A back-of-the-envelope calculation making the rounds on social media this week has reframed one of the most-hyped AI products of 2026: Meta’s Muse agent. The math, first floated by Daytona CEO Ivan Burazin and amplified by Box CEO Aaron Levie, puts a jarring price tag on Meta’s ambitions for Muse — roughly $2.8 billion in infrastructure just to serve 100 million users. Facebook alone counts close to 3 billion. The gap between those two numbers is now the subject of intense debate among engineers, cloud architects, and investors trying to figure out whether agentic AI at consumer scale is even economically sane.
The estimate, which Levie shared in a widely circulated post crediting the calculation to Burazin, states plainly: “@auchenberg calculated that Muse needs 65,000 CPUs and 75PB of DRAM to serve 100 million users. That’s $2.8B in infra for one app.” That single sentence has become a flashpoint for a bigger argument about whether Meta, OpenAI, Google, and Microsoft are building AI products whose compute appetite scales faster than any plausible revenue model can follow. Yahoo Finance and other outlets picked up the figure as it spread, and finance-focused platform Stocktwits has tracked it as a trending ticker-adjacent story tied to Meta’s stock-watchers.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Muse Actually Is, and Why the Math Matters
Muse is described in the reporting as Meta’s agentic AI platform, a personal-agent product built to handle multi-step tasks rather than simple chat replies. That distinction is the whole reason the compute math looks so different from a standard chatbot deployment. A conventional large language model query is largely stateless: a prompt goes in, a response comes out, and the session more or less ends there. An agentic system like Muse is expected to hold context across many steps, call tools, retain memory, and keep working in the background on a user’s behalf. All of that requires far more active memory and far more persistent compute per user than a typical chatbot session, which is exactly why the DRAM figure in Burazin’s estimate — 75 petabytes — is doing so much of the heavy lifting in the final cost number.
It is worth being precise about what is confirmed and what is not. The 65,000 CPUs, 75 PB of DRAM, and $2.8 billion figures are an external estimate circulated by Burazin and amplified by Levie — not a disclosure from Meta itself. Meta has not published an infrastructure budget specific to Muse, nor has it confirmed a 100-million-user target or timeline. Separately, Evercore analyst Mark Mahaney has been reported as expecting Muse to reach 100 million users within six to 12 months, which appears to be the basis for why the 100-million mark became the benchmark used in the viral calculation. None of this means the underlying arithmetic is wrong, but it does mean readers should treat the $2.8 billion figure as an informed outside estimate rather than an audited Meta budget line.
The Core Numbers Behind the Viral Estimate
Strip away the social-media framing and the calculation boils down to a capacity-planning exercise that any infrastructure engineer would recognize. If an agentic AI product needs to keep a meaningful chunk of each user’s session state resident in memory, and if that state is measured in gigabytes rather than kilobytes, the DRAM bill scales linearly with user count in a way that compute-only workloads usually do not. Below is a breakdown of the headline figures as reported.
| Metric | Reported Figure | Source of Estimate |
|---|---|---|
| Target user base used in calculation | 100 million users | Evercore analyst Mark Mahaney’s reported 6-12 month projection |
| CPUs required | 65,000 | Ivan Burazin (Daytona CEO) estimate, shared by Aaron Levie |
| DRAM required | 75 petabytes | Ivan Burazin estimate, shared by Aaron Levie |
| Estimated infrastructure cost | approximately $2.8 billion | Ivan Burazin estimate, shared by Aaron Levie |
| Facebook’s reported user base (comparison point) | approximately 3 billion | Widely cited figure used in the viral comparison framing |
The comparison to Facebook’s roughly 3-billion-user base is the part of the story that turned a niche infrastructure calculation into a mainstream talking point. If $2.8 billion buys compute for 100 million Muse users, and Meta’s own flagship social network already serves something close to 30 times that many people, the extrapolated cost of running an agentic AI layer across Meta’s entire user base would run into the tens of billions of dollars, assuming costs scaled linearly, which engineers are quick to point out they likely would not — economies of scale, shared infrastructure, and more efficient model architectures would all push the real number down. Even so, the directional signal is what has people talking: agentic AI, as currently architected, is dramatically more resource-hungry per user than the social feeds and messaging products that made Meta’s existing infrastructure economics work.
Why Agentic AI Breaks the Old Infrastructure Playbook
For more than a decade, the dominant cost driver in consumer internet infrastructure was storage and bandwidth — photos, video, and messages that could be cached, compressed, and served cheaply at the edge. Generative AI flipped that model once already, shifting the bottleneck to GPU-hours for inference. Agentic AI is now flipping it again, this time toward memory capacity and persistent compute state. An agent that needs to “remember” a user’s ongoing tasks, maintain tool-call history, and reason across multiple steps cannot simply discard its working state between requests the way a stateless web server does. That working state has to live somewhere fast enough to read and write cheaply, which in practice means DRAM rather than disk or even flash storage.
This is the same dynamic that has already driven coding-agent platforms toward heavier infrastructure footprints than their chatbot predecessors. Developer tools built around persistent agent sessions, such as Muse Code, Claude Code, and Cursor’s agent mode, have all had to grapple with the same tradeoff: richer, longer-running agent context improves task completion but multiplies the per-session memory footprint. What is different about Muse, as a consumer product aimed at a nine-figure user base rather than a professional developer audience, is that the same memory-hungry architecture now has to be justified against a far thinner average revenue per user. A $20-a-month coding subscription can absorb a heavier compute bill per seat; a free or low-cost consumer AI agent aimed at hundreds of millions of people cannot absorb the same per-user cost without a very different business model behind it.
Meta’s Broader AI Spending Context
The $2.8 billion Muse estimate does not exist in a vacuum. It lands in the middle of a year in which Amazon, Meta, Microsoft, and Google have collectively been reported pushing combined AI capital expenditures toward the $725 billion mark for 2026 alone, a spending pace that CNBC’s technology desk has covered extensively throughout the year. Meta’s own AI infrastructure build-out has been one of the more aggressive among the major platforms, with Mark Zuckerberg previously framing the company’s AI spending as a bet with no fallback plan. Viewed against that backdrop, a $2.8 billion figure for a single product line — even one aimed at 100 million users — is a comparatively small piece of a much larger capital commitment Meta has already made to AI infrastructure broadly, including data center buildouts, custom silicon, and the compute reserved for Meta’s other AI products under its wider enterprise AI platform push.
That context matters for how seriously the $2.8 billion figure should be taken as a constraint on Meta’s plans. A company already committing well over $100 billion a year to AI infrastructure is not necessarily deterred by a multibillion-dollar bill for one agentic product; it may simply be absorbed into existing data center capacity, existing custom chip investments, and existing DRAM supply contracts that Meta has presumably already negotiated at a different scale than a standalone estimate assumes. What the viral calculation does usefully highlight, however, is the per-user economics problem: even inside a company with deep enough pockets to fund the infrastructure, someone eventually has to answer whether each of those 100 million users generates enough revenue, engagement, or strategic value to justify their individual slice of a $2.8 billion bill.
How the Estimate Spread: From a Developer’s Post to a Trending Story
The path this figure took from a niche calculation to a top trending technology story is itself instructive about how AI infrastructure debates now move through the industry. Ivan Burazin, CEO of developer-infrastructure startup Daytona, produced the original back-of-envelope math. Aaron Levie, the Box CEO known for his frequent public commentary on enterprise AI economics, amplified it with his widely shared post crediting the calculation and summarizing it in a single, quotable line. From there, the figure jumped from developer and tech-executive social circles into financial media coverage tracked by outlets including Yahoo Finance, and into retail-investor-facing platforms like Stocktwits, where it became a talking point for people watching Meta’s stock in the context of its AI spending story.
That spread pattern — a technical estimate from a credible infrastructure voice, amplified by a well-known tech executive, picked up by financial media, and then absorbed into retail investor discourse — has become a recurring feature of how AI infrastructure skepticism reaches a mainstream audience in 2026. It mirrors how earlier capex-skepticism narratives around Zuckerberg’s $145 billion AI spending bet and the industry-wide capex figures gained traction: a specific, surprising number is easier to argue about on social media than an abstract concern about “AI spending being too high.”
Competitive Landscape: How Rivals’ Agent Products Compare
Meta is far from alone in pushing agentic AI products toward consumer scale, and the infrastructure math underlying Muse raises the same questions for its direct competitors. OpenAI, Google, and Microsoft have all shipped agent-style products in 2026 that carry similar architectural assumptions — persistent memory, multi-step task execution, and tool use — even if none of them has had a comparable viral cost estimate attached to their user base yet. The table below lays out what is publicly known about how these products position themselves, without assigning unverified cost figures to any of them.
| Company | Agent Product | Positioning | Infrastructure Cost Disclosure |
|---|---|---|---|
| Meta | Muse | Personal AI agent for consumer tasks | Not disclosed by Meta; $2.8B figure is an external estimate |
| OpenAI | Agent-based ChatGPT features and enterprise agent tools | Consumer and enterprise task automation | Not publicly itemized per product |
| Gemini-based agent features | Workplace and consumer agent tasks | Not publicly itemized per product | |
| Microsoft | Copilot agent capabilities | Enterprise-first agent deployment | Not publicly itemized per product |
What stands out across this landscape is that Microsoft and Google have both leaned more heavily toward enterprise-first agent deployment, where per-seat pricing is high enough to plausibly cover the heavier compute footprint that agentic systems demand. Meta’s Muse, by contrast, is positioned as a consumer product aimed at hundreds of millions of users, many of whom may never pay anything directly for it. That positioning choice is precisely what makes the Burazin-Levie calculation sting: it is one thing to spend billions on infrastructure for an enterprise tool with a $20-or-more monthly price tag per seat; it is another to spend billions on infrastructure for a consumer feature with no clear, comparably sized revenue stream attached to each user.
Historical Context: AI Infrastructure Costs Have a Pattern of Surprising Everyone
This is not the first time a specific infrastructure number has forced a public reckoning about AI’s economics. When large-language-model training costs first became public knowledge, the multi-hundred-million-dollar price tags for training frontier models shocked an industry accustomed to thinking of software as a near-zero marginal cost business. Inference costs told a similar story a year later, as companies realized that serving millions of chatbot queries daily could cost more than the training run itself over time. The Muse calculation represents a third wave of that same surprise, this time centered on the memory and state-retention costs of agentic systems rather than on raw model training or basic inference.
Each of these waves has followed a similar arc: an eye-popping number goes viral, the affected company either confirms, denies, or stays silent, and within a matter of months the industry quietly absorbs the cost into existing capital expenditure plans while engineering teams work to bring the per-unit cost down through better hardware utilization, model efficiency gains, or architectural changes. There is little reason to expect the Muse story to play out differently, especially given that Meta has not responded publicly to confirm or dispute the $2.8 billion figure at the time of this report.
Market Impact: What Investors and Analysts Are Watching
For Meta’s stock watchers, the Muse infrastructure story adds another data point to an already crowded debate about whether the company’s AI spending is being met with proportional returns. Evercore’s Mahaney has been cited as projecting Muse could reach 100 million users within six to twelve months, a bullish adoption timeline that would itself represent a significant achievement for a product still in its early rollout phases relative to Meta’s billions of existing Facebook, Instagram, and WhatsApp users. But rapid adoption cuts both ways in this story: faster growth toward 100 million users means the $2.8 billion infrastructure bill, if accurate, arrives sooner rather than later.
Financial commentary trackers including Stocktwits have flagged the story as relevant to Meta’s near-term narrative precisely because it offers a concrete, falsifiable claim rather than an abstract capex concern. Investors who have grown accustomed to Meta citing large AI infrastructure numbers in earnings calls now have a specific, independently calculated figure they can compare against whatever Meta eventually discloses about Muse’s actual operating costs, whether in a future earnings call, an investor day presentation, or a technical disclosure. That comparison point, more than the raw dollar figure itself, may end up being the most lasting consequence of this week’s viral calculation.
What Meta Has and Hasn’t Confirmed
It bears repeating because it is the single most important caveat in this entire story: Meta has not confirmed the $2.8 billion figure, the 65,000-CPU estimate, the 75-petabyte DRAM estimate, or a specific 100-million-user target for Muse. All of these numbers originate from an external calculation performed by Burazin and shared publicly by Levie, based on assumptions about how much memory and compute an agentic system like Muse would plausibly require per user. Meta has also not confirmed that Facebook’s user base sits precisely at 3 billion in current reporting, though that figure is widely used in industry discussion as a general benchmark for the platform’s scale.
This matters because the framing of “Meta needs $2.8 billion for 100 million users, but Facebook has 3 billion” is a rhetorical device built on an extrapolation, not a disclosed roadmap. Meta could serve 100 million Muse users at a very different cost than the estimate suggests, depending on model architecture choices, caching strategies, how aggressively it compresses or evicts idle session state, and how much of the compute is shared with other Meta AI products rather than dedicated solely to Muse. Readers should treat the headline number as a provocative estimate that has sparked a useful industry conversation, not as an audited figure from Meta’s own books.
The Engineering Questions This Raises for Agentic AI Broadly
Beyond the specific dollar figure, the Muse calculation has reignited a broader engineering conversation about whether current agentic AI architectures are fundamentally too memory-hungry to scale to billions of users without major redesigns. Several approaches are already being explored across the industry to address exactly this problem. Tiered memory systems that keep only the most active user sessions in expensive DRAM while evicting idle sessions to cheaper, slower storage are one path. More aggressive context compression, where an agent’s working memory is periodically summarized rather than retained in full, is another. A third approach, already visible in smaller open-weight agent models, is simply shrinking the per-session memory footprint through more efficient model design rather than relying on brute-force DRAM capacity.
These are the same tradeoffs that teams building coding agents and developer tools have been wrestling with for the past year, and the parallels to Meta’s own prior AI infrastructure disclosures are notable. Meta’s push toward custom AI infrastructure, including its data-center robotics and cabling automation efforts, suggests the company is already aware that traditional infrastructure scaling approaches will not cheaply support agentic products at billion-user scale without significant efficiency gains first.
Five Predictions for What Happens Next
- Meta will likely stay silent on the specific $2.8 billion figure in the near term, neither confirming nor denying it, consistent with the company’s past pattern of not itemizing per-product infrastructure costs in public disclosures.
- Expect rival labs at OpenAI, Google, and Microsoft to face similar viral cost estimates for their own agent products within the next two to three quarters, as the Burazin-style back-of-envelope calculation becomes a template other engineers apply to competing platforms.
- Memory-efficiency improvements, rather than raw DRAM capacity increases, will become the most-cited solution path in engineering discussions, echoing how inference-cost concerns a year earlier drove a wave of model-efficiency research.
- Analysts including those at Evercore will likely revise their 100-million-user timeline for Muse up or down depending on how Meta’s next earnings call frames AI product adoption metrics, which could shift the entire basis of the viral calculation.
- DRAM and memory suppliers may see renewed investor attention tied to agentic AI’s memory-heavy architecture, adding another demand driver on top of the AI training and inference-driven memory shortages already reported throughout 2026.
Why This Story Resonates Beyond Meta
The reason a single social media calculation turned into a cross-outlet trending story is that it crystallizes an anxiety that has been building across the entire AI industry for much of 2026: the suspicion that agentic AI, for all its promised utility, may carry a cost structure that does not scale the way consumer internet products historically have. Social networks, search engines, and messaging apps all benefited from the fact that serving the billionth user cost only marginally more than serving the millionth. The Muse calculation suggests agentic AI may not follow that same curve, at least not yet, and that uncomfortable possibility is what has kept engineers, investors, and journalists debating a single viral post for days.
Whether the $2.8 billion figure proves directionally accurate or ends up revised sharply in either direction once more information becomes available, the underlying question it raises is not going away: can any company, even one with Meta’s balance sheet, make the economics of agentic AI work at the scale of a mainstream consumer product, or will agentic AI remain a premium, enterprise-skewed category for the foreseeable future while consumer deployments stay limited to lighter-weight, less memory-intensive versions of the technology.
Frequently Asked Questions
What is Meta Muse?
Muse is described in reporting as Meta’s agentic AI platform, built as a personal AI agent capable of handling multi-step tasks rather than simple one-off chat responses.
Where does the $2.8 billion figure come from?
It originates from a calculation by Daytona CEO Ivan Burazin, shared publicly by Box CEO Aaron Levie, estimating that serving 100 million Muse users would require 65,000 CPUs and 75 petabytes of DRAM.
Has Meta confirmed this cost estimate?
No. The $2.8 billion, 65,000-CPU, and 75-petabyte-DRAM figures are an external estimate. Meta has not published an infrastructure budget specific to Muse or confirmed a 100-million-user target.
Does Facebook really have 3 billion users?
The figure is widely cited in industry discussion as a general benchmark for Facebook’s scale, but it was not independently verified in the reporting behind this story.
Why does an AI agent need so much DRAM?
Agentic systems maintain persistent session state, memory, and tool-call context across multi-step tasks, which requires far more active memory per user than a stateless chatbot query.
Is 100 million users a confirmed Muse target?
No. Evercore analyst Mark Mahaney has reportedly projected Muse could reach 100 million users within six to twelve months, but this is an analyst projection, not a confirmed Meta target.
How does this compare to other AI agent products?
OpenAI, Google, and Microsoft all have agent-style products, but none has had a comparable public cost estimate attached to a specific user target, and most have leaned toward enterprise pricing models that can better absorb heavier compute costs.
What happens if the estimate is accurate?
If directionally accurate, it would suggest agentic AI products carry a fundamentally different and more expensive cost structure per user than earlier generations of consumer internet and chatbot products, raising questions about how such products get monetized at scale.
