Anthropic profit forces OpenAI price cuts

**SPEAKER_1** (0:00)
Anthropic's annualized revenue exploded from roughly $1 billion to $47 billion in a highly compressed operational window. And in response, OpenAI is actually considering drastic cuts to its token pricing just to win back enterprise customers.

**SPEAKER_2** (0:16)
Which is, it's a weird paradox, right? The entire hardware and software stack for AI is basically treated as this giant furnace for burning venture capital, yet Anthropic just posted a projected $500,000,000 operating profit. And they did that while simultaneously paying SpaceX over $1,000,000,000 per billing cycle just for computing power.

**SPEAKER_1** (0:37)
Right. So that brings up the defining question here. How did a company that was supposed to lose money indefinitely suddenly turn a profit?
And why is it forcing the most recognized name in AI into a defensive pricing war right before they both try to go for an IPO?

**SPEAKER_2** (0:51)
Yeah, the profit part definitely caught everyone off guard. I mean, the consensus economic model for frontier AI basically assumed variable costs would just stay chained to revenue forever.

**SPEAKER_1** (1:00)
Right, like the more you sell, the more you burn.

**SPEAKER_2** (1:02)
Exactly. If you're paying a company to generate text or code for you, the assumption was always that the company was paying something very close to that amount just to keep the servers humming in the background. The margins were just this theoretical concept for some future date when hardware magically got cheap.

**SPEAKER_1** (1:20)
But we need to look at what actually changed under the hood. So, Anthropic more than doubled its revenue in a single financial reporting period, hitting $10.9 billion.

**SPEAKER_2** (1:30)
Yeah.

**SPEAKER_1** (1:31)
But obviously, revenue alone does not equal profit. The metric that flipped the entire equation for them is the compute ratio.
They dropped their compute spending from $0.71 per revenue dollar down to $0.56.

**SPEAKER_2** (1:44)
Which, I mean, a $0.15 reduction sounds really small until you apply it to $47 billion of annualized revenue.

**SPEAKER_1** (1:50)
Yeah, the math gets crazy fast.

**SPEAKER_2** (1:51)
It does. When you're operating at that level, dropping your compute ratio by that much alters the fundamental physics of your business. It means their inference gross margin surged from 38% to over 70%.
And just to clarify for you listening, inference is the cost of actually running the model to answer your prompts, as opposed to the billions they spend training the model in the first place.

**SPEAKER_1** (2:14)
But you don't just find an extra 30% margin hiding in the couch cushions. It really comes down to the unit economics of token production.
Because a token is just a fragment of a word or a piece of code, the basic unit these models use to read and write. Anthropic found a way to make generating those individual tokens significantly cheaper. And the primary mechanism driving this is prompt caching.

**SPEAKER_2** (2:38)
Caching is essentially the system deciding to remember something rather than reading it from scratch every single time. If you feed a model a massive instruction manual or an entire proprietary code base, it usually has to process the entire document from start to finish every single time you ask it a new question about that same document.

**SPEAKER_1** (2:55)
Which is incredibly inefficient.

**SPEAKER_2** (2:57)
Highly inefficient. The compute required to do that attention mapping over and over again is just astronomical.

**SPEAKER_1** (3:03)
So prompt caching lets the system hold that massive document in active memory. When you reuse that context, it costs Anthropic a fraction of the computational power to process your request.

**SPEAKER_2** (3:15)
Yeah.

**SPEAKER_1** (3:16)
They pass some of those savings on to the developer, obviously, but they keep enough of it to heavily pad their margins.

**SPEAKER_2** (3:22)
Yeah. The other factor padding those margins is that users are actively migrating to their premium most expensive model tier. So they're generating significantly more revenue per request anyway.

**SPEAKER_1** (3:33)
Though generating those requests requires physical infrastructure on a scale that is honestly hard to comprehend.
We are talking about Anthropic's deal with SpaceX to secure access to over 220,000 NVIDIA GPUs at the Colossus One Facility. The cost for that access is $1.25 billion per billing cycle.

**SPEAKER_2** (3:52)
Which is just a wild number. A single data center requiring that level of capital expenditure totally alters how we think about corporate infrastructure.
They are locking down 300 megawatts of computational capacity.

**SPEAKER_1** (4:06)
Just for one facility?

**SPEAKER_2** (4:07)
Yeah. That is the kind of continuous power draw you usually associate with a medium-sized city.
They're routing all of it into rows of servers just to keep these specific models available.

**SPEAKER_1** (4:18)
I do have to push back on calling this a sustainable profit though.

**SPEAKER_2** (4:21)