After 200+ Days of AIGC Fervor, Investors Reach Three Broad Consensuses

By Elune | Edited by Jiumu | Curated by Eason

On July 19, Meta, Facebook's parent company, released Llama 2, a large language model free for both research and commercial use — billed as the strongest open-source alternative to GPT-4. This would change the previous situation where numerous global large models were built on Llama but constrained by its non-commercial licensing.

As the AI market landscape shifted, venture capital attention locked in once again. While people debated whether the singularity of human-level AI was near and whether the AIGC era had fully arrived, something slowly changed as AIGC charged ahead.

First, on the question of how long AIGC fever could sustain itself, the investment community gradually splintered. Some argued that investing pursues returns and certainty, while the monetization capability of general-purpose large models remained unclear. The market was cooling; caution was warranted.

Others held the opposite view, believing AIGC development had only just begun and would heat up further next year. Today's AIGC still operated mainly in text; multimodal large models hadn't yet emerged. By year-end, OpenAI's breakthroughs in image generation might further unleash imagination.

The divergence wasn't just in attitude — it also showed in cold data that belied the market's high attention. According to relevant data, ChatGPT's visit growth rate dropped from 131.6% to 2.8% between January and May this year. And in terms of actual action, the scarcity of investment deals stood in stark contrast to the enthusiasm flooding investors' social media feeds.

New things always emerge alongside "polarized" attitudes — this seems to have become a natural law. In the 200-plus days of AIGC's continuous fermentation, what consensus have investors reached? And where lie entrepreneurs' opportunities?

First New Voice interviewed multiple investors, attempting to clarify from the present moment what has settled during AIGC's rapid advance and what shifts have occurred in guidance, hoping thereby to contribute positively to industry development.

01

The AIGC Wave Opens Opportunities for the Next Era

AIGC's explosive popularity set the investment world aboil as well.

According to estimates by QbitAI's think tank, the AIGC market is expected to exceed one trillion RMB by 2030.

Publicly available data shows that in 2022, China's AIGC industry saw over 500 investment events with total investment exceeding 90 billion yuan. According to incomplete statistics from Tianyancha and First New Voice, from January to June 2023 (as of June 27), domestic AIGC industry financing totaled 4.959 billion yuan across 46 rounds.

Rewind to late 2022. Fang Zhenghao, managing partner at Xiaomiao Langcheng, noticed AIGC attracting niche attention in tech and investment circles. By March 2023, ChatGPT broke into mainstream consciousness. "My social media feed was practically flooded with related information daily — people's excitement and underlying anxiety about AI reached unprecedented peaks," Fang recalled.

As synthetic video and image generation achieved breakthroughs in consumer applications, "the feeling that a new paradigm of productivity transformation had arrived" grew stronger in Fang's mind. He believes AI will play increasingly important roles across vertical domains and industry applications.

Bai Zeren, investment vice president at Linear Capital, shared similar sentiments. "AIGC represents a very long-term opportunity. Analogous to the internet, the future trend will certainly be AI penetrating every scenario like capillaries." Optimistic about this AIGC wave, he believes substantial investment opportunities will follow.

"We very much look forward to seeing more innovation and transformation," said Wang Xiao, founder of Unity Ventures. Behind the rapid adoption of applications like ChatGPT in this new AI wave lies the emergence of a new generation of AI capabilities represented by intelligent emergence.

"From now on, whether working or founding a company, make sure you're involved with AI." Lu Qi, former Microsoft corporate vice president, Baidu COO, and founder of MiraclePlus, took a firmer stance: "AIGC is no passing fad. 'Fad' implies opportunism — that gravely underestimates AI's impact on world development."

The opportunities of the next decade, even the next era, have already begun unfolding.

02

Watching Large Models Without Investing, Pouring Real Money into Vertical Models and Application Layer

In the 200-plus days of AIGC's fermentation, investors reached some consensus, mainly in three areas:

Consensus 1: Compute Infrastructure Holds Certain Opportunities; Large Models Are a Rich Person's "Game"

Within the AI new-wave ecosystem architecture comprising compute infrastructure layer, model layer (foundation models, open-source models, self-built vertical large models), and application layer, certain deterministic opportunities have emerged.

First, AI development has triggered explosive expansion in compute demand. That the compute infrastructure layer holds deterministic opportunities has become consensus across Chinese and American capital markets.

Secondary market performance corroborates this view. From late October 2022 to July 17, NVIDIA's stock price surged from $123 to $464 per share. From early 2023 to date, domestic AI companies at the compute infrastructure layer such as Cambricon and Sugon have maintained resilient stock prices.

Second, large models will bring enormous transformation to R&D and application paradigms. Some investors believe the prospect of AI technology dominated by large models genuinely reducing costs and improving efficiency has ignited entrepreneurs' excitement. This is why the "hundred-model war" could occur.

Baidu launched "Wenxin Yiyan," Alibaba released Qwen, iFlytek introduced its Spark large model, and Meituan, Baichuan, Unisound and others joined the large model race. By some counts, as of July, China had over 80 large models with parameter scales exceeding 1 billion.

Today, the hundred-model war has evolved from intensifying to gradually stabilizing. Consensus has formed at the model layer: this is a game for big "players" — referring to both entrepreneurs and investors.

Large model inference and training impose direct demands on chip computing power and GPUs, and the model layer requires very strong technical team support, making capital investment enormous.

Take OpenAI as an example. According to relevant statistics, a single GPT-4 training run costs approximately $63 million, requiring 1.8 trillion massive parameters. This doesn't even include data collection, RLHF, and other costs.

At root, top-tier technical talent and chips both require burning money. "At this stage, the core factor is fundraising capability — whoever has stronger capital strength has greater probability of success," Fang noted. Domestic startups face compute bottlenecks, and compared to overseas, large models still lag. On top of compute and high-level technical talent, the competition is over whose R&D investment proves more effective and who performs better technically.

For institutional investors, the same applies.

"Investment opportunities at the model layer currently can only continue as a game among players with substantial capital," Fang observed. For investment institutions without particularly large management scale, if they didn't position early before the industry heated up, the current timing is no longer suitable for participating in large model investment.

Last year would have been a relatively good timing for large model positioning; this year is no longer a good window for most early-to-mid-stage investment institutions.

Beyond enormous costs, multiple factors drive investors' cautious approach: optimal timing windows, commercial monetization capability, and more.

"If it were me, I wouldn't choose to invest in large model-related projects this year," said Shi Mao, founding managing partner of Changlei Capital. Having missed the optimal timing window for investing in underlying large model tracks, he observed a considerable gap currently between the model layer and application layer, and large model technology's monetization capability remains unclear.

Notably, vertical models with clear demand and landing scenarios have attracted capital attention.

Early this year, Xiaomiao Langcheng reached internal consensus: no investment in large models, but maintain attention on industry-specific medium-to-large models at the hundred-billion level. "Compared to large model companies already visible above water, startups will have more opportunities in segmented vertical domains. Because after landing in specific industries, startups can more easily accumulate relatively high-quality datasets."

This aligns with the view of Allen Zhu, managing partner at GSR Ventures. GSR Ventures is among the domestic early-stage institutions most active in vertical AIGC, and Zhu has publicly stated that for most entrepreneurs, they should be "scenario-first, data-king" — train their own vertical models rather than blindly worshipping general-purpose large models.

Consensus 2: At the Application Layer, "Old Forces" in Certain Vertical Domains Have Relatively Good Opportunities

One fact stands out: investors are largely watching large models without investing, pouring more real money into the application layer.

"We also closely follow progress and changes in large models themselves. Considering current market competition dynamics and capital barriers, when deploying capital we tend toward application layer opportunities and new infra opportunities," Bai Zeren expressed. Linear Capital cares more about how new technology lands in industries to more effectively solve industrial problems and create enormous commercial value for industries — this is Linear Capital's consistently unchanging investment logic.

"We encourage all portfolio companies to consider whether their business could potentially integrate with AIGC in the future. At minimum, from a company management perspective, they should definitely consider how to use AI to improve internal human efficiency," Bai Zeren added.

As for what specific deterministic opportunities exist at the application layer, investment circles haven't yet reached consensus. However, multiple investors indicated that from the enterprise perspective, established players in To B vertical domains hold distinct advantages.

Fang Zhenghao compared enterprises entering AI in sequence as "old forces" versus "new forces." "Old forces" refer to the first wave of AI companies born around 2016's deep neural network breakthrough, including the AI Four Dragons and certain startups. AI companies emerging in recent years constitute the new forces.

"'Old forces' in certain vertical domains have mastered customer needs and scenarios, while also being able to embrace AI technology iteration first," in Fang's view. Such companies represent relatively certain development opportunities at the application layer.

Other investors expressed similar views: beyond new startups, there exists a batch of existing AI companies in various vertical domains at the application layer, having developed for six to seven years to date, that may become protagonists among application-side AI companies in the coming period. "Because they hold customers and scenarios, they'll have greater competitive advantage."

Wang Xiao, founder of Unity Ventures, stated that AI will transform all industries, including SaaS, tool software, and previous-generation AI companies, all of which can leverage this technology iteration for structural upgrading. "Xiaoduo Technology, which we invested in back in 2015, recently launched XPT, an e-commerce vertical model based on large language model technology. It will leverage large models and accumulated industry data to empower more e-commerce business scenarios with better solutions."

Consensus 3: Investors Focus on Teams and Commercialization Landing

In this AIGC wave, investors mainly bet on people and direction; invisible backgrounds have become important considerations.

"What future does the founder see, and what role does he hope to play in that future? We look for resonance points between the future the entrepreneur depicts and what we ourselves see. Then from the entrepreneur's growth background and technical level, we judge whether he can truly achieve this." This is the core logic Jinjian Zhang, founding partner at Oasis Capital, employs in his investment process.

Beyond betting on people, in specific evaluation factors, various institutions tend to emphasize projects' commercialization capability.

From Linear Capital's perspective, commercialization capability mainly manifests in three aspects: barriers to entry — such as sufficient technical differentiation, good adaptation to specific scenarios, or scenarios themselves requiring domain knowledge; rapid productization — ability to quickly integrate LLM capabilities into products on pain-point issues; and forming effective data and feedback loops.

Unity Ventures, which has backed star projects like Airdoc and Tungee, shares similar views with Linear Capital. Wang stated that beyond founder considerations, investments should also evaluate whether there are clear landing scenarios and needs that can truly reduce costs and improve efficiency for customers.

Xiaomiao Langcheng's investment strategy focuses on selectivity and concentrated bets: actively deploying on the 5% of unmissable projects, while approaching the 95% from a hunter's perspective, patiently conducting research. "The core of investment is expressing understanding of something with heart. If you understand, you should invest long-term and continuously."

03

Different Concerns: Technology Development Speed, Valuation, Business Model

The previous two waves of AI occurred in 2012 and 2016 respectively.

In 2012, the AlexNet developed by Professor Geoffrey Hinton and two students, based on deep learning, won the ImageNet Large Scale Visual Recognition Challenge. From then on, deep learning established AI's underlying technical foundation.

In 2016, AlphaGo defeated world Go champion Lee Sedol in a human-machine match, rapidly igniting global AI venture capital fervor.

One unavoidable fact: in the previous two AI booms, 90% of startups were unprofitable. And only early-entry investors made money.

One CVC investor noted that before generative AI emerged, investment enthusiasm for AI was extremely low, because these companies' commercial performance fell far short of investor confidence. "Massive customization, data cleaning and preparation work, extensive model parameter tuning — every business scenario was practically a non-standardized project. Combined with unreasonable industry talent cost structures, this resulted in less than 1% of companies being profitable across the previous two AI waves."

The current third AI wave has been described by Kai-Fu Lee as progress "from islands to continents." Compared to the previous two, this wave makes generalization possible, building a new world through cross-domain capabilities. Once powerful models are supported by sufficient data, in suitable scenarios, AI will create productivity surpassing human capability.

Of course, concerns exist in this process.

Linear Capital's concerns for enterprises include: insufficient team execution capability, unable to flexibly trial-and-error in rapidly changing technical and business environments; overly shallow scenario entry, unable to form effective closed loops, leading to red-ocean competition in the future.

Xiaomiao Langcheng has two concerns: First, the rapid development of open-source models and algorithms lowers technical access barriers, potentially rendering leading AI companies' past technical investments ineffective. Under intense homogenized competition, the business models investors expect may ultimately fail to materialize.

Second, although current AI has some generalization capability, achieving high accuracy still requires learning and parameter tuning for each scenario. Customers have substantial customization service demands. That is, even if main modules are already generalized at the model level, numerous functional plugins still require customization, ultimately forcing startups to provide customized services and fall into the predicament of difficult scalability.

In this wave, how should entrepreneurs seize opportunities? Multiple investors offered advice.

"If there were a Creator of the world, He has already fired the starting gun," Jinjian Zhang, founding partner at Oasis Capital, believes. In major waves, entrepreneurs should actively embrace them — not mapping out plans, but sprinting when the starting gun fires, entering industries as quickly as possible.

"Before the Industrial Revolution, people had no surplus productivity, hence no commodities, hence no commodity circulation. After the Industrial Revolution, commodity circulation emerged, and only then did the transportation and retail industries develop. Now entering the AI era, in theory, all Fortune 500 companies could be rebuilt," Zhang expressed.

Fang Zhenghao advised entrepreneurs to utilize capital market windows to complete fundraising, while not burning cash frantically after securing funding — definitely plan before acting. "Because the timing for major explosion in AI application endpoints has not yet arrived, entrepreneurs need to discern which among these represent genuine opportunities for startups, then refine their products and business, to leverage this wave to run further."

In investors' view, all industries have development cycles. True entrepreneurs are those who can withstand temptation and testing through industry cycle fluctuations, never giving up, possessing powerful resilience.