Mythos, Muse, and the Opportunity Cost of Compute | Stratechery by Ben Thompson

ELI5/TLDR

For twenty years, big tech worked because serving one more user cost effectively nothing. AI breaks that. The chips are so expensive and so scarce that every GPU Microsoft gives to Azure customers is a GPU it can’t give to Copilot, and vice versa. The new cost isn’t “how much to make one more unit” — it’s “what did I give up to make this one instead of that one.” Thompson walks through Anthropic’s new model Mythos, Meta’s new model Muse, and argues that Meta may quietly have the cleanest path to winning consumer AI because it has no enterprise business to starve.

The Full Story

The old rule, and why it broke

Thompson starts with Doug O’Laughlin’s argument from Fabricated Knowledge: aggregation theory — the framework that explained Google, Facebook, Amazon, the App Store — rested on something deeper than any one company. It rested on the fact that in digital businesses, the cost of serving one more user was effectively zero.

Economists call that marginal cost. Bake it into a factory and it’s obvious: land and machines are fixed, but raw material scales one-for-one with output. Software businesses mostly escape that. The output is digital. The electricity bill and the payroll are large but flat — they don’t really tick up with each new user.

AI looks like software on paper. The output is digital, the chips run at full tilt regardless. So marginal cost is still roughly zero.

“No one with AI chips is making marginal cost calculations in terms of utilizing them. They’re going to be used. Rather, the decision that matters is what they will be used for.”

That last sentence is the pivot. The constraint has moved from marginal cost to opportunity cost.

Opportunity cost, unpacked

Opportunity cost is the thing you give up by choosing. If you spend an hour watching a film, the cost isn’t the ticket — it’s the other hour you could have spent reading. For Microsoft, compute is finite this quarter. CFO Amy Hood essentially admitted on the earnings call that Azure missed its growth number because Microsoft fed its GPUs to Copilot and GitHub Copilot first, then its internal AI teams, then whatever was left went to external Azure customers.

“If I had taken the GPUs that just came online in Q1 and Q2 and allocated them all to Azure, the KPI would have been over 40.”

Microsoft chose its own products because their margins and lifetime value were higher. Every hyperscaler now runs a version of this triage: Amazon juggles retail, AWS, Anthropic, and OpenAI; Google juggles GCP, Anthropic, and its own consumer apps. There is simply not enough compute to serve everyone, so somebody is always getting bumped.

Mythos — and why Anthropic wants it scarce

Anthropic announced Mythos, its most advanced model, by waving the danger flag. Project Glasswing pitched it as a tool urgent enough to reshape cybersecurity because it can find software vulnerabilities faster than almost any human.

Thompson is mildly allergic to the marketing style — “disaster porn as marketing tool” — but concedes the underlying point: whether or not Mythos is actually dangerous, a future model will be. Fine. The more interesting question is why Anthropic is gatekeeping access.

Two answers, both about opportunity cost:

Compute triage. Anthropic is already starving. Users on X spent the weekend arguing Claude has been “dumbed down” — probably because Anthropic is rationing. Giving fixed-price subscribers access to a bigger, hungrier model would make things worse.
Pricing power against distillation. Chinese labs — DeepSeek, Moonshot, MiniMax — have been quietly training their own models on Claude’s outputs. Distillation is a legitimate technique; labs use it internally to make smaller versions of their own models. But when a competitor does it to you, they get your capabilities at a fraction of your training cost. If Anthropic can keep Mythos behind a small, expensive gate, fewer distillation attempts succeed, and open-source alternatives stay one step behind.

Thompson adds a subtle twist. Stopping distillation isn’t just about protecting margins. It’s about keeping rivals from being able to offer a “good enough” alternative that would otherwise free up compute Anthropic wants for itself.

Muse, and Meta’s weird advantage

Meta launched Muse Spark the same week — the first model out of its reorganized superintelligence lab. Not state-of-the-art, but in the game. More importantly, it exists, which vindicates Zuckerberg’s decision to tear Meta’s AI effort down and rebuild it.

Thompson’s most interesting argument is about Meta’s strategic position. Every other big player is torn between enterprise workloads (high margin, paying by the token) and consumer workloads (huge user base, harder to monetize). OpenAI is the sharpest version of this tension: ChatGPT is the most popular consumer AI product in the world, but enterprise customers using Codex for agent workflows pay real money per token, so OpenAI will keep tilting compute toward them.

Meta has no cloud business. No enterprise customers to disappoint. It already has the best consumer advertising machine ever built, so monetizing AI usage is a solved problem. That means Meta can pour its compute into consumer AI without giving anything up — the opportunity cost is near zero.

“Meta may actually face less competition in winning the consumer space than it might have seemed a few months ago, simply because that is their primary focus.”

Thompson’s advice: open-source Muse, the way Meta open-sourced Llama. It hurts rivals more than it hurts Meta, because it drags down the pricing power of frontier labs and eats into their compute budgets.

Demand versus supply

So is aggregation theory dead? Thompson isn’t ready to sign the obituary. Distribution and transaction costs are still free in AI — the two preconditions for an aggregator to exist. The winner in consumer AI should still be whoever builds the most compelling product, because that demand generates the revenue to buy the compute.

OpenAI is making the opposite bet: that locking up supply early gives it a durable lead over Anthropic. Anthropic has countered by snagging a chunk of Google’s TPU allocation — a reminder that compute can be bought if you have the cash flow.

“My bet is that owning demand will ultimately trump owning supply, suggesting that the underlying principles of aggregation theory lives on.”

Then he immediately hedges: if more compute is what makes products better, maybe supply is the game after all.

Key Takeaways

The real constraint in AI is no longer marginal cost (it’s still zero) but opportunity cost — every GPU deployed here can’t be deployed there.
Hyperscalers and frontier labs are all making the same triage: first-party products, then strategic partners, then external customers.
Anthropic has two reasons to gatekeep Mythos: protect its own compute and raise the cost of distillation for Chinese rivals.
Meta’s lack of a cloud or enterprise business is unexpectedly a feature — it can point all its compute at consumer AI without giving up higher-margin revenue.
Thompson’s bet: demand still beats supply in the long run, which means aggregation theory is bruised but alive.

Claude’s Take

Thompson is doing what he does best here — taking a piece of MBA vocabulary that most people skim past (opportunity cost) and using it to reframe an entire industry. The move is quiet, but it reorganizes a lot of noise. Why is Azure missing numbers? Opportunity cost. Why is Anthropic rationing Mythos? Opportunity cost. Why does Meta look oddly well-positioned despite being late? It has nothing to ration.

The weakest part of the piece is the hedge at the end. He spends 3000 words arguing demand will beat supply, then shrugs that maybe supply determines product quality after all. It’s an honest shrug — but it also means the central prediction is softer than it sounds. If OpenAI’s compute lead lets it ship better agents, then “owning demand” becomes circular.

The Meta argument is the one to watch. It’s underweighted in most AI commentary because Meta feels like the lumbering giant that missed the boat. Thompson’s point is that the boat hasn’t sailed on consumer AI — it’s still being built — and Meta is the only player who can steer without sacrificing something else.

Score: 8/10. High-density strategic analysis, one clean reframe, honest about what it doesn’t know.