What about costs?

“It ain’t that kind of movie, kid”

They make it up on volume.

See 6.

Here's what I found most interesting in the Customer Stories:

> 4.3 Non-billable hours recaptured per lawyer weekly, in an independent study of 31 firms

-Legora

> 24→5 days Working days for new feature time to market

-Rakuten

> 2x Quarter over quarter productivity gains (estimated 8.1 million hours annualized)

-Salesforce

The rest are pretty fuzzy/fungible (percentage of employees "using" Claude etc.). But these claims potentially show real, measurable productivity value (assuming they're true and measured accurately).

> (assuming they're true and measured accurately).

Precisely, because TTM doesn't tell the full story about the potential shifts in quality of the work. I don't refer to vibeslop here, to be precise, just the additional followup work required on those features released faster, which always happens but with AI it may or not happen more often.

TTM for sure is a bad indicator. Earlier this year at $DAYJOB, they attempted to switch all of the devs to a fully LLM based workflow: LLM writes the ticket, gets stakeholder input to refine the ticket, reads the ticket, does the work, submits the PR, etc. Their case study was a single project that did that process soup to nuts. The estimated time was a couple months, but they ended up going about, and I quote, "2x-3x" faster. So obviously they expected everyone to go faster.

The problem is that the way it was described to the other devs, they were able to sit all of the stakeholders down to really flesh out the idea and produce a working and detailed description of what needs to be done. A process that the company struggles to get done in general. I'm also convinced myself that they did not include the time to get the specifications written, a time we typically include in our TTM.

So of course the example's TTM was shorter, but that's always true when you start with detailed specifications. I can't speak to quality of the feature, since they hadn't even finished rolling it out to customers by the time they told us that's how we're operating. With promises of 2x-3x more features. And we continue to struggle to get stakeholders to nail down specifications and struggle to get features complete as a result. In fact, now it seems even worse than ever.

If the cost/time to code is compressed to near zero, the rest of the SDLC comes under pressure.

My vantage point as a technical seller at the margins of this space: Outside of software development, the productivity argument is a challenge to CFOs / Finance / Procurement. The ship has sailed when it comes to friction free AI purchases. Buyers need to articulate why productivity will adjust the trajectory of top line or bottom line metrics for the business.

I presume the productivity gains was coming to the realisation that a lot of staff at said (any!) company are running around looking busy rather than actually doing anything useful.

> 2x Quarter over quarter productivity gains (estimated 8.1 million hours annualized)

This is a great encapsulation of why productivity gains are an imperfect measure. 2X is a massive improvement. Yet IMO you can barely tell a difference as a user of the various SFDC offerings (Salesforce, Slack, Tableau, etc). So where is all that work going? A generous interpretation is that it is in the form of small, somewhat invisible improvements that add up over time, but that seems unlikely give then nature of what people prioritize at work...

Are there people still doubting the productivity gains of AI? We're way beyond that point.

> Customers under $100K: ~40 million

All claude code subscribers are classified in this category it seems.

Total customers >1M trailing twelve month revenue: ~2150. I suspect a vast majority of their revenue comes from limited list of customers with >100M revenue.

So there are 10x more people who see more value in their netflix subscription than in a $20 claude sub, interesting

ChatGPT was first mover and has the non-engineering AI chat market. My teen kids say ‘Chat’ to refer to all of AI, and they mean ChatGPT. The rest might as well not exist to them.

I think we're all in a bit of an echo chamber. Our job is to keep up with this stuff, and we work with people who are ALSO keeping up with this stuff.

To my friends (mechanics, CPAs, musicians, police officers), girlfriend (RN), and other people in my personal life... it's ChatGPT. (It's a neat tool to generate funny pictures or ask a question).

Now that I think about it, I have NEVER heard anyone mention Anthropic / Claude outside of work.

Maybe not interesting? I wouldn't expect the number of subscribers to a bespoke "AI Service" to rival Netflix. I realize on this forum we're probably getting lots of use out of them in our personal lives that justifies the $20/mo but we're also the people who pay for search engines and VPSs. Outside of work and school normal people don't do anything that requires general purpose AI.

Myself included. Why get a Claude subscription if I can't use pi with it?

For my private use a free tier is enough. There is no point to pay $20 for Claude subscription.

Obviously? More people want to watch TV than people who want to build things with AI.

Didn't the consensus expect an initial market cap considerably higher than that?

Basic market cap is based on 13 billion shares, but the fully diluted market cap is $1.4T because there are another 5 billion shares on top of that.

A flat line from their last raise 6 months back. Are they panicking today for some reason?

  Anthropic Valuation ($B, Basic)
  ------------------------------------------
  Ser E Mar '25 | #                   $61.5B
  Ser F Sep '25 | ####               $183.0B
  Ser G Feb '26 | ########           $380.0B
  Ser H May '26 | ################## $965.0B
  IPO   Nov '26 | ################## $968.7B