San Francisco's AI Startup Scene: What's Really Happening in 2026

Key Insights

• Frontier labs are consolidating dominance across coding, legal, finance, and bio domains through intensive post-training datasets and reinforcement learning optimization
• Market segmentation is shifting: Competition now focuses on finding domains frontier models cannot handle, creating new business opportunities
• Test-time compute is the new frontier: Models like GPT-5.6 achieving 750 TPS show that inference optimization—not just scale—drives performance gains
• Silicon Valley's startup ecosystem is dominated by founders in their early 20s, with valuations 5-10x higher than comparable Korean startups
• Korea's chip and infrastructure sectors are gaining investor attention as VCs recognize underexploited growth potential

---

The Evolution of Frontier Labs' Strategy

After spending three weeks in San Francisco's Hayes Valley neighborhood, conversations with founders, entrepreneurs, and venture capitalists revealed a clear market pattern: frontier labs (OpenAI, Anthropic, Google DeepMind) have shifted focus from pursuing general artificial general intelligence (AGI) to dominating specific high-value domains.

The domains they're targeting directly—coding, legal, finance, and bio—share a critical characteristic: they generate massive post-training datasets where verification occurs entirely within digital environments. For coding, legal work, and financial analysis, frontier labs are reportedly purchasing high-quality, non-public datasets and building specialized training pipelines through reinforcement learning from human feedback (RLHF) and reward-based value learning.

The strategy is ruthless in its focus: areas where performance gaps still exist—robotics, healthcare, longevity research—are being left to partner startups and domain-specific labs, suggesting a consolidation wave is likely ahead.

The Post-Training Infrastructure Arms Race

What became clear during meetings with engineers at frontier labs: the real competition isn't in algorithms or pre-training anymore. It's in post-training pipeline optimization.

The post-training process operates on two axes:

1. Training infrastructure: Building systems that efficiently consume compute resources while running reinforcement learning loops
2. Dataset quality and volume: Acquiring or generating domain-specific training data that converts expert knowledge into verifiable problem sets

Frontier labs are reportedly pouring "everything" into this domain—benchmark optimization loops, inference model (SLM) optimization, and what the industry calls "benchmark maxing": reverting to checkpoints and restarting if performance drops at good intermediate points.

This represents a departure from earlier rhetoric around AGI. Instead, performance gains are being driven by deepening reinforcement learning across increasingly narrow domains—a pattern that questions whether this constitutes true general intelligence or specialist domain mastery at scale.

Test-Time Compute and Inference Efficiency

A recurring theme in conversations with researchers: future model improvements may not require larger models or bigger training datasets, but rather increased test-time compute allocation per problem.

Recent announcements like GPT-5.6 achieving 750 tokens per second (TPS) underscore this shift. Researcher accounts suggest that OpenAI's solution to the Erdős problem wasn't a fundamentally new model architecture—it was simply applying massive test-time compute to existing GPT-5.5 through inference optimization.

This has profound implications: if test-time compute can be optimized through speed improvements, today's models could unlock significantly higher performance without architectural breakthroughs. This reframes the competitive landscape from "who builds the best model" to "who can run inference fastest and cheapest."

Silicon Valley's "Everyone's Gone Crazy" Startup Landscape

The most striking observation: startup founders are predominantly in their early 20s, with even late-twenties founders considered "old" by local standards. Many enter Y Combinator immediately after university graduation, and an increasing number skip college entirely.

Y Combinator's selection process has become a de facto credential system. With nearly 200 teams selected per batch and only one-minute presentation slots at Demo Day, the process now signals "this founder is reasonably promising" rather than "this is a guaranteed winner." The organization itself treats Y Combinator acceptance as an investment in people, not products—rapid pivoting is expected and encouraged.

Founder valuations reflect this bet-on-talent strategy. Companies with minimal Annual Recurring Revenue (ARR)—as low as 3-4 billion Korean Won—are receiving pre-seed valuations of 300-400 billion Korean Won. Engineers departing from Meta, Anthropic, or ByteDance—earning 5+ million USD annually—immediately command 100 billion Korean Won enterprise values when founding startups, justified by investor consensus that acquihire activity and high acquisition valuations create profitable outcomes even for failed companies.

Korea's Overlooked Advantage in Infrastructure

Venture capital interest in Korea is concentrated in one area: chips and infrastructure software—a complete contrast to the agent and consumer AI focus dominating US startup attention.

When asked about Korean companies, investor questions focused specifically on whether former Samsung and Hynix engineers were founding startups, whether Korea had competitive chip innovation pipelines, and what infrastructure software was being developed. This reveals a gap: while Korea lacks large-scale data center development compared to the US, its chip manufacturing expertise and nascent infrastructure software sector are seen as underexploited growth opportunities.

The disconnect is significant: VCs recognize that NVIDIA, Samsung, and Hynix have not cornered the market—rather, the sector is viewed as only beginning to scale, with "much more room for growth" despite apparent market consolidation.

The Academic-to-AI Software Divide in Biology

One of the most revealing insights came from observing how biology is being approached through two competing paradigms.

Traditional biotech researchers (like Nobel laureate Jennifer Doudna) maintain that AI cannot innovate in biology—it can only summarize. This perspective reflects decades of domain-specific methodologies where each biological problem requires a unique network architecture, dataset, and training approach.

AI software engineers, by contrast, are applying the "foundational model" approach: gathering large amounts of general protein and biodata, training Transformer models, and allowing the model to discover solutions through in-silico methods.

The irony: OpenAI's tool "OpenCRISPR" already found proteins functionally equivalent to Cas9—the protein Doudna herself discovered—entirely through AI-driven protein data analysis. Yet many traditional biology labs remain unaware these methods exist, or skeptical they could work.

This gap represents what may be a significant business opportunity: AI software engineers entering biology don't need domain experts' buy-in to succeed. They just need access to data and computational resources.

---

Conclusion

San Francisco in mid-2026 reflects a market in sharp transition. Frontier labs are consolidating power through post-training infrastructure and test-time compute optimization. Young, early-stage founders are commanding unprecedented valuations based entirely on perceived talent and execution capability. And Korea's chip and infrastructure expertise—long considered commodity sectors—are suddenly strategic assets in the eyes of Silicon Valley capital.

The message for Korean founders: The window for consumer AI services innovation may be opening globally, but speed and global connectivity are now critical. The AI Frontier's upcoming demo days and Silicon Valley correspondent updates will track where this next wave emerges.

---

원문출처: EP 102. 샌프란시스코의 교훈: "사람들이 다 미쳤어요"

powered by osmu.app