Anthropic is a Power Plant
Anthropic is not a software company.
It’s not SaaS anymore. It’s a power plant.
And once you see that, everything else about the AI industry makes sense.
The Thing Nobody’s Saying Out Loud
Software has this ridiculous advantage: you add users, the cost doesn’t move. Add another Gmail user? Cost to serve approaches zero. Add one more Stripe transaction? Basically free. That’s the whole advantage of software, marginal costs collapse to nothing.
AI inference is the exact opposite.
Every Claude API call costs real electricity. Not metaphorically. Literal megawatt-hours. The model’s already trained. The hardware’s already there. But every token you generate pulls power from the grid. That’s a variable cost. That’s industrial.
Think about it: Anthropic probably processed hundreds of billions of tokens last year. Each one needs electricity. Running 24/7. Getting more expensive as they add load.
That’s a utility problem.
Why Everyone’s Wrong About the Competition
Anthropic has Claude 4.7. OpenAI has GPT-5. Google has Gemini. All the models are hitting “good enough” at this point. Nobody’s winning or losing because of model quality anymore. It’s table stakes.
The question is: who can actually serve inference at scale without the grid melting?
That’s a completely different game.
Here’s what’s happening with SaaS growth: you want to scale 10x? Spin up more servers, distribute the load, add instances in the cloud. Cool, you’re 10x bigger. Capital-light. Maybe some infrastructure costs, but nothing serious.
Inference scaling is brutal:
Want to do 10x more tokens? You need 10x more power. Want to serve both coasts? You need power contracts in both regions. Want faster inference? Newer chips, more power, more capacity.
That’s not capital-light. That’s a data center business. That’s utility-scale capital spending.
And the grid doesn’t just expand because you want it to. The US electrical grid has regional constraints. You can’t run 500 megawatts of AI inference in San Francisco. The grid doesn’t have headroom. You go to Texas. Or Oregon. Or somewhere with spare capacity. And guess what? AWS and Microsoft already negotiated the power contracts there.
You’re fighting over scraps.
Where the $7 Billion Actually Went
Anthropic raised $7 billion. Everyone assumes it’s for training.
It’s mostly not. Claude’s already trained. That money is going to inference infrastructure. Data centers. Power contracts. Grid interconnections. Geography expansion.
This is a $7 billion infrastructure play, not a software play.
Which means here’s how their growth actually works:
You can’t grow revenue faster than you can add power. That’s the constraint now. Not demand. Not product. Power.
If it takes 18 months to lock in new power capacity (which it often does), your revenue growth is capped at whatever infrastructure you deployed 18 months ago. That’s not a feature constraint. That’s physics.
This is why Sam Altman keeps talking about energy partnerships. Not because it’s trendy. Because they’re desperate. They have demand they can’t fulfill because the grid can’t handle it.
The Part That’s Actually Unfair
AWS, Microsoft, Google. They’ve been doing this for 20 years.
They already own long-term electricity contracts in all the cheap-power regions. They’ve built redundant grid connections. They have relationships with utilities that Anthropic doesn’t even know exist.
When Anthropic tries to build new capacity, it’s competing for the same power that AWS is already consuming. AWS has way more leverage.
This is the real asymmetry. Power relationships.
Microsoft’s building its own AI chips not because it wants to own AI, it doesn’t. It wants to control how much power each inference costs. Google’s doing the same. They’re not trying to win. They’re trying to not run out of power.
OpenAI’s in the same boat as Anthropic. Great model. Huge demand. Can’t scale because the grid isn’t there.
Geography’s Back, Which Sucks
For 20 years, cloud computing meant location didn’t matter. Build anywhere. Data goes everywhere. Servers are fungible.
AI inference flips that. You build where power is cheap and plentiful. Texas? Wind capacity, deregulated grid, open headroom. Oregon? Hydro. Virginia? Existing data center infrastructure and reasonable grid space.
You don’t pick those places because the talent is there. You pick them because the power is there.
Anthropic can’t scale inference evenly. It scales where power exists. That’s the constraint.
Geography’s back. And if you wanted to stay in California, sorry.
What Investors Should Actually Be Looking At
Everyone asks about ARPU. DAU growth. LTV. CAC.
Stop. Wrong model.
Question: how much power can they procure per year, and at what cost?
If Anthropic signs a 500 MW contract over three years, and each MW supports $X in revenue, that’s your growth ceiling. Demand is irrelevant if you can’t fulfill it.
Revenue forecasts for AI companies should be power forecasts. How much new capacity can they lock in? How fast can they deploy it? How much does it cost?
What This Actually Means
So here’s the thing: Anthropic isn’t losing at AI. It’s losing at the power game.
Claude works. The model’s great. But AWS, Microsoft, and Google already own the power infrastructure. They have decades of relationships. Better rates. Faster moves.
Anthropic has money and a good model. In a power game, that’s not enough.
The “AI race” is actually a grid modernization race. Whoever can secure the most reliable, cheapest power and convert it into inference wins. Training a smart model is commodity now. Controlling the megawatts is what matters.
Cloud providers have a decade-long head start.
Also: if you’re building something on the Claude API, their constraints are your constraints. If they hit power limits, you hit rate limits. If they can’t scale, you can’t scale. You’re downstream of their infrastructure problem.
The Full-Stack Capitalist Take
Anthropic isn’t an AI company. It’s an energy company that runs models.
The market prices it like software. Growth multiples. TAM expansion. SaaS curves.
Wrong. Price it like a utility. Megawatt capacity. Cost per MW. Revenue per megawatt. How fast can they lock in power?
Those are the metrics.
For Anthropic: being smart about AI isn’t enough anymore. You need to get sophisticated about utilities, grid operations, long-term power procurement. That’s a completely different skill set. That’s what wins.
For the AI market overall: the race is over. AWS, Microsoft, Google won it. Not because they built better models. They didn’t. They control the infrastructure and the power.
For founders: understand that your supplier (Anthropic, OpenAI, whoever) is constrained by physics, not demand. Your API limits reflect power constraints, not adoption. Plan around that.
The AI revolution is being constrained by a boring commodity - electricity - that’s increasingly scarce and geopolitically competitive.
And if you don’t control the grid, you lose.