🟡 Generative AI Heading For (Short Term) Disappointment

The generative AI wave may have taken off with ChatGPT last November, but the first ripple started in June 2020 with the release of GPT-3. Back then, the model’s performance already made quite a stir, especially within the programming and engineering community. Tweets that talked about prompting GPT-3 to build a working React app and other similar ones lit up my timeline.

Hundreds of “GPT-3 apps” popped up overnight. But perhaps because of the larger timing – we were (in hindsight) still in the early months of a global pandemic and had the future of the world to worry about – nothing took off, except for one: GitHub Copilot. A few months after GPT-3’s release, Microsoft licensed exclusive use of the underlying model, and an experimental initiative to build an AI-assisted programming product on GitHub was underway.

In mid-2021, GitHub Copilot was released in technical preview for a small number of hardcore developers. In June 2022, it became generally available as the first real commercial product based on the GPT models – five months before ChatGPT. From GPT-3’s release to GitHub Copilot becoming good enough to charge users $10 per month took two years.

This timeline is important. It took two years before a single dollar came into the bank account, under perhaps the best of circumstances with no noise, no competition, and no heightened expectations.

Fast forward to now, a mere nine months after ChatGPT stirred up all the AI hype for the rest of the non-programming world, we have the so-called Magnificent 7 (Apple, Microsoft, Alphabet, Amazon, Nvidia, Tesla, Meta) carrying the entire US stock market, accounting for more than 100% of the S&P 500’s gain. (More than 100% implies that the rest of the S&P 493 declined.)

This massive gain is largely caused by the extremely elevated expectation of what generative AI can contribute commercially. And this expectation is setting up for some big disappointments in the 2nd half of this year. It is a disappointment that can be easily avoided if public market investors understand the basic rhythm of developing, testing, previewing, generally releasing, and eventually making money from a software product. It is a disappointment that 3 of the Magnificent 7 are starting to warn the market about.

Managing the Short, Medium and Long Term

Two weeks ago, Microsoft, Alphabet, and Amazon all released their quarterly earnings. Each earnings call had snippets that, in one way or another, is conducting expectations management without losing the narrative energy that this AI hype is bringing to all three companies.

Amy Hood, Microsoft’s CFO, gave probably the clearest answer to analysts by walking through literally some of the most common milestones of releasing a software product, as if she was talking to a group of junior engineers during their first week on the job. Here’s what she said:

“As you know, we've -- last week, we announced pricing [for Microsoft 365], then we'll continue to work through the paid preview process and get good feedback.

Then we'll announce the general availability date, then we'll get to the GA date. Then we'll, of course, be able to sell it and then recognize revenue. And that is why I continue to say that I am just as excited as everyone else about this, and it should be more H2 weighted…

But I do think this is really about pacing. And of course, we've still got to get our Security Copilot and some of the Dynamics workloads priced and released. And we'll continue to work toward that.”

(Note: Microsoft’s fiscal year starts on July 1st, so H2 in this case means the first half of the 2024 calendar year.)

Let’s try the same product release timeline comparison and use GitHub Copilot as the reference point. Microsoft 365 Copilot was announced in March, four months after ChatGPT. Four months later, in July, the product got a price: $30 per user per month. If most of the revenue recognition from Microsoft 365 Copilot won't happen until early 2024, as Amy suggested, then it will have taken this product roughly one year from inception to real money in the bank.

This timeline may seem abnormally fast compared to GitHub Copilot’s two years, but Microsoft 365 Copilot is benefiting from the experiences learned from GitHub Copilot, so it should be faster. On the other hand, Microsoft 365 is a much bigger, heavier enterprise-grade product, so this one year execution timeline is still quite astounding, especially for a company of Microsoft’s size. Yet, even this impressively fast timeline disappointed the market. This disappointment was clearly foreseeable, because the market’s expectation, shown in the multiples assigned to $MSFT before earnings, was even more hyped than the best case scenario for a major product release rhythm, like what Microsoft 365 is trying to achieve.

As “disappointing” as Amy Hood’s answer may be, Microsoft’s generative AI products are actually starting to have prices, with a fairly clear roadmap to recognize revenue in the near future. Alphabet and Amazon, on the other hand, are both avoiding the short-term and managing the medium and long-term.

Don’t take my word for it. Here’s what Sundar Pichai said when asked by an analyst about any uptick in customer spending on Google Cloud Platform due to AI:

“It is an exciting moment overall in Cloud because there is definitely a lot of interest from customers on AI, and they definitely are engaging in many more conversations with us. So I would say, without commenting on the short term, but when I think about it long term, I view the AI opportunity as expanding our total addressable market and allows us to win new customers.”

On Amazon’s earnings call, a similar question was asked with regard to AWS, and here’s what Andy Jassy said:

“I think when you're talking about the big potential explosion in generative AI, which everybody is excited about, including us, I think we're in the very early stages there. We're a few steps into a marathon in my opinion. I think it's going to be transformative, and I think it's going to transform virtually every customer experience that we know. But I think it's really early. I think most companies are still figuring out how they want to approach it.

They're figuring how to train models. They want to -- they don't want to build their own very large language models. They want to take other models and customize it, and services like Bedrock enable them to do so. But it's very early. And so I expect that will be very large, but it will be in the future.”

If we apply the same product release timeline comparison, then it’s clear that no matter how many generative AI applications that either Alphabet or Amazon has cooking in the kitchen, none of them is good enough for its own pricing preview yet. Thus, if any of them were to reach general availability one day and bring in real dollars, that day is at least two years away. (The only minor exception is Amazon CodeWhisperer, an AI-assisted programming product similar to GitHub Copilot, which does have pricing information. It is much easier to price a product when a similar one has already existed in the market for a year.)

For Amazon, it is not clear yet if the company is even interested in building standalone generative AI commercial applications; it may be content to leverage AWS’s cloud market dominance and its home-grown GPUs to be the AI infrastructure for other companies. That’s a perfectly good strategic position to occupy. As for Alphabet, that interest is clear given all the generative AI it is injecting into its various Google Workspace products, a direct competitor of Microsoft 365.

As investors, we are better off keeping the reality and basic rhythm of building, pricing, and releasing software products in mind, rather than falling into the trap of hype and storytelling.

Overestimation vs Underestimation

“Most people overestimate what they can achieve in a year and underestimate what they can achieve in ten years.”

This is a quote often attributed to and popularized by Bill Gates. The original inspiration of this saying is another quote first uttered in the 1960s by a Stanford computer scientist, Roy Amara:

“We overestimate the impact of technology in the short-term and underestimate the effect in the long run.”

If we replace “technology” with the more specific “generative AI” in what Amara said, then it perfectly captures what has happened since GPT-3, GitHub Copilot, and ChatGPT, as well as what will likely happen in both the near and far future.

For the 2nd half of this year, the near future, we are likely headed for more disappointments and enter into a phase of disillusionment from the overestimation of generative AI. As GPU shortages ease by the end of 2024 or early 2025, generative AI will emerge from a period of underestimation to fulfill its potential in the long run.

The market is the collective wisdom and emotions of the crowd. And most of the crowd don’t understand, or care to study, the basic rhythm, processes, and timelines of how a software product gets made and sold. For the ones who do, it’s not hard to see what looks overestimated and what looks underestimated. With that insight alone, there is an abundance of opportunities to make a killing.