Anthropic filed confidentially around 1 June 2026 at a reported valuation near $965 billion. OpenAI is targeting a listing around September 2026 at a reported $730 to $850 billion, with some commentary placing the number above $1 trillion. The two companies, taken together, propose to bring close to $2 trillion of pure-play artificial intelligence onto public markets within a single calendar year. The disclosed economics, on the available reporting, are striking in a way that deserves plain language.
OpenAI is on a recurring revenue run rate of roughly $25 to $30 billion as of the first quarter of 2026, growing at a pace that has few precedents in software. The same business is projected to lose around $14 billion in 2026 alone and a cumulative figure approaching $44 billion through 2029 on the analyst consensus. Anthropic's revenue trajectory targets $20 to $26 billion in 2026, and the company is loss-making, with the precise figure undisclosed. Both companies are organised as public-benefit corporations, a structure that obliges directors to weigh stated public-interest commitments alongside shareholder returns.
The headline is the strangeness of a $20 billion revenue company that loses $14 billion a year being asked to clear a near-trillion-dollar mark. The honest reading is that the listing prices, like the SpaceX listing before them, are narrative pricing on top of a real and accelerating business. The job of the prudent reader is to separate the part that exists from the part that is promised, and to be honest about which is which.
What is actually selling
The revenue that exists is enterprise software revenue, in the form that the corporate market understands. It is paid seats on ChatGPT and the enterprise tier of Claude. It is application programming interface usage by software companies embedding the models in their own products. It is bespoke deployment contracts with banks, insurers, law firms, government agencies and the larger consumer brands. The growth is real, the customers are named, and the renewal rates published by the larger enterprise buyers indicate that the product is sticky.
The revenue that is promised is something else. It is artificial general intelligence and the assumption that the model that defines that frontier will be developed by one of these two firms. It is the assumption that the firm that develops it will be permitted to charge a meaningful price for it. It is the assumption that the regulatory and geopolitical environment in 2030 will resemble the environment in 2026 closely enough for the long compounding to play out. These are not unreasonable assumptions, but they are assumptions. The disciplined investor pays for the first revenue stream and is honest that the second is a lottery ticket with real but unknown odds.
Separate the part that exists from the part that is promised, and be honest about which is which.
The public-benefit corporation, plainly
Both OpenAI and Anthropic are structured as public-benefit corporations. The legal language is dense. The practical meaning is that the directors are permitted, and in some readings required, to weigh stated mission commitments alongside the financial interests of shareholders. In a normal corporation, a board that turns down a profitable contract on ethical grounds opens itself to litigation. In a public-benefit corporation with a stated safety mission, the same decision is more defensible.
For an ordinary investor, the implication is not that the structure is bad. The implication is that the structure introduces a second principal, the mission, alongside the shareholder. In the SpaceX listing, the second principal is the founder. In these listings, it is a stated public interest. In both cases, the minority shareholder owns less than the headline figure suggests, and the disclosure of that fact in the filings should be read in full before any order is placed.
The infrastructure read-through
The direct beneficiaries of the AI capital expenditure cycle are Nvidia, Oracle, CoreWeave, Microsoft, Amazon and the broader hyperscaler stack. The reverse implication is that the same names carry concentration risk if the capital expenditure expectations reset. The OpenAI listing in particular, with Microsoft holding roughly 27 per cent, is a partial proxy for the read-through to the largest single beneficiary. A sober reading of the sector should treat the listings, the infrastructure names and the capital expenditure cycle as a single position rather than as three independent ones.
What the S-1 filings will need to disclose
The load-bearing document in each listing will be the S-1, and the items that deserve close reading are foreseeable in advance. The first is the segment-level disclosure of gross margin. Enterprise software at scale has historically delivered gross margins in the seventies. The artificial intelligence businesses, on the published commentary, operate at materially lower gross margins because of inference costs. The trajectory of that figure across the next eight quarters is the single most important variable in the path to profitability. A model gross margin that expands ten points over eight quarters supports a very different valuation than one that holds flat or compresses.
The second is the customer concentration disclosure. A business in which the top ten customers account for a small share of revenue is fundamentally more resilient than one in which a handful of contracts, including the closely-watched federal government contract on the OpenAI side, account for a large share. The S-1 will disclose the concentration in standardised form, and the investor who reads it before the order opens will be better placed than the investor who reads it after.
The third is the related-party disclosure. Both companies have intricate relationships with their principal infrastructure partners, including the Microsoft equity and compute arrangement on the OpenAI side and the multi-cloud arrangement on the Anthropic side. The terms of those relationships, including the revenue-share mechanics and the right-of-first-refusal provisions on next-generation accelerator allocation, are material to the durability of the unit economics. The disclosure will be lengthy. It will repay the time it takes to read.
The fourth is the use-of-proceeds and the secondary component. A raise dominated by primary issuance, with proceeds allocated to compute capacity and research, signals an extension of the operating runway. A raise with a material secondary component, in which existing holders sell into the offering, signals a partial liquidity event. The distinction matters for the alignment between the new public shareholder and the existing insider base, and the S-1 will state it plainly.
The competitive picture beyond the two leaders
The narrative around the listings has consolidated around OpenAI and Anthropic, but the competitive picture is wider than the two names. Google DeepMind continues to ship frontier models at a cadence that, on the published benchmarks, places it among the leaders. Meta has invested at a scale that is, in absolute terms, comparable to either of the listed names, and is pursuing a different strategy through the open-weights distribution channel. The Chinese frontier laboratories, principally those associated with Alibaba, ByteDance and a small number of better-resourced national champions, are publishing models that compress the technical lead that the United States laboratories enjoyed in 2024 and 2025. The competitive question, looking out to 2028 and beyond, is whether the frontier remains a small oligopoly or whether the diffusion of capability widens the field.
For the investor in the two listings, the implication is that the bull case rests on a continued narrow frontier and on the pricing power that the frontier confers. The bear case rests on diffusion, on the commoditisation of the underlying model layer, and on the migration of value to the application layer rather than the model layer. Neither case is settled by the listing. The reader who treats the two scenarios as live possibilities and sizes the position accordingly will be better positioned than the reader who treats the bull case as inevitable.
The specialist segment is the other dimension of the competitive picture. Vertical artificial intelligence businesses in coding, in legal, in life sciences and in customer support are each generating meaningful enterprise revenue on top of one or more of the foundation models. The economics of those businesses are typically stronger than the foundation model businesses themselves, because they are closer to the customer and because they capture the workflow rather than the inference. A complete portfolio view of the artificial intelligence cycle is more nuanced than the two-listing trade.
A prudent participation
LUMINAIRE does not publish price targets. Three observations, however, are worth carrying into the listings.
First, the right size is the high end of the volatility budget of a diversified portfolio, not the centre. Both companies are growing fast and losing money fast. The path to durable profitability is plausible and not guaranteed.
Second, the two listings will not be priced in isolation. The arrival of two record-scale offerings within weeks of each other places real pressure on institutional appetite. A valuation cut on one is a credible scenario and should not be read as a verdict on the broader sector.
Third, the reader who participates is paying for a future they can read about in advance. The S-1 filings, when they become public, are the load-bearing document. The threads are not.
Readers who want to model the path to profitability for themselves can use the AI IPO Burn-Rate and Path-to-Profit Model on CalculatorIQ, which takes recurring revenue, growth rate, gross margin, operating expense growth and a target valuation, and returns the implied break-even year, the cumulative funding gap and the multiple versus the reported mark. The institutional treatment of the valuation, the public-benefit corporation governance and the infrastructure read-through is carried on Cabier. Both are linked below. Nothing here is investment advice.