Is the mania phase finally starting for AI?

What happened

Several high-profile OpenAI announcements have dominated headlines in recent months:

• OpenAI’s fundraising has sent its valuation soaring.
• A $300bn agreement was unveiled with Oracle, under which OpenAI would purchase compute capacity.
• Nvidia announced plans to invest up to $100bn in OpenAI. Nvidia would provide cash to OpenAI in exchange for equity, and OpenAI would buy the latest Nvidia chips for new datacenters.

These spectacular deals, which may look like circular funding, have fueled growing concerns about the emergence of a bubble.

Impact on our Investment Case

Sketchiness rather than commitments

Recent announcements were full of ambitious claims, but hard commitments were harder to spot. In our view, each deal contains blind spots that warrant caution.

For the Oracle partnership, the key questions are around funding and actual need: can OpenAI afford it, will it have sufficient use for such capacity, and can Oracle realistically deliver datacenters on time given supply chain and construction constraints?

The Nvidia announcement feels more like promotional noise than substance. For now, there are no binding commitments, only letters of intent, and timelines remain vague. More concerning are similarities with vendor-financing patterns from the dotcom bubble: Nvidia funding customers to buy its products, who then leverage them to buy more (e.g., CoreWeave - not in our portfolio -, or private players like Lambda).

At the same time, Meta’s hiring spree at inflated salaries echoes bubble-era behavior, with the added bonus of witnessing employees already returning to OpenAI, presumably with even-higher compensation.

OpenAI as a systemic risk?

The real issue is OpenAI’s spending trajectory, which seems detached from financial common sense: $100bn planned for backup servers and $350bn for cloud services over five years, supported by projected revenues of $200bn annually (greater than Meta in 2024). The company also expects $115bn in losses before reaching profitability. The question is whether investors will continue funding at this pace without clearer evidence of a path towards profitability: the company is currently losing money on every request, even from higher-tier plans, and increasingly complex requests only accelerate this trend.

Valuations are accelerating just as competition intensifies. OpenAI was valued at $160bn in October 2024, $300bn in March 2025, and is reportedly targeting $500bn in the current financing round, which was set to allow employees to cash out. Meanwhile, China is moving quietly but aggressively with a cost-efficient approach that could undermine this capital-intensive model while building an alternative independent ecosystem.

Rising valuations

This enthusiasm is at the moment clearly visible in hardware, as Nvidia is the main benefactor of this spending spree along Broadcom (both in our portfolios).  Forward P/Es in semiconductors have remained aligned with the software-oriented segments since the start of the “ChatGPT trade”, meaning semiconductors are at a historical valuation premium. 

Although earnings revisions for Semiconductors have ticked up recently, partially justifying these valuations for some investors, we believe the dependency on OpenAI makes the outlook fragile. In addition, past cycles shown in the chart below give little reason for optimism about sustainability.

On a stock level, Oracle appears extremely expensive relative to its history (top chart), with its exposure to OpenAI creating major concentration risk, especially in light of the recent rally. Nvidia, in contrast, looks relatively cheaper when factoring in EPS growth (bottom chart - missing data corresponds to negative values not showing on a logarithmic scale), though its fortunes are also tied closely to OpenAI’s spending trajectory.

Our Takeaway

Our focus has always been on pure-players, which naturally creates a mid-cap bias: the weighted median market cap of our strategy is about $30bn, compared with ~$1.4tn for the Nasdaq 100. We hold no Magnificent 7 stocks except Nvidia,which has been in the portfolio since the strategy’s 2015 launch, and which is arguably the “purest” large-cap player. This also limits our involvement with the generalist chatbot players dominating headlines.

In addition, since 2022, we have steadily increased exposure to data (critical for model customization) and applications, where we target productivity, security, and healthcare. These segments offer both solid visibility and sufficient maturity to identify leaders.

As a result, the portfolio is positioned in software companies still at the early stages of their acceleration curves, many of which delivered strong earnings despite AI representing only part of their growth (e.g., Mongodb), while valuations have normalized compared to historical averages.

Still, we remain vigilant on valuations, and have adjusted exposure with some profit-taking over the summer. For instance, Palantir is no longer our top holding, although we have not exited the position completely as there may be further upside left in this cycle. As a result, as of 26 September, the portfolio PEG stood at 1.75 vs. 1.91 for the Nasdaq 100 and... 2.19 for the MSCI World. At the same time, we are already preparing the allocation for the next cycle, ensuring the portfolio remains one step ahead.