OpenAI's IPO Is a Regulatory Filing First

The market conversation around OpenAI's impending IPO has already collapsed into the valuation number — a trillion dollars, give or take — and the model horse race that number supposedly reflects. Nate Jones, writing and podcasting at AI News & Strategy Daily, offers a more useful frame: the real bet isn't which lab has the smarter model, it's whether OpenAI and Anthropic can build the operational layer around raw intelligence fast enough that enterprise customers rent the whole system rather than assemble their own.

Jones calls that layer the "harness" — the files the model can see, the tools it can invoke, the permissions structure, the memory, the evaluation logic, the routing between cheap and expensive models. "A model gives you intelligence," he argues. "A harness gives you work." OpenAI's 2025 Codex coding agent — distinct from the original Codex model the company deprecated in 2023 — is his clearest example: it's not impressive because the underlying model is clever, it's impressive because the model operates inside a system that can read a repository, run tests, inspect errors, and move through the actual loop of software development.

The business logic here is sound. Jones is right that as token prices trend downward across frontier providers — a directional claim that holds, even if the rate and durability of that decline vary considerably by model tier — raw intelligence becomes harder to defend as a moat. Commodities get commoditized. The value migrates to whatever sits above the commodity, which is the workflow harness.

Where Jones and I diverge is on what happens next. He frames the forward deployed engineering move — OpenAI embedding engineers inside enterprise customers to map workflows and customize harnesses — primarily as a product strategy. I read it as a regulatory classification problem that neither lab has publicly reckoned with.

The S-1 Is a Disclosure Document Before It Is a Pitch Deck

An S-1 filing is, at its legal core, a Securities Act document designed to give prospective investors a complete picture of material risks. For most technology IPOs, the regulatory risk section is boilerplate — a few paragraphs about hypothetical future legislation and ongoing FTC scrutiny. OpenAI and Anthropic cannot write boilerplate regulatory risk sections.

Consider what SEC staff will actually be reviewing. OpenAI is structuring its conversion from a capped-profit LLC to a public benefit corporation — a governance form that has never been tested at this scale under SEC disclosure rules. Anthropic is already organized as a Delaware public benefit corporation. Neither structure maps cleanly onto standard securities disclosure frameworks, because both companies have fiduciary obligations that extend beyond shareholder returns in ways that will require novel disclosure language. The SEC's standard materiality tests were built for companies with a single bottom line.

Layered on top of that: OpenAI's core products are simultaneously subject to FTC review, under active scrutiny in the EU under the AI Act's General Purpose AI provisions, and the subject of at least three active congressional hearings this session. Under SEC Rule 10b-5 and the general antifraud provisions of the Exchange Act, any known, pending, or reasonably foreseeable regulatory proceeding that could materially affect the business must be disclosed. The S-1 cannot wave at "evolving AI regulation" and move on. It will need to describe, specifically, what the EU AI Act's GPAI transparency and copyright provisions mean for model training costs and product liability. It will need to characterize FTC interest in OpenAI's partnership structures — Microsoft, in particular — and what remediation scenarios look like. It will need to address the Copyright Office's ongoing proceedings on AI training data, which have direct implications for the licensing costs buried in the cost-of-revenue line.

What these companies will almost certainly try to compress into footnotes: the specifics of any FTC civil investigative demands already issued, the actual terms of their compute agreements (the monetization reckoning around compute dependency is material information for any investor), and the precise scope of data retention practices that are simultaneously their competitive advantage and their regulatory exposure.

The SemiAnalysis estimates that have circulated — figures suggesting heavy users of OpenAI's and Anthropic's $200 monthly plans may be consuming API-equivalent value in the range of $14,000 and $8,000 respectively — are instructive here, but require qualification. The methodology underlying those numbers, including which API endpoints and what usage assumptions drive the calculation, has not been independently verified, and SemiAnalysis has not published the full underlying model. What matters for the S-1 is not the precise figure but the disclosure question it raises: if the $200 consumer plan is a deliberate subsidy designed to accelerate usage while labs race inference costs down, that pricing strategy has to be characterized accurately for investors. Is it a customer acquisition cost? A research subsidy? A competitive moat investment? The accounting treatment determines what it looks like in the prospectus.

Forward Deployed Engineering Is Not Just Consulting. It Is Also HIPAA, FINRA, and FedRAMP.

Jones's point about forward deployed engineering is structurally correct: labs can't learn a customer's internal context from the outside, so they send people in. The lock-in, he argues, comes not from the model but from the workflow now wrapped around one company's harness architecture.

But the regulatory dimension of embedding lab engineers inside regulated industries is something the market conversation has almost entirely ignored. When OpenAI or Anthropic places engineers inside a hospital system to build workflow harnesses that route clinical decisions through AI systems, HIPAA's business associate provisions attach — not just to the data the model sees, but to the engineers who see it. When the harness connects to a financial institution's approval workflows, the question of whether the lab is a third-party service provider under FINRA Rule 3110 or the OCC's third-party risk management guidance becomes a compliance question for the bank, not just a contract question for OpenAI. Defense contractors operating under CMMC (Cybersecurity Maturity Model Certification) requirements face certification obligations that extend to technology vendors whose engineers have access to covered defense information.

None of this makes forward deployed engineering impossible. Large consulting firms navigate these frameworks routinely. But OpenAI and Anthropic are not positioning themselves as consulting firms — they're positioning as software platforms that happen to also send engineers. That framing matters, because software platforms and professional services firms face different regulatory classifications, different liability structures, and different disclosure obligations. The S-1 will need to characterize which one these companies actually are, and the enterprise contract opacity that has characterized Anthropic's deals so far will not survive SEC scrutiny at the level of specificity a public offering requires.

Jones is right that the labs' goal is to make the customer's workflow inseparable from their harness architecture. He's right that this is stickier than model performance. What he underweights is that regulatory frameworks in healthcare, finance, and defense are specifically designed to prevent any single vendor from achieving exactly that kind of embedded dependency without corresponding oversight, audit rights, and exit provisions. The enterprise customers who will be most valuable to these labs — the hospitals, the banks, the defense primes — operate in environments where "we can't switch because our workflow is built around your harness" is not a defensible answer to a regulator. Their compliance teams know this. The lab's forward deployed engineers are going to encounter it in the field, repeatedly.

What the S-1s Will Actually Reveal

The financial metrics Jones identifies as meaningful — whether heavy users are getting cheaper to serve over time, whether gross margin improves with scale, whether enterprise customers are buying software or buying labor — are the right questions, and they're precisely the questions the SEC will pressure these companies to answer with specificity.

I'd add the regulatory risk section line count as a leading indicator of institutional honesty. A company that genuinely believes its regulatory exposure is manageable writes two pages. A company managing simultaneous FTC scrutiny, EU AI Act compliance obligations, open-source competitive pressure from models like DeepSeek that carry their own data provenance questions, and congressional attention on both safety and copyright writes considerably more.

The Senate Commerce Committee's AI working group is currently drafting disclosure legislation that would require AI companies above a certain revenue threshold to file quarterly model capability assessments with a designated federal agency — a provision that, if enacted, would create ongoing SEC-adjacent disclosure obligations that no current tech company faces. That proceeding, and the timeline it creates, is what I'll be watching when the S-1s file.

Jones frames the IPO as "the first public test" of whether labs can own the work layer. He's right on the business logic. But for OpenAI and Anthropic, the IPO is also the first time they will be legally required to tell the public — under penalty of securities fraud — exactly how exposed they are to the regulatory frameworks that could rewrite the business model before the harness strategy has time to compound.

The harness is the product. The S-1 is the disclosure. Both of them will be revealing, but only one of them has a lawyer reviewing every word for what it cannot afford to admit.

---

Samira Barnes covers technology policy and regulation for Buzzrag.