49. What Can China Learn from Israel’s Tech Innovation? | A 10-Year Retrospective from a Hard-Tech VC

Summary

• 杨光’s 10-year retrospective on Yaotu Capital shows its core strategy evolving from “competing with major US and European players for Israel’s frontier projects” to “taking inspiration from Israel and returning to China to back top teams with the right people-market fit.” The representative case: in 2016-2017, the firm encountered AI-chip company Habana in Israel but could not invest; it returned to China, mapped the talent pool, found 2 Chinese executives from AMD, and invested in their company, Enflame Technology, in September 2019. 3 months later, Habana sold to Intel for $2B. That became Yaotu’s differentiated edge.
• Israel is a global sourcing market for major technology companies, excels at going from 0 to 1, and more than 80% of startups exit through M&A. Mobileye was founded in 1999, did not win its first customer, Volvo, until 2008, and sold to Intel for $15.3B in 2017. Nvidia acquired Mellanox for $6.9B in 2020: Intel, Microsoft and Xilinx had bid only in the $5B-plus range, while Nvidia offered $1.5B more. The deal was widely viewed as an overpay at the time, but its foresight became clear in the era of large-model clusters.
• Hard-tech investing gives you “one pull of the trigger”: the businesses are capital intensive, can take 10 years, and even the best projects may return only 30-50x, so a 20-plus-company fund must hit 2 or 3 winners to make money. The founder test requires leaving the conference room: during diligence, one CEO said after drinking, “If I cash out RMB80M, why would I start another company?”—revealing a low ceiling for financial ambition. A VC wants someone aiming higher: build a RMB4B or RMB5B company and personally pursue RMB1B-RMB2B.
• Extreme competition downstream in China is the best friend upstream technology can have—one of 杨光’s key lessons. The company called “Fidelity Semiconductor” in the original, later shortened to “Fuda,” started by challenging Apple with 15W wireless charging and was repeatedly pushed by Xiaomi, Huawei and others to raise power levels; China now has 120W wired and 80W wireless charging. LiDAR was a case of “getting the technology path right but the market call wrong”: Yaotu backed Innoviz in 2017 on the solid-state route but underestimated China’s EV and intelligent-vehicle markets. In retrospect, it should have invested in one Israeli company and one Chinese company—local for local.
• The way to digest the 2021-2022 bubble is transaction design: persuade founders to accept a down round, compensate the previous investors, and in extreme cases sell old shares cheaply and lend the after-tax proceeds back to the company for survival capital. The Valley of Death is typically “commercialization below expectations plus financing constraints.” GPU and AI-chip companies Yaotu backed were once viewed as overvalued; some now generate RMB1B-RMB2B in revenue, and public filings show 6 companies preparing applications, but roughly 60%-70% did not make it through the cycle. He expects embodied intelligence to cool over the next 1-2 years.
• China’s M&A exit market is held back by pricing and mechanism: overseas acq-hire deals may pay $100M for 50 people, while domestic buyers sometimes acquire at cumulative funding raised, leaving founders with nothing and early investors merely breaking even. Regulatory requirements around PE multiples and net-profit commitments also suppress acquisition prices. Differential pricing is improving, but early investors still lack a good exit channel.
• The AI paths are diverging: the US is spending without regard to cost to chase SOTA, while China is pursuing affordable, broadly accessible AI; 杨光 believes Israel may fall behind in AI. The non-US, non-China model companies he cited include Israel’s AI21 Labs, Canada’s Coherent and France’s Mistral. DeepSeek’s effects include reducing chip requirements to support a 671B-parameter model that covers a large number of use cases, accelerating private deployment by state-owned enterprises, and releasing model-layer profits to applications and end users. The host added that more than 100K Qwen derivatives exist; 杨光 agreed that open source and a stronger generation of young domestic talent have lifted China’s confidence in AI.
• Geopolitics has reshaped fundraising and deployment: after Huawei and ZTE came under stringent sanctions in 2019, Yaotu reduced overseas investment and shifted toward China, global Chinese teams and companies expanding abroad. Trump’s policy proposed in January this year was referred to in the original as the “America First Investment Policy.” US capital faces restrictions on investing in China, although non-US dollar capital may still reallocate to Chinese assets. Today, 70%-80% of Yaotu’s effort is AI-related, with infrastructure accounting for roughly 60%-70%. Its most memorable moment of the decade was running in Jerusalem and Tel Aviv in the early days while debating whether to raise RMB20M or RMB50M for the first fund.

Deep dive

1. A Non-typical Dual-Currency Fund: Three Stages in 10 Years

• Host 卫诗婕’s framing: 杨光 (Eric) began his investing career at Israeli fund Infinity Group, co-founded Yaotu Capital with Jerry in 2015, and has spent the past decade focused on early-stage hard tech. The strategy moved through 3 phases: competing with major US and European players for Israel’s frontier projects; investing alongside Israel’s best teams and helping them enter China while copying the talent-matching playbook back into China; then shifting toward a more industrial, commercialization-driven approach after the Israel war and rising US-China tensions, while continuing to track Chinese founders globally.
• State before substance: Eric still travels constantly because he worries that once he leaves the front line, he will struggle to keep pace with rapidly changing technology. Anxiety about the speed of technological change is the undertone of the entire conversation.

2. Omron, Microsoft and SK: Two Technology Calls at the Start of a Career

• Born in 1984, 杨光 entered Shanghai Jiao Tong University in 2002 for a combined undergraduate and master’s program and studied Japanese. In 2007, he worked on facial recognition at Omron in Kyoto—“AI 1.0.” Back then, the work required graduate researchers and large teams for years; today, a middle-school student might be able to write a face-detection algorithm.
• In 2008, he worked on Windows Mobile at Microsoft. Resistive screens with styluses were already an “aha moment” at the time, but most of the phone customers he served “are no longer around today.” After experiencing the first revolutionary iPhone, he felt the Mobile division was on a sinking ship.
• After graduation, he joined South Korea’s SK Telecom. The world was still thoroughly globalized then; today, geopolitics has reshaped the semiconductor, energy and other operations of foreign companies in China.

3. Infinity and the F-16 Pilot: People Who See the Big Picture

• 杨光 returned from Seoul to Beijing in 2011 and joined Infinity through a friend. Infinity partnered with Suzhou Yuanhe, now known as Yuanhe Holdings, and placed its first fund in 杨光’s hometown, Suzhou. The 2 Israeli founding partners’ hiring criteria were engineers who were curious about new technology and, ideally, had overseas work experience.
• One of the founders had been an F-16 fighter pilot and one of the top special-mission pilots. 杨光 believes people like that see the world from a higher vantage point and are good at seeing the big picture; the founders were already bullish on China in 2004-2005.
• Before joining Infinity, he knew almost nothing about Israel. Once there, he found Israel at least at that time more advanced in the originality of many technologies than Japan and South Korea, the countries he knew best. Israel’s underlying belief is that if you dare to imagine something, you can go build it.

4. Capsule Endoscopy and a Hongkou Office: Two Ties to Israel

• 杨光 recounted a story he had heard: a missile-guidance engineer listened to a family member complain about the pain of gastroscopy and decided that, having solved more complex problems, he could solve this one too. He founded a capsule-endoscopy company built around a flash, a wireless-transmission module and a camera. Given Imaging later sold for roughly $1B.
• Yaotu deliberately chose an office in Hongkou rather than Lujiazui. On the eve of World War II, Shanghai issued more than 20K visas and helped more than 20K Jews enter the city; the Jewish Refugees Museum is nearby. 杨光 sees the location as a symbol of China-Israel cooperation and cultural ties.
• The new Lujiazui skyline outside the window and old Puxi represent a fusion of old and new. 杨光 believes that without joining a Jewish institution, he and his partner would have found it difficult to immerse themselves fully in technology during the internet boom of 2010-2015.

5. Infinity’s Style: Technology Calls and Talent Matching

• The comfort zone of Israeli VCs is judging technological change, not innovating around business models; Israel’s small market offers limited soil for business-model innovation. While many dollar funds at the time were heavily exposed to consumer and mobile internet, Infinity built a deep early-stage technology portfolio in China.
• An engineering background helps investors read papers and deep technical reports and speak the same language as industry experts. Non-engineers can develop the skill, but they need a strong methodology and mentorship; during the first 1-2 years in the industry, technical training materially affects learning speed.
• The methodology is sequential: start with TAM and future adoption rates, then break down the technology path; assess industry talent, technology owners and the talent pool; finally, find the founder with the strongest people-market fit.

6. Centec Networks: First Methodology Lesson and Darkest Hour

• The first project 杨光 and Jerry worked on at Infinity was switch-chip company 盛科网络, now listed on the STAR Market. Its founder had worked on chips at Cisco and returned to China in 2006-2007 with a single bag, betting that the rise of Huawei, ZTE and H3C would create demand for domestic switch chips.
• The post-investment reality was brutal: the company had limited cash, spent roughly 2 years developing a chip, and still could not find a switch maker willing to use it as a guinea pig. The team later built white-label switches and even equipment for internet cafes, doing whatever it could to survive.
• The Snowden revelations later pushed China to emphasize “secure and controllable” technology, now known as Xinchuang. Core equipment at government agencies, state-owned enterprises, military-industrial groups and research institutions could not use US chips, giving domestic switch-chip makers an opening. 盛科网络 was one of the best-prepared companies at the time.
• The counterexample was a core-chip company for fiber-to-the-home: China Mobile and China Telecom only began pushing FTTH a year after the company shut down. Technology choice and team quality both matter, but ultimately an entrepreneur still has to outlast the market window.

7. Judging Founders: The Truth Comes Out After You Leave the Room

• 杨光 acknowledges that founder resilience was “extremely difficult to assess” early on. Younger employees are often better at technical research, while experienced partners are better at judging a person’s ability to absorb setbacks, respond to details and draw on life experience.
• A US LP with 30 years of experience advised investors not to eat boxed lunches in the office, but to take a 1-hour walk and talk. A relaxed setting encourages people to describe the challenges of the past year and the underperforming companies in their portfolios. 杨光 concluded that conversations with founders should also go beyond the conference room.
• A negative example came from a company under diligence. After hearing that a founder in another portfolio company had cashed out RMB70M-RMB80M and then put RMB10M into a new venture, the CEO said over drinks: “If I cashed out RMB80M, why would I start another company?” 杨光 saw the comment as evidence of a low ceiling for financial ambition.
• If a founder owns 35% and only wants to make RMB80M, the implied ceiling for the company is roughly RMB150M. A VC taking 8-10 years of risk wants the company to reach RMB4B or RMB5B, with the founder personally pursuing RMB1B-RMB2B. The target is someone who aims higher.

8. Hard-Tech Economics: One Pull of the Trigger, 10 Years on the Bench

• Hard tech is constrained by capital intensity, long R&D and commercialization cycles, and limited fund sizes. A chip company must pay for top talent, EDA tools and tape-outs; a single tape-out can cost $10M-$20M.
• The front end may require 3-5 years of R&D, followed by another 3-5 years of commercialization. That is why the semiconductor industry says you must “sit on the bench for 10 cold years.”
• Most hard-tech companies may be worth RMB20B or RMB30B after listing, not RMB300B. After dilution through multiple rounds, even the best projects may generate 30-50x returns. A fund with more than 20 investments often needs 2 or 3 high-return winners to make money; successful mobile-internet cases can have 100x or 1,000x upside.
• 杨光 quoted a classmate working in AI: even if the company ultimately fails, investor money may still retain computer talent through high salaries, keeping engineers from moving into finance or consulting because of low pay. Over the past decade, Chinese engineer compensation has risen to roughly 70%-80% of overseas levels, reinforcing an engineering culture.

9. First Impressions of Israel: Scruffy, Frugal and Driven by Pessimism

• 杨光 expected Israel to be as fancy as Shanghai, Seoul or Tokyo. In reality, infrastructure was ordinary; outside a few CBD high-rises, much of the country resembled the urban-rural fringes of China.
• Some serial entrepreneurs who had sold several companies and were already wealthy still wore ordinary T-shirts and drove fuel-efficient Japanese or Korean cars. The explanation 杨光 relayed was that Israel has no oil, fuel must be imported, and material possessions could be destroyed in a war, so there is little reason to invest resources in conspicuous consumption.
• They appear optimistic but are pessimistic underneath: if Israel loses a single war, the country may cease to exist. So when they have an idea, they move quickly, hoping not to miss the opportunity over the next 3-5 years.
• 杨光 has met people still founding companies in their 60s and an entrepreneur giving a presentation while already suffering from Parkinson’s disease. Professors, students, soldiers and former comrades build companies together; successful founders donate to schools and society, creating a powerful culture of innovation, education and social responsibility.

10. Mobileye: 9 Years Without a Customer

• This origin story was relayed to 杨光: a computer-vision professor at Hebrew University met Akio Toyoda at a conference in Japan and was told not to focus only on e-commerce applications, but to put cameras in cars and alert drivers when they were drowsy or made mistakes. The professor returned home and founded Mobileye.
• Mobileye initially focused on algorithms, but chip compute was weak around 1999-2000. Even a strong algorithm would have unacceptable latency, so the founders later partnered with chip companies and eventually had to build chips themselves.
• Mobileye started in 1999 and did not win its first customer, Volvo, until 2008—a 9-year wait. 杨光 believes Israel has many stories of waiting years for a market, and that they encouraged later founders to pursue forward-looking directions.

11. From Bargain Hunting to the China Story: Yaotu’s Differentiated Playbook

• Infinity initially looked for cheap, underappreciated Israeli technology and brought it to China for validation. One semiconductor pilot-production technology was deployed in Suzhou and eventually produced a listed company. But the model required valuable technology that had not yet been discovered, plus a Chinese team capable of cross-border technology transfer, making it difficult to replicate.
• Yaotu changed its strategy in 2015. It did not seek control, but invested alongside Israel’s best companies and helped them enter China: “I’ll give you $3M and connect you with several major Chinese customers.” Israeli companies kept their core business in the US and Europe; Yaotu supplied the China story and smart money.
• The host added the Horizon Robotics story: Li Auto used Mobileye early on, then shifted to China’s Horizon after cooperation issues. 杨光 sees the case as evidence of the rise of Chinese companies and the Chinese downstream market.
• The deeper return was access to a network of exceptional founders. Great entrepreneurs would point to emerging mainstream technology paths, after which Yaotu would return to China to find the right team. Capital was limited, so most of the money still went into Chinese companies.

12. Unable to Invest in Habana, Then Backing Enflame in China

• In 2016-2017, Yaotu encountered Israel’s first AI-chip company, Habana Labs. The company wanted to challenge Nvidia with an innovative architecture, but Yaotu could not get into the deal.
• The team returned to China and mapped the talent pool, finding 2 Chinese executives from AMD Shanghai and AMD Toronto. They founded Enflame Technology. Both sides believed a more innovative architecture built with less capital might challenge Nvidia and AMD. Yaotu invested in September 2019.
• In December 2019, Habana sold to Intel for $2B. Its founders had previously sold 2 companies, one to Marvell and one to Amazon, and had deep relationships with major overseas technology companies as well as a clear read on AI demand.
• 杨光 believes Israeli founders tend to be sophisticated about both the direction of a venture and the timing of an exit. Had the deal been delayed until 2021-2022, Habana might not have commanded a higher price.

13. Mellanox: A $6.9B Bidding War

• 2 months before the acquisition was announced, 杨光 learned in Israel that Mellanox had 4 buyers: Nvidia, Intel, Microsoft and Xilinx. The other 3 bid only in the $5B-plus range; Nvidia offered $1.5B more and won at $6.9B. Many people called it an overpay at the time.
• In April 2020, there was no Transformer or large-model boom, and the market could not foresee demand for 100K-GPU or 200K-GPU clusters. Mellanox’s network chips connected large numbers of servers. They had primarily served the smaller supercomputing and HPC market, but AI later expanded the addressable market dramatically.
• After the Israel war, Nvidia executive 杰森 mentioned in an internal letter that the company had 4K employees in Israel, with the core team coming from the group retained after the Mellanox acquisition. Mellanox’s founder lost his youngest daughter and her fiancé in the October 7 terrorist attack; 杰森 wrote to express his condolences to the founder and affected employees.
• The host cited former guest 姚欣’s view that Mellanox was one of Nvidia’s most important acquisitions. 杨光 agreed, highlighting Nvidia’s ability to identify value and see ahead of the market.

14. Why Israel Fell Behind in Autos and Consumer Electronics

• Technological vitality needs a market to sustain it. After Mobileye, automotive innovation in Israel slowed partly because German automakers were slow to iterate and slow to embrace new technology. Fierce competition among Chinese automakers, by contrast, created demand for better autonomous driving and smart cockpits, pushing upstream innovation.
• 杨光’s engineer-supply thesis is that Israel, China and the US all have ample coders and engineers, while Europe and Japan lack the dividend from a young engineering workforce. Software and AI may ship a major new version every 3 months, which European and Japanese markets struggle to absorb.
• An investor’s knowledge does not come from being especially clever, he says, but from “standing on the shoulders of giants.” Speak with 3-5 of the best people in an industry within 1-2 months; if they all say a technology is close to maturity, your understanding improves rapidly.
• He admires 余凯 for turning autonomous driving into a real business after many people initially called the idea impossible or even viewed him as a fraud.

15. The AR/VR Tuition Bill: Must-Have and Converging Technology Paths

• Magic Leap triggered an AR/VR boom in 2015. Yaotu systematically studied chips, waveguides, control chips and controllers, and invested in 2 companies.
• The waveguide company remains one of the industry leaders. 杨光 believes it “should be able to enter Meta’s supply chain,” but it has never reached mass commercialization; the ring-controller company has disappeared.
• The problem with the ring was dependence on the AR/VR market scaling and uncertainty around the control method. The eventual interface could be a ring, a controller, gestures or something else. Vision Pro relies mainly on gestures, but other methods may still emerge.
• Waveguides are essential to the display stack. As long as downstream giants continue investing in R&D, a company with scarce technology and customer relationships can survive. Controller companies face much higher commercialization risk before the downstream market takes off.
• The direction that was validated was AI. Habana, Mellanox and similar companies were early examples of Israeli founders correctly anticipating future demand for compute.

16. The Israeli Shopping List of Big Tech

• Intel’s early dual-core processors, as well as switch chips from Cisco and Broadcom, were linked to Israeli R&D or acquisitions. After acquiring a team, major companies would place it in a local R&D center and continue acquiring a second and third company.
• Israeli labor costs are roughly 70% of Silicon Valley’s, while technical ability and innovation are strong. Talent turnover is also relatively stable versus Silicon Valley, encouraging large companies to build core R&D capabilities in Israel.
• This was part of the internal logic behind investing in 盛科网络 in 2011: Israeli switch-chip companies had successfully sold to Cisco and Broadcom, so a Chinese company would either be validated by the market or eventually attract a buyer.

17. Yunbao Intelligent: Reclassifying “Nice to Have” as “Must Have”

• After Nvidia acquired Mellanox, Yaotu decided it had to find a domestic company with similar technology as quickly as possible. The team searched the relevant talent pool but could not find a suitable Chinese team.
• When Broadcom later launched a Mellanox-like product, Yaotu inferred that Broadcom must have Chinese chip talent internally. 杨光 noted that Nvidia’s 黄仁勋, AMD’s 苏姿丰 and 陈立武, who later went to Intel, are Chinese; his original point was that “many” people in Silicon Valley’s chip sector are Chinese, not that everyone is.
• Yaotu eventually found several Chinese executives who had left Broadcom and invested in the company they founded, Yunbao Intelligent, alongside Tencent and Sequoia in its angel round. The company was roughly 4 years old, and 杨光 viewed it as one of the domestic teams with the best chance of replacing Mellanox.
• The thesis had 2 legs: around the emergence of BERT, model parameters exceeded 1B, requiring multiple servers to train jointly and increasing demand for network chips; meanwhile, roughly 30% of cloud providers’ CPUs were consumed by network virtualization, encryption/decryption and protocol processing. SmartNICs could take on those workloads, freeing CPU and increasing rentable capacity. ChatGPT brought the importance of network chips to a much wider audience.

18. Founding in 2015: 31, No Track Record and Overflowing Confidence

• 杨光 and Jerry were the same age, born just 3 days apart, and founded the firm after working together for 4 years. The reasons included overseas partners at Infinity lacking a deep understanding of China and the need to combine the investment methodology with China’s industrial ecosystem. The year 2015 also brought the mass-entrepreneurship and innovation campaign, VC 2.0 and the overseas micro-VC wave.
• They believed O2O and P2P financing were too easy and necessarily contained a bubble, but with sentiment running high, they should still seize the opportunity. Overseas micro-VCs at the time managed $50M, and Yaotu originally hoped to build a fund of similar size.
• They underestimated fundraising difficulty. At Infinity, they did not interact with LPs and focused only on the investment-manager job, so they assumed capital could be raised in 2-3 months. At 31 in 2015, neither had a semiconductor or software investment that had fully matured; without a meaningful track record, convincing investors to back the next generation of hard tech was extremely difficult.

19. Fundraising Through Israel Study Trips: From EMBA Groups to Industry Leaders

• The only approach that worked was taking technology-oriented entrepreneurs to Israel and converting them into personal LPs. 杨光 left Infinity on June 30, 2015, and began building Yaotu on July 1. Its first RMB fund and dollar fund were not fully established until 2016-2017, with sizes of RMB180M and $20M, respectively. In between, the partners made more than a dozen trips to Israel and led more than a dozen groups.
• The turning point came from blunt feedback by an Israeli founder: the meetings were not responsive enough, the questions were not sharp enough, and there was no next step afterward. Yaotu changed its audience and began inviting industrial companies including OPPO, vivo, Sunlord, O-film and TRO.
• For Israeli companies, these Chinese industrial customers could generate real orders. For Chinese companies, Israel provided a differentiated technology gateway. From 2015-2019, the Chinese market still looked up to overseas technology, and many people were less interested in domestic technology.
• Seeing 20-30 companies in a week and visiting Jerusalem and other locations together made relationships easier to build. Yaotu could then take Chinese founders to phone and auto makers to validate industrial connections. Compared with taking money directly from a CVC and being forced to pick sides, taking Yaotu’s money offered customer access without alignment pressure.

20. A Shabbat Dinner: Perhaps a Child Not Yet 10

• 杨光 once attended a Shabbat dinner at the home of a Teva executive. The host asked a son who might not yet have been 10 to share what he had learned from the Bible that day: one person visited only the desert in southern Israel and saw a barren land; another traveled from south to north and saw oases and abundance.
• The child’s takeaway was that he had just seen news of Hamas launching missiles and rockets, and many people said Israel was unsafe, but he had still gone to school and studied with his classmates that day. The issue could not be viewed one-sidedly.
• Entrepreneurs were struck by the form of Shabbat: from sunset Friday to sunset Saturday, no work, phones or email—only family time and deep conversations with children. 杨光 sees it as evidence of the culture’s emphasis on family ties.

21. “Fidelity Semiconductor” in the Original, Later “Fuda”: Betting on 2x Apple’s Number

• Around 2018, Apple added wireless charging support to the iPhone SE, prompting Yaotu to study the market. Android makers said wired charging had already reached 30W and Apple’s 7.5W wireless charging was meaningless; they would try it only if a Chinese supplier could reach roughly 15W, or 2x Apple’s level.
• Among several candidates, Yaotu chose the third: revenue was still limited, but both founders came from TI, the CEO had run TI’s power-management product line, and the CTO was also highly experienced. The company was about to run out of angel funding, so Yaotu provided a bridge loan and persuaded it to abandon accessories and focus on challenging the smartphone-chip market.
• Xiaomi later decided to partner with and invest in the company, after which Yaotu added to its position. The company built the first 15W wireless-charging chip and entered the Xiaomi 9; OPPO, Huawei and others followed. Huawei demanded 30W, and Xiaomi then demanded 50W, creating a power race that lasted for years.
• The company moved from being a vendor to a supplier capable of pricing to customer specifications. OPPO and Samsung later invested in or adopted its chips. 杨光 notes that China now has 120W wired and 80W wireless charging, while Apple made limited progress beyond 15W wired and 7.5W wireless for years—a clear example of downstream Chinese competition driving upstream innovation.

22. DVS and Systematic Positioning: The Industrial-LP Flywheel

• The lead on DVS, or Dynamic Vision Sensor, came from a question Chinese phone makers asked while visiting Israel. A standard camera captures roughly 25 frames per second, while DVS can reach 200-2,000 frames per second, reducing motion blur and improving stabilization.
• Research showed that the technology originated with a professor at the Swiss Federal Institute of Technology Lausanne. His students later founded companies in Switzerland, France and Shenzhen. Yaotu invested in the Shenzhen company referred to in the original as “瑞思致敬.”
• Apple had filed related patents but had not used the technology as of the interview. 杨光 said Chinese phone makers had already decided to adopt it and did not need to wait for Apple to validate the market.
• Yaotu’s transformation was to enter through Israel and validate in China. Wireless charging did not come from an Israeli project; it came from observing Apple’s moves and combining them with demand from Chinese manufacturers.
• In areas of high conviction such as AI, Yaotu has built a systematic portfolio across data-center compute, networking, optical-transmission chips, silicon photonics and memory. That system-level positioning and industrial connectivity are why many listed companies are willing to become LPs.

23. An Underwater Project: Bottom-Up Deal Sourcing

• Elite founders can often be found through a top-down framework. Younger founders require tracing through industry and academia, reading papers and monitoring activity in open-source communities, then reaching out directly.
• A young investment manager who graduated from Fudan’s computer-science program believed he could not directly locate executives from Alibaba or ByteDance, so he went to places where engineers gathered, including Zhihu and GitHub. He wrote articles explaining Habana’s value and shared open-source research on GitHub.
• Engineers who saw the research sometimes contacted him to discuss the sector. Some were still exploring whether to start a company when Yaotu had already held a first-hand conversation with them. That is bottom-up deal sourcing.
• Organizationally, 杨光 favors trusting and empowering younger people. Once research and analysis are complete, the firm should not demand excessive consensus from everyone; younger colleagues should be allowed to trust their own judgment.

24. Solid-State LiDAR: Right on the Technology, Wrong on the Market

• Yaotu invested in Innoviz in 2017. After speaking with roughly 10 overseas companies and 3-4 Chinese companies, the team asked founders to analyze each technology path, supply-chain maturity and dependence on internal versus external capabilities. It concluded that solid state was closer to the long-term direction, with stronger front-vehicle fit and higher barriers.
• At the time, Hesai and RoboSense were mainly building mechanical LiDAR as alternatives to Velodyne for Robotaxis, and the team had not fully recognized the OEM market. A Robotaxi can mount a mechanical sensor on the roof, but front-installed systems for ordinary consumers and a useful life of roughly 10 years require greater durability.
• In hindsight, Yaotu believes investing in the 2 Chinese companies could have generated higher returns. The flawed assumption was that Israeli companies were closer to German and US OEMs. In 2017, autonomous-driving kits cost roughly $5K-$10K and were mainly suited to premium models from BMW, Mercedes-Benz and Audi, while Chinese domestic brands were mostly priced in the RMB100K range.
• They underestimated the speed of China’s EV market and downstream competition. The revised playbook is to invest in one Israeli company and one Chinese company to serve overseas and domestic markets respectively. Innoviz entered BMW and Volkswagen but missed China’s growth; Chinese automakers can develop 1 or 2 new models in roughly 18 months, while BMW and Volkswagen still need 3-5 years.

25. A Fight for Survival Among 100-Plus Brands: Involution Creates Early Adopters

• More than 100 Chinese auto brands compete on range, intelligent driving and smart cockpits. An increasing number of models above RMB150K are adding LiDAR. General Motors and Ford bought Cruise and Argo, but after Tesla pulled ahead, they lacked the same momentum for continuous iteration.
• Chinese automakers may install LiDAR before the algorithms are fully ready, then upgrade them over the air once the software matures. 杨光 acknowledges that fancy autonomous driving on RMB100K-RMB150K cars may not make economic sense, but automakers do not want to be left behind.
• China’s advantage is the number of brands and intensity of competition: automakers must keep searching for features that truly attract users. Fierce downstream competition is friendly to upstream technology companies because it drives aggressive adoption. 杨光 believes Japan and Germany have more stable competitive structures and therefore lack the same adoption pressure.

26. An Open-Book Exam: VC Differentiation After the State Picks the Track

• State think tanks and policy guidance have already laid out many supported sectors. RMB funds typically invest along those lines, making hard-tech investing resemble an open-book exam.
• Differentiation comes mainly from 2 things: finding the best team in a given sector and helping it commercialize. 杨光 believes it is difficult to turn a second-tier team into a first-tier team through resources alone, but helping an already excellent technology team win customers and reach the front of the industry is achievable post-investment value.
• “Out-of-syllabus” questions still exist. SaaS may lack obvious policy support, but it can still have market value when combined with AI to improve productivity. Market-oriented funds should retain first-principles and contrarian thinking and look for opportunities in sectors that are too cold or too hot.

27. 2019 Changes Everything: Sanctions, America First and Global Chinese Talent

• After Huawei and ZTE came under stringent sanctions, Yaotu felt the geopolitical shift and increased its exposure to domestic substitution. Overseas investment fell and the focus moved to China. The pandemic and war also made it harder for Israeli companies to enter China.
• Dollar funds still exist, but overseas investments now focus mainly on Chinese teams and companies led by Chinese founders that sell into overseas markets; the global entity may be incorporated in Singapore or another international jurisdiction.
• 杨光 believes Chinese people are sufficiently strong in hard tech. Much of the hardest technical work overseas is done by Chinese and Indian engineers, while Chinese teams have greater affinity with Yaotu’s own background.
• The original mentioned Biden’s executive order and Trump’s “America First Investment Policy,” proposed in January this year. The policy restricts Chinese investment in US critical technology, restricts US investment into China and limits Americans’ ability to start companies in China. 杨光 believes that even with a Cayman structure, Chinese funds will struggle to participate in mainstream US deals.
• Yaotu still tracks Israel, but no longer invests in many companies there each year—possibly just 1. The center of gravity is increasingly China, or helping Chinese companies expand overseas.

28. The Dollar Fund Playbook and Chinese-Style Deep Tech

• The host quoted 米磊: in the dollar-fund era, the market looked down both on RMB funds and on funds investing in hard tech. 杨光 believes dollar funds are still worth studying: back imaginative ideas, provide enough time, and remain purely return-driven.
• But the dollar playbook must be combined with the RMB fund’s deep industrial understanding. Overseas deep tech or hard tech more often refers to breakthrough technologies such as quantum computing, brain-computer interfaces and controlled nuclear fusion, where failure rates are high and capital intensity is even greater than in semiconductors.
• A large share of Chinese hard tech is application-driven technology: it solves engineering problems, carries relatively controllable risk and adapts along the way to the needs of China’s leading customers.

29. From Polar Night to Midnight Sun: The Bubble’s Bill Is Paid Slowly

• Hard tech went from years on the bench to a frenzy in 2022, rapidly lifting the mark-to-market value of Yaotu’s portfolio. 杨光 was not especially happy: valuation gains may have outrun the milestones supporting the fundamentals, and founders can easily become overconfident.
• Many companies later entered a valuation digestion period. If revenue doubles over the next 2 years, valuations from 2 years ago may gain fundamental support; if not, both financing and IPO plans will face pressure.
• Hot sectors such as embodied intelligence also pulled attention away from industrial and service robotics. Some of those companies may generate 10x or 20x the revenue of humanoid-robot companies while trading at much lower valuations. If humanoid-robot investments disappoint, capital may reassess these overlooked businesses.
• 杨光 expects embodied intelligence to cool over the next 1-2 years. No market in China stays hot forever; an IPO is also like thousands of soldiers trying to cross a single-plank bridge.

30. Down-Round Transaction Design and the Valley of Death

• For high-valued companies, Yaotu urges founders to push commercialization as hard as possible, spend only where it matters and seek an up round or at least a flat round. If the fundamentals have not materialized, valuation must be reduced.
• If the previous valuation was RMB2B and is now reset to RMB1.5B, the founder can compensate the previous investors by a few percentage points to get them to accept the down round. New investors are then more likely to accept RMB1.5B as a reasonable valuation.
• In an extreme case, the founder can sell old shares at a very low price and lend the after-tax proceeds back to the company. For example, a new investor invests at a RMB2B valuation while also buying RMB200M of shares with a nominal value of RMB200M but paying only RMB20M; the founder then lends the funds back to the company, creating survival capital.
• 杨光 says the “death stock” typically combines commercialization below expectations with financing constraints. Yaotu helps design the transaction structure and connect customers to support the company through an endurance race; without a clear IPO path, a company valued at RMB2B-RMB3B will find financing extremely difficult.

31. Cycle Survivors and a FOMO Self-Check

• GPU and AI-chip companies invested in 2019-2020 were marked to as high as $2B in 2021-2022. The market thought the prices were too high at the time; several years later, some generate RMB1B-RMB2B in revenue, and public information shows 6 companies preparing filing materials. Valuations and PS multiples may look reasonable in hindsight.
• Not every company survived. 杨光 estimates that roughly 60%-70% did not make it through the cycle. The same pattern could recur in humanoid robots and embodied intelligence; winter is the coming-of-age test for startups.
• 杨光 admits that when new energy was hot in 2022, Yaotu invested in several companies, some of which underperformed expectations. Sentiment made the judgments less rational.
• In familiar sectors, an investor’s bar rises steadily. In a new sector, insufficient knowledge can make the bar too low. Every new industry requires accumulated learning and humility.

32. M&A Exits: The Gap Between Acq-Hire and Acquisitions Priced off Funding

• Overseas markets, especially Israel, have an acq-hire culture: the buyer wants the team and technology and may pay $100M for 50 people. Investors who put in $30M have a chance to make money, and founders can monetize with dignity. Gutting a startup team through direct hiring can damage a major company’s reputation.
• More than 80% of Israeli startups exit through M&A, 杨光 says. Founders do not need every capability required to go from 0 to 100; strong technology, a good product and the ability to attract top talent may be enough to get acquired.
• In China, M&A is often used to add profit and is constrained by the CSRC, exchanges, PE multiples and net-profit commitments. Buyers therefore struggle to pay up for loss-making teams and technology. In some cases, the price is based on cumulative funding: a company that raised RMB200M is acquired for RMB200M, allowing investors to recover principal while leaving founders with nothing.
• Overseas markets are fully competitive, and institutional investors bear responsibility for buying too high or buying the wrong asset. Whether Nvidia overpaid for Mellanox is ultimately judged by market prices and investor trading. China has introduced some differential pricing—for example, 1x for the last round, 8% annualized for middle-round investors and 2x for early investors—but early investors still lack a good exit channel. When a company can list directly or raise smoothly, selling some old shares may offer a higher return.

33. Where Long-Term Capital Is Going: Dollar Retreat and Hong Kong Reallocation

• US policy restrictions mean many US LPs can no longer invest in China, weighing on fundraising and investment by Chinese dollar funds. Some capital has shifted to projects founded by Chinese people in Southeast Asia or the US; Yaotu still wants to back excellent Chinese companies.
• Hong Kong stocks have recently improved. Non-US dollar capital may reallocate from the US into Chinese assets, with Hong Kong stocks often the first step. Some technology stocks, including Horizon Robotics, trade at 70%-80% of their A-share valuations, while the overall valuation level remains attractive.
• Entering China’s mainland primary market and investing in VC requires LPs to be more optimistic about the next 10 years. Yaotu wants investors firmly bullish on China and willing to commit capital for the long term, including long-duration institutions such as Hong Kong-based Yitu Capital.
• 杨光 remains bullish on China’s technological competitiveness. A unified domestic market is an advantage in consumer electronics, autos and AI applications; products forged in China’s fierce competition remain competitive overseas, although access to the US market is another question. Yaotu will seek alpha within high-growth beta sectors, arguing that its small scale still gives it a chance to find 10%-20% growth opportunities.

34. DeepSeek’s Three Effects

• 70%-80% of Yaotu’s current effort is AI-related, with infrastructure accounting for roughly 60%-70%; application areas include robotics, AI for Science and software. 杨光 is excited about AI, though not because excitement had been absent—every year brings a new hot sector. AI matters because it is epochal.
• Yaotu did not invest in DeepSeek, but companies such as SiliconFlow benefited directly. For chip companies, supporting large numbers of open-source and closed-source models previously required substantial software work. Now, being able to support DeepSeek’s 671B-parameter model may cover roughly 80% of application scenarios.
• DeepSeek also accelerated private deployment. State-owned enterprises previously struggled to use large-model APIs hosted in the cloud; with open source, companies with strong IT capabilities can deploy models themselves, while weaker companies can hire application-software providers. This has revived a group of software companies.
• On the technology curve, 杨光 believes OpenAI’s reasoning capability is strong and that inference is moving closer to human thought. Multimodal models have not yet reached a comparable level to DeepSeek, and VLA brain models for embodied robots may still need 2-3 years. At the current pace of iteration, the relevant industries should remain active for at least several more years.

35. Open Source Rewrites China’s AI Economics: Qwen and Profit Release

• The host added that before DeepSeek exploded onto the scene, the overseas version of Qwen already had influence among global developers, with more than 100K derivative models and an earlier lead over Llama. The market had not broadly recognized the open-source trend then; DeepSeek created much stronger confidence.
• 杨光 noted that DeepSeek released distilled versions of both Llama and Qwen at launch. Qwen’s small models already had an advantage, which was strengthened after training by R1 as the teacher model; DeepSeek’s reasoning capability complemented Qwen’s small-model capabilities.
• He believes models can reinforce one another and drive down costs, bringing more downstream ecosystems into the market. Profit at the model layer will flow to applications and end users: in the past, a To B application calling OpenAI’s API might have passed half its revenue to the model company; open source allows the application layer to capture more value.
• The external signal is that capital and talent blockades did not stop China’s AI progress. China did not see a large wave of overseas talent at the level of OpenAI or Google Gemini return home, but domestically trained master’s and doctoral graduates—many of them very young—still produced results.
• Tencent Yuanbao demonstrates the value of productization. Product experience, server capacity and the WeChat Official Account knowledge base can combine to deliver a user experience that may be better than DeepSeek’s.

36. The China-US-Israel AI Divide: SOTA Versus Affordability

• 杨光 believes Israel may fall behind in AI because foundation models require a large market and substantial capital. Outside China and the US, he cited Israel’s AI21 Labs, Canada’s Coherent and France’s Mistral, but believes long-term competitiveness will remain concentrated in China and the US.
• The US is spending without regard to cost to chase SOTA. xAI, OpenAI and Google Gemini continue pushing toward the technological frontier; even after DeepSeek’s shock, xAI is building a 200K-GPU cluster.
• China’s approach is more pragmatic, emphasizing affordable AI and low-cost applications. 梁文锋 and DeepSeek may also want to reach AGI, but the first step is to make the technology usable, reduce costs and support a broader range of applications.
• 杨光 believes both paths can work if they ultimately lead to AGI. China may be taking the more ingenious route, but it will still require sustained investment.

37. Resource Waste in the Bubble and Staying Calm

• 杨光 believes too many teams have entered embodied intelligence, and concentrated capital is creating duplicated effort. It might be more effective for different teams to build the brain, body and cerebellum than for hundreds of teams to pursue the same full-stack configuration.
• If a DeepSeek-like moment arrives—a company uses less capital to build and open-source a brain model—the excessive capital invested earlier could become wasted industry resources. A sector with overwhelming consensus is not necessarily easy to make money in.
• Yaotu generally does not chase hot sectors; it prefers to position before the heat arrives. It looks for opportunities in sectors where it has deep knowledge and does not try to capture every opportunity. 杨光 also believes a moderate bubble can help an industry develop.

38. The Defining Moment of 10 Years: Running in Jerusalem

• The most memorable image is of the 2 partners running in Jerusalem and Tel Aviv during the firm’s early days, debating whether the first fund should raise RMB20M or RMB50M. They were young, optimistic and fearless through ignorance, believing that once the fund launched and a few investments were made, the brand would take off.
• Looking back, without that optimism, fearlessness and momentum, it would have been difficult to start the business. Today, even with more experience and more contacts, they might not be able to retain the same degree of optimism when facing fundraising, exits and the industry’s other challenges.
• The decade’s achievements include a more complete team, a group of long-term dollar and RMB investors, and support for nearly 100 CEOs. 杨光 believes the past 10 years brought a major shift in the VC investment paradigm, with some former industry stars leaving after missing the opportunity.
• His remaining confidence comes from having captured several opportunities and seen the principles he defended validated. That is enough to keep him in the game. However difficult it gets, it will not be harder than the first 2 or 3 years of building the firm.