I was today days old when I learned Anthropic filed to go public today — the same day I'm writing this — with its IPO paperwork landing at the SEC and bankers reportedly targeting as much as $100 billion raised at a valuation between $1.8 and $2 trillion. That would make it one of the largest IPOs ever, bigger than SpaceX's $86 billion debut back in June. The ticker is expected to be “ANT” on the Nasdaq, with trading targeted to start the week of November 16. And buried inside that filing is the closest thing to a real, audited answer to a question I've had since I started paying for Claude Max: what does Anthropic actually spend to keep someone like me online?
Here's what the S-1 says, what my $200/month plan has to do with it, and why “what does my subscription actually cost them” turns out to be a live industry argument, not a settled number.
The Numbers From the S-1 Itself
Revenue for the first half of 2026 came in at $16.3 billion, up 1,368% year over year — the kind of growth curve that explains why bankers are chasing a $2 trillion number in the first place. The scarier headline is the net loss for the same six months: $80.7 billion. But that number is doing something easy to misread if you stop at the headline — about $72 billion of it is a non-cash accounting charge, a fair-value adjustment on convertible bonds, not cash that actually left the building. That distinction matters a lot for how worried the loss figure alone should make you.
The number worth actually anchoring on is compute. Infrastructure and compute cost Anthropic $8.9 billion in H1 2026, versus $2.3 billion in H1 2025 — nearly 4x more in raw dollars. But as a share of revenue, compute fell from 201% of revenue a year ago to 55% today, which is the real story: the company is still spending enormously on chips, but it's growing into that spending rather than drowning in it further. Separately, and more ominously, Anthropic disclosed roughly $417 billion in future compute commitments — contracts it owes providers down the road, a real risk factor the company flagged about itself, not something a critic invented. Funding to date sits at $132 billion across 19 rounds, including up to $20 billion-plus from Amazon and up to $40 billion from Google — both of whom are also the cloud and chip providers Anthropic buys that $8.9 billion of compute from, worth sitting with for a second: some of its biggest investors are also its biggest vendors. Source: Anthropic's S-1 registration filed with the SEC, as reported by Reuters and Bloomberg.
That same investor-is-also-a-vendor shape is the subject of a companion post I wrote about Anthropic's circular compute deal with Rumble and the dot-com-era pattern it echoes — worth reading alongside this one.
My $200/Month Plan Is the Ground-Level Version of the Same Story
I pay for Claude Max, the $200/month tier, mostly because I write and ship a lot of code and burn through the regular Pro plan's limits by lunchtime. Max 20x gives roughly 20 times Pro's usage inside each 5-hour session, plus a separate weekly cap on top of that. According to Anthropic's own published documentation, fewer than 5% of subscribers ever actually hit that weekly ceiling — meaning the plan is sized around people like me, who use a lot, not around the small sliver of users who try to max it out every single day. Source: Anthropic's Claude Max plan documentation.
So What Does It Actually Cost Anthropic to Serve Me?
This is the part I wanted a real number for, and it turns out there isn't one — not an audited, Anthropic-disclosed, per-subscriber figure. What exists instead is a live, public argument between two estimates. The scary version: at official API list pricing (Opus 4.5 runs $5 per million input tokens and $25 per million output tokens), a subscriber who fully maxed out their Claude Max allowance every month would be consuming something like $8,000 worth of tokens at retail price. The rebuttal: list price isn't real compute cost — it's a retail markup on top of actual infrastructure spend, the same way a restaurant menu price isn't the ingredient cost. One widely-cited compute-cost analysis estimates real-world inference cost runs around 10% of list price, which would put the true cost of a fully maxed-out Max subscriber closer to $500–$800 a month — meaning Anthropic is likely losing money, by that estimate around $300 a month, only on the most extreme power users, not on a typical $200/month subscriber who never gets near the weekly cap. Sources: Anthropic's published API pricing at anthropic.com/pricing; third-party compute-cost estimate methodology via SemiAnalysis.
I want to be honest about what that means for this post: there is no verified number for what I, specifically, cost Anthropic to serve. Both the $8,000 panic number and the $500–800 rebuttal number are estimates built on public list pricing and third-party compute-cost modeling, not something Anthropic has confirmed about any real subscriber. Anyone telling you a precise per-user profit or loss figure with total confidence is guessing, same as I am here — just with better-labeled math.
Why This Matters When You're Reading an S-1
Here's why this is more than a fun math exercise: it's the exact question investors reading this same S-1 are going to be asking about Anthropic's entire subscriber base, not just me. A company growing revenue 1,368% year over year while compute costs fall from 201% to 55% of revenue is a business getting more efficient at serving its users — but “more efficient” and “profitable per $200/month subscriber” are two different claims, and the filing answers the first one with real numbers while leaving the second one to estimates like mine. That gap is worth remembering the next time a headline about a $2 trillion AI company and a $200 subscription plan shows up in the same sentence.