An Tran Solutions
VI
An Tran Solutions
Back to Blog
newsAI infrastructureAI costsAnthropic

Investors Just Asked Anthropic a Question Usually Reserved for Power Plants

September 8, 20266 min readby An Tran
On this page

For three years, the tech industry has taught you one way to read AI news: look at the leaderboard. Which model tops the benchmark, which one is smarter, who just beat whom by a few points. This week is no different. GPT-6 Astra, Claude Fable 5.1, intelligence scores dancing across the headlines.

But the people putting the most money into this industry have stopped asking about intelligence. They're asking something else entirely. Something that would make you do a double take if you heard it in a software company's board room.

They're asking: how much revenue does a gigawatt of electricity produce?

What happened on September 4

Anthropic is preparing for what is expected to be the largest share offering ever. According to Reuters on August 15, 2026, investment banks are discussing a listing valuation of up to roughly $2 trillion, a figure built on a revenue projection of $190–200 billion in 2028.

On growth alone, there is little to criticize. Anthropic's annualized revenue went from about $9 billion at the end of 2025 to $47 billion in May 2026, and according to Axios on August 17 it passed $65 billion in July. In Q2 2026, the company reported its first quarter of operating profit, roughly $559 million.

Then came September 4. The Information reported that prospective investors were asking Anthropic for far more granular figures than a standard financial statement provides. Two metrics were named specifically:

  • Revenue per token
  • Revenue per gigawatt of compute

Read that again. This is what's being asked of the fastest-growing software company in history.

Why "gigawatt" in an IPO process is a big signal

The entire valuation magic of the software industry rests on one property: near-zero marginal cost.

You write the software once. Your first customer costs millions to serve. Your millionth customer costs almost nothing. That's why a SaaS company can be valued at 40 times revenue without anyone blinking: nearly every additional dollar of revenue drops straight to the bottom line.

AI doesn't work that way. Every answer burns real electricity. The millionth customer costs exactly as much as the first, plus whatever energy prices have done in the meantime. Marginal cost isn't zero. It's measured on a power meter.

When investors ask for "revenue per gigawatt," they are saying something out loud: we no longer value this as pure software. That is a factory metric: output per unit of installed capacity. It's how you look at a power station, a refinery, heavy infrastructure.

Reuters also spells out the corresponding risk: this valuation applies a revenue multiple to projections two years out, and it falls apart if GPU and training costs don't fall as expected. For comparison, Palantir trades at around 53 times its expected 2026 revenue; Cloudflare and SpaceX at around 41.6 times.

The key point: the market isn't doubting whether AI generates revenue. $65 billion answered that. The market is starting to question the cost structure behind that revenue. And that is a completely different question.

The evidence: Anthropic has been measuring itself in gigawatts all along

The interesting part is that investors didn't invent this unit. They're simply borrowing the language the labs themselves were already using.

Look at how Anthropic announces its commitments. None of them is described in "number of models" or "number of parameters." All of them are measured in power:

  • October 23, 2025: Anthropic expanded its partnership with Google Cloud, gaining access to up to one million TPU chips and, per the official announcement, "well over a gigawatt of capacity" coming online in 2026. Deal value: tens of billions of dollars.
  • November 12, 2025: Anthropic announced a $50 billion plan to build data centers in the US, starting in Texas and New York. The reason it gave for choosing partner Fluidstack: its ability to deliver gigawatts of power quickly.
  • 2026: a further expansion with Google and Broadcom for multiple gigawatts of next-generation compute.

A software company announces how many customers it has. Anthropic announces how many megawatts it has. The language shifted long ago. It just took until September 4 for investors to formally demand answers in that same language.

The quiet winner doesn't sell intelligence. It sells the ability to deploy power.

If you need one more piece of evidence that value is shifting, look at a company almost nobody outside the industry has heard of.

Fluidstack was founded in 2017 in Oxford. It doesn't train models, has no chatbot, and has no benchmark to show off. Its job is building and running data centers, faster than anyone else.

In December 2025 the company was valued at $7.5 billion. According to TechCrunch, by April 2026 it was negotiating a round at $18 billion. And at the start of this September, that round closed at $1.5 billion led by Jane Street, holding the $18 billion valuation. Forbes on September 3 called it the "tiny startup helping Google take on Nvidia."

Its value more than doubled in nine months. Not because it invented something smarter. Because it can plug things in faster.

The most telling detail: Jane Street, one of the largest buyers of compute on the market, didn't go shopping for compute this time. It bought equity in the company that builds compute. When the heaviest user decides to become an owner, that usually signals it believes the thing will stay scarce for a long time.

That same week, Nvidia confirmed it would buy Hugging Face for $12.9 billion. The money is still flowing toward infrastructure, not applications.

Where I used to be wrong

I need to own something I got wrong for the past two years, because it bears directly on advice I used to give.

I used to argue that AI costs would fall so fast they weren't worth factoring into a business case. On unit price, that's still true: the price per million tokens has dropped by orders of magnitude.

But I ignored the other half: a lower unit price doesn't mean a lower bill. When each call gets cheaper, people make far more of them. Multi-step reasoning models consume many times more than a single-turn model. Autonomous agents run in the background all day. The result is that total spend goes up even as the unit price goes down.

That's exactly why investors are asking for "per gigawatt" instead of "per token." They learned this lesson before the rest of us.

What this means for a business like yours

You're not listing on Nasdaq. But if you're buying, or planning to buy, an AI feature for your website, sales process, or customer support, the logic above applies to you directly.

First: don't plan on the assumption that AI will be free. Many "AI integration" proposals today leave operating costs completely blank, as if the thing runs itself once it's built. It doesn't. Every customer chat is an expense. Ask your vendor to state the estimated cost per 1,000 interactions, and how that cost changes if traffic grows tenfold.

Second: calculate your own "revenue per token." This metric isn't just for the labs. For each AI feature, do one simple division:

(Revenue this feature actually moves each month) ÷ (Monthly model-call cost)

If you can't answer the numerator, you don't have a business feature. You have a showpiece. If the ratio is under 5x, that feature is fragile in the face of any price increase.

Third: put AI where the money is. Because marginal cost is real, you shouldn't spread AI evenly across your website. Concentrate it on the touchpoints where a better answer turns into an order: product-selection advice, handling objections when a visitor is about to leave, qualifying and routing leads. A greeting chatbot in the footer burns power steadily and converts nothing.

Fourth: design so you can swap providers. While the whole industry's cost structure is being repriced, locking yourself into a single provider is an unnecessary risk. The architecture should separate business logic from model calls, so that changing models is a configuration change, not a rewrite.

That's how I approach AI integration for websites and workflows: operating costs estimated in the proposal itself, AI placed at the step that generates revenue, and success measured in business outcomes rather than how impressive the demo looks. If you're unsure where to start, a single conversation will be faster than trial and error.

On September 4, while the world debated which model was a few points smarter, the people about to put hundreds of billions of dollars into this industry quietly asked an electrician's question. Intelligence is converging: labs ship comparable models within days of each other. What's diverging is cost. That is the variable to watch, and it's the one that decides whether your AI feature survives next year.

Sources

Related articles