in infrastructure because they cannot stand by. If demand fluctuates, the losses are on the capital expenditures of large companies, not on startups. The improvement of infrastructure will only provide a fertile ground for building new ideas.

Diana: Economist Carlota Perez has studied technological revolutions multiple times and divides technological trends into two phases: installation and deployment. The installation phase is accompanied by heavy capital expenditures and frenzied infrastructure, which indeed looks like a bubble. Just like the chat model craze in 2023, everyone was extremely excited and frantically purchasing GPUs and building gigawatt data centers.

We are now transitioning from the “installation” to the “deployment” phase. This is excellent news for startup founders because they do not need to engage in costly data center construction but can build the next generation of universal applications during the deployment phase. Before 2000, the heavy asset investments of telecom giants led to overcapacity, even resulting in many unused fiber optic dark lines. But that’s okay; the internet eventually became a massive economic driver. This means that future giants like Facebook or Google may not have emerged yet, as they often come to life during the deployment phase of technological revolutions.

Garry: Speaking of infrastructure, the current situation is even somewhat exaggerated. The limitation we face is not technology but power generation capacity. A company called Starcloud first proposed building data centers in space and was initially ridiculed. But eighteen months later, Google and Elon Musk have also started to follow this direction.

The reason for this shift is simple: our land-based power generation and construction capabilities cannot keep up with demand; we simply do not have enough land and power to support future needs. In this environment, placing data centers in space has become a reasonable way to relieve pressure.

Jared: Upon reflection, YC already has several companies addressing these issues. For example, Boom and Helion are working on energy shortages, and another company called Zephyr Fusion, founded by senior engineers from national laboratories, discovered that placing fusion reactors in space is actually cost-effective. This may be the only way to achieve gigawatt-level space energy in the next decade.

4. Shift in Entrepreneurial Paradigms: Vertical Models and Talent Democratization
   Harj: Additionally, I’ve noticed an increase in interest in founding model companies this year. While only a few companies can directly challenge OpenAI, within YC, more and more people are starting to build smaller-scale models. These projects often target edge computing devices or speech models for specific languages. This reminds me of the early days of YC when the knowledge of building startups began to decentralize and democratize. When a scarce skill becomes a common skill, an explosion occurs. A decade ago, founding OpenAI required an extremely rare talent combination: top research thinking, strong engineering skills, and sharp business acumen.

Garry: What you describe is precisely the combination of Ilya, Greg, and Sam.

Harj: Such teams are indeed extremely rare. But ten years later, today, there is a large influx of talent with research backgrounds, engineering backgrounds, and financing experience. This means we will see more application-oriented AI companies rise and vertical models for various specific tasks.

Diana: I agree. The popularization of reinforcement learning (RL) is also creating a snowball effect. There are now many open-source models available that can be fine-tuned for specific environments and tasks. This means you can fully utilize open-source models to train the best medical models, whose performance can even surpass general models like OpenAI.

In fact, I have seen quite a few examples. One YC startup fine-tuned their model using the best medical dataset they collected, resulting in performance that completely exceeded OpenAI in benchmark tests in the healthcare field, using only an 8 billion parameter small model. This proves that with vertical domain data and post-training infrastructure, startups can indeed build highly competitive products.

5. Annual Review: Stable Systems and Streamlined Efficiency
   Garry: Over the past year, what has truly impressed you or stood out?

Diana: This is interesting. We did an episode on Vibe Coding at the beginning of the year, which received a lot of views. I remember we viewed it more as a behavioral pattern observed among founders. But I was surprised to see it evolve into a huge industry category. Now many companies are emerging in this field, such as Replit and Emergence.

Harj: Varun Mohan has now gone to Google. He released a video related to anti-gravity that looked very cinematic, with Varun operating at the keyboard while Sergey Brin closely followed him. I am actually curious if they used Nano Banana or similar video generation tools, as that video looked too perfect. But considering Google’s budget, it’s normal to produce high-production-value videos. In any case, now people are not just discussing space data centers but also Vibe Coding.

While my previous comment was somewhat sarcastic, as far as I know, Vibe Coding is still not fully reliable. You cannot rely solely on “vibe” throughout the entire coding cycle; as of the end of 2025, it still cannot deliver 100% stable production code.

Jared: What has surprised me about 2025 is the degree to which the AI economy has stabilized. When we recorded the episode at the end of 2024, we felt we were still in a period of extremely rapid change, even some turbulence, with no one knowing when the next huge shock would come or where startups and the overall economy would head. But now, we have entered a relatively stable AI economic system: the division of labor among model, application, and infrastructure layers is clear, and it seems everyone has profit opportunities. How to build an AI-native company based on models now has a relatively mature operational manual.

Harj: This is actually a downstream chain reaction. Although the models themselves have gradually improved this year, there has not yet been a major technological breakthrough that completely disrupts everything. We previously discussed that finding entry points for entrepreneurship seems to have returned to a normal level of difficulty. As long as you can persist for a few months, there is likely to be a new major feature release, which will give rise to a series of new ideas and building opportunities.

Garry: I agree. And I am not surprised by the so-called “AI 2027” report. That pessimistic article once predicted that society would begin to collapse in 2027, but later they quietly revised the prediction time while keeping the shocking title. I have always been skeptical of the narrative of “technology taking off rapidly.” Even according to scaling laws, technological growth is logarithmic, meaning it will gradually encounter limits and the speed will be slower than expected.

This is actually good news. Humans are inherently resistant to change. We previously analyzed a report from MIT that mentioned 90% of enterprise AI projects fail. The fact is that most enterprises haven’t even figured out their basic operations, let alone apply AI. This human resistance to change has instead become a buffer against the excessive penetration of new technologies.

While I support pushing for rapid technological advancement, in this case, taking it slow might be a good thing. Under the joint action of logarithmic scaling and human resistance to change, society, culture, and government will have enough time to digest and respond to this powerful technology without falling into a state of chaotic loss of control.

Harj: One more thing surprised me. Around this time last year, we discussed some startups achieving $1 million in annual revenue with just the founder and no employees, raising Series A funding. I thought this model would continue to expand, such as reaching $10 million ARR while still not hiring. But the reality is that these companies later went back to start building real teams. While the current company sizes may be smaller, the playbook remains largely unchanged. The bottleneck still lies in the time required to recruit talent.

Garry: While AI has improved efficiency, the rapid growth of business has made the demand for high-end execution talent even more urgent. Companies like Harvey and Open Evidence secured large amounts of venture capital early on, using capital as a defensive moat. Now we see a second wave of AI-driven companies emerging, such as Lora and Giga, and competition remains fierce. Some companies have burned through significant capital fine-tuning models without translating that into actual competitive advantages, with only investors profiting because they own more shares of your company.

Harj: Regarding team size, there are currently two viewpoints: one believes AI will make everything more efficient, thus requiring fewer people; the other believes AI lowers production costs but also raises customer expectations, necessitating hiring more people to meet the growing market demands.

This year seems to lean more towards the second scenario. Those leading AI startups are still hiring actively as before. They are limited by the availability of outstanding personnel capable of executing well, rather than by creativity.

Garry: I agree. The era of one person running a trillion-dollar company has not yet arrived, but it will ultimately move in that direction. I don’t think this will happen in 2025 either. However, we will see more stories of teams of fewer than one hundred people generating hundreds of millions in revenue. For example, Gamma has only 50 employees but achieved $100 million in annual recurring revenue (ARR).

This reflects a wonderful “anti-signaling” trend. Previously, everyone liked to flaunt how much money they raised or how many people they hired; now, people are starting to showcase how high their revenue is and how streamlined their teams are. Well, that’s all for today; happy new year to everyone, and see you next time.