What retail investors need to know before jumping into Anthropic’s IPO

You wait years for a record-breaking IPO and then two look to be turning up within four months of each other.

Back in June, SpaceX staged the largest initial public offering in history, with the company valued at $1.8tn at flotation — and with a record 20 per cent of shares allocated to retail investors.

Anthropic is targeting October for its launch — and some investors expect its debut valuation to reach $2tn, nudging the Claude maker into the top spot and Elon Musk’s AI and rockets group into lowly second place.

But is it an attractive investment opportunity?

Let’s address some important questions. First, what will the growth trajectory be? There isn’t much financial data available at this stage — we’ll have to wait for the prospectus when it’s published. What we do know is that revenues are expected to be in the region of $100bn this year — up from about $10bn in 2025.

But it’s worth interrogating those numbers more closely.

When you develop an AI agent, you pay Anthropic monthly for “tokens” to access its system and enormous computing power. During the past year, computing costs per task have shot up — semiconductor prices and energy costs have all risen sharply and Anthropic has passed these on. But it means a chunk of that impressive revenue growth is simply programmers paying more to advance existing projects.

Though not unusual for a new software company, that growth — even with my caveat — is still impressive. Some believe it will carry on for years. Certainly, the market opportunity for AI seems huge.

Recommended

Anthropic estimates that AI is capable of covering more than 80 per cent of tasks in fields including management, business and finance, computing, architecture, law, arts and the media, among others. The company’s strength lies in its Claude coding models and developer tools that can take on this work.

Don’t hang your head in despair and start worrying about your job yet. The theory is fine; the practice so far is very different.

What often seems to happen is that the person who has lost 80 per cent of their routine execution tasks now has to spend that time reviewing and debugging the AI tool’s work, to catch all the mistakes it makes and step in where it struggles with messy real-life complexity.

Soon, the finance department is on the phone, asking what happened to the expected savings from their AI investment.

Clearly, AI will get better. But this raises another issue. Can Anthropic maintain its competitive advantage?

Many people who have decided to develop an agent to handle basic processing tasks will have chosen Claude. But OpenAI’s models or Chinese alternatives, while offering more modest functionality, charge a fraction of the price. And we saw last year, with the surprise launch of DeepSeek in China, how quickly things can move in the AI space.

The next big question is over shareholder returns. What proportion of Anthropic’s revenues will end up as cash earnings? Information here is sparse. I’ve seen claims that a quarter of revenues may be available as cash.

That’s not high in the world of technology companies. Look carefully at any figures announced to see if they include stock-based compensation for employees. This is a cost of doing business in my book, even if not shown that way in the accounts.

Those rising costs I mentioned earlier could squeeze profits if tougher competition means Anthropic can no longer increase its fees. And more costs are looming. Cloudflare, whose technology can block AI crawlers from scraping websites, says it’s evolving a “pay-per-use” model that would force AI companies to pay publishers to use their content for search answers and in other ways that create value. The free and easy way of training AI models may be ending.

Are there other business risks? Erm, yes.

Anthropic has had a public row with President Trump. The US government imposed export controls on its Mythos 5 model on national security grounds, after the model used deception and fake identities to cover its tracks in cyber security stress tests.

Much of this will be shrugged off by enthusiasts. However, shareholders should always ask questions about these sorts of risks. Courts, sooner or later, protect consumers from damage, especially financial loss. Technology companies have seen few damage claims against them until more recently, but the numbers are rising.

Take the $942mn recently awarded against Meta in New Mexico (which it is contesting). Or take the trial that started in the US this week regarding claims that the company deliberately designed Facebook and Instagram to encourage compulsive use among children. The penalties there could reach nearly $200bn if Meta loses.

I will also take a negative view of this IPO if only a very small percentage of the company’s shares are floated and major exchanges give them a larger weighting than their actual public float rate.

Any attempt to rig the market by producing artificial scarcity in the short term to lift the price on flotation should give a company and its advisers a bad name. As we saw with SpaceX, a premium on launch doesn’t tell you where the price might be two weeks later.

Finally, the most important question: why is the company coming to market?

The costs of taking these AI models to the next stage seem to be inflating fast, and Anthropic may need new capital to keep its lead in the race. Managers hoping for some cash for themselves too may fear a rush of similar IPOs, worrying that this window is neither large nor likely to fund everyone.

Before the tech bubble of 2000, many investors desperate for a stake in the internet bought Freeserve and AltaVista. Big mistake! The companies that made the most from the internet floated after the bubble burst: Google in 2004, Facebook in 2012.

Just because AI is a promising area doesn’t mean you have to rush to invest.

Simon Edelsten is a fund manager at Goshawk Asset Management.