AI applications: food for thought and investor feedback

If only we had an ETF tied to the growth of total value of AI usage…

Over the past few weeks, word broke out that investors should buy AI applications. After rounds and rounds of discussions with investors, it turns out that there is no ideal AI application target to buy for investors.

Technically speaking, there is no pure-play AI-native AI application company that is publicly listed today. But one has to admit that we see the application of AI across the sector at various companies. Below we lay out the few types of AI applications plays and why investors have had their fair amount of pushback:

1) The first type is big tech firms who do AI applications. It’s hard for AI applications to drive stock prices in the short term. This hardly works (at least for now) because these big tech firms also conduct a range of other business segments. Names that come to mind include the internet companies, for example.

2) The second type is software / SaaS companies. While most of these companies leverage AI to enhance their existing product offerings, as long as scaling law remains valid, the SaaS fear or overhang re how LLMs may continue to disrupt SaaS will persist. As a result, it would be hard for multiple to expand and investors may have to just earn their fair share of turn via EPS/sales growth. However, the mainstream investor base rarely has the patience to earn return based on EPS/sales growth. Further, there is always this “what if” that one day LLMs may iterate their way to actually reduce value pool currently captured by software companies. On balance, investors are still hesitant around this space. Similarly, some investors believe the recent valuation multiple recovery with US SaaS names is not sustainable for this reason.

3) The third type is most of these recently-listed names who serve AI to mostly non-tech enterprise customers. They help enterprises on-board AI via front-end deployment, agent product set-up and more. They are not nearly labor-intensive as what some investors may fear. However, broadly speaking, based on the investor feedback we receive, these companies’ loss-making status, unclear differentiation and unproven business model tend to deter investors.

4) The fourth type, although not too common in investor perception, is the LLMs themselves. Admittedly, some of the best AI applications are made by large language model companies themselves, whether it’s coding agentic products or multi-modal generation tools. These are actually more AI-native. However, as the foundational layer seems to see more competition, investors seem to have lost confidence in betting on any single horse here.

5) The last type lies in the private space. Some of these names have their own debates while some others are simply relatively unknown to the public market investors. We will leave this to a separate report topic later on.

As can be seen, none of the above is ideal today from investors’ perspective.

Public equity investors have been forced to do “private equity” lately

Public equity investors are not used to investing at so early stages of companies.

If only one could buy the whole downstream to benefit from the continual growth in end demand (which we believe is still healthy to say the least), some investors comment. Some food for thought: if investors wanted to be invest in the rapid rise of manufactured automobiles from 1920s, would they have done well by having exposure to more than 100 auto OEMs at the time till the sector consolidated into the Big Three (Ford, GM and Chrysler) in 1950s? Perhaps this is like VC investors investing in every startup that does the same thing because they couldn’t tell which one will survive till the end. Unfortunately, public investors are not used to this mindset – this is another observation we have these days. It also explains how some public valuations have trended lower than private ones recently.

AI is just a piece of new technology; it is not a new industry by itself

It’s key to let the header sink in.

Frankly, even “internet” sector was never a sector. Internet changed every industry and changed everybody’s life. Similarly, AI will do the same. Therefore, one has to ask, is there really such thing as pure-play AI-native application? The application of AI may turn out to be incumbent players across sectors getting stronger via AI. Or it may be better tools to achieve the same outcome via AI vs the old ways.

LLM may not have to be separate from the application layer

We don’t know to what extent LLM layer has to be separate from the application itself. We observe that top AI application companies are turning into neo-labs by training their own models and building custom inference stacks instead of renting intelligence from traditional frontier labs. The idea is that post-train will be fine-tuned to suit a specific use case and above the foundational layer, and every AI-native application company will do its own model (think Cursor making Composer). On the same note, Harvey, a legal AI company, recently introduced Tenet, its first proprietary, in-house AI model built on top of Kimi K3 base model.

We foresee at one point there will be a clear difference between foundational layer model and vertical-specific models. To that extent, we don’t expect foundational layer models to do everything (more later on in a separate topic).

The debate on how much further AI will go

We observe an interesting debate lately among investors and beyond.

Some believe that while models will continue to scale, most of white-collar use cases can be covered (cheaply) by current models already. Therefore, the next big hit model won’t help AI penetration accelerate.

By contrast, some others believe that current AI models can do a lot already but there is ample room to iterate and improve. For example, long-range tasks still have a long way to go and AI still needs time to execute more comprehensive problems unsupervised.

In terms of what may lie ahead to accelerate the demand, the nearest secrete sauce may be harness to accelerate white-collar (as model prices go down even faster), then edge computing, and then physical AI (robotics etc.).

Risks

Risks related to the slower progress of AI penetration and liquidity downcycle to drain these high flyers of capital.

本报告分析师：

王宪策 Nathan    

SFC HK执业证书编号：BKS605

王晴

SFC HK执业证书编号：BUL117

宋琦

SFC HK执业证书编号：BTF521

免责声明

本公众号涉及的内容仅供参考，并不构成对所述证券的推荐、邀约或邀约邀请。同时亦不构成任何合约的成立，不构成买卖任何证券或交易的建议， 亦不构成投资建议、财务、法律、税务或其他意见。本公众号对于因此信息和内容或者据此进行投资而直接或间接引起的任何损失不承担任何责任亦不作出任何保证。

本公众号中所发布的信息、观点以及数据有可能因发布后的市场走势或其他不可预见之因素而影响其有效性。文中的意见预测和估计反映了我们截至制作本文之日期的评估，本公众号并不承担更新信息、观点以及数据的责任，而同时在对相关信息进行任何更新时亦不作另行通知。

本公众号及其关联机构可能会对于上述信息、观点的标的公司所发行的证券进行交易，还可能为这些公司提供其他服务。本公众号及其关联机构可能购买、持有或出售相同或相关产品。

版权及相关权益归海通国际及相关内容提供方所有。未经海通国际书面许可，任何人不得以任何方式或方法修改、复制、分发或转载官方微信中之任何内容。

《研究报告》免责声明

本公众号不是海通国际研究报告的发布平台。本公众号涉及的研究报告信息由海通国际证券有限公司于香港特别行政区发布，内容从于发布之时从认为可靠的渠道获得的信息或基于海通国际已发布的研究报告编制而成。研究报告信息中的观点及评级的发布取决于报告作者所知悉的各种市场因素及公司内在因素，以及海通国际对其之盈利预测。目标价格是基于一系列的假设和前设条件，投资者需在准确了解相关研究报告中的全部信息的基础下，方可对研究报告所表达的观点形成比较全面的理解。本公众号仅为对相关研究报告部分内容的引用或者复述。受法规技术或其它客观条件所限，无法同时提供观点、评级的依据或盈利预测及目标价格背后的假设等相关信息。因此其内容可能无法完整或准确表达相关研究报告的观点或意见。海通国际并不保证本公众号涉及的研究报告所包含的信息、观点以及数据的准确性及完整性。如欲了解完整观点，应参见海通国际研究网站(www.equities.htisec.com)所载完整报告，或与海通国际销售人员联系。

投资涉及风险，海通国际并没有考虑投资者的具体投资目标，财务状况等的特定需求，任何人不得依赖研究报告包含的信息、观点以及数据的准确性及完整性作出投资决定。