Signals Inbox·July 28, 2026·AI Infrastructure

Is Meta becoming a cloud company?

Meta is not an AWS-style cloud company, but it is taking credible steps toward selling AI infrastructure to outside customers—and that could become a meaningful business surprisingly quickly.

Q1Why are people calling Meta a cloud company now?

Meta’s cloud push is now serious enough to investigate, although the business itself has not launched.

The idea moved from possibility to active preparation in roughly seven weeks. At Meta’s latest shareholder meeting, Mark Zuckerberg said a cloud business was “definitely on the table” and revealed that outside companies approach Meta almost every week asking to buy compute or access its models through an API. Reports then said Meta was designing a service that could sell raw AI capacity and hosted models. Soon after, Reuters reported preliminary talks with Anthropic over a compute agreement worth as much as $10 billion over two years.

Meta has also recruited Dave Brown, one of Amazon Web Services’ most senior compute executives. Brown spent nearly two decades at Amazon and helped run areas including EC2, machine-learning infrastructure, Bedrock and SageMaker. He is expected to work under Meta’s infrastructure chief on its expanding data-center operation.

The sequence is unusually clear. Meta acknowledged regular customer demand, began shaping a product, entered talks with a possible anchor customer and hired someone who understands how AWS turns servers into services people can buy.

We are still looking at a plan rather than an operating business. But Meta is now taking several steps that would make little sense if cloud sales were only a passing idea.

Q2What would count as Meta becoming a cloud company?

For this article, Meta becomes a cloud company when outside customers can repeatedly buy its compute or hosted AI services under dependable commercial terms.

A huge data-center fleet alone does not qualify. Meta has operated large computing systems for years, but almost all of that infrastructure has supported Facebook, Instagram, WhatsApp, advertising and its internal AI work.

The first meaningful level would be a wholesale AI infrastructure business. Meta could reserve thousands of chips for a small number of AI laboratories, charge them under multiyear contracts and offer hosted access to models such as Muse Spark. That would place Meta in the same broad market as CoreWeave and Nebius.

Becoming another AWS would require much more. Enterprises expect databases, storage, cybersecurity, networking, monitoring, disaster recovery, developer tools, billing controls, regional availability and support teams. Meta currently has nothing close to that public catalog.

The useful question is narrower than whether Meta will recreate AWS. We need to determine whether external AI infrastructure can become a permanent and material Meta business.

What would count as Meta becoming a cloud company?

Q3Does Meta already have the infrastructure of a cloud giant?

Yes. Meta currently has the physical scale of a hyperscaler, even though nearly all of that capacity serves Meta itself.

Meta’s newest Canadian data center will become the 33rd facility in its global fleet. The company says it has broken ground on ten data centers within the past two years. Its latest disclosed developments include a five-gigawatt expansion in Louisiana, a one-gigawatt campus in Canada, a one-gigawatt project in Indiana and a 168-megawatt leased facility in India. Those four sites alone represent about 7.17 gigawatts of planned or leased capacity.

For perspective, CoreWeave recently passed one gigawatt of active power across its entire cloud. Meta’s Louisiana project is designed to reach five times that amount at one location, although active and planned capacity are not directly equivalent.

The spending is moving just as quickly. Meta invested $72.22 billion in capital expenditure last year. Its current guidance for this year is $125 billion to $145 billion. At the midpoint, that is an 87% increase in twelve months. Meta also had $103.77 billion of leases that had been signed but had not yet started, mostly involving data centers, colocation facilities and network infrastructure.

Meta is also building deeper into the hardware stack. It has hundreds of thousands of its own MTIA accelerators in operation, plans four new chip generations within two years and expects to deploy more than one gigawatt of custom silicon before moving into several gigawatts.

So the infrastructure question is settled. Meta already knows how to finance, design and operate hyperscale computing systems. The hard part is turning some of that machinery into a product an outside customer can trust.

Q4Is Meta really sitting on spare AI compute?

Probably not in the simple sense. Meta currently says it needs its capacity, while construction schedules can still create temporary pockets of unused compute.

Zuckerberg has pushed back against the idea that Meta already owns a large pool of idle machines. His cloud comments were conditional. If Meta builds ahead of internal demand, and another company offers an attractive price for that temporary capacity, selling it could make sense.

Meta’s own purchasing behavior supports that explanation. CoreWeave says Meta recently signed a new $21 billion infrastructure commitment with it. Meta is also bringing tens of millions of AWS Graviton cores into its compute portfolio and ended last year with $131.05 billion of contractual commitments covering third-party cloud capacity, servers, networks, data centers and other hardware.

A company can buy and sell compute at the same time. A finished inference cluster in one country cannot always replace a delayed training cluster elsewhere. Different chips support different models, while power availability, network design and data-location rules further limit what can move between sites.

Meta may also want more capacity than it needs under its expected scenario. Building extra headroom protects it if AI usage grows faster than planned. Renting some of that headroom until Meta needs it would turn an expensive insurance policy into revenue.

The clearest reading is that Meta is managing timing mismatches rather than admitting to a company-wide glut. That can support a cloud business, but the service may remain uneven unless Meta eventually builds capacity specifically for paying customers.

Q5Does Meta have real cloud customers today?

No. Meta currently has no disclosed base of companies paying to use its general AI infrastructure.

The Anthropic discussions remain preliminary. The reported value is a maximum rather than a signed commitment, and either side could still change the scope or abandon the deal.

Meta does have customers and developers using hosted products. Businesses rely on the WhatsApp Cloud API, developers have tested the Llama API, and selected partners can access Muse Spark through a private API preview. The recently launched Meta Business Agent Platform also lets companies build and deploy customer-service agents using Meta’s infrastructure.

Those products show that Meta can host software for outside organizations. They do not prove that a customer can reserve 10,000 accelerators, select a region, negotiate an uptime guarantee and run its own model on Meta’s hardware.

Zuckerberg’s claim that companies contact Meta almost every week is encouraging. Regular inquiries suggest real demand. Today, though, Meta has inbound interest and experimental services rather than a repeatable infrastructure business.

Q6Would the Anthropic deal change Meta’s business?

A full $10 billion Anthropic contract would validate Meta’s cloud idea, yet it would barely change Meta’s overall revenue mix.

The reported maximum is spread over two years. An even split would produce $5 billion in annual revenue, equivalent to about 2.5% of Meta’s $200.97 billion annual sales.

The number looks more important beside Meta’s activities outside advertising. Family of Apps generated $2.58 billion of other revenue last year, while Reality Labs generated $2.21 billion. Together, they produced $4.79 billion. One fully utilized contract of the reported size could exceed all of Meta’s existing non-advertising revenue.

It would still be small beside an established cloud. AWS generated $37.6 billion in its latest quarter, including $14.2 billion of operating income. Meta’s hypothetical $5 billion of annual cloud revenue would equal only 13% of one AWS quarter, or roughly 3% of AWS revenue at its current annualized pace.

The deal would matter most as proof. It would show that a sophisticated AI company is prepared to place important workloads inside Meta’s infrastructure. After that, investors and customers would look for a second buyer, then a third. A cloud business becomes convincing through repetition.

How large would a $10 billion Anthropic agreement be?

Q7Is Meta trying to beat AWS or the AI neoclouds?

Meta’s realistic target today sits in the AI neocloud market, where customers buy large blocks of accelerator capacity and hosted models.

AWS, Microsoft and Google sell broad technology platforms. Microsoft Cloud generated $54.5 billion in its latest quarter, while Azure revenue grew 40%. AWS produced $37.6 billion. Google Cloud now sells an integrated stack covering infrastructure, models, agents, data management and security.

Meta would take years to match that product range. It has no public database business, enterprise productivity suite, cloud-security portfolio or large marketplace of third-party software.

The neocloud market offers a much easier entry point. CoreWeave generated $2.08 billion in its latest quarter, had $99.4 billion of revenue backlog and recently passed one gigawatt of active power. Its customers buy infrastructure built specifically for demanding AI workloads.

Meta already understands those workloads. It owns large clusters, designs chips, operates high-speed networks and runs AI inference for billions of people. A small group of large customers could absorb meaningful capacity without Meta building thousands of conventional cloud products.

The competitive overlap is already visible. CoreWeave and Nebius shares fell sharply when reports of Meta’s cloud plans appeared because both sell the kind of AI capacity Meta is considering. Meta is also a major CoreWeave customer, so it understands the service from the buyer’s side.

Meta may eventually add storage, developer tools and enterprise services. For now, the practical opportunity is wholesale AI compute, not a general-purpose replacement for AWS.

Q8Can Meta offer cheaper or better AI compute?

Meta may become very competitive on selected inference jobs, but there is currently no public evidence that its cloud would be cheaper overall.

The strongest argument comes from Meta’s custom hardware. It has deployed hundreds of thousands of MTIA chips for advertising and content-recommendation inference. Newer versions will expand into generative-AI inference, with four chip generations scheduled across a two-year period.

Meta is scaling several hardware families at once. Its Broadcom partnership supports an initial deployment exceeding one gigawatt of custom silicon. Its agreement with AMD covers as much as six gigawatts of Instinct GPUs, while its Nvidia partnership adds another long-term source of training and inference hardware.

This mix lets Meta match hardware to specific tasks. A purpose-built inference chip can be cheaper and more power-efficient than a general-purpose GPU when it repeatedly runs the type of model it was designed for.

Outside customers complicate the picture. Their models may use different software, memory configurations or networking patterns. They also need isolation, monitoring, technical support and predictable availability. Those services add costs that do not appear when Meta runs its own carefully controlled workloads.

Meta is making its chips compatible with widely used software such as PyTorch, vLLM and Triton, which should reduce the friction. Still, no public pricing, independent benchmark or service agreement currently shows a cost advantage over AWS, Azure, Google Cloud or CoreWeave.

We can make a strong claim about Meta’s internal efficiency. We cannot yet extend it to a cloud product nobody has been able to test.

Q9Does Meta’s own business help or hurt a cloud launch?

Meta’s own apps make cloud expansion safer, although they also limit how much capacity outsiders can trust Meta to reserve.

A new infrastructure provider normally needs enough customers to keep expensive machines busy. Meta begins with a guaranteed customer: itself. Its apps served an average of 3.56 billion people each day in the latest quarter, while revenue reached $56.31 billion. Ad impressions increased 19% and the average price per ad rose 12%.

Meta can move compute toward recommendation models, advertising systems, Meta AI, video ranking, content moderation or new consumer products when outside demand slows. That makes its infrastructure much less dependent on one external buyer than a typical neocloud.

The drawback appears when capacity becomes scarce. Improving advertising performance across Meta’s enormous revenue base may generate more value than renting the same hardware. A small gain in ad targeting, recommendations or engagement can be worth billions when applied across Facebook and Instagram.

Cloud customers will worry about priority. They need to know Meta will not reclaim their chips whenever an internal product becomes popular or a new model requires more inference capacity.

Long-term contracts can solve part of that problem by reserving hardware and setting penalties. Eventually, Meta may need separate clusters or entire sites dedicated to outside customers. Without that separation, Meta’s internal scale remains both the cloud business’s safety net and its biggest source of uncertainty.

Q10Is Llama giving Meta a shortcut into cloud?

Yes. Llama gives Meta a large developer audience before the company has built a traditional cloud distribution machine.

Llama passed one billion downloads more than a year ago. Companies, governments, researchers and independent developers can run the models on their own hardware or through other cloud providers.

That popularity has largely benefited somebody else’s infrastructure business. AWS offers Llama through Bedrock, Microsoft distributes it through Azure, and Google makes it available through Vertex AI. Those companies handle the hosting, billing, security and enterprise relationship.

Meta’s Llama API began to reverse that arrangement by letting developers access a first-party hosted version. Muse Spark is also available to selected API partners, while the Meta Business Agent Platform gives companies a hosted way to build agents connected to Meta’s consumer and business products.

The path is straightforward. A developer starts by calling a Meta model through an API. As usage grows, Meta could sell reserved inference, fine-tuning, private deployments or raw capacity beneath the model.

Llama’s distribution does not automatically produce cloud revenue. Many users choose it precisely because they can host it elsewhere. Still, Meta already owns something new infrastructure providers spend years trying to create: a large group of developers familiar with its technology.

That gives Meta a credible opening, particularly if it offers a simple and competitively priced way to run Llama and Muse models directly.

Q11Can Meta win enterprise customers without acting like AWS?

Meta can win a handful of giant AI customers now, but broad enterprise adoption would require a business it has never built.

An AI laboratory buying thousands of chips knows exactly what it needs. The contract can be negotiated directly between technical and infrastructure teams. Meta does not need a worldwide sales organization to serve a few customers of that size.

Traditional enterprises work differently. A bank, manufacturer or hospital may need data-residency controls, certifications, disaster recovery, integration support, predictable billing and help moving old systems. Purchasing decisions can involve security teams, lawyers, technology consultants and several layers of management.

Microsoft’s $627 billion of commercial remaining performance obligations shows how deep those relationships can become. Customers make multiyear commitments across Azure, Microsoft 365, security, databases and business software. Meta currently sells advertising and communication tools to businesses, but it rarely runs their central technology systems.

The recruitment of Dave Brown matters here. Meta needs people who understand capacity planning, service design, customer guarantees and cloud economics. Brown brings experience from the company that created the modern public-cloud market.

Even with the right executives, Meta would probably start with AI-native companies that care primarily about chips, speed and price. Reaching ordinary enterprises would require several years of product development and trust-building.

Q12Could cloud meaningfully reduce Meta’s dependence on advertising?

No. Cloud cannot meaningfully diversify Meta yet because advertising still supplies 97.6% of its annual revenue.

Meta generated $196.18 billion from advertising last year out of $200.97 billion in total sales. Paid WhatsApp messaging, Meta Verified, hardware, software and every other revenue source combined contributed only $4.79 billion.

To push advertising below 90% of revenue, while keeping ad sales unchanged, Meta would need approximately $21.8 billion in annual non-advertising revenue. That is about $17 billion more than it currently generates.

Even a new cloud business producing $10 billion every year would leave advertising at roughly 93% of Meta’s total revenue. Cloud would become a useful second activity without changing Meta’s basic economic identity.

Reaching genuine diversification would probably require several large infrastructure customers, paid model usage, business-agent subscriptions and other AI services growing together. Relying on one wholesale customer would simply replace some advertising concentration with customer concentration.

Cloud can still matter before it transforms the revenue mix. It could absorb unused capacity, improve purchasing power and create a direct way to monetize Meta’s AI models. Calling it an advertising hedge today gives a small and unproven business too much weight.

Q13Can selling compute improve the return on Meta’s AI spending?

Yes. Selling idle capacity could lift returns when that compute would otherwise sit unused.

Meta’s annual capital expenditure has moved from $72.22 billion to a current expected range of $125 billion to $145 billion. At the midpoint, the increase is nearly $63 billion in a single year.

Data centers rarely fill at exactly the planned pace. Buildings, power connections and chips arrive on different schedules. Meta may have a completed cluster waiting for an internal model, a facility with more power than one region currently needs or hardware that suits an external workload better than its immediate internal queue.

Renting those resources generates revenue while Meta continues paying depreciation, interest, leases and electricity. A multiyear contract can also make the economics of a new site easier to predict.

The calculation becomes less attractive when Meta expects to need the capacity soon. AI hardware improves quickly, so a customer may want the newest chips for several years while Meta would prefer to upgrade or reclaim them. Fixed prices can also become painful if power and component costs rise.

Meta’s latest spending guidance was increased partly because component prices and future data-center costs were climbing. A cloud contract must cover those costs while compensating Meta for giving up flexibility.

Selling compute works best as targeted capacity management. It becomes dangerous when Meta signs long commitments mainly to reassure investors that its infrastructure spending will produce immediate revenue.

Q14What could derail Meta’s cloud push?

The cloud push could stall quickly if Meta’s own AI products absorb the capacity or if outside customers demand stronger guarantees than Meta wants to give.

Internal demand is the most obvious risk. Meta is rolling Muse Spark across its apps, introducing more AI tools and preparing several generations of inference chips. Billions of daily users can create enormous compute demand even when only a small share regularly uses an AI feature.

A second risk is customer concentration. Wholesale AI contracts are huge, but there are relatively few buyers capable of signing them. Losing one major customer could empty a large cluster and remove billions of expected revenue.

Meta could also discover that operating a customer-facing service is harder than running infrastructure for its own engineers. External workloads bring new security obligations, support requests, performance disputes and service penalties.

The hardware itself may create another constraint. Meta’s custom chips look attractive for its own models, while customers may continue to prefer Nvidia systems that are easier to move between providers. Meta could end up owning efficient capacity that outsiders value less than expected.

Finally, the current cloud interest may reflect a short period of unusually tight AI supply. Amazon, Microsoft, Google, CoreWeave, Oracle, Nebius and several newer providers are adding capacity aggressively. If supply catches up, prices could fall before Meta has established a permanent customer base.

Meta can walk away more easily than a pure cloud company because its apps can use the infrastructure. That lowers the financial risk. It also makes the cloud strategy easier to abandon.

Q15What would prove that Meta has become a cloud company?

We would call Meta a cloud company once outsiders can buy its compute repeatedly and the revenue becomes visible in Meta’s accounts.

One large private contract would prove that Meta’s infrastructure can support an outside customer. Several unrelated customers would prove that Meta has built something repeatable.

Public availability would strengthen the case. Developers should be able to see what hardware and models are offered, choose a region, understand the price and receive clear support terms.

Service guarantees are equally important. A cloud customer needs reserved capacity, uptime commitments, security controls and a clear process when something fails. Selling whichever machines happen to be free would be closer to opportunistic leasing.

Financial reporting would provide the strongest evidence. Meta currently reports only Family of Apps and Reality Labs as business segments. A separate revenue figure for infrastructure or hosted AI services would show that management considers the activity important enough to measure.

Evidence that would prove Meta has become a cloud company

Q16Is Meta becoming a cloud company?

Partly. Meta is becoming an AI infrastructure seller, while the evidence still falls well short of a full cloud transformation.

Meta already has the difficult physical pieces. It operates at hyperscale, is rapidly adding gigawatts of capacity, designs its own chips and can finance projects that would be impossible for most new cloud providers.

The commercial side is still at the starting line. There is no disclosed infrastructure customer base, no public compute catalog, no pricing and no cloud revenue in Meta’s accounts. Its current APIs remain narrow services attached to Meta’s models and applications.

The most likely outcome is a hybrid business. Meta will keep most of its infrastructure for Facebook, Instagram, WhatsApp, advertising and consumer AI. When it has the right hardware available at the right time, it will also sell large blocks of compute or hosted model access to selected outside companies.

That could eventually produce tens of billions of dollars in revenue and make Meta a serious competitor to AI neoclouds. Becoming another AWS looks far less likely because Meta has little reason to build hundreds of general enterprise services.

So the title is directionally correct, but broader than the evidence supports. Meta is opening part of its internal computing machine to the market. It has not yet become a cloud company in the way Amazon, Microsoft or Google have.