The smartest thing that online bookseller and e-tailing giant Amazon ever did was let you pay for its enormous IT expenses as the company was growing at a crazy pace. You did this by renting compute and storage capacity on the nascent Amazon Web Services cloud. It is not a “public cloud,” it is a private compute utility that people rent capacity on and a massively profitable one at that.
I do not know precisely where the annual Amazon IT budget and the operating profits from Amazon Web Services crossed each other. Amazon’s IT spending (which included content creation) was 8 percent of revenue in 2001, the peak of the Dot-Com Boom, and shrank to 5 percent of revenue in the following year and was around 4 percent of revenue up until 2005. At that point, the investments to create the AWS cloud started to rise, and the money coming in for AWS was ploughed back into infrastructure expansion and as far as I can tell, AWS did not really make an operating profit until late 2014.
If you assume that the underlying non-AWS businesses at Amazon average somewhere around 5 percent of revenues for their internal IT budgets, then the crossover where AWS operating profit was bigger than that internal IT spending happened around 2019. And it has been more than paying the Amazon IT bills since then.
Genius, really. Enlightened self-interest defined.
So, yeah. There is no surprise that Meta Platforms, which is the only hyperscaler that is not also a cloud builder, has been looking at providing some kind of retail cloud service to customers who want access to its hardware and software infrastructure.
In its Q3 2025 financial call last October, founder and chief executive officer Mark Zuckerberg admitted that the company had been mulling over cloud services in a more substantial way than just selling tokens generated by its Llama API service, which was launched in April 2025. Meta Platforms had hoped at the time that its suite of Llama 4 models – Scout, Maverick, and Behemoth – would be among the best models on the planet and that customers would flock to them.
Zuckerberg danced a little bit, and described this move to adding cloud services as more of a scenario where Meta Platforms overbuilds capacity and has to find customers to use it so it doesn’t lose huge amounts of dough. Take a listen:
“There’s just a lot of demand for other new things that we build internally, externally. Like almost every week, people come to us from outside the company asking us to stand up an API service or asking if we have different compute that they could get from us. And we haven’t done that yet, but obviously if you got to a point where you overbuilt, you could have that as an option. And then, the kind of very worst case would be that we effectively have just prebuilt for a couple of years, in which case of course there would be some loss and depreciation, but we would grow into that and use it over time.”
At the company’s annual shareholder meeting at the end of May, Zuckerberg said that selling cloud services was “definitely on the table,” but did not elaborate much beyond that.
Today, Bloomberg is reporting that Meta Platforms is actually getting together a battle plan to more formally sell cloud services and directly take on AWS, Microsoft Azure, and Google Cloud.
This is ironic considering that Meta Platforms has been buying capacity to run some of its workloads from AWS and Azure for many years to supplement its own infrastructure capacity in its own datacenters. Last August, Meta Platforms inked a deal with Google Cloud to rent $10 billion of capacity, and has agreed to buy $21 billion in capacity from king of the neoclouds, CoreWeave. That makes Meta Platforms CoreWeave’s biggest customer, where previously it was Microsoft offloading OpenAI AI training and inference workloads that was CoreWeave’s largest customer.
Now, it looks like Meta Platforms wants to sell access to an API running its more recent Muse Spark AI model, which is not open source like the Llama 3 and Llama 4 models were – no more free lunch, people, and we’re raising the price on the beer, too. It is not clear what models Meta Platforms might load on its infrastructure, but clearly it will do whatever key customers want. If it does have excess capacity on its Nvidia and AMD GPU clusters, it can pretty much run any model its customers want it to run. That will not necessarily be the case on its homegrown MTIA accelerators, which have a fairly aggressive roadmap between now and the end of next year.
The Bloomberg report also says that Meta Platforms is considering offering more basic infrastructure as a service – meaning virtual machines and containers that customers would create instances on like they do on other clouds, and that this whole effort is called Meta Compute.
There are apparently three executives who have a hand in steering the Meta Compute effort: Santosh Janardhan, who is head of global infrastructure and co-head of engineering at the social network, seems to be riding point. Janardhan was a database and storage admin at PayPal in the wake of the Dot-Com boom and then moved over to YouTube to do the same job for nearly four years, and then moved to Meta Platforms in 2009 to have the same job again. Janardhan rose up through the ranks, becoming a director of production engineering and site reliability, then took over software and production engineering before becoming the chief infrastructure exec four years ago. Daniel Gross, part of the Superintelligence Labs unit that is creating the new models, also has a hand on the tiller at Meta Compute, as does Dina Powell McCormick, who took over as president and vice chairman of Meta Platforms back in May 2023 after being a partner at Goldman Sachs for sixteen years.
With all of the hyperscalers and cloud builders complaining that they cannot get the capacity they need for their AI efforts, it is a bit of a wonder how Meta Platforms thinks it will have excess capacity. It seems to me that the company has figured out that the only way that it can make money from its enormous investment in AI infrastructure and AI model building is to become a neocloud. And we think that given the size of its datacenter investment, this cloud layer was always part of the strategy. If Meta can be at the front of the GPU line and keep Nvidia and AMD hopping because it has its own accelerators – just like AWS, Microsoft Azure, and Google Cloud – then it needs to leverage that advantage to make money. Being a neocloud does this.
There is precedent for this. Back in November 2024, in Anthropic And OpenAI Show Why Musk Should Build A Cloud, I explained why Elon Musk needed to take better advantage of the enormous amount of money that was spent to build clusters to train its Grok AI models. And early this year, lo and behold, he turned the Colossus datacenters in Memphis into a neocloud and sold capacity to both Google and Anthropic to pump up SpaceX’s financials ahead of its initial public offering a few weeks ago.
As we said back then at the end of the “Musk must cloud” story, there will be a day – we are not sure what year that day will be in – when Nvidia also has to become a cloud and cut out all of the middlemen so it can keep its profits high after its revenues inevitably level off.
Nvidia spends three-fourths of its payroll on creating software, and because of the very high prices it can charge for its GPUs and ancillary gadgetry surrounding them, that software is largely free. But if hardware prices drop and margins follow it, Nvidia will have to make it up selling software – perhaps starting with fees for support for the new open source Nemotron models that it released this year. Nvidia is the only company that can afford to give its AI models away, and it is one of the few companies that can get compute engines at cost.
Meta Platforms cannot get compute engines at cost, and it cannot afford to give its AI models away anymore, especially in a world where Google, AWS, Anthropic, and OpenAI are most definitely not giving their models away.
