Last week, news hit the market that artificial intelligence (AI) research and development startup Anthropic could become a listed company as early as October, with a touted $2tn valuation.

This is a truly meteoric rise for a company very few had heard of even a year ago. However, market sentiment towards the AI complex has become more mixed of late, with more focus on competition between models than we’ve ever seen before.

Fear and greed swinging in both directions

Anthropic and OpenAI are the world’s largest and most advanced large language model providers (LLMs). LLMs can be considered the engines of generative AI applications. By this we mean AI that can generate things on prompts from a user. These two companies account for the lion’s share of AI revenue and sit at the epicentre of the AI trade.

A major debate facing investors is where value for them will accrue in the AI sector: among the ‘picks and shovels’ companies that don’t provide AI but make it possible, the data centre operators, the model layer, (companies like Anthropic and OpenAI that build and train the AI itself) or the application layer (the businesses that plug into those models to build everyday products, from customer service tools to coding assistants)?

Since early this year, the release of products by Anthropic that compete at the application layer, combined with an unprecedented surge in spending on data centres, has triggered a rotation out of tech software in favour of technology hardware and AI enablers, companies involved in the build-out of AI. As faith in long-standing competitive advantages shrank, so did valuation multiples of established winners. Specifically, software and data stocks, steady compounders of returns for much of the last decade, took a beating.

Public investors instead expressed AI enthusiasm through infrastructure beneficiaries, ranging from semiconductors and networking equipment to power and cooling providers. As demand expectations surged, backlogs expanded, margins improved, and valuation multiples, such as forward-looking price-earnings ratios, rose.
However, of late, there’s been a slight change in tone; there’s a little more caution about the hardware names, and more a nuanced appreciation of the risks and opportunities facing the software-as-a-service (SaaS) sector, which has outperformed the widely cited PHLX Semiconductor index since the end of June. 

The software response

Now that there is more data and customer experience on AI usage in the rear-view mirror for us all, we’ve had a little time to see strategies, products, costs, and experiments. Every series of results leaves new clues, and management commentary takes us beyond the hard data.

Take models: much has been said this summer about the performance of Chinese models, their uptake, and their relatively low cost. As the cost of AI to enterprise users – businesses and other organisations, rather than individual consumers – has risen, large consumers of tokens, the primary unit through which AI usage is measured and often monetised, in software development are getting creative and attempting to wrestle back control. One such user, digital music, podcast and video service Spotify, has built an internal tool called Chirp, which it uses instead of directly using Claude Code or ChatGPT’s Codex. As Spotify Co-CEO Gustav Victor Soderstrom says:

“Chirp lets our engineers switch models mid-task and route every job to the best available price performance, including open-source models that we host ourselves, so we are never locked in. It also shares the context across not just different models and different developers, but across the company. So, we don't pay for the same reasoning twice, and we don't lose our own data.”

American educational technology company Duolingo’s CEO Luis Von Ahn has talked about reducing the cost of its AI-enabled Video Call feature from $0.30 to under $0.01 per call over the last year, mainly by using cheaper open-source models:

“The reality is the quality's indistinguishable for many applications that we use. For example, serving a call – like a conversation practice to somebody who can only speak 100 words of a language… Your model doesn't need to be as good as a philosopher.”

We’re seeing similar commentary from others like restaurant delivery service DoorDash, cloud-based IT platform ServiceNow, and cryptocurrency exchange Coinbase.  This ability to toggle to lower-cost models is important because it protects the margin structure of software companies that were at risk of seeing their value leak to the AI labs. 

Many software names continue to post solid numbers that defy fears of immediate disruption. Talk of AI driving faster product, and better and more services at lower costs, is perhaps also beginning to land. Hosting marketplace provider Airbnb CEO and Co-Founder Bryan Chesky said during its Q2 earnings call, earlier this month:

“We've reduced the time from concept to launch by as much as 60% and compared to the same six months last year, we've increased the number of features and improvements we shipped this year by nearly 80%.”

Back in February, global software company Atlassian positioned itself as an AI beneficiary with increased growth in active users and seats from its AI services. However, at the time, the market remained convinced that AI was going to reduce the requirements for software users and in so doing reduce software billings and revenue. Now the same message is landing alongside accelerating revenues. CEO and Co-Founder Michael Cannon-Brookes had this to say:

“I think there's going to be more developers in the world in five years' time than there are today. The cost of building technology is going down. The amount of technology we're going to build is going to go up…the number of people building technology that we don't necessarily call developers is also growing. So, we're seeing people in finance and in HR and in marketing using a lot of these tools to build technology of various different forms.”

Following Q2 results in early August, the stock is up over 60% across the month.

Anthropic takes the lead

Share of US businesses with paid subscriptions to competing AI models, platforms, and tools

The AI labs

The Anthropic IPO will represent one of the first opportunities for public investors to value a frontier lab (leading AI research organisations) directly, rather than through the surrounding ecosystem of companies that enable frontier labs to do what they do. Investors will effectively be betting that the model layer can achieve sufficient revenue growth and ultimately profit to justify valuations that now surpass the largest listed software companies.

While the picks-and-shovels companies may be agnostic about the end-model winner, the frontier labs’ commitments to purchase hundreds of billions of dollars of data centres has supercharged their growth outlook in the medium term. That means any hint of reduced investment and enthusiasm hurts this AI-sensitive cohort. That triggers volatility across names from $1tn Korean memory stock Samsung at the top of the benchmark to UK small-cap industrials, because of their high correlations to the AI investment theme.

Much like the Grand National

This is not to say that we’re seeing the beginning of the end for the frontier labs. They are growing revenue at an astounding pace, while innovating. According to financial technology platform Ramp, despite growing interest in open-weight models, only 6% of AI-spending businesses currently use the inference platforms most associated with them, rising from 4.5% of businesses at the beginning of this year.  The point is that at this early stage, AI is a highly dynamic market where we expect, like the Grand National, there will be frequent changes in leadership. 

For example, at the beginning of Summer 2025, Alphabet was an AI loser, before a sharp change in perception shifted the story to a full-stack beneficiary. Until 2 weeks ago, US Human Capital Management software provider Workday was firmly in the sights of the SaaSpocalypse – the current parlance for a potential destruction of software companies’ business – before news of a potential $50bn+ take-private by private equity firm Silver Lake Capital gave those waving the white flag pause for thought. 

When high valuations meet high expectations, it doesn’t take a large fly in the proverbial ointment to knock one off course. Equally, there may be bargains to be had in a group of high-quality stocks where the market has become too pessimistic about the risk of AI disruption.