The first-mover contest is a first-mover liability

The prevailing frame on the Anthropic-OpenAI listing sequence treats going first as a strategic prize. I think it is closer to a strategic tax. Anthropic, pricing in October at a reported $965 billion post-money, absorbs the valuation discovery risk on behalf of a category — and OpenAI, arriving in Q1 2027, gets to price against whatever comp Anthropic has by then established, good or bad.

That is the frame. What actually decides how these two list is not the ordering. It is three questions the roadshow decks will try to skate past: how much of Anthropic's $47 billion ARR (annual recurring revenue, the annualised run-rate of subscription and API revenue) is gross versus net of Amazon Web Services distribution fees; whether the capital-efficiency gap between the two companies survives a public S-1's disclosure rigour; and how either business defends inference-layer pricing against Meta's free Llama weights. The first-mover framing is the story the pod circuit is telling. The three questions are the story the S-1 will be forced to tell.

The anchor cuts both ways. The "first-mover sets the pricing floor" argument has an appealing symmetry to it, which is usually a warning. Cerebras gave back 10% on day two of its listing. Facebook traded below its IPO price for approximately a year. Uber did the same. Snap's chart is a reminder that setting the anchor and setting a useful anchor are different exercises. If Anthropic prices well and trades up through November, OpenAI's Q1 window looks generous. If Anthropic prices thin or breaks below issue, the anchor Anthropic sets is the one that shuts OpenAI's window.

The second-mover is not neutral to the first-mover's outcome — the second-mover is a call option on it. That is a valuable position, not a compromised one. OpenAI's later date reads as institutional patience, not tardiness.

The capital-efficiency gap is the most important number nobody is pricing. Anthropic reports roughly $47 billion ARR at about 3,500 employees. That is approximately $12–13 million of revenue per head. OpenAI reports approximately $25 billion ARR at about 8,000 employees — approximately $3.1 million per head. The ratio is roughly four to one.

Numbers that wide are not rounding differences. They are structural claims about cost base, product mix, and how each company converts inputs into revenue. Institutional investors pricing growth-at-scale stories weight revenue-per-employee heavily because it is the crudest available proxy for what post-listing operating margin might look like once capex-through-P&L stops being the dominant line item.

The gap is so large it raises the obvious secondary question: what are OpenAI's roughly 4,500 extra employees doing that Anthropic's are not? Consumer distribution, ChatGPT support, safety and policy staffing at consumer scale, and a much larger product surface (image, video, voice, agents, hardware partnerships) are all plausible answers. Some of that cost is optional; some of it is the thing the equity story is made of. The S-1 will need to tell that story convincingly enough that the ratio reads as strategic breadth rather than operational drag.

And then there is the accounting question that could reset the denominator. Anthropic's largest cloud partner is Amazon, which has committed up to $8 billion across cloud and equity. A meaningful share of Anthropic's revenue flows through Amazon Bedrock — AWS's managed marketplace for third-party foundation models. When an API call to Claude is billed to an enterprise customer via Bedrock, AWS takes a distribution fee. The question the public S-1 must answer is whether Anthropic's reported $47 billion ARR figure is gross of that fee or net of it.

The distinction is not academic. If a material portion of the ARR is Bedrock-routed and reported gross, then net revenue — the number every P/S (price-to-sales, market cap divided by revenue) multiple should actually be built on — is smaller than the headline. How much smaller depends on the Bedrock share and the AWS take-rate, neither of which is public. But it is entirely possible for the disclosed net figure to reset the effective P/S the IPO is priced at by 15–30%, which is not a rounding error at a $965 billion post-money anchor.

Amazon is also a related party. The S-1 will have to disclose related-party revenue as a percentage of total, and it will have to do so in a form audit-committee-signed and lawyer-reviewed. That disclosure is where the "extraordinary" revenue-per-employee number either becomes the equity story or becomes the risk factor. I would not want to be the equity analyst who built a model on the pre-filing figure and discovered the net-of-Bedrock number in the amended S-1.

Meta is the variable the roadshow cannot make go away. Both Anthropic and OpenAI sell inference — the ongoing cost of running trained models to answer real queries, as distinct from the one-time cost of training them — as a premium managed service. Meta gives Llama weights away. This is not accidental. It is a deliberate strategy to commoditise the layer both listing candidates depend on for margin.

The counter-argument is real: enterprise buyers pay premiums for compliance, SLAs (service-level agreements, contractual uptime and performance guarantees), safety tooling, and the reliability of a managed API. Bare Llama weights do not come with any of that, and the operational cost of self-hosting a frontier open-weight model at enterprise scale is substantial. Many enterprise buyers treat Claude and Llama as complements or as distinct product classes, not substitutes.

The first-mover contest framing treats the ordering as the strategic variable. The ordering is the least interesting variable in the story.

That defence works until it doesn't. Every quarter Meta ships a stronger open model, the premium enterprise buyers will pay for managed access to a frontier proprietary model narrows. Both S-1s will have to spend risk-factor pages on this, and both roadshows will be asked the same question: at what open-weight capability level does your pricing power meaningfully compress? Neither company has a comfortable answer, because the honest answer is "we don't know, and it depends on Meta's roadmap."

Read the FutureSearch forecast as a warning, not a comp. FutureSearch models first-day market caps for both companies landing at approximately $1.08–1.10 trillion — a near-identical draw. It then projects OpenAI's 90-day cap at roughly $0.86 trillion, a meaningful pullback attributed to loss-disclosure digestion.

Two things worth noticing. The first is that the "contest" is already modelled as a tie at listing, which suggests the market is not pricing much genuine first-mover advantage — it is pricing a category, not a winner. The second is that a projected 90-day drawdown of roughly 20% on OpenAI is not a stumble; it is the model saying institutional holders take three months to absorb the loss trajectory before sizing final positions. If that is right, OpenAI's later window is timed roughly correctly. Anthropic's window is where the discovery pain happens.

I would treat the specific numbers as a model, not a market. But the shape of the forecast, draw at listing, drawdown after, is directionally consistent with how every prior large-cap loss-making tech listing has behaved. Which is the pattern the pod circuit's "contest" framing is ignoring.

What I would actually watch. The disclosed net-of-Bedrock revenue figure in Anthropic's amended S-1 is the single most consequential number in either filing. It resets or confirms every multiple the sector will be priced from for the next twelve months. Anthropic's related-party revenue concentration disclosure is second — a high number sharpens the Amazon-dependency risk factor into something a fund's IC (investment committee) has to formally accept. Third is whichever company first commits, on paper, to a defensible answer on Llama-driven pricing compression. That answer will not be comfortable, but the company that provides it will get graded on candour; the company that dodges will get graded on evasion.

The "first-mover contest" framing collapses each of those questions into a horse-race story that is easier to narrate on a podcast than to underwrite in a term sheet. Anthropic goes first because its cap table and Amazon relationship pushed it to file first, not because going first is strategically superior. OpenAI arrives later because its consumer scale, headcount, and losses need more roadshow runway, not because arriving later is strategically superior. Neither is a prize position. They are the positions the two companies are in.

If Anthropic prices well and holds the anchor through year-end, the sector gets an orderly OpenAI listing at a firm comp. If Anthropic breaks issue or the Bedrock disclosure resets the multiple, OpenAI's window narrows and its final valuation compresses regardless of the merits of its own business. That is not a contest OpenAI is losing by going second. It is a contest neither company controls, because the variable that decides both outcomes is a set of accounting and competitive disclosures that only one of them has to make first.

Footnotes