Unique Research · Unique Viewpoint · April 25, 2024
Background: On January 24, 2024, Unique Capital and partners hosted the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners. This piece is compiled from the guest roundtable.
Guests
- Xiaoyu Xu — Managing Partner, AMINO Capital
- Lu Zhang — Founding Partner, Fusion Fund
- Jun Qiu — Overseas Partner, Chinese Media Capital (华映资本)
- Lewis Hong — Founding Partner, FP Solutions VC
- Jierong Pan — Amazon Web Services (moderator)
Q: Please briefly introduce your firm and the sectors you focus on.
Xiaoyu Xu (AMINO Capital): Our investment thesis has always been big data. We see AI’s future as big-data deep learning + deep industry integration (product thinking) — AI does not invent brand-new domains so much as disrupt existing ones one by one.
Example: we invested in Replit in 2018; as a unicorn it can stand in for GitHub and for engineering teams setting up an IDE, so you no longer download a dozen teams’ and GitHub’s codebases — generate code on OpenAI, then run the product on Replit, even substituting for part of Google Cloud.
“Generative AI is real” — it solves concrete problems and does not open a wholly new competitive arena. AMINO’s job is to back products that overturn traditional competitive arenas. Product thinking is therefore our investment spine. To date AMINO has nearly 20 companies valued above $1B.
Jun Qiu (Chinese Media Capital): Founded in 2008, we manage about ¥10B, with offices in Beijing, Shanghai, Singapore, and Silicon Valley; nearly 280 portfolio companies across software/hardware, enterprise services, TMT, and hard tech; nine funds so far.
I graduated from Peking University CS, then studied AI at USC — my then-focus was what is now called Agents; my papers were mostly in Trip AI and Agent venues. After Cisco and Broadcom in Silicon Valley, then Tencent, I entered investing — over a decade of VC experience. Given that background and CMC’s AI layout, we have backed names such as AI chip Biren and AI annotation Stardust Data.
Lu Zhang (Fusion Fund): I am founding and managing partner of Fusion Fund. We are Silicon Valley–based and invest early in U.S. healthcare, Enterprise AI, and industrial automation startups. Founded 2015; more than 80 U.S. companies — supporting early growth and continuing to add in growth stages, helping founders with capital and industry integration.
My background was materials science and engineering research at Stanford; then founder; after exit, investor. We hope a deep read of the startup ecosystem helps early founders better.
Lewis Hong (FP Solutions VC): Co-founder of FP Solutions. Sectors include space, new energy, automation, automation+AI; our angle leans industrial-grade AI applications because we see many hardware-related opportunities in the U.S.
I am a former SpaceX executive; my partner is ex-Tesla, involved in the Shanghai Gigafactory and 4680 cell development. As an early cohort out of SpaceX and Tesla, our goal is to find the next SpaceX and Tesla.
Q: What is the biggest variable AI brings to value creation? For AI-infra-like startups, how should they position product and ecosystem to build lasting moats?
Lewis Hong: AI goes in two modules — process efficiency, and amplifying human capability. In the past year we have seen very direct new opportunities — industrial software that accelerates hardware engineers’ development and iteration inside embedded systems. Those are clear, high-attention demonstration projects.
Lu Zhang: AI Infra has long been a focus in our AI layout. When I discuss the AI stack I use “icing on the cake” — it looks attractive and draws attention, but the icing needs a solid cake underneath, especially when shipping AI apps. “Architecture is the key” — how you choose and build infrastructure is critical to using AI efficiently.
Our lens is simple: what are the hard challenges of using AI today? First, compute — how to cut GPU demand. Second, energy — power burn is very high. Third, communication efficiency — CPU–GPU communication and latency. Fourth, data privacy. Against these four, many infrastructure technologies respond.
Many founders obsess over optimizing the model itself — but “model optimization” is really usage efficiency: lower energy, lower compute, lower communication cost. Those are large AI Infrastructure opportunities. Infra innovation at this layer also does not flip every time models iterate. We have been laying out since 2018, including edge compute for communication; from chip to cloud to serverless edge we have strong companies embedding into industry.
Xiaoyu Xu: We enter via product. As investors AMINO invests in businesses; most companies already have design partners — which is why “generative AI is real.” In 2018, before ChatGPT, our Replit already had an online IDE with tens of millions of users — millions of engineers checking in data and code. That is infrastructure to us: data is paramount.
On a data platform, if the product is good enough, cloud services and a business model follow. Anthropic’s recently publicized margin issues are actually normal.
From our own history: in 2014 AMINO invested in face-recognition API company Orbeus, acquired by Amazon in 2015. One of Orbeus’s three modules was API calls — even at millions of calls, per-call profit was tiny and gross margin unacceptable. Orbeus took a product-thinking path similar to today’s OpenAI: launched PhotoTime, tried experiments including a Facebook mini-game that scored friends’ faces — accumulating data. With that, PhotoTime shipped ahead of Google Photos.
Many of our LPs were Facebook executives who watched Google Photos succeed at the app layer. Because Orbeus shipped PhotoTime early, it could run alongside Google Photos, and Amazon acquired the ReKognition API in 2015, reconstituting it as AWS ReKognition — valuable on AWS because AWS could offer the API free to paying cloud customers. Whatever the algorithm, you eventually need a product; that is why data-as-infra and cloud-as-infra both matter.
Hardware also matters. Recent disclosures put NVIDIA A100 inventory in the hundreds of thousands to millions — but stockpiling compute that way does not solve everything. At Davos, scientists noted that even with such compute, video generation remains hard; training on a 30-minute YouTube video at current compute is unrealistic, and that still lags NLP. New compute methods may break today’s bottlenecks.
Q: Are AI Agents the next big opportunity after foundation models? In which industries is commercial closed-loop easiest?
Lewis Hong: We have not gone deep on Agents; our angle is more industrial — human–computer interaction lifts and potential applications in the new workforce.
Did you follow CES early this year? The rabbit AI hardware drew huge attention — a new hardware platform built for AI. I see that as a dominant trend; we are excited about new imagination for hardware development.
We invested in a next-gen mobility + AI project: a new wheelchair for elderly and mobility-impaired users. The team partners with MIT, Carnegie Mellon, Pittsburgh and other top AI schools to build apps that help with daily tasks — e.g. opening a door from a wheelchair becomes feasible with AI/agent support. From a hardware angle this is the most exciting part of AI; AI also lets all hardware develop on a new-generation platform that carries endless imagination.
Jun Qiu: My PhD research in the U.S. was on intelligent agents. AI has had at least two phases. In the first, AI was a big bucket absorbing many directions — hard to define cleanly; agents were one domain. After deep learning, almost every area — vision, NLP, agents themselves — is being reshaped by DL.
The biggest variable AI brings, I think, is the technical path. Paths have fixed and changing points; fixed points should be viewed from ML/DL. Today’s agent tech differs a lot from what I studied early on — do not treat it as a standalone module.
Lu Zhang: Universal AI-agent tech still needs time to mature. We focus more on vertical AI applications — some of which can be called AI agents.
Example: one portfolio company focuses on pharma — think a ChatGPT specialized for pharma, retrained and optimized on pharma data for professionalism, precision, and consistency, so the agent becomes a professional product for that industry. Similar logic applies to finance, insurance, etc.
As in the scenes people imagine: ask for a Q3 long/short strategy industry report; based on public and internal data it can fully automate the report. On B2B and vertical industry apps, agents still have large opportunity. Training/optimization is also not as hard as imagined — consumer edges have many random edge cases; industry use is more concentrated and concrete.
Healthcare overall is a vertical I watch closely. Under this mega-trend, healthcare has a natural advantage: massive high-quality data and diverse scenes. I hope more founders pay attention to this vertical.
Q: With foundation models expected to improve cross-modal understanding and controllability — but paths still unclear — what capabilities should founders strengthen? When will multimodal large models have their real GPT moment?
Jun Qiu: Multimodal is quite complex. Innovation today concentrates at the foundation-model layer; we look for application-layer opportunities. Text→code→image/video is relatively more mature; cross-modal still has many openings — e.g. the hot embodied intelligence space, which will be a large research and opportunity area.
Because related work is still mostly foundational, application-layer room is limited. Difficulty can rival building a foundation model itself. On cross-modal we still need a dialectical, objective view.
I think the GPT moment is already close, with possible large leaps — but crossing into other modalities (e.g. intelligence lifts, healthcare) still needs time.
Q: For founders with a global market and go-abroad ambition, what advice? Which capabilities matter most to harden in 2024?
Lu Zhang: I warmly welcome and encourage more Chinese founders to lay out for global markets — even explore the U.S. AI innovation today is consumer AI on one side and enterprise AI on the other; for vertical B2B AI apps, the U.S. is still the world’s most mature market.
If you explore the U.S., my simple advice: relearn the market’s rules, partner efficiently and smartly with local teams. Start from zero — do not simply transplant what worked in China. The markets and business models differ fundamentally; the opportunity is still huge. Hope to see more outstanding Asian and Chinese entrepreneurs explore the U.S.
Lewis Hong: Entering a new market requires deep lived experience of localization. China has hardware advantages; the U.S. has innovation advantages. Combine both and accurately see the local specific problem.
Jun Qiu: In the AI wave, foundation-model opportunity is huge — but many founders underestimate the bar. This stage of foundation models and GenAI is not like internet/mobile internet: you cannot just ship a webpage or app and commercialize, nor simply program a machine — you program against a foundation model. You need real cognition of foundation-model architecture and capability; cutting in only from the top of the app layer rarely yields an effective model.
Example: people say private data + LLM is easy — every company has private data. But how to bind private data to the base model has many paths; how to store it, in what format and dimensions, is hard. Underestimating that lowers GenAI startup success rates.
Most founders need to recognize this, invest heavily in talent, go deep technically into the foundation model, carefully explore concrete scenes, and combine precisely with data — only then will trained models show strong adaptability and effectiveness in real use.
— END —