There’s a lot of debate about valuations and where we go from here. I’m going to break this down using a very simple example: my wife, Alex, trying on different outfits. My goal in this post is simple: use an anecdotal example to estimate AI demand going forward, which will help us better understand valuations.

At the macro level, the general consensus is that AI is going to continue growing and driving value but there are concerns that stocks are overvalued. To make this more concrete, look at a couple different examples of investor sentiment across the industry and then look at industry growth rates.

Investor Sentiment

• Janus Henderson conducted a survey where they found that “Two-thirds of investors (67%) are concerned about a potential AI bubble or AI-driven market correction in the near term. Over a longer horizon, sentiment becomes more constructive: 46% of investors expect AI to have a modest positive impact on market returns over the next five years, while a smaller but more optimistic segment (15%) anticipates a major positive impact.” (source)
• Nvidia’s PE ratio is 33 times LTM earnings. 21 times forward earnings

We can agree or disagree with Janus Henderson’s survey, but it’s evident in the market valuations that we’re seeing. Nvidia is trading at 21 times forward earnings. Using heuristics, that’s not a bargain and it’s not overvalued.

Industry Growth Rates

We’re not going to generate alpha by using heuristics. Let’s look at industry growth rates to develop a better understanding of how quickly AI demand is increasing.

Nvidia

• “Record revenue of $81.6 billion, up 85% from a year ago” (source)

OpenAI

• “Search usage has nearly tripled in a year, and our ads pilot reached more than $100 million in ARR in under six weeks.” (source)
• “Codex now serves over 2 million weekly users, up 5x in the past three months, with usage growing more than 70% month over month.”

Anthropic

• “Anthropic generated $4.8 billion in sales in the first quarter. Its quarterly revenue is now growing faster than [<u>Zoom</u>](https://www.wsj.com/market-data/quotes/ZM) did during the pandemic, and Google and [<u>Facebook</u>](https://www.wsj.com/market-data/quotes/META) in the run-up to their initial public offerings. It is set to turn an operating profit of$559 million in the June quarter.” (source)

There’s so much demand that Anthropic is only operational 98% of the time.

Current Valuations

Investors have frameworks for valuing visible revenue. They have a much harder time valuing latent demand from use cases that are technically possible but not yet embedded into daily behavior. It’s simply too technical and confusing.

For example, walk a portfolio manager through the tech stack for silicon photonics or the assembly and packaging process and let me know how much money they’re willing to invest. That technical complexity is causing them to use heuristics to make investment decisions.

Demand Going Forward

So far, everyone is pricing what they can see (e.g. historical growth rates). They’re not pricing what they can’t see. There are almost an unlimited number of use cases for AI. Your personal token consumption could plausibly increase by orders of magnitude but there are three main drivers causing you not consume more tokens:

1. tokens are expensive
2. integrations: agentic features are still very new and haven’t had time to work into everyday applications
3. many solutions, applications, games, etc. have not been created yet because we’re still learning how to develop with AI

An Example: Alex’s Outfits

Let’s use Alex’s outfits for guidance as to how her personal token consumption will significantly increase. She always tries on several different outfits before choosing one and coordinates with the girls on what looks best. But what if she also had a fashion-focused bot that could help her.

Let’s assume she sends a prompt like “Hey Chat, I’m going out to eat in West Village tonight. I want a chic, cool vibe. How does this outfit look?”. This prompt and it’s corresponding image is approximately 113 tokens - 28 input text tokens and 85 image tokens using low detail on gpt-4.1-mini (source).

Estimating Alex’s Incremental Demand for Tokens

Let’s start with the smallest possible version of the product. Alex uploads one low-detail selfie and asks: “I’m going out to eat in the West Village tonight. I want a chic, cool vibe. How does this outfit look?”

In my API test, the text prompt plus low-detail image consumed roughly 113 input tokens. That is intentionally conservative. It does not include multiple images, video, closet memory, retailer integrations, Pinterest inspiration, tool calls, or generated outfit alternatives.

For a rough compute estimate, transformer inference is often approximated as:

I’ll use gpt-oss-120b as a proxy because it is a publicly documented MoE model with 117B total parameters and 5.1B active parameters per token (source.)). For MoE models, active parameters matter because each token is routed through only a subset of the total model.

On the hardware side, an H200 GPU computes 67 TFLOPS of FP32.

Calculate the total amount of GPU seconds for each of her messages to the model.

Then compute the vision encoding step for the image.

Calculate the prefill and decoding steps.

Alex’s very simple query with 113 tokens is going to use ~0.1820 H200 seconds of compute. Extrapolate that to 100 million users daily.

Note: Operations are parallelized across several GPUs but I’m not standing up a Vera Rubin rack to run real simulations. For the data center folks out there that are fuming about the simplifications I made, let’s consider this back-of-the-envelope math close enough. Horseshoes and hand grenades math is all we need.

Takeaways

This solution for outfit recommendations is already being built and deployed by several different companies. This isn’t hypothetical. This is real demand coming online at light speed. That’s 5,055 hours of incremental GPU demand. And we didn’t even scratch the surface for this use case.

Here are several features to our application that we haven’t incorporated but these are features Alex wants needs.

1. Let’s index her closet so the bot can see all of her clothes and it can provide recommendations for other outfits to wear
2. She has a lot of inspiration on Pinterest. Let’s link the public profiles of other creators whose style she likes
3. She loves buying clothes. Let’s also link her bot to her favorite clothing brands and boutiques so the bot can recommend new clothes to buy (I can’t believe I’m streamlining her ability to buy clothes).
4. She doesn’t send one picture to her friends/bot to get feedback. We need front views, side views, back views. I think I’ve seen her take aerial views before. Now we’re sending several pictures instead of one.
5. Actually, she wants to send several pictures plus a 4 second video of her doing a spin.

Depending on how much of this workflow moves from text to images, video, tool calls, and generation, the compute demand is no longer 1x. It could easily become 1,000x or 10,000x. That’s one girl, one outfit, one day. Now multiply that by 150 million American women. Circle back when you’re done extrapolating and let me know what you think growth rates are going to be going forward.

Videos

In our demand calculations, we haven’t factored in video. No one is factoring in video. The numbers become ridiculous. In fairness, currently AI videos are mostly slop. But, the models will only get better. Just to put into perspective, how big video is.

• Netflix is 15% of internet traffic
• YouTube is 12%
• Xbox live is 3%

source

Let’s not conflate the numbers. A video is created once and streamed many times. It’s not like that’s a 1:1 relationship between AI generation and consumption. But, how many AI-generated videos will need to be created to create one usable 5 second scene? Maybe to make a 30 minute TV show for 3 year olds it will take 3,000 minutes of AI-generated content. What’s demand going to look like when subagents are generating 100 five-second scenes independently and then choosing the best scene to use.

Counterexamples

We’ve already seen exponential step-ups in demand in other industries. Two great examples are:

1. Architectural Shift: ASICs and Purpose-Built Silicon a. In the early days of Bitcoin, mining was done entirely on consuming GPUs because they were excellent at parallel cryptographic hashing. Demand exploded, and prices soared. However, as the math stabilized, companies built ASICs (Application-Specific Integrated Circuits) —silicon chips hardcoded to do exactly one mathematical operation and nothing else
2. Software & Algorithmic Efficiency Gains a. In the late 1990s and early 2000s, server CPU demand was projected to grow exponentially to handle enterprise database workloads. Instead, engineers developed massive breakthroughs in indexing, query compilation, virtual machines, and distributed data frameworks (like Hadoop).

Those are fair points. Here are a couple concrete examples that will help us use existing infrastructure more efficiently.

• Tokens will definitely get smarter. For example, “GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence. It also uses significantly fewer tokens to complete the same Codex tasks, making it more efficient as well as more capable.” (source)
• Maybe Llama.cpp makes a step-up improvement in performance and more inference is done locally

Those points and improvements don’t change the trajectory going forward.

• Demand for tokens explosive
• Demand for networking is explosive: I want my bot to communicate to every service on earth (e.g. the weather channel, clothing retailers, restaurants, travel sites, my bookie, etc.)
• Demand for real-time compute is explosive
• Demand for energy is explosive
• Demand for innovation is explosive to offset increasingly more complex use cases

As an investor, I only care about two things:

1. That demand remains strong
2. And that I understand where demand will be served from so I can apply a proper valuation framework to companies

Conclusion

The AI investment question is not whether token costs fall or if growth will continue. A mirror selfie is not a trivial example. It is the entire AI demand thesis in miniature. A previously offline human judgment becomes a multimodal compute workload. Add memory, personalization, video, agents, and commerce integrations, and the demand curve stops looking like software usage and starts looking like infrastructure consumption.

You can come up with a similar use case as this for basically everything. The only question left is which companies are best positioned to capture this demand going forward and which ones are misunderstood and undervalued.