Anthropic opens its IPO roadshow at $70 to $78 a share, with a $26 billion placement from its own suppliers
825 million shares at $70 to $78, an implied valuation of up to $1.4 trillion, and a private placement from Google, Nvidia, AMD, Microsoft and TSMC. Enterprise AI buying just acquired a public benchmark.
What happened
Summary of reporting by AnthropicAnthropic began marketing its initial public offering on 10 November, offering 825 million Class A shares at $70 to $78 each. At the top of the range that implies a fully diluted valuation of up to $1.404 trillion. Pricing is expected on 18 November, with the shares due to start trading on Nasdaq under the ticker ANT on 19 November. Around a fifth of the offering is reserved for retail investors. Reporting: TechCrunch (https://techcrunch.com/2026/11/10/anthropic-sets-ipo-price-range-of-70-to-78-per-share-to-reach-an-implied-valuation-of-1-404-trillion/) and CNBC (https://www.cnbc.com/2026/11/10/anthropic-sets-price-range-of-70-to-78-in-upcoming-ipo.html).
The amended registration statement sets the offering at 825,000,000 shares: 801,351,354 offered by Anthropic and 23,648,646 by existing stockholders, plus a 30-day underwriter option on up to 123,750,000 more. Filing: https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm. Alongside the float sits a concurrent private placement of about $26 billion at the IPO price, roughly $10 billion each from Google and Nvidia, with AMD, Microsoft and TSMC also taking part. The placement is in non-voting Class N shares, unregistered in the offering and outside the usual lock-ups.
Control stays with the founders, who retain 50.1% of the voting power. Demand looks heavy enough that the Financial Times reports bankers weighing a valuation of $1.5 trillion or more. Anthropic's own roadshow launch note is at https://www.anthropic.com/news/roadshow-launch.
The Azrty take
Anthropic's float puts frontier AI's cost of capital on the record. Treat model portability as a procurement requirement, not a nice-to-have.
Frontier AI just put part of its cost of capital on the record, and that changes what happens at your next renewal. Anthropic opened its roadshow on 10 November at $70 to $78 a share, an implied fully diluted valuation of up to $1.404 trillion, with pricing expected on 18 November and trading on Nasdaq as ANT from 19 November (TechCrunch, CNBC, Anthropic). The amended registration statement behind those numbers sets the offering at 825,000,000 shares: 801,351,354 offered by Anthropic and 23,648,646 by existing stockholders, with a 30-day underwriter option on up to 123,750,000 more (SEC Form S-1/A-2). A lab that has to defend that valuation in public markets will defend its pricing in private ones too. Do not assume token prices fall because a vendor listed. Expect contract terms, committed spend and deprecation schedules to tighten instead.
The capital stack tells you who you are really buying from. The concurrent private placement raises about $26 billion at the IPO price, roughly $10 billion each from Google and Nvidia, with AMD, Microsoft and TSMC also taking part. It is sized in non-voting Class N shares, which carry no votes except where the law requires one, are not registered in the offering and are not bound by the underwriters' lock-ups; they convert to Class A only on transfer to an unaffiliated third party (SEC filing). That puts a hyperscaler, two accelerator vendors and a foundry on the cap table of the model vendor they supply. When the model, the chips, the cloud and the fab are equity-linked, procurement independence erodes quietly. The practical test is simple: can you move a workload to a second provider within a sprint without rewriting the application?
Governance is the third thing to read. The founders keep 50.1% of the voting power, so public shareholders and enterprise customers get disclosure rather than influence (SEC filing). What you can still negotiate is operational. Anthropic's model documentation commits to at least 60 days' notice before a publicly released model is retired, and the current table shows what that means in practice: claude-sonnet-4-5-20250929 was deprecated on 30 September 2026, retires on 30 November 2026, and claude-sonnet-5-5 is the named replacement (Anthropic model deprecations). The company also commits to preserving the weights of every publicly released model for at minimum the lifetime of the company (deprecation and preservation commitments), which protects the record, not your uptime: requests to a retired model fail. Demand looks heavy: the Financial Times reports bankers weighing a valuation of $1.5 trillion or more. Capital is concentrating around a handful of labs, and single-sourced buyers will feel that first.
How Azrty would approach it technically: put an OpenAI-compatible gateway such as FastLLM Proxy in front of every model call, whether it lands on hosted providers or on your own LLM servers, so routing, budgets and access control live in one place. Pin exact model versions in configuration, never floating aliases. Give each team a hard monthly budget and a fallback chain: one frontier hosted model as primary, a second hosted provider behind it, and a self-hosted open-weight model for steady workloads on your own GPUs, which is where Kuvryn AI fits. Provider switches then become configuration changes judged by an eval set, not rewrites. The policy shape we deploy looks like this.
routes:
agents.default:
primary: { provider: anthropic, model: "<pinned-sonnet-version>" }
fallback: [openai/<pinned-model>, self-hosted/llama-3.3-70b]
budgets:
team: { monthly_usd: 20000, hard_stop: true, alert_at: 80 }
policy: { pin_versions: true, eval_gate: "promptforge/switch-benchmark" }
For UAE and GCC teams the next 90 days are procurement work, not market watching. Take the published numbers into renewals, ask for deprecation and version-support commitments in writing, cap minimums at your observed consumption, and keep a share of inference volume on a provider you can swap without code changes. Our team in the UAE does this deployment and support work. The matching risk is local capacity deals that copy take-or-pay economics without matching demand certainty. Hedge the stack first, commit second.
What to do now
- Diarise 18 November (pricing) and 19 November (ANT starts trading on Nasdaq). Before any AI renewal this quarter, add vendor concentration to the risk register and answer one question: how many model providers can you reach today without a code change?
- Put an OpenAI-compatible gateway (FastLLM Proxy) in front of every model call this month. Pin model IDs in the config, set a hard monthly budget per team, and configure a fallback chain to a second hosted provider plus one self-hosted model.
- Build a 50-case eval set in PromptForge or your existing harness and make it the gate for any provider switch. Remove floating "latest" model aliases from production code and pin SDK versions in your lockfile. Track each provider's published deprecation table against your dependencies.
- Audit AI contracts for committed-spend and take-or-pay language before signing anything new. Cap minimums at your trailing three months of observed consumption and keep at least 20% of inference volume on a provider you can swap within a sprint.
