Field report NW-68 / K3 · sheet /deals

M&A & DealsFiled from the territory

Anthropic's $1.4 Trillion IPO Rests on Customers Burning Millions on Tokens

Anthropic targets a $1.4 trillion IPO valuation as customers like Grafana Labs burn tens of millions a year on Claude tokens — some engineers outspend their own salaries.

Grid
NW-68 / K3
Surveyed
November 19, 2026
Reading
5 min read
Correspondent
Sophie Lindqvist
Meet 4 engineering leaders whose budgets explain how Anthropic has gotten so rich
Fig. NW-68 / K3Meet 4 engineering leaders whose budgets explain how Anthropic has gotten so rich — AI-generated

Field notes

  1. Anthropic targets a $1.4 trillion valuation in its IPO this week; 11,900 business customers spent $100,000+ last year, 250 spent $10 million+.
  2. Top 5% of Claude Code users average about $1,500 a month; non-Claude Code business customers average under $75.
  3. Grafana Labs' AI token bills run into the tens of millions a year, about two-thirds on Claude.
  4. Harness used 265 billion AI tokens in one week across 500+ engineers, with a soft $125-a-week per-engineer budget.
  5. Theory Ventures' Bryan Bischof hit six figures of personal token spend in one month, mostly on Claude.

Anthropic is targeting a $1.4 trillion valuation in its initial public offering this week, and the spending records of four engineering leaders explain how the company got there: business customers are pouring sums into Claude tokens that, in some cases, exceed an engineer's salary.

The numbers in Anthropic's IPO prospectus frame the stakes. Customers who did not use Claude Code (Anthropic's AI coding agent) spent less than $75 a month on average. The top 5% of Claude Code users averaged around $1,500 a month. The company counts 11,900 business customers that spent $100,000 or more over the previous year, and 250 that spent $10 million or more.

What does this change for the AI market's economics? Months after the "token-maxxing" craze — the push by companies to consume ever more tokens, the units of AI output — some buyers appear unbothered by the bills. Others expect engineers to occasionally spend more on tokens than they earn in wages. Anthropic's growth, in other words, rests on customers who treat AI spend as a bargain relative to labor. That claim deserves scrutiny, and the buyers themselves supply most of the evidence.

What does Grafana's tens of millions buy?

Grafana Labs, an observability startup last valued at $9 billion, spends at a rate of tens of millions of dollars a year on AI coding tools, chief technology officer Tom Wilkie said. About two-thirds goes to Claude; the rest mostly goes to OpenAI's Codex.

"We're not worried about the spend," Wilkie said. "We're not worried about controlling it, just yet."

The spending shows returns in specific cases. One team spent tens of thousands of dollars in tokens in a single month to improve the performance of a Grafana service, and the optimization cut six figures from the company's Amazon Web Services bill. A non-engineer routed customer feedback to Anthropic's Fable model, spent more than $30,000 on tokens, and landed a cluster of successful product changes.

Controls exist, but loosely. After a data analyst accidentally ran up tens of thousands of dollars by leaving a model running, Grafana now grants access to its most powerful models in 24-hour increments. Wilkie frames the guardrails as deliberately light: "We really want to be careful with the chilling effect of budgets and access control to models because you just don't know where the next big piece of innovation is going to come from."

Note what is not said: no aggregate figure for token spend against total savings. The anecdotes are strong; the accounting is absent.

How high can a single user's bill go?

Bryan Bischof, head of AI at Theory Ventures, a San Francisco venture firm, personally racked up six figures of token spend in one month, mostly on Claude. He said he could easily pass $1 million if he spent more time running simulations. His team of six engineers builds custom software for investors, talent leads and intelligence staff.

His team sometimes uses open-weight models (AI models whose parameters anyone can download and run), which cost far less than Anthropic's top-end options. Yet Bischof often prefers the premium tools. "If a very intelligent model is going to dramatically reduce the cycle time, then that's absolutely worth it," he said. The firm may also build its own always-running AI agents, similar to Meta's Muse and OpenAI's Dots.

The caveat is obvious: a $1 million ceiling described by the spender himself is a forecast, not a result.

Why do startups favor Claude over Codex?

At Menos AI, a startup building AI-powered research software for investors, cofounder and chief AI officer Chris Yang said his engineers use Claude Code or Claude models through the coding tool Cursor. Loyalty is a matter of capability, not procurement. "Because Claude has been evolving so, so well in terms of coding capabilities, I don't even have much opportunity or willingness to try out Codex," Yang said.

The company's investor software consumes around 90% of its token costs, routing user inquiries to different models depending on difficulty. Yang calls himself one of the firm's "hero users," spending $3,000 to $4,000 a month on Claude tokens while running several AI agents at once.

"We do spend a lot of expenses on the Claude Code use, but when we compare it to a human resources hire, it's still a big benefit," Yang said. "It's still much cheaper than human labor."

That comparison is the industry's core pitch — and its core unverified assumption. No wage figure accompanies the claim.

Where does the spending plateau?

Harness, a software development startup worth $5.5 billion with more than 500 engineers, consumed 265 billion AI tokens in a recent week, said Kiran Gangadharappa, senior director of software engineering. Claude Code and Cursor rank as the most popular tools among its staff.

"We definitely see it improving and taking a bigger and bigger chunk of the manual work of writing code," he said. The heaviest consumption comes from gigantic one-off tasks, such as rebuilding existing software architecture or conducting security research.

Unlike Grafana, Harness has tightened cost discipline. Moving work to Sonnet, one of Anthropic's cheaper models, cut spending, as did a soft budget of $125 a week per engineer. Gangadharappa expects AI expenses have reached a "steady" level. His justification is blunt: "Agents never get tired."

For Anthropic's prospective IPO investors, the pattern across these four buyers matters more than any single bill. Two companies impose no meaningful limits; two are already economizing. The prospectus numbers show a small cohort carrying enormous spend — and whether that cohort keeps paying, or starts behaving like Harness, will decide how durable the $1.4 trillion valuation proves.

via feeds.businessinsider.com (Source)

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Correspondent covering money, funding and market shifts around corporate strategy and chief-executive leadership at CEO Frontier.

479 articles

Filed under

Adjacent sheets

« Previous articleNext article »

End of sheetNW-68 / K3 · November 19, 2026