Meta’s $700B AI Bet Gets Its First Proper Sales Test With Muse Code
Meta has committed nearly $700 billion to the AI arms race. If Muse Code cannot win real engineering work from Claude Code and Codex, that spending starts to look less like ambition and more like a very expensive hobby.
Meta has committed nearly $700 billion to future AI-related spending. That is not a strategy. That is a bill screaming for a business model.
Mark Zuckerberg’s newest answer is Muse Code: Meta’s beta AI coding agent, released on August 5, built to plan changes, write code and validate results across big software repositories. It is aimed squarely at OpenAI’s Codex and Anthropic’s Claude Code.
Good. Because Meta does not need another clever model demo. It needs customers willing to hand it money.
The core story: $696.3 billion needs an answer
Meta disclosed almost $700 billion in future spending commitments tied to AI data centres, cloud computing and associated infrastructure. The number comes in two ugly chunks: $349.3 billion in non-cancellable contractual commitments, mostly for third-party cloud arrangements, servers and network equipment; plus $347 billion in leases that have not begun.
Add them up and you get $696.3 billion.
That is the number people should be staring at when they look at Muse Code.
Until recently, Meta’s AI story was fairly easy to understand. Better models help it show better ads, keep people engaged longer and improve recommendations across Facebook, Instagram and WhatsApp. Those are excellent businesses. But they do not automatically justify an infrastructure tab this enormous.
Muse Code is different. It is a direct shot at paid enterprise software.
Meta says the agent can tackle complete engineering jobs across large codebases. Rather than simply suggesting the next line in an editor, it can break larger work into sub-agents operating in parallel in isolated worktrees. Zuckerberg said Meta tested it by having it build six game features at once without collisions.
That is a much more valuable proposition than autocomplete. If it works reliably, a development team can turn a backlog item into a reviewed pull request faster, with fewer hands involved in the boring middle.
And boring middle is where businesses spend a fortune.
Why coding agents are the first AI product that can really charge rent
The AI industry has spent years selling possibility. Coding agents sell a much cleaner proposition: reduce the time between a product decision and a working release.
That is why Anthropic’s Claude Code became strategically important, why OpenAI kept adding serious capability to Codex, and why Meta has now piled in. This is not about making programmers redundant by Thursday. It is about capturing the budget currently spent on software engineering, testing, maintenance, internal tools and the mountains of fiddly work around all of it.
The useful question is not, “Can an AI write a login page?” Any half-decent model can have a crack at that.
The useful question is whether it can understand a messy, old, commercially important repository; follow the company’s conventions; inspect dependencies; make a change without breaking three unrelated systems; test its own work; and leave enough of an audit trail that an engineer will sign off on it.
That is the game Muse Code has entered.
Meta is trying to make cost its wedge. Alexandr Wang, who runs Meta Superintelligence Labs, has positioned Muse Code as a strong option particularly on price. That matters because coding agents are hungry beasts. The more capable they become, the more they inspect, reason, call tools, run tests and make repeated attempts. Usage can become expensive quickly.
A cheaper capable agent does not merely save a few bucks. It changes who gets to use it. It moves from the innovation team’s experiment to something a 100-person engineering group can deploy without the CFO developing a rash.
Meta has an advantage — and an awkward handicap
Let’s not pretend Meta is a scrappy underdog here. It has money, computing capacity, distribution and a giant engineering organisation that can punish-test products internally.
It also has a problem that OpenAI and Anthropic do not carry in quite the same way: buyers do not naturally think of Meta as their trusted enterprise software supplier.
Meta owns attention. It owns social distribution. It is formidable in advertising technology. But the chief information officer of a bank, insurer, hospital network or industrial business does not wake up hoping to put more Meta software inside the firm.
That trust gap is not fatal. Microsoft crossed versions of it over decades. Amazon went from bookshop to cloud backbone. But it is real.
This is why product quality matters more than the launch video. For Muse Code to become a proper business, it needs to be dependable where companies actually bleed money: legacy code, permission controls, security review, compliance, integration with existing developer workflows and predictable costs.
It also needs a clean answer to the question every sensible security leader will ask: what exactly can this thing touch?
That question became sharper in August, when Meta said one of its AI models hacked another company during cybersecurity testing. Reuters reported that the incident added to concerns over how developers contain increasingly capable systems, following similar disclosures involving Anthropic and OpenAI.
The lesson is not “avoid AI agents.” That ship has sailed.
The lesson is that giving an agent the power to inspect, edit, execute and browse is not a normal software procurement decision. It is closer to hiring a very fast junior employee who can work all night, has access to your tools and occasionally does something startlingly literal.
The overlooked angle: Meta is not late if it wins on economics
Everyone loves declaring winners early in technology. Usually because it saves them the trouble of thinking.
Anthropic and OpenAI got real momentum in coding. That is undeniable. But early leadership in a new software category does not guarantee the economics of the category belong to the pioneer.
Meta does not need to be loved by developers on day one. It needs to offer a better value equation for enough workloads.
Think about what happened in cloud computing. The first question was capability. The second was reliability. The third, and eventually the biggest, was cost and control. Once a product becomes infrastructure, buyers become brutally practical.
Coding agents will head the same way.
If Muse Code performs close enough to the leaders for routine engineering tasks, but costs materially less, a lot of rational companies will test it. Not because they have suddenly become Meta fans. Because engineering budgets are real money and software leaders are measured on delivery speed.
That could make the market much more competitive than the current chatter suggests. The winners may not be the labs with the flashiest benchmark chart. They may be the ones that give operators predictable output, permissioning, logs, integrations and a bill that does not turn into a hostage negotiation.
There is another twist. Meta’s $696.3 billion commitment means it has more urgency than most. Urgency can create stupid decisions. It can also create sharper products.
The market should hope it creates the second one. Competition between three heavily funded labs is far better for customers than being locked into whichever AI provider got there first.
What this means for you
If you run a company, do not buy an AI coding agent because a founder on X says it is magic. Run a controlled trial starting this week.
Pick one ugly but contained workflow: fixing low-risk bugs, writing internal admin tools, improving test coverage or clearing a known backlog of small product requests. Do not start with payments, customer identity, production infrastructure or anything that can ruin your month.
Then measure four things for 30 days:
1. Cycle time: Did work get from ticket to accepted pull request faster? 2. Human review load: Did your senior engineers spend less time doing drudge work, or more time cleaning up agent nonsense? 3. Defect rate: What escaped into production, and what was caught in review? 4. Fully loaded cost: Include model usage, tooling, security review and the engineer time required to supervise it.
Do this across at least two tools where practical. Do not assume the brand with the best reputation is automatically best for your codebase or your budget.
Most importantly, keep humans accountable for merge decisions. An AI agent can draft, test, investigate and propose. It should not get to quietly redefine your operating risk because someone wanted to save a salary.
For founders and investors, the bigger takeaway is simpler: stop treating AI spend as proof of AI value. Meta’s nearly $700 billion in commitments is only impressive if it produces durable revenue, better margins or a defensible product position.
Muse Code is not the whole answer. But it is the right sort of question: can Meta turn colossal infrastructure expenditure into software that customers choose, use and pay for?
That is the test now. Everything else is theatre.