Anthropic files for a Nasdaq listing: $11.6B quarter, a first operating profit, and $782B of compute leases

The filing settles two arguments at once. Enterprise willingness to pay is no longer in doubt: revenue reached $11.6B in the quarter, nearly 15 times the year-earlier figure, existing customers spent about nine times more than a year before, and revenue per employee hit $6.6M. But the unit economics are industrial, not SaaS. The same document discloses $782B of long-term lease obligations for compute Anthropic does not own, against just $482M of its own capex in the first half of 2026. For UAE and GCC organisations the useful reading is not the listing. It is the shape of the commitments your AI vendors now carry, because those commitments set their pricing, their minimum terms and how much usage risk they push onto you.

One concrete example of how that funding is structured: Broadcom arranged a $63B vehicle with Bank of America, Blackstone, Citi and Morgan Stanley (https://investors.broadcom.com/news-releases/news-release-details/bank-america-blackstone-citi-and-morgan-stanley-announce) to finance Anthropic's chip purchases. The pattern matters more than the number. Compute gets built by partners, financed by third parties, and leased back to the model vendor, which keeps the vendor's own capital spending low and its fixed obligations high for years. A second example is the power line in the filing: Anthropic acknowledges a growing reliance on natural gas for data centre power. In our reading, that points to grid and fuel supply, not accelerator supply, as the binding constraint on capacity. Anthropic confirmed the filing here (https://www.anthropic.com/news/s1-sec); the registration statement itself is on SEC EDGAR (https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/0001628280-26-067168-index.htm), while The Decoder (https://the-decoder.com/anthropic-opens-its-books-for-wall-street-showing-record-revenue-and-mounting-obligations/) and The Verge (https://www.theverge.com/ai-artificial-intelligence/1018476/anthropic-ipo-details-s1-filed-sec) carry the detail.

Power is where this becomes a regional story. In the GCC, grid capacity, gas allocation and cooling water shape AI programmes far more than model choice does. Power contracts and data centre capacity are strategic assets, and they should be planned for deliberately during procurement rather than discovered halfway through a rollout.

How Azrty would approach it: put one OpenAI-compatible gateway in front of every model endpoint, hosted or self-hosted, so routing, budgets and access control live in one place. That is what our FastLLM Proxy does in front of your own LLM servers and 80 hosted providers. Serve steady workloads from your own GPU estate with Kuvryn AI, and keep portability in application code with our open-source go-ai-sdk, which covers 40 providers. Meter everything, because when existing customers spend nine times more year on year, your bill is a direct function of usage, and budgets belong in the gateway, not in a spreadsheet.

import os
from openai import OpenAI

every SDK call goes through one gateway: providers become routes, not code changes client = OpenAI( base_url=os.environ["LLM_GATEWAY_BASE_URL"], api_key=os.environ["GATEWAY_API_KEY"], )

resp = client.chat.completions.create( model="default", # the gateway maps this route to a provider of your choosing messages=[{"role": "user", "content": prompt}], )

meter per team, per model, per week; alert at 80% of budget, hard stop at 100%

print(resp.usage.total_tokens)

Where teams get it wrong: reading this as a stock story, and then signing the first multi-year token commitment put in front of them. The obligations in that book run for years while usage forecasts move in quarters, so the monthly minimum and the non-cancelable share are the clauses that decide your bill. The governance detail is the other part to read twice. Dual-class shares, confidential ranked successor lists for each co-founder and a public-benefit-corporation charter mean control can sit with aims other than shareholder return. That is a vendor risk line item. If mission-critical workflows run on a frontier model, pin versions, keep a second provider warm behind the gateway, and keep the workload placement decision yours.