Anthropic’s potential IPO could become the biggest test yet of how public markets value AI. Here’s how it could influence upcoming commoditization, next open-source models, and what founders build in the years to come
Since ChatGPT’s launch in late 2022 set off the current AI investment cycle, the valuations of frontier AI companies like OpenAI and Anthropic have been set almost entirely in private funding rounds.
A small group of investors has written those cheques: sovereign wealth funds such as Singapore’s GIC; crossover funds such as Coatue and D1 Capital (Series H, Series G); hyperscalers such as Amazon and Microsoft; and chipmakers such as Nvidia, which committed up to $10 billion. Together, they have agreed on valuations that would have sounded absurd when that cycle began: Anthropic went from about $623 million at its 2021 Series A to $965 billion in May 2026, with one investor reportedly pledging $5 billion just to get a meeting with the company’s CFO.
However, none of those numbers have yet been tested by public markets, but that is about to change very, very soon. That is about to change. Anthropic confidentially filed for an IPO on June 1, 2026, and Reuters has now reviewed its draft prospectus. According to the Wall Street Journal, investors expect the company to be valued at around $2 trillion and to raise up to $100 billion, while Reuters puts the target at more than $2 trillion, over twice its May private valuation. For the first time, a frontier-model company will be priced every day by thousands of investors based on quarterly results.
At GapMinder, we invest at Seed and Series A, far from trillion-dollar order books. But what happens at the top of the stack always flows down. The multiple public markets accept for Anthropic will shape how growth investors price AI companies, how early-stage funds think about exits, and which kinds of AI startups get funded in the next cycle.
This article looks at how the IPO could influence the market, how to think about the value of a frontier lab, what commoditization really means for the model layer, and what we believe founders should focus on as they raise this autumn.
Anthropic could give public markets their first major AI pricing test
The scale of what is being attempted is hard to overstate. The largest tech IPO San Francisco has ever seen was Uber, which listed in 2019 at $82.4 billion. Lyft, Slack and Airbnb, the other members of that 2019–2020 wave, each debuted below $100 billion. A $2 trillion Anthropic would be worth roughly 24 Ubers on day one.
Measured against the entire history of US tech listings, the number is even more interesting. By one tally based on Jay Ritter’s IPO data, the 3,365 US tech companies that went public between 1980 and 2025 had a combined debut value of about $4.07 trillion. Apple, Microsoft, Nvidia, Google, and Facebook are all in that number. Anthropic alone is targeting roughly half of it.
The path to this point has been almost entirely private. Anthropic’s Series A in 2021 valued the company at around $623 million. It raised a $30 billion Series G at a $380 billion valuation in February 2026, then a $65 billion Series H at a $965 billion valuation in May, and filed confidentially for an IPO on June 1.
Each of those rounds was negotiated between the company and a small group of investors who wanted in. An IPO is different: it is the first moment when the market is asked what it’s willing to pay for AI exposure, and then asked again every trading day thereafter.
The context around it is also important to look at. SpaceX, which absorbed xAI earlier this year, set the previous record in June at $1.77 trillion. Its shares jumped on debut, gave back most of the rally within weeks, and have since recovered to trade above the offer price. OpenAI, meanwhile, has pushed its own listing to 2027, and Anthropic’s own date has reportedly slipped from October to November.
Source: Anthropic Just Passed OpenAI in Revenue. While Spending 4x Less to Train Their Models
From private valuations to the public-market test
Private and public markets ask different questions of the same company. A private round is a point-in-time bet: investors price long-term potential, scarcity of access and the chance of future market dominance. If you believe a company can own a category, paying up for a limited allocation is rational, and nobody marks you to market the next morning.
Public investors reassess constantly. They compare the company to every other asset they could own; they react to each quarterly print, and they can sell. Scarcity premiums shrink once anyone with a brokerage account can buy the stock. Guidance, not vision, sets the price.
The leaked prospectus, reported by Reuters, gives public investors their first audited view of Anthropic. The headline figures:
- Revenue growth. 2025 revenue rose twelvefold to nearly $4.6 billion. Momentum has since accelerated: second-quarter 2026 revenue reached $11.5 billion, and annualized run rate passed $65 billion by the end of July, up from $9 billion at the end of 2025.
- Margins and profitability. Anthropic reported an operating loss of $8.06 billion in 2025. The headline net loss of about $42 billion looks alarming, but roughly $34 billion of it was a non-cash accounting charge tied to the company’s own rising valuation. More importantly, Anthropic says it turned an adjusted operating profit in Q2 2026 and expects to repeat it in Q3.
- Compute expenditure. Compute and infrastructure cost $7.33 billion in 2025, about 58% of total operating expenses, and the company has committed to $518 billion of future cloud, compute and infrastructure obligations.
- Concentration. Two customers accounted for nearly a quarter of 2025 revenue, and many large customers are not locked into long-term contracts.
- Path to cash flow. Bankers are reportedly underwriting the deal against 2028 revenue of $190–200 billion. Some analysts do not expect positive cash flow before 2028.
Another technical point is that Anthropic reports revenue from cloud resellers such as AWS, Google, and Microsoft on a gross basis, counting total end-customer spend as revenue and booking partner payouts as costs. That flatters the top line relative to peers who report net. Public investors will adjust for it, and so should anyone benchmarking a startup against Anthropic’s revenue multiple.
What revenue multiple will the market accept?
Mature software companies trade at single-digit to low-double-digit revenue multiples; hypergrowth outliers have briefly traded at far higher multiples.
At $1 trillion, Anthropic is priced on a revenue multiple that looks demanding for software but modest for a company growing several times over in a year. At $2 trillion, investors are paying for 2028 today, and the stock will be highly sensitive to any quarter that suggests growth is slowing or compute costs are rising faster than revenue.
Public investors will judge Anthropic on durable growth, visible margins, and credible cash flow, and those expectations will filter down to every later-stage round. So, in practice this can make founders:
- Separate run rate from recognized revenue. Report both, explain how you calculate ARR, and be clear about what is contracted versus usage-based.
- Show gross margin after inference costs. Know your model and compute cost per customer, and show how it has moved over the last few quarters.
- Track revenue concentration. If your top two or three customers account for a large share of revenue, show a plan to diversify and put key accounts on longer contracts where you can.
- Build a path to profitability you can defend. Investors do not expect early-stage companies to be profitable, but they do expect a credible story for how unit economics improve as you scale.
- Benchmark against the right peers. Frontier-lab multiples are not a reference point for an application-layer startup. Compare yourself to companies with a similar business model and margin profile.
Anthropic vs OpenAI: how do you value a frontier AI company?
Anthropic and OpenAI are the first two AI-native companies to operate at this scale, and they have taken strikingly different routes to get there. That makes them the natural comparison set for anyone trying to price the category.
A year ago, OpenAI was well ahead. It ended 2025 with more than $20 billion in annualized revenue, compared with about $9 billion for Anthropic. By the second quarter of 2026, the positions had flipped: Anthropic brought in $11.5 billion that quarter, against $6.7 billion for OpenAI.
The main reason is that the two companies sell to different customers. Anthropic focused on businesses and developers, which account for roughly 80% of its revenue. Companies spend far more per account than individual consumers do, and their spending grows as they integrate AI into more of their work. Coding shows this clearly: Claude Code reached $1 billion in annualized revenue about six months after launch, and trackers estimate it was close to $15 billion by August 2026.
OpenAI took the consumer route. ChatGPT is the world’s best-known AI product, with around 900 million weekly users, but only about 5% of them pay. That gives OpenAI enormous reach, but also the cost of serving hundreds of millions of people for free.
Source: Mavaos Research
Why does compute make AI economics different from SaaS?
In traditional software, once the product is built, adding one more customer costs almost nothing. That is why mature SaaS companies typically keep 70–90% of revenue as gross margin.
AI works differently. Every question a user asks, and every task an agent runs, consume computing power that costs money. The more a customer uses the product, the higher the bill. As a result, AI companies average gross margins of around 52%, according to ICONIQ, and Bessemer puts the typical range at 50–60%.
Frontier labs also have to pay upfront. Each new generation of models costs more to train, and companies must book computing capacity years in advance. Anthropic has committed $518 billion to cloud, computing and infrastructure. Think of it as signing a very long lease before knowing how many tenants will move in. If demand keeps growing, that capacity becomes an advantage few rivals can match. If demand slows, it becomes a high cost to carry.
This is the main question public investors will ask: how much of today’s growth will turn into profit?
- Two things work in Anthropic’s favour. It reported an adjusted operating profit in Q2 2026, and most of its revenue comes from businesses, which tend to stay longer and pay more than consumers.
- Two things work against it: compute costs rise with usage, and cheaper models are putting pressure on prices, as the next section shows.
What exactly are investors paying for?
When someone buys a frontier lab at $1 trillion or more, they are not buying a single asset. They are buying a bundle, and each part of it has a different shelf life:
- The model itself. Today, the scarce asset. But leadership at the frontier is measured in months, and each generation has to be re-earned.
- Access to compute. Increasingly a moat in its own right. Securing gigawatts of capacity across AWS, Google Cloud, Microsoft, and others is something very few companies can do.
- Proprietary data. Usage data, feedback loops and domain-specific data that improve models over time.
- Customer relationships. The hundreds of thousands of businesses already building on the platform and the switching costs associated with integrations, evaluations, and security reviews.
- The ecosystem. Developer tools, agents, integrations, and the partners building on top of them. Claude Code is a good example of a product that integrates the model more deeply into daily work.
Our view is that the model alone is the least durable part of that bundle. The parts that compound are compute access, distribution, and the ecosystem. That distinction matters a great deal when we look at what happens as models become commoditized.
The commoditization of AI
Commoditization is often misread as “every model becomes the same”. That is not what is happening. On the hardest reasoning and agentic tasks, closed frontier models still hold an edge, and practitioners still recommend keeping them in the mix for the hardest, highest-risk, or most ambiguous work. What is happening is subtler, and more important for founders.
For a growing share of everyday workloads, several models are now good enough. When that is true, you’ll often find the following:
- Pricing power drops, because buyers can credibly threaten to switch. The price of top-tier intelligence has already fallen fast: GPT-4 launched at $30 per million input tokens in March 2023, and GPT-4o cost $5 per million input tokens about 18 months later, an 83% drop.
- Switching becomes easier as APIs look increasingly similar, and routing platforms such as OpenRouter let developers access models from multiple providers through a single interface. Changing providers becomes a configuration change, not a rebuild.
- Value migrates to other layers of the stack, where differentiation is harder to copy.
Open vs closed: mapping the main players
The line between proprietary and open models is now one of the most important strategic choices in AI, and the main players have taken clear positions.
The US labs have mostly gone closed. Anthropic sells its models only through its API and cloud partners, and released Opus 5.5 days before its prospectus leaked to counter OpenAI. OpenAI’s flagship GPT-5.6 line launched in July 2026, although it also offers smaller open-weight gpt-oss models under the Apache 2.0 license.
Google plays both sides, with proprietary Gemini models alongside open-weight Gemma 4, also under Apache 2.0. The biggest change has come from Meta, once the champion of open models: it has released no new open-weight Llama since April 2025 and moved its frontier work to a closed model line called Muse.
Europe’s main contender stays open. Mistral has released its Mistral Large 3 family under Apache 2.0 and is positioned as the vendor-backed open option for enterprise deployments.
China now leads the open frontier. DeepSeek released V4 and V4 Pro in April 2026 under the permissive MIT license, and V4 Pro scores 80.6% on SWE-bench Verified, the strongest published open-weight result on that coding benchmark.
Alibaba’s Qwen team has the fastest release cadence, shipping Qwen 3.5 in February and Qwen 3.6 in April, then Qwen3.8-Max, a 2.4-trillion-parameter model, in August, with a promise to release its weights. Zhipu’s GLM-5.3 and Moonshot’s Kimi K3 currently top the open-weight rankings on the Artificial Analysis Intelligence Index.
What stands out when we are looking at these developments? First, the open frontier is increasingly Chinese, while the largest US labs have moved further toward closed models. Second, the gap on well-defined tasks has narrowed sharply.
On narrow coding benchmarks, the best open models now sit close enough to closed frontier models that benchmarks alone no longer settle a buying decision, often at a fraction of the price. DeepSeek V4, for instance, launched at $1.74 per million input tokens.
If the model becomes a commodity, where does the value go?
These terms get used interchangeably, but they mean different things for a startup’s risk profile:
- Proprietary models are accessed through an API. You get top capability, and the provider handles infrastructure, but you depend on their pricing, roadmap, and terms.
- Open-weight models let you download and run the trained weights, subject to a license. Licenses vary a lot. DeepSeek V4 is licensed under MIT, and Gemma 4 under Apache 2.0, both free to fine-tune and use commercially with no royalties. Others carry custom terms: Kimi K3 uses its own license, and Qwen’s larger models require a revenue-sharing agreement for providers earning more than $50 million a year.
- Truly open-source AI is a stricter standard. Under the Open Source Initiative’s definition, a model must share its weights, its full code, and enough information about its training data for a skilled person to recreate a substantially equivalent system. Most models marketed as open source are really open-weight; by that standard, even Meta’s Llama did not qualify, because it kept its training data private and restricted commercial use.
The common thread is that value comes from being embedded in the environment where customers actually work.
We explored this in our earlier piece, Product–Market Fit in the Age of AI: product-market fit in AI is a system-level question. The product needs access to the customer’s data, alignment with existing workflows, and a clear role in daily operations. A product that plugs into the finance team’s approval chain or the support team’s ticketing system is much harder to replace than one that simply wraps a model.
Why can being model-agnostic be an advantage?
For many startups, the smartest position is not to bet on one model but to be able to use any of them. A model-agnostic architecture lets you route each task to the cheapest model that does it well, move quickly when a better model ships, negotiate with providers from a position of strength, and offer customers who need it an open-weight, self-hosted option.
It takes investment in evaluation, routing, and abstraction layers, and the hardest tasks may still need a specific frontier model. But as the frontier keeps moving, the ability to swap the engine without rebuilding the car becomes a competitive advantage in itself.
“The first question we ask is: if the underlying model becomes twice as good and ten times cheaper tomorrow, does this company become stronger or irrelevant? The defensible AI businesses are the ones where better models improve the product, while the real moat sits elsewhere — in proprietary data, workflow ownership, distribution, customer relationships and the feedback loops built around usage.”
— Dan Mihaescu, Partner at GapMinder
It’s fundraising season: what investors look for in AI companies
The Anthropic IPO will not change what makes a good early-stage company. It will, however, sharpen the questions investors ask. When the most valuable AI company in the world is judged on margins, compute intensity, and customer concentration, Seed and Series A investors will apply the same lens earlier.
Here is what we look for at GapMinder, and what we think founders should prepare for.
Measure cost per successful outcome, not just cost per token
Cost per token is an input metric. It tells you what the model costs, not what the customer gets. The metric that matters is cost per successful outcome: a resolved support ticket, a reconciled invoice, a merged pull request, a qualified lead.
A more expensive model that succeeds on the first attempt can be cheaper per outcome than a cheaper one that requires three retries and human review. It also connects your unit economics directly to the value you create, which pricing, gross margin, and customer ROI all depend on.
Founders who can show how their cost per outcome has fallen over time, and why it will continue to fall, give investors a far clearer picture than a token bill.
Know where open models make sense and where you need the frontier
Most products should not run on a single model. A useful rule of thumb:
- Open or smaller models tend to win on high-volume, well-defined tasks such as classification, extraction, summarisation and routing. They also make sense when data residency, on-premises deployment or cost predictability are hard requirements, which is common with European enterprise and regulated customers.
- Frontier proprietary models remain necessary for the hardest, most ambiguous or highest-risk work: complex reasoning, long agentic workflows and tasks where an error is expensive.
The strongest teams we meet treat this as a portfolio decision and can explain, task by task, why each model is in the stack.
Closing thoughts
Whatever price Anthropic lists at, the IPO will do something no private round could: put a live, public number on what frontier AI is worth. If the market embraces a $2 trillion valuation, it will be a signal that investors believe AI can become a high-margin platform business. If the stock struggles, the message will be that growth alone is no longer enough and that compute costs, concentration, and competition have to be priced in.
Either way, the direction for founders is the same. Models will keep getting better and cheaper. The companies that win the next phase will not be the ones with access to the best model, because everyone will have that. They will be the ones that turn models into measurable outcomes, embed themselves in how their customers work, and build value that grows as the cost of intelligence falls.
If you are building an AI or B2B SaaS company in Central and Eastern Europe and raising this season, we’d like to hear from you.
