The Case for Meta Enterprise Platform

Meta seems to be announcing something everyday these days. Yesterday, the company announced “Meta Enterprise Platform”. From the official blog post:

Meta Enterprise Platform will use our strengths that few other companies have: advanced models, leading agents, large-scale infrastructure, and years of working closely with many businesses. Initially, we will focus on bringing our full technology stack, including the Muse agent, Meta Business Agent, Muse API, Muse Code, and more to businesses and developers to help them grow.

More interestingly, Meta hired CJ Desai away from MongoDB to run their enterprise platform. Desai had been MongoDB’s CEO for eleven months, and was president of product and engineering at Cloudflare before that.

For a company that is quite adept at monetizing consumer attention and a history of failure at getting much traction in enterprise, I think it is fair to say the consensus around this announcement leans towards skepticism. The below tweet captures such understandable skepticism well.

Ben Thompson, while remaining enthusiastic about Meta’s AI related investments, also said that he “hates” this. I don’t want to be contrarian for the sake of it, especially because consensus is often right. However, I do believe there are compelling reasons why Meta is pursuing the enterprise opportunity here.

The simplest reason is probably that Meta needs a “second” customer. Meta is committing “tens of gigawatts” against a demand forecast for a set of first-party products whose adoption curve likely has a high margin of error. Muse seems to have found at least the initial product-market fit as a consumer agent, but I have no idea, and I suspect Meta doesn’t either, whether it ends up a 500 million user product or a two billion user product (also in what timeframe), or how much compute an agentic product will consume per user once people actually let it do things at scale. Moreover, we also don’t know the forward pricing curve of computing itself. Even small margin of error (on either side) can really compound here in forecasting Meta’s aggregate compute need in out years.

With only first-party products, being wrong is expensive in both directions. If Meta underbuilds and they can’t serve the product they just found product-market fit for, the users will drift to competing AI agents/products.

If Meta overbuilds (which is both Alphabet and Meta’s stated strategic preference if they must choose one), the surplus will depreciate through the ad business’s margins. A second demand sink lowers the expected cost of being wrong on the high side as building aggressively then starts looking the more of an optimal strategy. I do believe Meta’s first party products will remain the best use of compute, but the optimal strategy when you are underwriting a gargantuan capex build out may not be 100% 1P products. As regular readers are perhaps well aware by now, I am much more nervous about 3P hyperscalers (e.g. Amazon) who are much more reliant on selling compute without the reservoir of compute needs from 1P products. But the opposite extreme may also be sup-optimal.

Meta already has upside flex since it rents in from CoreWeave, Nebius, Google and Oracle when it needs more than it owns. However, it still does lack a downside flex. A 3P channel is the missing half which will let Meta rent in on the upside, and rent out on the downside.

Moreover, one of the challenges with pursuing the consumer agent opportunity is that Muse may well be a category defining business in 2030, but it is not going to contribute billions this year or, I suspect, next. It’s partly why companies such as OpenAI may find it difficult to make their consumer agents available to everyone. In fact, rumors suggest that the consumer agent OpenAI will launch today will only be available to Pro subscribers first. The fact that OpenAI’s annual run rate reportedly reached $70 Billion on the back of enterprise momentum will make it even more challenging for them to divert compute to consumer agents, especially when they still require outside capital to fund their business.

While Meta is very profitable business, I think Zuckerberg too wants to keep raising capex, and perhaps would like to tap the debt or equity markets to fund such capex. Even for Meta, it may be asking investors too much to fund a slow grind of consumer agents (relative to enterprise revenue momentum) and he likely needs the canvas of opportunities to look wider than the street currently believes. As OpenAI’s own rapid pivot and success in enterprise suggests, enterprise is the fastest way to widen it.

There is also an underappreciated risk that I actually worry about. In my recent post on Anthropic, I noted:

“The end-2026 exit run-rate ($100 Billion, the low end of what the FT says investors expect) over year-end inference watts gives a yield of ~$49 Billion per inference gigawatt-year. If such yield sustains till 2030, Anthropic would be a steal at the rumored $2 Trillion valuation at IPO.”

Given OpenAI is also perhaps racing to similar level of revenue run rate by this year, it is fair to say that the labs are making money hand over fist, at least on the inference slice of the fleet.

Now think about what that means for the incremental capacity coming online over the next few years. It seems plausible to me that if more than half of it ends up with OpenAI and Anthropic, and if yield per inference gigawatt stays anywhere in the $30-40 billion range, then the labs can afford to bid up the price of compute to a level anyone other than labs will find difficult to match. If the marginal inference gigawatt earns Anthropic $36 billion a year, it can pay $20 billion a year for it and still be fine. If Meta's marginal gigawatt earns, say, $15 billion a year, Meta can't. To be clear, I don’t really believe that this is what’s going to happen, but I would also like to be a bit humble about this possibility given that I would probably be also skeptical this time last year if you mentioned to me that OpenAI and Anthropic would be generating the kind of inference revenue per GW they’re generating in 2026.

Moreover, I don't think it's only Meta's problem. If the incremental capacity keeps flowing to two labs, the hyperscalers' backlogs will get more concentrated in those two counterparties over time, which hands the labs a lot of power over the hyperscalers, and the same thing rolls down through the chip vendors and the rest of the value chain. Nobody in that chain should want a duopoly at the frontier.

That’s why I believe Zuckerberg wants revenue per inference gigawatt to come down to a level that leaves room for everyone else near the frontier. I’m not suggesting he wants to drive the labs’ returns to zero; that would make no sense for anyone, including Meta, which needs its own returns on capex. But he needs the yield to compress enough that a Meta can still get its hands on the marginal gigawatt. Meta doesn’t have to win the enterprise for this to work, but it has to be a credible alternative at a lower price which caps what the labs can charge, and which caps what they can bid. Even a modestly successful enterprise business has strategic value well beyond Meta’s own P&L.

But what is Meta’s enterprise go-to-market? How can they possibly have a shot here? Well, in this cycle, the model is the go-to-market. Neither Anthropic nor OpenAI got to their current run-rates by building a decade of enterprise field sales. They got there because they had a frontier model developers wanted, and, between them, they distributed that model through Bedrock, Vertex and Foundry.. Such an option did not exist when Google Cloud started grinding its way into the enterprise. There was no marketplace where a third party would sell your “cloud” for you. There are now at least three. A hyperscaler whose backlog is concentrating in two labs also has every reason to put another frontier option in front of its customers, if only to dilute those labs' bargaining power. However, Muse served through Bedrock runs on Amazon's gigawatts and the downside flex I described earlier only materializes to the extent enterprise inference actually lands on Meta's own fleet, which I suspect is the point of having a direct platform at all. As a result, I believe the marketplaces are the on-ramp for the model to get it in front of the developers, but Meta Enterprise Platform is where the larger and more agentic workloads end up over time.

Of course, the whole enterprise strategy relies on Muse Spark staying close enough to the frontier that the price advantage means something. If it doesn’t, none of this may matter. Watermelon as well as post-watermelon models will have to keep up with the labs to gain the confidence from the enterprise customers that there is no “Llama-4” type fiasco lurking here.

Look, everybody knows Meta isn't an enterprise company. CJ Desai knew it too, and he still chose to walk out of a public CEO seat literally a day before his own investor day to take the job. The team Meta has been assembling also makes me think Meta is much more serious here than market probably believes. A former enterprise chief reporting straight to Zuckerberg, an ex-AWS S-team member owning the infrastructure product, Daniel Gross running capacity strategy and supplier partnerships at Meta Compute, and Dina Powell McCormick working the sovereign financing angle…this doesn’t seem like a half-hearted meek attempt at the enterprise. I would rather Meta take this swing NOW when the model is the go-to-market and the hyperscalers have a reason to help, than wait until the labs have locked up the next twenty gigawatts and the question of who gets to compete at the frontier has already been answered.

Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 70+ companies. The daily is just how I think out loud between the Deep Dives!

Current Portfolio:

Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.

My current portfolio is disclosed below: