The AI Inflection Point in the GPT Wave: Who's Betting Big? Who's Benefiting?

On February 28, ChinaVenture hosted a closed-door salon in Shanghai on the "AI Singularity," focusing on how ChatGPT is reshaping entrepreneurship and investment in artificial intelligence. Zheng Can, managing director at Linear Capital, and Fangbo Tao, founder and CEO of Mindverse — an angel-round portfolio company of Linear Capital — attended and shared their perspectives. Zheng noted that Linear had already been tracking and investing across multiple AIGC scenarios; many of the use cases being discussed today existed before, but today's large models have made their accelerating generalization visible to everyone, bringing with it massive infrastructure demands as well as new imagination and opportunity. The following is a distilled transcript of the event.

By Zhang Lijuan

Source: ChinaVenture

One hundred million monthly active users in 60 days — OpenAI's ChatGPT burst onto the scene and made AI's current wave of innovation familiar to the general public. Since then, Google, Microsoft, Baidu, and others have joined the race.

So beneath the well-known,既定 opportunities in AI, who is thinking and who is acting? Some investors have already moved; others are watching. Some entrepreneurs are bullish; others are proceeding cautiously.

Of course, the steady stream of industry figures like Huiwen Wang, Bowen Zhou, and Jie Tang entering the arena has added fuel to the AGI fire that ChatGPT represents, drawing enthusiastic backing from investors. With unicorns already established in AI sub-fields like speech and image recognition, cheering on AI innovation's forward march has become the obviously correct position regardless of one's private views.

But the problem is that not every technology wave's dividends are truly captured by the previous generation's giants, nor are startups guaranteed to seize innovative opportunities. As AI rapidly becomes a general-purpose technology empowering every industry and enters the era of general-purpose large models, who will actually grasp this chance to pull ahead?

Against this backdrop, on February 28, ChinaVenture held a closed-door salon on the "AI Singularity." Supported by Lightspeed, this installment of ChinaVenture's 2½ Series salon brought together over a dozen guests who offered industry perspectives worth contemplating on AI's current trajectory and investment directions.

Invest or Not, Everyone's Flying the Flag

At the salon, investors from both RMB and USD funds were observing AI-related opportunities and waiting for the right moment to act.

The only difference, as Oriental Fortune Capital partner Wang Bing put it, is that RMB funds tend to favor underlying infrastructure — clear business models with core moats and well-defined returns are worth betting on.

USD funds, by contrast, are more willing to take risks and make concentrated bets. Liu Lan, partner at Redpoint China, noted that for USD investors, wagering on large models is imperative; they hope to catch up or even surpass through greater investment, especially with entrepreneurs like Huiwen Wang and Bowen Zhou who possess exceptional fundraising ability and team-building capacity — they're worth a shot.

Yet attendees also lamented that with so many companies rushing toward GPT, even Baidu — seemingly China's most determined player — must wait to see how Ernie Bot performs post-launch. Numerous startups will need to find opportunities in "vertical" directions, discovering their advantages through differentiation to form a genuine domestic ecosystem.

As Zhang Qing, managing director at CICC Capital, put it: whether AIGC or AGI ends up more like the metaverse — hotly hyped two or three years ago but rarely landing — or more like the internet and mobile internet eras that spawned giant after giant and carved out broad avenues, has only just begun to unfold.

Even at the current AIGC stage, this isn't a new story for practitioners or investors. How to justify valuations? What about regulatory compliance? How does the business actually land? All remain to be further validated.

After all, throughout AI's development, speech and vision offered relatively limited imagination. Today's ChatGPT has already exceeded many people's expectations. Whether this "miracle" will match or surpass the space the metaverse promised, even transforming entire industries, awaits a true brute-force breakthrough.

Ultimately, AI returns to core elements: compute, algorithms, data, and sensing. These four advance in coordinated, spiral fashion. Only by building the highest barriers across all four within a given timeframe can dominance be achieved. Why does OpenAI seem unmatched? Because it reached a certain stage of balanced development across these dimensions. A beautiful future lies before investors and entrepreneurs — but one with significantly raised barriers.

Zheng Can, managing director at Linear Capital, elaborated that Linear had previously focused on and made multiple investments in AIGC scenarios. Many of today's touted AIGC applications already existed; what large models have done is make their accelerating generalization visible to all, bringing massive infrastructure workloads alongside new imagination and opportunity.

With "money, talent, chips, and data" as key factors, "money" and "talent" are relatively manageable — there's always money to be found, and people know where to dig for talent, though whether they can actually extract it and at what cost varies. "Chips" aren't terribly concerning either; two or three years should yield progress. The greater worry is data. Data cleaning demands are high, yet commercial returns are relatively modest. Only with quality Chinese datasets and future interconnection with other models can further innovation and development continue.

Liu Lan of Redpoint China, citing her firm's investments in two compute chip companies, observed that China's massive data volumes enabled AI security and other niche startups to grow. But this wave's core has shifted to algorithmic models. Because this matters enormously for China, even with large models offering the highest payoff odds, participation is essential.

Only with China's own algorithmic models can talent at major companies, in startups, and at universities be more effectively mobilized, forming mature business models. With so many practitioners already pouring in and a百花齐放态势 emerging, the logic of betting on models first and applications second can guide continued investment.

Cai Wei, partner at Lightspeed, noted that in the previous AI entrepreneurship wave, many investors placed bets and saw returns, but also hit an "AI innovation wall." It wasn't until 2021, as large model applications spread, and 2022, when generative models produced excellent applications like Jasper, plus Lightspeed's U.S. investment in Stability.AI, and now ChatGPT's rising capabilities, that this wall was broken — strengthening his conviction to keep investing in this space.

After all, once NLP produces emergent effects, will images follow? Will future large models become vertical-domain models directly, or will multimodal large models unify everything? Any shift would alter AI's developmental trajectory and spawn new commercializable models, empowering more industries — precisely where many opportunities lie.

Liang Junzhang, founding partner at Kunzhong Capital, has since last year been discussing with portfolio AI companies how to leverage this new wave to upgrade and strengthen their businesses. He's thus focused on whether more companies can produce targeted industry solutions for sector-specific applications, and on screening targets among the dizzying array of end-side applications according to his investment habits.

Moreover, given China-U.S. differences, in China's context B2B companies typically can't solve industry pain points through product alone; they must provide complete solutions and services. Companies already operating this way can push forward on their existing industry know-how, customer accumulation, and channel advantages.

Conversely, B2C may see new companies emerge, whether in gaming or entertainment — the metaverse layer warrants attention. Just as China always seems to produce gaming companies, this is worth anticipating.

Shen Libin, CEO of Leyan Technology, was blunt: the U.S. has many startups built on OpenAI, but whether China will have this wave of entrepreneurial opportunity remains unknown. What's more visible is that large model companies currently have two opportunities: open source and vertical specialization. After all, general models can't be directly applied to vertical domains — there's still a time window.

SaaS enterprises that skillfully employ large model technology, especially in interaction-heavy domains, can build capability advantages amid homogeneous competition. Technology has windows; finding vertical applications through large model-like approaches might yield a six-month to one-year lead — though two years of unassailability seems too unlikely. Thus more opportunities appear reserved for companies already at scale within industries, not newcomers.

Digging deeper: the devil is in the details. In customer service, for instance, can interaction generate round-specific scripts tailored to each store's operational strategy, tightly binding technological innovation with industry know-how to more easily form one's own moat?

Do or Don't, Discerning Real from Fake Opportunities

At the salon, guests vigorously debated why OpenAI wasn't the product of market-driven VC investment. OpenAI began as a non-profit researching AI's impact on humanity, with no commercial KPIs. Any commercial VC requires clear commercial landing targets from its investments.

Thus, when OpenAI produced GPT-3, Microsoft saw its potential and — wanting to advance its own cloud services — ultimately invested. Even so, Microsoft's investment was business-like: 80% of the money flowed back to Microsoft as its own revenue, with share prices rising significantly on the news.

So how should more startups leverage this wave to grow?

Wang Bing of Oriental Fortune Capital noted that OpenAI had likely been training models with unlimited budgets for roughly four years. With the U.S. not open-sourcing, domestic companies face a considerable journey to figure things out independently.

Three barriers exist: hardware — manageable for leaders but difficult for most startups; talent — beyond pure academic barriers, even experienced people need accumulation; and data — 95% of academic literature is in English, and its accuracy matters enormously.

Why did the "AI Four Dragons" ultimately stop making money? Wang analyzed that it wasn't their specialized capabilities but rather low barriers. Whether general or specialized AI, high technology and data barriers are key to profitability.

Li Kejia, partner at Heart Capital, offered perspectives as both tech entrepreneur and investor. From user and data perspectives, OpenAI remains in a league of its own, strengthening closed-source large models. As costs for developers plunge dramatically, "ecosystem (parasitic) entrepreneurial opportunities" built on OpenAI remain the highest-certainty bet. OpenAI reinforces this through capital, launching accelerator Converge — offering not just funding but special incentives including licensing discounts and early access to GPT-4 and other new technologies.

Open-source large models like Meta's LLaMA remain worth watching and investing in. As latecomers,叠加 certain safety and compliance issues, open and closed source will push the entire industry forward in competitive chase — familiar from three decades of chip architecture and operating system development.

For entrepreneurial opportunities beyond self-developed large models, one might hypothesize: if a product's main differentiation ultimately lies in AI itself, then verticalization plus middleware (large model and domain model training and hosting for developers) will likely win. But at the application layer, AI demand will show long-tail characteristics, making horizontalization more probable. Over time, we should see more traditional moats re-established, returning to business and efficiency fundamentals — including new types of moats that Heart Capital may also see take hold.

Xie Yujuan, partner at Chuangdongfang Investment, emphasized that future ChatGPT application and product ecosystems must focus on industry regulation. Especially in certain industries prioritizing proprietary rights, how hardware ecosystems integrate with software will determine practical product and technology path choices. Which industries will produce ChatGPT unicorns requires analyzing where core resources reside in the value chain, and whether ChatGPT can help achieve breakthroughs.

At the application layer specifically, how to form cost-effective solutions based on application scenarios ultimately starts from the customer perspective — how products or solutions can close loops at the business level while delivering superior cost-performance. Overall, from a business model闭环 angle, who ultimately pays in the industry chain, plus domestic and foreign regulatory tendencies, will significantly influence startups' choices of underlying algorithm models and hardware.

Fangbo Tao, CEO of Mindverse, admitted he hadn't expected ChatGPT so soon. Even with his industry background and early AGI attention, he'd estimated 5–19 years as an optimist. Its arrival in one year brings pros and cons.

From the previous perspective, much past AI work may become meaningless and need reconstruction — software may be entirely rebuilt one day. So GPT should be viewed as a new "brain resource pool": ChatGPT is a cognitive resource, the CPU of the new era.

Today's AIGC differs from AGI; AGI could颠覆 entire industries. Tao even called it a bigger opportunity than mobile internet. China may not have it now, but he hopes it will — success need not come through him personally. The path: first build the large model layer well, so Chinese entrepreneurs can look upward and better deploy middleware layers above, placing large models into applications in targeted ways to generate better applications through orchestration.

Zhang Lei, founder and CEO of Stardust Data, pointed to an overlooked question: does China have AI infra companies? The answer is only foreign ones do. Domestic practitioners focus too much on methodology, which is公开; it's the non-public elements that contain more know-how and barrier potential. Additionally, 95% of datasets are generated in the English-speaking world; Chinese-language datasets are currently relatively scarce.

Thus Zhang, reflecting from his own position, considered what AI angle to pursue. Data direction, he felt, is overlooked yet sufficiently important, requiring certain algorithmic background: 90% of data work can be automated, but those who dialogue with algorithms need strong data strategy. As models iterate, data must iterate too — all worth attention and effort.

Deng Yulong, founder of Xiguang Technology, speculated that chat tools representing social aspects and video/CV may face major disruption. Why? WeChat scaled quickly not just through massive QQ user backing but through mobile chat, voice, and other key features — social disruption is inevitable, only the timeline varies.

For startups, whether OpenAI or Baidu builds the underlying framework matters less than platform usability. The focus should be on one's own space to excel, doing one's part well; at sufficient scale, one can leverage large models to build industry-specific models.

Wang Zhiwu, CEO of Yuanjing Technology (a subsidiary of Tianyu Shuke), noted his persistent focus on virtual digital humans, an early vertical adopter of ChatGPT capabilities. The initial thought was using small NLP models to animate virtual humans for tourism and cultural applications.

But application revealed virtual digital human latency of 6–8 seconds, plus直播时上下文对不上, "speaking nonsense with a straight face." The solution: using ChatGPT's model to train NLP, large models training small models — enabling continued深耕 in verticals like virtual customer service and live-streaming virtual anchors.

Yu Wei, founder of Ruiqi Technology, candidly noted that enabling enterprises with AI was extremely difficult during his Microsoft tenure, largely because technology wasn't mature enough — speech recognition and NLP provided to clients were insufficiently accurate, while business complexity ran high. Only through deep场景 immersion, product refinement, and technology iteration could true落地 be achieved.

Thus Yu focused on how good ChatGPT actually is. Ruiqi Technology's recent testing found that while ChatGPT as a general model can do much, gaps with professional industry models remain substantial. Only by going deep into industries and truly fusing AI capabilities with business can AI technology落地, solving key efficiency and productivity problems in B2B.

Perhaps, after all is said and done, it requires domestic practitioners joining forces to truly seize this AGI opportunity and manifest a more百花齐放 ecosystem.

The good news: in recent days, beyond Huiwen Wang and Bowen Zhou leaping back into entrepreneurship, heavyweights like Jie Tang and Wang Xiaochuan are on their way. Baidu's Ernie Bot is set for March release. China's AI, like the early spring weather, is showing new vitality.