Anthropic vs Meta: Which AI Bet Has More Downside?

Why similar valuations hide different risks

A company's valuation isn't a direct readout of how risky its AI strategy is. Two firms can trade at comparable enterprise values while relying on different sources of current revenue, carrying different costs for experimentation, and facing different consequences if a new product underperforms.

Meta's established business is advertising, where AI is already useful. Better content understanding, ranking, and ad tools can improve targeting and advertiser results. Meta is also building models and products that could become new revenue drivers, but it doesn't need every one of those bets to land for the ad machine to keep working.

Anthropic's established business is mainly selling access to its models, mostly through API usage. That revenue is tied more directly to model quality and unit economics. If customers can switch to a cheaper model that still performs well - or build something comparable themselves - Anthropic's revenue becomes harder to defend. First-party products could reduce that dependence, but their scale and economics are still less established.

This isn't just a "legacy company vs. startup" contrast. Meta's AI work is expensive and uncertain, while Anthropic's current business already has paying customers and meaningful growth. So when things go wrong, the downside doesn't hit the same place. If Meta's new AI products don't work out, shareholders still have advertising earnings - though they've paid for major infrastructure and research along the way. If Anthropic's models lose their edge, the API business itself could be worth less because customers can compare models by performance and price.

That's what "room to be wrong" really means here. It's not a promise that Meta's spending is safe or that Anthropic can't build durable products. It's a way of asking whether the existing business can hold up the valuation if the next stage takes longer, costs more, or generates less revenue than expected.

The useful comparison isn't just AI capability. It's the relationship between the core business, the AI investment, and the fallback value if the investment underperforms.

What is Meta actually betting on?

Meta plans to spend an estimated $135 billion to $145 billion on capital expenditure in 2026 - about 3.5 times its 2024 level. Meta doesn't spell out how much of that supports advertising versus MSL and other AI work. One set of estimates, using two different approaches, puts roughly 60% of this year's capex toward AI-related bets, rising to around two-thirds of a forecast $212 billion in 2027.

Those estimates aren't a precise accounting split. Meta says infrastructure capacity can be redirected between model training, existing products, and new services. Its models can also contribute to advertising before any separate product launches with standalone revenue. For instance, improvements to content processing or ranking can make ads more effective even if users never touch an "AI product" at all.

There are also costs that sit outside the capex headline. Data-center joint ventures and third-party cloud commitments may support AI workloads, but the public numbers don't make their exact use fully clear. So the line between maintaining the core business and funding the new bet stays fuzzy - sometimes one dollar of infrastructure can do double duty.

The big open question is whether Meta's new products can earn returns that justify all this spending. Meta hasn't fully established the economics of products such as Muse. Its CEO has also said Muse has to become more than ten times cheaper to run before it can be profitable at scale. That's a real execution risk: you can build infrastructure ahead of demand, and model capability doesn't automatically turn into a sustainable business model.

Meta does have an advantage, though: it can use its existing advertising business as both distribution and monetization. If AI makes ad ranking more effective, the payoff may show up in the core business instead of a new subscription or standalone product. Using 2027 consensus earnings as a starting point, one estimate suggests that after adding back projected Reality Labs losses and estimated MSL and AI costs, the advertising business could generate around $155 billion in EBIT. Against an enterprise value near $1.5 trillion, that implies roughly 10x EBIT for the core business. This isn't a reported segment valuation - just a rough analytical exercise.

Meta's risk is large spending with uncertain returns. The cushion is that the spending sits on top of a major business that can still benefit from AI even if a standalone AI product never becomes a major profit center.

What does Anthropic need its spending to prove?

Anthropic's AI investment looks less like capex because much of its compute is rented and shows up through the income statement instead. In one estimate from the source analysis, the share of compute serving paying customers fell from 44% in the first half of 2025 to about 25% in the third quarter of 2026. The rest went to research, training, and free-tier users - at the same time the company reportedly faced more demand than its available capacity could handle.

That allocation amounts to two linked bets. First, training and research need to make future inference more efficient, so each unit of compute used to serve a customer generates more revenue. Second, model improvements have to keep demand strong and support first-party products that reduce reliance on usage-based API sales.

There is evidence of better efficiency. Over the period covered by the analysis, quarterly revenue reportedly grew about 29 times while paying customers served by compute grew around five times. Put simply: revenue per unit of serving compute rose sharply. But that doesn't capture the full cost of research, the price of rented compute, or how long it takes to turn better models into durable products.

The cost of pushing toward the frontier has also risen quickly. Reported R&D expense increased from about $1.9 billion in the first half of 2025 to roughly $9.7 billion a year later. Benchmark progress accelerated, but progress per R&D dollar fell by about 60% over the period, according to the analysis. Revenue per R&D dollar improved from $0.74 in the fourth quarter of 2025 to $1.91 in the second quarter of 2026, then slipped to $1.78 in the third quarter as research spending grew faster than revenue.

A weaker ratio in one quarter doesn't automatically mean research is failing. Research spending is lumpy, and new capacity can take time to translate into sales. The real question is whether returns recover once that capacity is deployed. There's also a second reason for Anthropic to keep investing: if the newest models drive most revenue, falling behind doesn't just hurt today's business - it can spill into future products too.

Anthropic's spending has to show up in both better economics and durable demand. Model progress alone isn't enough if each improvement costs more to produce than customers are willing to pay.

How durable are the companies' core revenues?

Meta sells advertising outcomes, not access to a specific model. If its ranking and recommendation systems improve, advertisers can get better returns and Meta can capture more value through its auctions. A customer can't easily replace Meta's internal ad system with a cheaper open model and keep buying the same audience on the same platform. That makes the core ad business less directly exposed to "model switching."

Anthropic's API revenue is different. Customers pay based on usage, and sophisticated buyers can calculate model cost per task. If a cheaper model performs well enough, swapping providers can be straightforward. Some customers may also build or adapt their own models once the math makes it worth the effort.

The source analysis points to enterprise examples that show the incentives at work. DoorDash's AI spending reportedly rose quickly before leveling off, while its leaders kept asking whether additional spending was returning enough. Microsoft's CEO has also talked about customers asking whether they can own model weights and manage costs. Those questions don't guarantee customers leave Anthropic. But they do show that strong returns don't remove pressure to reduce the bill.

There are counterweights. Anthropic's customer base is broadening, and better models can enable new tasks faster than customers can replace older models with cheaper options. The API is also a practical way for businesses to use advanced models without building an entire research and infrastructure operation. Still, revenue concentration in recently released models makes continued performance commercially important. The analysis also notes that a meaningful share of Anthropic revenue comes via cloud marketplaces, where customers may find it easier to compare and switch.

All of that creates compounding pressure: big bills lead to more scrutiny, cost-per-task comparisons surface cheaper alternatives, and even simple integration changes can make switching easier. That doesn't mean API revenue is doomed. It does mean it's harder to value as a stable, long-term stream if it depends on maintaining a clear performance lead.

Over the longer term, Anthropic's response is to build revenue that's less tightly tied to model usage. Subscriptions or first-party products can help, especially if customers pay for outcomes rather than individual tokens. But this shift doesn't happen automatically. If usage-billed revenue dominates already, non-metered products have to grow meaningfully - and find effective distribution - before they can provide anything like the same base.

The central difference is what each company sells. Meta is paid largely for outcomes within a system customers already rely on. Anthropic is paid for model usage in a market where buyers can increasingly compare cost and performance.

What should investors monitor?

The cleanest way to track these companies is to connect operating signals to the assumptions that sit under each valuation. A single benchmark, a revenue-growth figure, or one capex number won't tell you whether the underlying flywheel is actually working.

For Anthropic, three indicators stand out:

• The share of compute serving paying customers. If it falls while revenue per unit of serving compute rises, research may be generating enough efficiency to justify shifting more capacity toward training. If the share falls without better serving economics, the company may just be absorbing more cost.
• Revenue per R&D dollar. It's imperfect - payback can take time and investments vary - but a sustained decline would raise a basic question: is spending turning into commercial value?
• The mix of subscription and usage revenue. A larger, more durable subscription or product base would reduce dependence on customers continually buying tokens from the current frontier model. The key isn't only product revenue growth, but whether it's repeatable and profitable.

For Meta, investors should focus on whether AI improves the ad engine and whether new products show workable unit economics. Muse's inference costs matter, but so does whether the product solves a problem people will pay to solve. If AI tools improve ad performance, some of that value may appear in advertiser outcomes and Meta's core earnings even without a separate AI subscription business.

You also have to compare the balance sheet and income statement carefully. Meta owns more of its infrastructure, which means upfront capital and more asset and utilization risk sitting on its books. Anthropic rents capacity, which can reduce upfront infrastructure commitments but increases exposure to provider pricing and operating costs. Comparing only capex can make Meta look more expensive; comparing only income-statement compute costs can understate the real value and risk of owning a fleet.

A practical quarterly review comes down to four questions: Is the core business still generating cash? Is AI improving the core business, or creating new revenue? Are incremental costs rising faster than returns? And if model advantage narrows, what customer value remains?