# AI’s $600B Question Addressed : Recovering $600bn invested into Gen AI using Ads

David Cahn of Sequoia Capital, in his famous blog titled [*AI’s $600B Question*](https://www.sequoiacap.com/article/ais-600b-question/), highlights that AI's explosive infrastructure growth has far outpaced revenue generation. The $600 billion question is: how will companies justify the massive CapEx and recover these costs? While many are stockpiling GPUs to avoid being left behind, I believe the answer lies in LLM ads—monetizing free users much like how Google turned its free search engine into a goldmine in the 2000s through advertising!

> AI agents will most likely start getting ads in 1.5 years time. However those ads will be so well integrated that they will be impossible to tell apart from real content. The ad revenue from these extremely high CTR ads will alone recover the $600bn invested into Gen AI models in 6 years time.

Recently I was talking with someone from [People+AI](https://peopleplus.ai/) on whether people should still start SEO companies. The discussion was around the adoption of Meta AI chatbot being the next step from ChatGPT and when "answering engines/AI assistants" will get ads.

This got me thinking, when will the ads come, what will they look like and how much money will they make? Will it be enough to cover the R&D costs that have been placed into AI models over the last decade.

## Why ads work?
Platform Ads in general work because the medium: Banners, Newspaper, TV, and Social Media have built some fundamental consumer trust. Now, a brand can quickly transfer some of that to itself. This greatly reduces the cost of customer acquisition. The CTR of the ad follows platform reputation or atleast perceieved platform reputation. And the CPM scales proportionally.
\
For example: Google Ads rely on the reputation of the first blue link, Facebook Ads rely on your behavior to click buttons on photos and videos, and so on.
\
The fundamental question to ask is why will people trust AI assistants. The answer is simple: **Reliable Assistants**. The more reliable the LLM is at responsing well to vague queries the more scope for ads in AI assistants.

## AI Assistants

From a long time, fiction people and techbros have imagined AI to present as a JARVIS style reliable assistant. And all AI advances have always marketted themsleves in a similar manner all the way from [Cortona and Alexa](https://www.youtube.com/watch?v=KxwjnuhNVIY) to [Google Duplex AI](https://www.youtube.com/watch?v=D5VN56jQMWM) in 2018.

People love and trust this type of interface. See below, LLM adoption between GPT3 launch on 28 May 2020 and ChatGPT interface launch ~2.5 years later.


The core ability of LLMs which corresponds to this reliable assistant like behavior is known as "tool calling" (idea being everything from your email to uber app is a tool that you use as a human). This is so important in the community that every major LLM provider offers "tool calling" natively for their models now.
\
And most importantly, the performance of the state of the art models is improving consistenly on all types of function calling.

<iframe height="600" style="width: 100%;" scrolling="no" title="SOTAvsTime" src="https://codepen.io/PenGPT/embed/qBzGbmP?default-tab=result" frameborder="no" loading="lazy" allowtransparency="true" allowfullscreen="true">
See the Pen <a href="https://codepen.io/PenGPT/pen/qBzGbmP">
SOTAvsTime</a> by Priyesh (<a href="https://codepen.io/PenGPT">@PenGPT</a>)
on <a href="https://codepen.io">CodePen</a>.
</iframe>

## Nature of Ads
The ads in AI assistant responses are most likely to be choice nudges along with convincing justifications of the assistant's preferences.


**Example Task:** Make dinner reservations for our sales lead with our 3 highest paying customers in the next week.

**Answers:**

|Persona |Personalized Answer | Advertisers Plugged In | Data Vendors Plugged In |
|---------|-------------------------|-----------------------------|---|
|Persona 1| I would recommend taking them to Chianti, Indiranagar. It has an intimate yet professional setting, with a menu filled with authentic Italian dishes like lasagna, risotto, and gourmet pizzas. **Shall I reserve a table for 5 at 6:30pm on Friday when everyone is free?** | Swiggy Dineout| Google Calendar |
|Persona 2|I would recommend meeting them at the Polo Club, Oberoi in Chruch Street as it is closer to their place and offering great discounts for corporate bookings. **Shall I make reservations for 6:30pm, book you a cab for 5pm and block everyone's calendar?** | Zomato Dine-in, Uber | ZoomInfo(Salesforce), Slack DMs |

## Presence of Ads - The Math

Just because something is technically feasible, does not mean its economically viable and sustainable for business. The main driver for that will be cost and ROI.

The most important thing for the long term sustenance of ads on any platform is how is ad revenue closely related to cost of serving the ads. And how strongly is the additional cost per transacting user for advertisers related to the conversion rate.
\
$$\Sigma_{adv=1}^{n}(\Delta \\ cost\_per\_transacting\_user) \\ \ge \\ 2*Inferncing\_Cost$$
$$\frac{\Delta cost\_per\_transacting\_user}{cost\_per\_transacting\_user} \lt \frac{\Delta Conversion}{Conversion}$$

Since most people will only interact with the AI assistants via phones, the only thing we need to look at is if the additional cost that can be charged to advertisers is justified for the inferencing cost of function calling with LLMs.
\
**Here is the bubble plot of cost vs performance for LLMs as it stands today:**

<iframe height="900" style="width: 100%;" scrolling="no" title="Function Calling Efficiency" src="https://codepen.io/PenGPT/embed/XWLwmrO?default-tab=result" frameborder="no" loading="lazy" allowtransparency="true" allowfullscreen="true">
See the Pen <a href="https://codepen.io/PenGPT/pen/XWLwmrO">
Function Calling Efficiency</a> by Priyesh (<a href="https://codepen.io/PenGPT">@PenGPT</a>)
on <a href="https://codepen.io">CodePen</a>.
</iframe>

So lets take a few models that could actually be useful and try to figure out their feasibility in production:

|Model Name| Avg. # Retries for reliability| Cost| Cost Per Transacting User| Advertiser ROI Needed| Product Cost| Feasability|
|-|-|-|-|-|-|-|
|Claude 3.5 Sonnet| 4| $2| >$5| >$15| $75+| Medium|
|GPT4o| 5| $2.5| >$7| >$21| $100+| Low Medium|
|GPT4o-mini| 6| $0.2| >$0.4| >$1| $10+|Medium|

While feasability seems pretty high by looking at this analysis, the accuracy performance numbers suffer greatly in practice, especially for smaller models like GPT4o-mini outside of controlled test environments.

**Footnotes:**
- All Ad ROIs are calculated using the [following benchmarks for ROAS](https://www.intensifynow.com/blog/facebook-ads-roas-by-industry-benchmarks/)
- All model pricing comes from [Artifical Analytics](artificialanalysis.ai)

## Timeline
We cannot realistically see ads coming into AI agent responses until there are models which are present in the red bounding box in the chart shown above. As there are no models in that zone currently, pursuing this as an open-ended problem does not make any sense. So lets look at scenarios that can play out:

### **Bullish Scenario**
- Given how fast the models are becoming cost efficient and how quickly knowledge distillation techniques are evolving, it is most likely possible that we see LLM ads as early as [Q4 of 2024](https://www.emarketer.com/content/perplexity-launch-ads-q4-marketers-hesitant-invest) but they are unlikely to be deeply integrated into agentic behavior.
- And we are already here to some extent. Purely AI generated campaigns can be found in the wild with large brands like Coke, Nike, and Samsung.
- A numerical analysis of trends(*model launch dates*) in the scatter plot can confirm that the most likely timeline is Q2 FY2025.

### **Realistic Scenario**
- From the past 1 year we havent seen any major breakthrough in open-world function calling performance as you can see in *SoTA vs time* plot above, so we might need some model level of drastic breakthroughs. Or a lot of thought process into defining the ad integration process.
- A analysis of academia and macro trends shows that the answer is most likely to be in 1.5-2 years time where LLM ads can be productionized well.