Discover how the AI economy stabilized in 2025, from model competition to startup strategies. Insights from Y Combinator leaders on LLMs, funding, and the fu...
AI Economy Stabilizes in 2025: What Startup Founders Need to Know
The artificial intelligence landscape has undergone a dramatic transformation. What once felt like constant upheaval and unpredictable shifts has evolved into a structured, predictable ecosystem. In this comprehensive analysis, we explore the most surprising developments in AI's evolution, from model preferences among startups to the broader implications for entrepreneurs building the next generation of AI applications.
Key Insights
- Model Preferences Shift Dramatically: Anthropic surged from 20-25% to become the #1 API choice among YC Winter 2026 founders, demonstrating a generational shift away from OpenAI's historical dominance
- The AI Economy Has Matured: Instead of chaotic disruption, the ecosystem now features clear separation between model layer, application layer, and infrastructure companies—all generating substantial revenue
- Model Commoditization Accelerates Competition: Startups are building orchestration layers that swap between Claude, Gemini, and ChatGPT based on specific tasks, pressuring model companies and benefiting application-layer startups
- Finding Ideas Returns to Normal Difficulty: The pace of innovation has slowed from explosive monthly announcements to incremental improvements, meaning startup ideation requires traditional hustle rather than waiting for the next breakthrough
- Infrastructure Constraints Drive Innovation: Power shortages and land constraints are pushing companies toward space-based data centers and fusion energy solutions, creating entirely new market opportunities
The Great Model Shift: From OpenAI Dominance to Multi-Model Era
For years, OpenAI's GPT models represented an overwhelming consensus among AI startups. When the Lightcone podcast began tracking model adoption among Y Combinator founders, OpenAI commanded over 90% preference. This dominance seemed unshakeable. Yet within a remarkably short timeframe—just the last three to six months of 2025—the landscape transformed completely.
Anthropic's Claude emerged as the unexpected winner in the Winter 2026 YC batch selection cycle, capturing the top position among founder choices. This wasn't a gradual climb. After hovering around 20-25% throughout most of 2024 and early 2025, Anthropic experienced what observers describe as "hockey stick growth" of over 52%. The numbers tell a compelling story: OpenAI's share dropped from dominance to minority status, while Anthropic ascended to leadership.
Why did this seismic shift occur? The answer reveals itself when examining where AI applications are generating the most value. Coding tools and agents have emerged as explosive growth categories, where AI-native companies are creating tremendous value for users. These specialized tasks demand specific model characteristics—and it turns out Anthropic's models excel in this domain. This wasn't accident or luck. As Tom Brown, Anthropic's research leader, has explained in previous discussions, this specialization was deliberate. Anthropic consciously designed their models with specific use cases in mind, optimizing for performance in targeted areas rather than pursuing generalized excellence.
For startup founders making technology decisions, this specialization has real implications. The "best model choice" varies based on your application. If you're building coding tools, semantic search, or agent-based systems, Claude often delivers superior performance compared to alternatives. This contrasts sharply with the previous era where OpenAI was the default choice for virtually all applications.
The model ecosystem has also welcomed strong new competition. Google's Gemini has climbed into third place, moving from single-digit adoption rates—approximately 2-3% just one year earlier—to commanding about 23% of choices in the Winter 2026 batch. This represents extraordinary momentum and validates Google's substantial investment in competing at the application layer, despite earlier concerns that model companies would squeeze application builders.
What makes this competition particularly interesting is how differently these models behave. Each has developed distinct "personality" characteristics that influence user preferences beyond raw capability metrics. OpenAI's ChatGPT carries what observers describe as "black cat energy"—sophisticated, slightly mysterious, demanding precise inputs. Anthropic's Claude feels more like an "optimistic golden retriever"—enthusiastic, helpful, eager to engage. Google's Gemini occupies middle ground, balancing multiple characteristics.
These personality differences matter more than pure benchmarks suggest. Users develop preferences based on interaction patterns, familiarity, and trust. Harj, one of the podcast hosts, switched to Gemini specifically because it demonstrated superior reasoning capabilities and provided more accurate real-time information through Google's grounding API and access to Google's search index. For research and information-seeking tasks, Gemini outperformed even specialized search tools like Perplexity, which prioritized speed over accuracy.
Yet another host remains attached to ChatGPT despite its technical limitations in certain areas. The reason: memory. ChatGPT's ability to remember previous conversations, understand user personality and preferences, and maintain consistency across interactions creates a "sticky" experience that competitors haven't fully replicated. For high-stakes tasks—purchasing a house, making complex decisions—this continuity provided tremendous value. The user filled ChatGPT conversations with extensive context about inspection reports, real estate agent communications, and decision criteria, and ChatGPT maintained this context throughout lengthy dialogues.
Model Orchestration: The New Startup Moat
As model capabilities have become more similar and competition has intensified, an unexpected pattern has emerged among sophisticated startup teams. Rather than remaining loyal to single model providers, founders are building "orchestration layers"—infrastructure that intelligently routes tasks to whichever model performs best for that specific purpose.
This behavior started at the consumer level. Individual users began opening multiple tabs simultaneously: one for ChatGPT, one for Claude, one for Gemini. They'd submit identical prompts to each model, then review the outputs before selecting the best response. In some cases, they'd feed the competing responses back into Claude, saying "Okay, the other guys said this, what do you think?" essentially forcing the models to check each other's work.
What started as individual workaround quickly evolved into enterprise architecture. Series B AI-native companies that previously claimed strong allegiance to OpenAI or Anthropic began implementing similar multi-model strategies at infrastructure level. Instead of depending on a single model provider, they abstracted away the underlying model layer entirely, building orchestration systems that swap models based on requirements.
The strategic implications are profound. When a new, more capable model launches, these companies immediately evaluate it for their specific use cases. If it performs better for certain tasks, they integrate it. They might use Claude for prompt engineering and context development, then transfer optimized outputs to GPT for execution. For highly specialized tasks, they might deploy Gemini 3 specifically for its unique strengths. They're not locked into any provider; they're building on top of the most valuable capabilities regardless of source.
One striking example: a healthcare-focused startup fine-tuned an 8-billion-parameter model using reinforcement learning and achieved performance that surpassed OpenAI's general-purpose models in their specific domain. The model's smaller size meant faster inference and lower costs, while specialization meant superior quality. This demonstrates an important realization spreading through the startup ecosystem: the era of "one model to rule them all" has ended.
However, this orchestration strategy reveals a critical challenge. When you build a specialized model fine-tuned for specific tasks using reinforcement learning, you've created an achievement—until the major labs release their next major version. A startup might launch with a Claude-beating model, only to find that GPT-4.5 or GPT-5.1 suddenly overwhelms their fine-tuning advantage. The innovation treadmill never stops. You must continuously iterate, continuously improve, continuously push forward. The moment you believe you've achieved sustainable advantage is the moment you're vulnerable.
Yet this constant pressure also creates opportunity. As long as AI labs remain fiercely competitive with each other—constantly releasing new models, pushing capabilities forward—application-layer startups benefit enormously. The abundance of increasingly powerful tools available as APIs means builders can focus on differentiation, not infrastructure development.
Why the AI Economy Has Stabilized—And Why That Matters
The most striking realization among AI observers is how much stability has emerged from what seemed like perpetual chaos. In late 2024, the industry existed in genuine turmoil. The ground felt unstable. Nobody could predict what would happen next, when the next major breakthrough would arrive, or how it would reshape the competitive landscape. The future seemed genuinely uncertain.
By 2025, that tumultuous feeling has largely dissipated. The AI economy has organized into recognizable layers, each with distinct characteristics and value propositions. Model layer companies—OpenAI, Anthropic, Google—operate in intense competition, constantly improving and releasing new capabilities. These are high-investment, high-risk businesses that require enormous capital and top research talent.
Application layer companies build on top of these models, using APIs to create specialized tools that solve specific problems for specific users. This layer has proven remarkably fertile, with successful startups like Lama achieving $100 million in annual recurring revenue with just 50 employees. These companies compete on user experience, specialization, speed, and the quality of their domain-specific knowledge.
Infrastructure layer companies—chip makers, cloud providers, companies building training infrastructure—generate consistent, predictable revenue by enabling everything above them. Nvidia's dominance seemed unshakeable, yet 2025 revealed surprising competition. AMD is gaining market share, Google's TPUs are proving highly effective, and the narrative that everyone universally chooses Nvidia has evaporated. This competition actually strengthens the ecosystem by ensuring abundant compute resources rather than dangerous concentration.
Within this stabilized structure, clear patterns for building AI-native companies have emerged. Successful approaches now include specialized models fine-tuned for specific domains, multi-model orchestration strategies, and tight integration with domain-specific data. There are validated playbooks. The mystery has been reduced.
This stabilization has profound implications for startup founders considering AI businesses. The gold-rush feeling of "first-mover advantage in AI" has diminished. Finding startup ideas no longer means waiting for OpenAI's next announcement that will magically unlock new possibilities. A few episodes earlier in the podcast series, observers noted that finding ideas had become easier than ever—simply wait a few months for a major announcement and new opportunities would emerge. That dynamic has reversed.
Finding ideas now requires returning to traditional startup fundamentals: identifying real problems, understanding customer needs deeply, and building solutions that deliver obvious value. The extraordinary tail winds that propelled early AI companies forward—where any competent founder could build something valuable simply by wrapping an API—have normalized. The work has become harder. The pace of innovation, which once felt explosive and unpredictable, has slowed into incremental improvements.
Infrastructure at the Edge: Solving Power and Space Constraints
While application-layer startups grapple with normal competition dynamics, a parallel infrastructure revolution is emerging—one driven by hard physical constraints. The infrastructure buildout required to support the AI economy is bumping against genuine, non-negotiable physical limitations: insufficient electrical power generation and insufficient land.
In summer 2024, a startup called Starcloud announced an ambitious goal: build data centers in space. The internet's reaction was overwhelming ridicule. Social media erupted with mockery and dismissal—"That's the dumbest idea I've ever seen." Yet within eighteen months, the narrative completely reversed. Google began implementing space-based infrastructure. Elon Musk adopted it as a primary talking point in virtually every interview. What seemed absurd became urgent.
The reason is simple: there isn't enough power generation on Earth. This hard constraint is driving extraordinary decisions. Boom Supersonic, originally focused on building supersonic aircraft, pivoted to providing power generation for AI data centers. They're using jet engines to generate electricity, yet even this approach faces constraints—the supply chain for jet engines is so overloaded that orders placed today won't be fulfilled for two to three years.
These aren't theoretical problems for distant future scenarios. Big technology companies are making critical infrastructure decisions right now based on three-to-five-year predictions. They must secure adequate power, adequate cooling, adequate physical space. California's regulatory environment, particularly CEQA (California Environmental Quality Act) restrictions that are often exploited to halt both environmental and housing initiatives, makes building new data centers on terrestrial ground increasingly difficult. The escape route has become literal: build in space where regulations are minimal and space is unlimited.
Y Combinator's recent investments reveal a coordinated attack on these constraints. Addressing the land shortage problem: companies building data centers in space. Addressing the energy shortage problem: fusion energy companies like Helion and Zephyr Fusion, which recently graduated from YC and closed a strong seed round at Demo Day.
Zephyr Fusion provides a particularly intriguing example. The company was founded by engineers in their 40s, individuals who have spent their entire careers working on tokamaks and fusion energy within national laboratories. One day they examined the physics, the mathematics, and the models with fresh perspective and realized something remarkable: implementing fusion energy in space would actually be profitable. Not someday, not in theory, but according to their analysis, achievable within a 5-10 year timeline.
If Zephyr Fusion and similar companies succeed, space-based fusion could provide the gigawatt-scale power generation that supporting massive AI operations requires. This would complete what observers describe as a "trinity" solving the data center construction problem: space infrastructure for the land constraint, fusion energy for the power constraint, and the application-layer innovation that actually uses the available compute.
This infrastructure race reveals something important about the current moment. While it might feel like there's an "AI bubble"—with concerns about round-tripping capital between Nvidia and technology companies, questions about sustainability—the situation more closely resembles previous technological booms that faced similar skepticism.
The Telecom Analogy: Why This Bubble Might Actually Be Different
When discussing whether AI represents a genuine economic opportunity or a speculative bubble, a useful historical comparison emerges: the telecom bubble of the 1990s. That era witnessed extraordinary overinvestment—billions, tens of billions, hundreds of billions of dollars poured into telecom infrastructure, much of it ultimately wasted. The telecom companies themselves largely failed to generate returns matching the capital invested.
Yet something remarkable happened as a consequence of that "wasteful" overinvestment. Tremendous excess bandwidth became available at relatively cheap costs. This abundance of cheap bandwidth enabled YouTube's existence. Without the telecom bubble, YouTube wouldn't exist—or it would have emerged far later, after natural demand-driven infrastructure buildout. The same argument applies to video streaming generally and the entire modern internet experience.
The AI situation appears analogous. Yes, there's massive capital investment happening. Yes, some of it is undoubtedly wasteful or misdirected. Yes, companies like Nvidia might face stock price corrections despite being long-term buys. Yes, Meta might be overinvesting significantly in CapEx and infrastructure.
But consider what this abundance of compute capacity enables. If you're an undergraduate in a dorm room building an AI startup, you don't have to build data centers. You're not Comcast. You're more like YouTube—you're a user of abundant infrastructure who gets to focus entirely on creating differentiated value. Meta's CapEx is Meta's problem, not yours. Whether Nvidia's stock price rises or falls doesn't matter to your ability to build great products.
The economic dynamics actually favor application-layer startups during periods of infrastructure overinvestment. When everyone is arguing about whether companies are "overbuilding" compute capacity, the outcome is abundant, relatively inexpensive compute available for anyone to access via APIs. This is precisely the situation that enables the next generation of transformative companies.
Economist Carlota Perez studied technology trends extensively and distilled her findings into two distinct phases. The first is the "installation phase," characterized by massive capital investment in infrastructure. There's a frenzy, excitement, and often overinvestment. People debate whether it's a bubble. The second is the "deployment phase," when everything becomes explosively abundant and practical applications proliferate.
The internet industry followed this pattern precisely. During the telecom bubble era (the installation phase), there was massive infrastructure investment, much of it seemingly wasteful. But that infrastructure provided the foundation for the explosive growth of the deployment phase, which gave us companies like Google, Facebook, Amazon, and Netflix. The future Facebooks and Googles in AI haven't even started yet. They'll emerge in the deployment phase, which we're entering now.
A New Batch of Model Companies Emerging—But Carefully
One surprise of 2025 has been the increased interest in founding model companies. This represents a significant shift from earlier patterns when most AI startups were application-layer companies building on top of existing models.
There are two distinct categories of new model companies emerging. First, the ambitious few attempting to raise enormous capital and go head-to-head with OpenAI. These attempts are rare and challenging—companies like Illo from SSI are outliers in attempting this path. Most investors view head-to-head competition with well-funded labs as extremely high-risk.
The far more common pattern involves building specialized, smaller models optimized for specific purposes. Within YC, increasingly sophisticated approaches are appearing: models optimized to run on edge devices, language models specialized for specific languages, models fine-tuned for particular domains using reinforcement learning techniques. These specialized models often outperform general-purpose giants in their target domain despite having far fewer parameters.
Why has model building become more feasible for startups? The democratization of AI knowledge and engineering capability. A decade ago, building a competitive AI model required an extraordinary combination of rare talents: cutting-edge researchers, talented engineers, and visionary business leaders with deep capital connections. Such combinations were vanishingly rare. OpenAI itself represented one of the few teams that possessed all these capabilities.
Ten years later, the situation has transformed. Training models remains specialized work, but it's no longer uniquely rare. Many researchers have moved from academic labs to startups. Engineering practices around model training have become better documented and more standardized. Startup fundraising and business building have become more transparent and teachable. The skill combination is still rare, but far more achievable than previously.
The rise of reinforcement learning represents another important shift. Open-source models fine-tuned using RL techniques for specific tasks are demonstrating remarkable results. One YC startup achieved an 8-billion-parameter model that outperformed OpenAI's general-purpose offerings in healthcare applications through careful RL fine-tuning.
However, this path contains a built-in challenge: the improvement treadmill. A startup might launch with the best specialized model in their domain. But when GPT-4.5 or GPT-5.1 or Claude 4 launches, they immediately threaten your specialization. You must continuously innovate, continuously improve, or risk becoming obsolete. Success isn't an endpoint; it's a permanent commitment to staying ahead.
The Surprising Success of "Vibe Coding" and AI-Native Development
One prediction from earlier in 2025 that proved surprisingly accurate involved the rise of "vibe coding"—using AI to generate code based on conversational prompts rather than explicit instructions. This was initially observed as a behavioral pattern among founders, an interesting quirk worth noting. It seemed niche.
Instead, it became an enormous category. Companies like Replit have scaled to impressive success by enabling vibe coding at scale. Emergence has grown rapidly. Even major companies like Google are investing in this space, as evidenced by Google's launches and leadership discussions about vibe coding.
Yet there's an important caveat that prevents vibe coding from being a complete story of AI automation replacing developers. By the end of 2025, vibe coding cannot yet deliver 100% fully usable, reliable production code at scale. You can't describe vaguely what you want and deploy directly to production without human review and iteration. The technology falls short of that promise—for now.
This limitation reveals an important pattern about AI-native development: it amplifies capabilities without replacing human judgment. Skilled developers who pair with AI tools become dramatically more productive. Someone unfamiliar with coding can't yet successfully use vibe coding alone. The technology enhances expertise; it doesn't eliminate the need for expertise.
This distinction matters for founder strategies. AI amplification of talented people creates sustainable advantage. Replacing talented people with AI hasn't yet become economically viable, and current trajectory suggests it won't happen soon.
Hiring Continues Despite AI "Efficiencies"
A striking observation from watching AI startups mature concerns what has and hasn't changed in their organizational structures. Around this time last year, an unusual phenomenon captured venture capitalists' attention: startups hitting $1 million in annual recurring revenue while maintaining tiny teams—just the founders, or founders plus one employee. Simultaneously raising Series A with these minimal teams created exciting narratives about AI-enabled efficiency.
One year later, these companies have scaled into their Series B phase and beyond. The anticipated outcome—"they hit $10 million ARR while still maintaining essentially no employees"—hasn't materialized. Instead, these companies brought their teams back and started hiring real teams again.
The reason illuminates an important dynamics about AI and work. Many founders initially believed AI would dramatically reduce the labor requirements for scaling businesses. This belief generated two possible outcomes: (1) everything gets solved by AI, fewer employees needed, or (2) AI reduces production costs and accelerates development, raising customer expectations, thus requiring more employees to meet those expectations.
The reality has tracked closer to the second scenario throughout 2025. AI-native companies are hiring just as many people as traditional companies building at similar scale. Customer expectations haven't decreased; they've increased. Competitors using AI are improving their products faster, which pressures all companies to maintain rapid improvement cycles. This constant competitive pressure requires hiring excellent people across product, engineering, customer success, and operations.
The bottleneck isn't ideas or even execution capability. Startups can access plenty of good ideas. The bottleneck is people who can execute well. Genuinely excellent operators, product leaders, engineers, and teams remain scarce and highly sought-after.
Some exceptional startups have demonstrated different patterns. Harvey, an AI-first legal platform, achieved extraordinary early success. Lama has achieved $100 million ARR with 50 employees—a genuinely impressive ratio. These represent achievable outcomes, not common ones.
Yet companies like Lagora, representing second-generation AI native businesses, are catching up to Harvey's market position while being significantly leaner. This suggests that first-generation companies like Harvey sometimes overinvested in fine-tuning and optimization that, retrospectively, didn't generate proportional returns. Second-generation companies learned from these patterns and took different approaches.
Similarly, Giga is competing effectively with Sierra, suggesting that specialized AI legal tools can achieve competitive positions through focused execution rather than massive fine-tuning investments. When a startup invests significant double-digit percentages of capital into fine-tuning and gains minimal benefit, the money ultimately transfers to investors as equity dilution rather than creating sustainable competitive advantage.
Looking Forward: The Scaled Indie Company Era
As 2025 concludes and patterns become visible, several predictions emerge for the coming era of AI startup development.
The fantasy of a single person running a multi-trillion-dollar company hasn't arrived—and probably won't arrive in 2026. The gap between current capabilities and that threshold remains substantial. However, a different and genuinely exciting pattern has become visible: companies with fewer than 100 employees earning hundreds of millions in revenue. This seems increasingly achievable and likely to become more common.
Lama's announcement of $100 million ARR with 50 employees captured something important about changing startup narratives. Previously, companies boasted through traditional metrics: "We've raised X amount of capital, and Y people work for us." The implicit narrative was "look how much money we're using, how big our team is, how much we're scaling in traditional terms."
The reversed boast is now emerging: "Look at all this revenue, look how few people work for us." This inverted metric captures a fundamental shift. Efficiency matters more than scale. Value-per-person matters more than headcount. The ability to do more with less has become the new status symbol.
The era where AI makes everything massively more efficient and everybody uses fewer people hasn't quite arrived. The era where AI enables much better work, raising customer expectations, and thereby justifying continued hiring is clearly the present reality. But the trajectory points toward that future scenario becoming increasingly real.
Conclusion
The AI economy has matured from chaos to structure in remarkably short timeframe. Clear layers have emerged: model companies competing fiercely, application companies building on top of abundant APIs, infrastructure companies enabling everything. Founding an AI startup no longer means waiting for the next revolutionary announcement—it means solving real problems with available tools. Model competition has become fierce enough that application-layer startups benefit tremendously. The infrastructure buildout, while ambitious, is enabling rather than limiting for founders. The next generation of transformative companies hasn't launched yet. The opportunity for undergraduates in dorm rooms building specialized AI solutions remains substantial. Focus on execution, assemble great teams, and solve real problems—the infrastructure will be there to support your ambitions.
원문출처: https://www.youtube.com/watch?v=cqrJzG03ENE
powered by osmu.app