Could the AI race push Meta to launch a cloud computing platform? (A thought experiment)
It's going to take a while for Big Tech to identify AI "killer apps" that actually earn a profit. In the meantime, could a cloud platform allow Meta to recoup its AI-related capex faster?
A thought experiment
This post is just a thought experiment. It’s not a prediction and it’s certainly not based on any direct knowledge of Meta’s strategy. This post was inspired by two facts:
Meta still stands out as a major “big consumer tech/media/commerce platform” that DOESN’T offer cloud computing.* The three platforms of this kind that do offer cloud computing are Amazon (AWS), Google (GCP), and Microsoft (Azure).
Meta (like many other companies right now) is taking a big gamble on hoarding expensive AI computing power that will need to be monetized ASAP.
* Apple would be another example, but that’s a post for another time.
Meta’s big bet
In case you haven’t heard, Meta is making an enormous bet on AI and AI infrastructure. Here’s what Mark Zuckerberg said on 18 Jan 2024:
“…we're building massive compute infrastructure to support our future [AI] roadmap, including 350k H100s by the end of this year -- and overall almost 600k H100s equivalents of compute if you include other GPUs.”
Raymond James estimates that these hundreds of thousands of specialized GPU (Graphics Processing Unit) chips are going to cost Meta close to $9 billion.
This hoard of GPUs is a big part of a big bet. Overall, Meta’s R&D expenses (as a percent of revenue) are currently roughly double that of Amazon, Google, and Microsoft:
Q1 2024 R&D expenses:
Amazon: 14%
Google / Alphabet: 15%
Meta: 27%
Microsoft: 12%
For historical comparison, Meta’s Q1 2021 R&D expense was 20% of revenue. While the leap from 20% to 27% in 36 months doesn’t strike me as constituting a risk-it-all existential moonshot, the magnitude of investment has been significant. And the AI arms race —with its voracious appetite for compute capacity and richly compensated technical staff— seems to be shaping up to be a giant money pit for tech competitors. For example, Zuckerberg recently cited the need for sheer electrical power as being yet another front opening up in the AI wars.
An emerging war of attrition among tech giants
Meta isn’t the only company digging an AI moat and filling it with GPUs (and maybe nuclear power plants). If we assume that we’re still in the early skirmishes of an AI battle royale, Meta is signing up for a long, costly war.
How long can Meta take to figure out how to monetize its GPUs?
Meta investors want assurance that R&D expenses (whether related to AI or not) will be levered into profits on some reasonable timeline. However, the “killer apps” of AI are still ambiguous and AI tech itself is far from fully baked. Tech giants like Amazon, Google, Meta, Microsoft, and OpenAI are probably still years away from figuring out how to actually make money from AI.
Meta needs a lot of things to go right in order to traverse the chasm from present-day R&D outlays to future payback.
Do I personally believe big bets on AI are justified? Absolutely, but there’s no getting around that bets are bets; they require risk and cost.
If you strip it down, Meta is betting that they can find at least one breakout AI “killer app” (whether internal, user-facing, or B2B) that starts to pay back the billions of dollars they’re investing before they have to start writing those investments off as losses (and possibly face more existential threats from competitors who were actually successful in finding AI killer apps).
Incremental AI-based improvements to Meta’s ad algorithms probably won’t cut it
Currently Meta makes around 97% of its revenue selling ads. Juicing this with AI is an obvious answer for how Meta makes its Big Bet pay off.
Meta is almost certain to improve its ad platforms and ad monetization using AI. AI could build better target audiences, AI could craft more effective ads, and AI could even create landing pages and what-not.
But is this really a “killer app?” Or is it just another incremental improvement in digital advertising’s long history of applying machine learning and AI to the problem of selling the most relevant consumer eyeballs to the highest bidders? In other words: would a better ads algorithm represent a step change or paradigm shift? Not really. Would it materially grow Meta’s TAM or unlock any new revenue streams (either on a consumer or B2B basis)? Not really. When a car company adds 50HP to next year’s model, they don’t expect to turn the car industry upside down.
Better ad algorithms using the latest advancements in AI are the new table stakes. Amazon will have them, Google will have them, everyone will have a better black box. There’s no reason to assume it would confer any special profit advantage to Meta and therefore little reason to assume that “making a better ads black box” is the reason Zuckerberg is sticking his neck out right now.
Could Meta rent out some of that compute power?
Instead of betting all its chips on its own ad business, could Meta launch a cloud platform? Such a platform could essentially “rent out” coveted GPU capacity to global businesses and startups.
Directly monetizing at least some of their GPU and specialized AI infrastructure is something that Amazon, Google, and Microsoft are already able to do via their own respective cloud platforms.
I couldn’t say what’s involved in establishing a cloud computing platform. I assume the investment levels and timelines are significant.
However, AI is new and GPUs are specialized. Meta could be nimble. They wouldn’t have to build a comprehensive general-purpose cloud behemoth like AWS with its dizzying array of over 200 services. Meta could offer a handful of specialized and differentiated AI-based cloud services that it could bring to market faster and at lower cost.
Setting aside the effort to stand up a cloud computing platform of any kind (glib as that may be), it seems that such a platform would shorten the “journey” that revenue would have to take in order to monetize GPU capacity. An improved ads algorithm is a more indirect way of monetizing GPU capacity than simply renting out the “raw” capacity via a cloud computing platform.
Lastly, it’s worth noting that improving its ad algorithms and launching a cloud computing platform are not mutually exclusive monetization options for Meta.
Is there an opening for a new kind of cloud platform?
This is where the thought experiment gets really interesting.
Let’s be imaginative and assume that AI will unleash new computing paradigms. Perhaps ones where computers and systems aren’t “programmed” or configured any more.
Instead of the burden being on the human user to carefully input instructions into narrowly-defined applications, perhaps AI-based systems will be more like responsive, versatile, general-purpose organisms that will obviate the need for step-by-step instructions and generate their own path to achieve whatever goal the user provides.
We can already see this with generative AI: instead of a human using a cursor to meticulously draw a cat in a highly specialized piece of drawing software, the human just says “draw me a cat” and generative AI figures out the steps for itself and draws the cat.
Just think about how different this is! There’s no cursor, no specialized drawing software, and no individual brush or pen strokes. It’s faster, it makes (mediocre) cat drawings even more abundant than they already are, it eliminates the need for specialized artistic skills, and it consumes orders of magnitude more compute resources.
We’re about to experience a radical change in how computers work and what differentiates effective computer applications (and infrastructure) from ineffective ones. I couldn’t say exactly what form this will take, but I would be willing to bet that Mark Zuckerberg agrees with that basic premise.
One of the most elegant illustrations of this revolution was tweeted by Dharmesh Shah, co-founder and CTO of Hubspot (who happens to be on Substack):
Now think about cloud computing, which has been around since 2006 when AWS launched its first services. That’s 18 years. All the current cloud platforms were designed from the ground up to break the shackles of the prior computing paradigm and disconnect computing power and storage capacity from the constraints of local machines and on-prem (or colo) infrastructure.
Revolutionary for its time, but not necessarily fit for purpose in an AI era.
If you believe AI is going to cause a tectonic shift in computing, then it stands to reason that cloud platforms carrying the baggage of the prior paradigm might be at a disadvantage when faced with a new competitor designed from the ground up for the new paradigm.
So in addition to monetizing their GPU hoard more directly, might there also be an opening for Meta to drive a wedge into the cloud computing market with a new platform designed exclusively for the AI era?
What would a cloud computing platform look like in the AI era?
Before I go further: yes, I know that AWS, Azure, and GCP (and all the rest) offer AI services, access to GPU capacity, and all that good stuff. I’m not predicting these platforms are going to die. What I’m suggesting is that like any incumbent company, the assets and innovations that lifted them to dominance may become a drag on their ability to pivot during a time of disruption. This has happened time and time again across many industries and there’s every reason to believe that today’s cloud computing giants may struggle to adapt to an AI era that totally redefines which kinds of infrastructure are valuable / differentiated and which are not.
So if we’ve drawn a line between “old” and “new,” what is it that might define “new?” If Meta were to launch a cloud computing platform, what might differentiate it in substantial, “paradigm-ey” ways?
Focus on application outputs, versus typical cloud platform obsession with meticulous user inputs and byzantine configurations (if you’ve ever used a cloud platform, you know of which I speak).
Native integration into Meta end users. Cloud platforms are mostly built on a “bring your own data” model (as well as a “find your own customers” model). This is a tough pill to swallow in the AI era because only a handful of companies in the entire world have enough data to train good standalone AI applications. Amazon, Google, and Microsoft, for example, all keep their end users and end user data (from services like Amazon Prime, Google Search, Gmail, or LinkedIn) completely firewalled from their cloud platforms. Might application developers be attracted to a Meta cloud platform that comes with built-in training data and a built-in global user base of billions?
Pricing. AI may rapidly erode the value of “bread and butter” cloud computing services that are usually consumed on an unbundled basis (even simple cloud applications can require dozens of discrete cloud services, all separately racking up consumption bills). Builders may come to expect extreme levels of abstraction and bundling due to AI’s ability to “eat” other software. This may affect pricing. I couldn’t say exactly how, but the incumbents will have trouble adapting to it, as it will mean killing their current flock of golden geese.
Those are just a few ideas. I expect someone somewhere has done a much deeper dive. I also expect the incumbents are already thinking about this and planning for it.
The “thousand monkeys” strategy
The last point I’ll make is that Meta is currently a single company looking for AI killer apps. If Meta were to launch a cloud computing platform, they would be incubating potentially millions of builders around the world. A cloud platform could be a way of “crowdsourcing” the discovery of AI’s killer apps and making sure they come to life in the Meta ecosystem versus Amazon’s, Apple’s, Google’s, or Microsoft’s (among others).
That’s all I got
If you got this far, thanks for reading. Are there other arguments in favor of this that I didn’t think of? What are the arguments against? What would an “AI-native” cloud computing platform look like?