Topics Pioneers Insight Method Author
Back

56. Silicon Valley Dispatches Over China's National Day Holiday — May There Be Bubbles Everywhere

·苔藓之火

Rendering the long image…

Topics AI Venture Funding and Bubble Risks AI Training Data and Evaluation Benchmarks Battle for Frontier AI Moats Early-Stage VC Investing Dynamics Chinese Founders Driving Global AI Expansion

Summary

  • Raymond thinks whether Silicon Valley is in a bubble matters far less than whether he can participate in the upside. He held 70 meetings in 20 days and saw big-tech employees start companies, founders invest, and capital and talent keep looking for their next move. On his first day, someone wanted to sell secondary shares in Anthropic bought in 2023 and reinvest in a former professor’s new lab: “Silicon Valley operates on a different clock from the rest of the world.”
  • New labs’ product validation is lagging the fundraising frenzy, but early-exit logic is still attracting capital. The figures he cited: 119 new labs are collectively seeking $94.3B from the market, and only 16 have disclosed revenue. Some investors are still using academic ties to get into low-valuation seed rounds, betting on follow-on financing or an acquisition. Raymond doesn’t understand that investment thesis, but supports researchers sharing in the wealth: “Studying shouldn’t mean being poor.”
  • Data companies are capturing post-training budgets as customers pay for expert teams, responsiveness, and end-to-end delivery. The industry view he heard is that coding may no longer be the focus in 2027: 90% is already covered, while the remaining 10% may have very low ROI, and the next domain is still up for grabs. A single data point can be worth $10K to tens of thousands of dollars, but projects worth several million dollars may still require delivery in 2 days. “RSI needs two things: GPUs and data”—and both come with heavy operational demands.
  • Raymond attributes the pressure on DeepMind to the combined effects of organization, compliance, and compute allocation. From the extra management layers after Gemini became a company-wide P0, to restrictions on user data and compute being prioritized for GCP, constraints beyond technical capability are affecting iteration. He heard that GCP may currently be worth 10x Research; meanwhile, he has watched different Chinese models take turns in third place.
  • YC’s shift to younger founders is supported by job-market and acquisition incentives, but B2B startups may lack industry experience. Junior roles and internships are shrinking, while startups can find funding more easily; a successful acquisition can even put founders straight into management at a big tech company. One recruiting founder he met knew AI well but had no clear handle on hiring pain points, like “a boy wearing his father’s suit—the whole thing looks a size too big.”
  • A more liquid M&A market supports Silicon Valley’s VC cycle, but an acquisition alone doesn’t prove investors got their money back. Raymond cited $159B in AI M&A over the past year and heard of an investor who put in $1M and ultimately exited for just $12K. Some people call finding a job an acqui-hire, so he now starts by asking: “What’s the structure of the deal?”
  • The growing base of Chinese founders is forming a cluster, but willingness to pay may still undermine commercial partnerships. A Midas investor estimated that Chinese founders may launch thousands of projects a year, with 50 to 100 people like 袁征 potentially emerging over time—a bet on a larger talent pool. On the other hand, after Raymond arranged for a Chinese big-tech company to see a B2B product demo, the response was to have its internal team build a cheaper copy. Repeated requests like that will exhaust suppliers’ willingness to cooperate.

Deep dive

1. Silicon Valley Runs on a Different Clock, Making the Bubble Question Beside the Point

  • On his first day, Raymond ran into a friend selling secondary shares in Anthropic bought in 2023, planning to put the money into a new lab founded by a former professor. The seller was already thinking about exits and reinvestment while others were still asking whether they could buy Anthropic: “Our clocks are about 3 years apart.”
  • Twenty days and 70 meetings left him with one clear impression: “Everyone has several things going on.” Big-tech employees were quietly starting companies; founders were quietly investing. It reminded him of China in 2014–15. Years later, people may regret not spending time looking at Meituan, ByteDance, Xiaomi, and Pinduoduo more than they resent companies that raised too much money back then.
  • Wall Street was asking where the next coding opportunity would come from, while people at labs were busy hiring, working late, and iterating; students, big-tech employees, and founders whose companies had been acquired were all looking for their next move. Raymond admits it feels restless, even bubbly, but still misses “the tailwind of an upcycle” (“上行期的推背感”). His wish: “May there be bubbles everywhere, and no business too hard to do.”

2. Most New Labs Have Yet to Prove PMF, but Early Investors Are Still Betting on a Handoff

  • Thinking Machines Lab is his cautionary example: it raised at a $12B valuation before it had a product and was once expected to raise another round at a $50B valuation by the end of 2025. Raymond says, as far as he knows, the round never closed and there has been no clear revenue news, even as secondary trades keep circulating.
  • The figures he cited underscore the gap between fundraising and validation: 119 new labs are collectively seeking $94.3B from the market, but only 16 have actually disclosed revenue. Most have no clear PMF, and some have no product at all. When people still call it a good sector, his response is: “I really don’t understand.”
  • Some investors’ logic is to use academic ties to get into angel or seed rounds and bet that professors at top universities will raise Series A or B rounds, creating an exit opportunity for earlier investors. Raymond questions that thesis, but supports universities letting failed founders return, and believes professors should have the chance to become billionaires. Giving people who study and do research a chance to build wealth is a positive feedback loop for society.

3. Data Companies Capture Post-Training Budgets Through High-Touch Services

  • Raymond sees little difference in the methods used by mature teams in China and the U.S.; the bigger gaps are in compute, funding, robust RL environments, and scalable infrastructure. Frontier models are spending more time and money on post-training, making data the most visible commercial opportunity of the past 6 months.
  • The industry view he heard is that coding may no longer be the focus in 2027: 90% is already covered, and the remaining 10% may have very low ROI. The next domain is still unclear, so data companies need to work with customers to assess demand, prepare data, and line up experts—making them more like “research partners” to model companies.
  • Training data can also shift from being a cost to a source of revenue, and may even be sold to competitors; Raymond knows of a Chinese model company that has done this. Valuable data points can sell for $10K or tens of thousands of dollars, but delivery is demanding: a project worth several million dollars may need results in 2 days, with the PM on call 24/7 and responding to changing requests at 2 a.m.
  • Customers are paying for the ability to organize experts, respond quickly, and deliver the full package. He also heard a rumor that one data company starts salaries at $500K and encourages employees to move close to the office. These businesses once fell outside the VC model because of limited scalability and upside, but now command significant AGI and RSI budgets: “One is GPUs, the other is data” (“一个是卡,一个是数据”).

4. DeepMind Faces Constraints on Organization, Compliance, and Compute Allocation

  • Raymond revisits the explanation for the earlier departure of guest Lewis: Google built models earlier but launched usable products later, and its personal agent iteration was slow. As Gemini’s founding PM, Lewis saw the project become a company-wide P0, after which more bosses and middle layers kept appearing and the work became a “telephone game.”
  • Seedance has stunned Raymond’s friends in the U.S. They attribute the gap to the high cost of compliance in the U.S., where using users’ faces for training creates major problems. They stress that Apple and Google really do not train on users’ private data.
  • Compute allocation also surprised him: according to people in the industry, most of the GPUs went to GCP, leaving DeepMind short. GCP may currently be worth 10x Research, so it gets priority when supply is tight. He also heard that Sergey Brin once arrived late to a meeting and had to stand—no one offered him a seat.
  • Raymond does not doubt Google’s technical ability, but competitors keep improving and the bar for the next SOTA release is rising. In his view, Kimi, DeepSeek, GLM, and other Chinese models have taken turns holding third place over the past year. U.S. investors visiting China in groups are driven more by concern about these capabilities than by curiosity alone.

5. YC Is Getting Younger, While Industry Experience Remains a B2B Startup Gap

  • Raymond likens the expanded YC to a university or business school: it has gone from 2 batches a year to 4, with close to 200 companies per batch and 700 to 800 a year. The average founder age was 28–30 from 2015 to 2022, then 28 in 2023, 27 in 2024, and 26 in 2025. He guesses it could reach 24 in 2026.
  • He sees this as a rational response. After Vibe Coding, big tech has cut junior hiring, many internship programs have shut, and new graduates are struggling to find work, while startups can more easily get funding and leeway. The learning curve is faster, and an acquisition can put a founder straight into a VP role—upending the traditional path of climbing the corporate ladder.
  • Raymond managed to meet the founder of a YC recruiting startup: young and smart, with experience working on TikTok recommendations, but prone to asking questions about recruiting that an industry outsider would ask. The product may be AI-native and save labor, but it did not seem to address a clearly defined hiring problem. It was like “a boy wearing his father’s suit—the whole thing looks a size too big.”
  • When the founder said the customers were mainly growth companies from seed to Series A with strong hiring needs, Raymond realized they sounded a lot like YC’s own graduates. He joked that they could “buy from one another, run up each other’s numbers, and improve together.” The underlying question: can young founders without corporate experience really understand the industries they plan to transform?

6. M&A Keeps the Exit Cycle Moving; Deal Structure Determines What Investors Get Back

  • By the figures Raymond cited, AI M&A totaled $159B over the past year, including Cursor at $60B and the recently closed $8.2B deal for World Labs. Public companies need a second growth engine, and stock can serve as currency; he cited a statistic that 33% of companies can exit within 10 years to explain why LPs get a positive signal and keep investing.
  • But acquisition claims can be exaggerated. Raymond met a Stanford student who had built a GPT for blind users and said he had been acqui-hired. When pressed, it turned out there was no cash or stock—just a job. He had only just finished his freshman year. Raymond wondered why even that experience needed to be packaged as an exit.
  • One investor told him, “80% of acquisitions are probably hype”: a portfolio company had widely publicized its acquisition, but an investment of $1M had yielded an exit of only $12K. Another startup was absorbed by a different startup in an all-stock deal, which also did not amount to making it. Raymond has changed how he responds to acquisition news; at a minimum, he now asks: “What’s the structure of the deal?”

7. Chinese Founder Clusters Are Growing, but Willingness to Pay Still Gets in the Way

  • A Midas investor estimated to Raymond that there may be thousands of Chinese-founded projects each year, mainly from people leaving big tech, students, and professors. 袁征 used to be an isolated example; as the talent pool grows, it could produce 50 to 100 people like him. Raymond has also seen more Chinese GPs emerge and recalls an earlier estimate that people of Chinese descent make up close to a third of the teams at 3 leading labs, while names written in pinyin account for about 15%, with an even higher share among senior researchers.
  • Over the National Day holiday, LPs were evaluating GPs, investors were looking at startups, and founders were trying to find customers. Raymond was struck by how hard Chinese entrepreneurs work. But after he arranged for a Chinese big-tech company to see a B2B company’s business and demo, the response was: “Let’s push our internal R&D to build a cheaper version.” He worries that if customers will not pay for the product, repeated requests for introductions and demos will eventually leave no company willing to cooperate.