AI Insights: Where are the opportunities for the next generation of AI startups? What are the pricing trends?
Today, in which aspects of the AI field are there good entrepreneurial opportunities? What will the AI products popular among users look like in the future? Recently, Bret Taylor, the Chairman of the OpenAI Board of Directors, and Kevin Weil, the Chief Product Officer of OpenAI, shared their latest views on the opportunity markets for AI entrepreneurship and the development direction of next-generation AI products during interviews. Another foreign technology writer, Kyle Poyar, discovered valuable AI pricing trends through the data of over 240 software companies. Let's take a look at their insights.
Three Opportunity Markets for AI Entrepreneurship
At the end of July, Bret Taylor, the Chairman of the OpenAI Board, was interviewed by the well - known podcast Lenny’s Podcast and talked about his latest views on AI.
He believes that there are three parts of artificial intelligence that will ultimately become quite meaningful markets. The first meaningful market is the frontier models or foundation models. In the end, only a few ultra - large - scale enterprises and truly large laboratories will remain in this area, just like in the cloud infrastructure and services market, because creating frontier models consumes a large amount of capital and requires well - funded large companies to support. Currently, all the startups that have attempted to do this have either merged or been acquired. There is almost no viable business model for startups in this area, and as an asset class, the value of these models will depreciate quite rapidly. Only by scaling up can investment returns be obtained.
The second meaningful market is tools, which here refers to data platforms. In the infrastructure services and cloud tools market, many large companies like Amazon and Azure have competing products in these fields. There are some truly meaningful companies, but many other companies have been replaced by the technologies of the infrastructure providers themselves. Many people may need your tools, but the problem is that if one of these large infrastructure providers introduces a competitor, why would people continue to choose you? It is a good market, but it is getting close to the end.
The third is the AI application market. There are companies like Sierra that help other companies use Agents to answer calls or chats to improve the quality of customer service; there are companies like Harvey that provide agency services for the legal, paralegal industries, antitrust reviews, contract reviews, etc. This is similar to the software and services market. They may be companies with higher profit margins because they sell products that can bring business results, rather than by - products of the modeling itself. They will definitely pay "taxes" to the model providers, which is why these model providers will ultimately grow to a very large scale, but their profits may be slightly lower. As time goes by, products are more important than technology.
At the same time, Bret Taylor also shared his unique insights on Agents. He said that he is very optimistic about the productivity improvement brought by artificial intelligence, but the current tools and products are somewhat immature and quite counter - intuitive. Therefore, the concept of self - reflection of Agents is very important. Letting AI supervise AI is actually very effective.
Imagine if you create an Agent that is correct 90% of the time, it is not that easy to achieve. However, it may be an easier problem to solve if you let another Agent detect errors in the remaining 10% of the time.
Four Signals of Next - Generation AI Products
On August 22nd, Kevin Weil, the Chief Product Officer of OpenAI, shared four key signals of next - generation AI products during an interview with "Moonshots".
Signal 1: Breakthrough in Reasoning, from "Knowing" to "Thinking"
Kevin Weil pointed out a misunderstanding at the beginning: People always think that AI is smart because it knows a lot. But the real change is that it starts to think.
It no longer gives you a ready - made answer, but can string together multiple steps and complete a full reasoning process on its own. For example, in the past, if you asked it "How many people can climb this mountain?", it would dig out a number from some source and give it to you; now, it will ask itself three questions: How high is this mountain? Is there a path? What's the weather like? Then it will give you an inference.
This may seem like a small detail, but in fact, it is one of the biggest changes in GPT - 5: Shifting from retrieving answers to constructing thinking processes. This ability is not achieved by piling up more data, but by the model itself evolving the ability of "Chain - of - Thought".
Kevin Weil emphasized that this kind of reasoning not only makes the model think more like a human, but also enables it to handle more complex tasks - that's why it can handle complex work such as financial analysis, scientific papers, and code logic.
Signal 2: Interface Re - engineering, Active Service Becomes the Standard
In most people's impressions, ChatGPT is a tool that waits for you to ask questions and then gives answers. However, Kevin Weil said that OpenAI's goal is to turn ChatGPT into a real intelligent partner beside you, rather than a tool waiting for your summons.
To make AI provide active services, it must first understand your habits and preferences. Kevin Weil mentioned several key abilities that are becoming the standard for next - generation AI products:
- Memory: The model can remember your name, preferences, and the things you didn't finish last time;
- Vision: It can generate pictures and videos, and create scenes and design content in real - time;
- Voice: It can have continuous conversations with you like a friend, rather than having an awkward question - and - answer session.
What lies behind this is not just pure technical strength, but experience design: It can understand the way you speak, perceive your work habits, and fill in the parts you didn't say.
In Kevin Weil's product plan, voice is a key turning point. He revealed that the team has tested a large number of details in voice conversations: intonation, rhythm, emotion, and even appropriate pauses and interruptions.
This is a design of "sense of co - existence". It allows you not to deliberately organize your language, can actively supplement when you hesitate, and can judge whether to continue speaking or switch topics based on your tone of voice.
This experience is changing the product form. In the past, AI applications were like independent tools, where you had to click in, select options, and input text. Now, when you open an application or an operation interface, AI is present by default: When you open your email, AI has already sorted out the emails you need to reply to today; when you switch to the meeting software, AI has already pulled up the topics you discussed yesterday but didn't follow up on; when you open a PPT, AI is helping you rewrite a slide with unclear logic.
This state doesn't mean that AI is stronger, but that it understands human nature better. Only when AI is everywhere instead of requiring a click to enter will humans really start using it.
Signal 3: Product Closed - Loop, Task Completion Rate Determines Value
During the interview, Kevin Weil also revealed an important signal: "Most users have started to not care how smart AI is, but only care whether things can really be done."
This reflects the maturity of the market: AI is changing from a "technology showcase" to a "productivity tool". The measurement standard is no longer the ability boundary, but the delivery quality.
What does it mean to "get things done"? For example, if you want to send an email for data analysis. AI doesn't just help you write the copy; it has to be able to mobilize the data in the CRM system on its own, apply your preferred email template, send it out, and keep records for your future tracking.
Behind this, it's not a single model working alone. It needs to call many external tools: documents, databases, API interfaces, email... A real intelligent agent doesn't do everything by itself, but can mobilize the capabilities in the environment and combine them to solve problems.
AI starts to create real business value, rather than just providing intellectual support. Whether it can "finish the job" depends on three things:
- Whether the tool call is stable (without interruption);
- Whether the process can be traced (able to check in case of errors);
- Whether the result can be written back to the system (the data can be read by other applications).
Kevin Weil pointed out an industry misunderstanding: "Many people think that connecting a model to a web page is an AI product. This is far from enough."
Signal 4: Global Implementation, Inclusive Ability Becomes the Dividing Line
Kevin Weil repeatedly emphasized in the interview: "What we care most about is whether the model can be used by people as soon as possible." - This is also OpenAI's global deployment strategy: Instead of waiting for the technology to be more perfect, it aims to popularize the current capabilities to more people as soon as possible.
He gave an interesting comparison of numbers: Today, there are about 30 million programmers in the world; but in the future, AI may enable 3 billion people to have certain programming abilities. Because the definition of "programming" has been rewritten. It is no longer about writing lines of complex code, but telling AI what you want in natural language.
In the next few years, the capabilities of top - tier large models may become increasingly similar. But the real difference will be reflected in other aspects:
- Who can achieve global deployment earlier?
- Whose pricing strategy is more flexible?
- Who can provide truly implementable products for different countries and different groups?
Behind this is a whole set of strategies including infrastructure, customer support, version adaptation, and language localization. Because the standard of global competition has changed: It's not about technology demonstration, but about practicality and stability.
Five Trends in AI Pricing
While AI is profoundly transforming the software industry, it also brings a problem: Although AI functions are powerful, due to value misalignment and cost pressure, traditional pricing methods are becoming ineffective, especially for AI - native products.
Recently, foreign technology writer Kyle Poyar collected data from over 240 software companies. Through these data, he concluded five trends in AI pricing:
1. Traditional Pricing Methods Face Challenges, and Hybrid Pricing Models Become the Mainstream
In the past, software pricing mainly had two models: seat - based charging and fixed - rate subscriptions. These models provided price predictability and were expected to bring sustainable recurring revenue (ARR).
However, due to value misalignment and cost pressure, these models are being replaced by hybrid pricing, which is a combination of subscription and usage.
The main reasons why hybrid pricing is so popular are as follows:
- It doesn't have a major impact on the original pricing system and can be integrated into the existing seat - based and subscription - based models;
- It creates a natural upselling path, allowing customers to "freely" try new products and then make profits as usage increases;
- It has considerable profit margins. By limiting usage, companies can control costs and minimize the risk of unprofitable customers;
- It is relatively predictable. By following traditional pricing models, buyers can estimate costs and control expenditures.
2. There Are Many Combinations of Hybrid Pricing, Each with Its Own Advantages and Disadvantages
As more and more AI products shift to hybrid pricing models, a new challenge emerges: There seem to be countless ways to construct hybrid pricing, but not every one is suitable.
Here, the author shared some common pricing methods and their advantages and disadvantages:
First, pay - as - you - go. Pay - as - you - go means no commitment and complete flexibility. This model is most effective when customers can reimburse expenses or include them in the operating budget. Otherwise, corporate procurement should be cautious!
Second, pay - as - you - go with a cap. This model reassures buyers by limiting potential usage/expenditure and is becoming more common in result - based models because the results are unknown in advance.
Third, usage - based packages. Customers commit to a certain amount of usage or a package, and it is usually "use - up - and - stop". The packages include various sub - models, including overage charging or declining models.
Fourth, platform fee plus usage. Charging a platform fee helps lock in customers while allowing them to enjoy premium features, high - quality support, etc. This method is very effective when pricing metrics are commoditized (e.g., SMS, computing, storage) or cannot reflect the full value of the product.
Fifth, platform fee (including usage) plus additional usage fees. This model is also called the three - part tariff model. It has a relatively high subscription fee but includes a certain amount of "free" usage fees. Providing a minimum usage amount helps attract customers and usually stimulates them to increase overall consumption.
Sixth, adaptive fixed rate. In this model, customers commit to a usage - based tier, but they can use the product freely during the contract period without incurring overage fees or needing to upgrade. When the contract is renewed, the tier will be adjusted up or down according to the actual usage. The disadvantage is that if the usage decreases, you still have to bear the corresponding costs.
Seventh, platform fee plus success bonus. In this model, pricing is presented in a more traditional subscription - fee form. If customers achieve a higher return on investment (ROI) than expected, they need to pay an additional bonus or commission.
3. Pricing Based on Results Is Not Applicable in Most Markets in the Short Term
When AI agents are positioned as "task - executors", it seems reasonable to price them according to the amount of work they complete (or the revenue related to that work).
It sends a strong signal that you have full confidence in your product and are willing to fully endorse it. At the same time, it also encourages suppliers to continuously invest resources to improve product effectiveness, thereby bringing more practical results to customers.
However, there are also many problems that cannot be ignored behind this model:
- Consistency: Different customers require different results, which leads to the need for customized results and thus a large number of customized contracts.
- Attribution: Can you convince customers to attribute the achieved results to your product? If customers cannot clearly see the role of your product in driving the results, they are unlikely to be willing to pay for it.
- Measurability: If the results cannot be accurately and timely measured, it is difficult to establish a transparent billing mechanism and a trust relationship.
- Predictability: Can you predict the results that your product will bring with a certain degree of accuracy? If the results fluctuate greatly and are unpredictable, enterprises will face huge financial risks.
4. The Value of Price Transparency May Be Overestimated
Public pricing allows you to meet the needs of some buyers (and the related search traffic) and control the narrative (i.e., actively define customers' understanding of the product value). In addition, it can screen out unqualified buyers and prevent them from wasting your team's time.
However, the reality is that although transparent pricing has its advantages, many enterprises have not fully adopted this approach.
This may involve complex pricing structures, differentiated quotation strategies, or concerns that price will become the focus of competition and weaken the value proposition. In addition, many software companies, especially startups and AI - related enterprises, have not fully sorted out their pricing strategies. Once the price is made public, it will be much more difficult to make subsequent adjustments.
Therefore, although the trend of transparency seems "inevitable", there are still many challenges and concerns in actual implementation.
5. AI Pricing Is Changing Rapidly, and Most Companies Are Not Prepared
As pricing decisions are increasingly becoming a strategic and complex task, enterprises need to invest corresponding resources in pricing work. There is a large amount of practical work to be done behind this, including in - depth understanding of cost structures, competitor dynamics, and customer - perceived value.
However, most enterprises still have deficiencies in two aspects:
- Personnel ability gap: There is a lack of talents with professional pricing analysis, value modeling, and market insight;
- Backward tools: They still rely on traditional Excel spreadsheets or outdated systems and cannot support real - time data - driven pricing decisions.
In other words, although the importance of pricing is increasing, many companies have not established a matching ability system to support this strategic transformation.
Special attention should be paid to avoiding falling into the so - called pricing "no - man's land" - the "gut - feeling" decision - making method in the startup period is no longer applicable, but a formal pricing mechanism and responsible person have not been established yet, resulting in a lack of clear ownership and strategic direction for pricing strategies.