Here we go.

Anthropic filed its S-1 on October 26, the formal start of its IPO. The financials will get the headlines, but read the founders’ letter at the beginning of the S-1. It describes how AI should be brought into the world as if from the AI company you would design from scratch: intelligence becomes abundant, and agents turn it into labor; safety comes before speed and profit; and a public benefit corporation and a long-term benefit trust protect the mission. And the letter admits that disruption, including to jobs, comes before the rainbows. This is ideal, in theory. That would be copacetic if Anthropic were the only AI company in the world (it’s not) and did not have to answer to regulators, rivals, and the cost of compute (it does).

Idealism Vs. The Real World

“Welcome to the machine” rings so true. Idealism always meets its match: reality. Anthropic’s trust and founder voting control is designed to stop that from happening, and it may hold. But the IPO machine that Anthropic is stepping into does not care what the founders’ letter says. Five gears in that machine will grind against the ideals:

  • Rivals own the data centers. Anthropic largely rents from Amazon and Google, partners that also sell competing model platforms, and its own filing names compute availability as a key risk. Agentic workloads will make this worse because the tools that agents call run on infrastructure owned by the hyperscalers. Founder control protects the mission from shareholders, not from the economics. These cost pressures can erode ideals.
  • Anthropic can decide if you deserve AI. A board insulated from shareholders can hold the line too well. Anthropic already drew that line with the Pentagon this year, refusing uses that its principles forbid. Defense, pharma, energy, and other industries must now ask whether they are next. This could result in loss of market share and the scale Anthropic needs to thrive.
  • Regulations override company ideals. Anthropic can write its own usage policy; it cannot write the law. A court order can compel it to hand over a protester’s transcripts, an export control can cut off a market overnight, and a statute can force it to serve an agency it would rather not. Everywhere Anthropic operates, the government’s ideals can force the company to comply.
  • Anthropic doesn’t make the safety rules for other frontier labs. Anthropic’s definition of safe is one vendor’s definition. OpenAI, Google, SpaceX, open-weight consortia, and labs in other countries all invest in safety, and none is obliged to match Anthropic’s standard. Safety claims also double as competitive weapons. Enterprises will decide which version works for them.
  • Customers may demand transparency. Anthropic is not transparent today. Its weights, training data, and methods are closed, and so are the harness and built-in agents that wrap its models. So enterprises take its safety claims on faith. The open-source community is pushing the other way, with open-weight models, open-agent frameworks, and tooling that anyone can inspect. Anthropic could open parts of its stack. The founders’ letter does not promise it.

You Need A Dual Vendor Frontier Strategy

One is not enough. The frontier market is still extremely fluid: The hyperscalers are building their own frontier model platforms, open weights keep closing the gap, and pricing and usage policies change by the quarter. When Anthropic and OpenAI announced their IPO plans, we told clients not to lock in. That’s still our advice, and so is building agents now, because:

  • The platform locks you in, not the model. Every call to a model runs through a harness, tools, memory, and connectors the vendor wraps around the model, and none of it transfers. Each agent you build tightens the hold. Read our blog post, Introducing Frontier AI Model Platforms — Because An AI Model Is Not A Business Model.
  • The frontier is a zone, not a point. Several frontier platforms, from OpenAI, Google, SpaceX, and Anthropic, sit inside it at any given moment, and which one is best depends on the task and the month. If all of your agents are built on one of them, you can’t take advantage when another pulls ahead.
  • Pick two frontier platforms and make one a hyperscaler. Both sit in the frontier zone, but one also owns the data centers and the infrastructure that your agents run on, which gives you a fallback when the other’s policies, pricing, or availability change. Anthropic or OpenAI plus AWS, Google, or Microsoft is the obvious pairing; SpaceX can fill either slot but not both. Run real work on both, with your own evals and kill switches, so your teams know how to move an agent before they have to.

Forrester clients: Schedule an inquiry or guidance session to solidify your frontier platform strategy.

Share