AI must now deliver

Read the market analysis and fund positioning

The hype surrounding AI is being put to the test. Data centres need to be set up, capacity needs to be booked and investments totalling billions require recouping. This also shifts the opportunities and risks within the portfolio. In this next phase, we are examining which companies can already present orders and revenue, and which still have to prove themselves.

In July, Alphabet's share price fell by around 6% following a strong quarter. Expansion costs impacted the share price, with record investment in data centres pushing free cash flow into negative territory for the first time. Meanwhile, at Meta, the free cash flow plummeted by 91% and the share price fell by almost 10% at times. By contrast, Microsoft and Amazon rose significantly because their expansion is already underpinned by cloud contracts and booked capacity. Same expenditure, contrasting reactions. On 26 August, Nvidia once again delivered exceptional figures: $96.2 billion in revenue, which is 106% more than in the previous year. Of this, $89 billion came from the data centre business. For the current quarter, the group is forecasting a further increase to around $108 billion. The key factor will be who can build the infrastructure in time, utilise it to full capacity, and convert it into cash flow.

Phase II has begun

In Phase I, the key factor was the amount that corporations intended to invest. Alphabet, Amazon, Meta and Microsoft are planning to invest a combined total of well over 700 billion dollars by 2026. Phase II focuses on the outcome of these budgets. Order books and utilised capacity show how far a company has progressed with implementation. For our portfolios, we therefore focus on providers whose order books are filling up as a result of this expansion. At the same time, we assess the new risks arising from the high capital requirements.

The chip is no longer the only bottleneck

This is because construction is taking longer than promised. Of the approximately twelve gigawatts of US data centre capacity that had been announced for 2026, which is roughly equivalent to the output of twelve large nuclear power stations, only about a third was actually under construction in spring. There is often insufficient access to reliable power supplies. At the same time, 2.06 terawatts of new generation and storage projects are awaiting grid connection on the electricity supply side. Not all of these projects will be realised, but the scale of the figure illustrates the bottleneck. Large transformers now have lead times of three to four years. Therefore, a single component costing tens of millions of dollars can hold up a data centre project worth billions. This creates opportunities for infrastructure providers such as NVIDIA and ASML, power generators Vistra and Constellation Energy, Vertiv in power distribution and cooling, and GE Vernova in turbines and grid equipment. What matters to us is that this expansion is already reflected in order books. We also assess how much of this growth has been factored into the valuation.

China has caught up, particularly in terms of open-source models and pricing. For example, Qwen is now one of the largest foundation model ecosystems on Hugging Face, while DeepSeek and Kimi are increasing competitive pressure by offering powerful LLM models at low prices. However, access to state-of-the-art computing power is still limited due to export controls and local chip shortages. In our portfolios, we are therefore focusing on Alibaba for Chinese software and cloud services. Furthermore, many new AI projects in Europe, the Middle East and India continue to rely on the same global supply chains for lithography, accelerators, networks, power and cooling.

The AI cycle extends well beyond the confines of the chat window

While AI is already being used in measurable ways today, it rarely takes the form of a chatbot. For the first time, the Pentagon has allocated $13.4 billion to autonomy and autonomous systems as a separate category in its 2026 budget. Ukraine is training recognition and targeting systems using millions of real combat images. In medicine, MSD (Merck & Co. in the US) and Moderna achieved two key objectives of a Phase III trial with a personalised mRNA cancer therapy on 19 August. These examples illustrate just how far our vision of new applications must extend beyond the major AI laboratories.

 Lessons from SpaceX's Initial Public Offering (IPO)

SpaceX went public on 12 June, initially rising to $225.64 before falling back to close to the issue price in the following weeks. The company's first quarterly results as a listed company revealed why the market is paying closer attention: the Starlink connectivity business generated $4.3 billion in revenue and $1.7 billion in operating profit in the second quarter. However, Starship and the AI business require substantial investment. Investment in the AI sector alone totalled $15.8 billion.

In the Ethna-AKTIV, we hold the SpaceX bond. Although it is an unsecured liability of the group as a whole, it is Starlink’s visible revenue base that counts towards the credit profile. By investing in the stock, investors are also funding Starship and the company’s AI ambitions. While this may pay off, the IPO has demonstrated how swiftly a promising business plan can be reassessed by the capital market once it sees the figures. Our decision to opt for the bond demonstrates our ability to distinguish between a company’s potential and the appropriate risk profile for the portfolio.

Next up are the AI laboratories

IPOs for Anthropic and OpenAI also appear to be on the horizon, although no firm dates have yet been set. Unlike the most speculative tranches of SpaceX's history, both laboratories are backed by rapidly growing tangible revenue. However, they are investing in computing capacities on such a large scale that gross margins, cash flow and financing have become key considerations for investors.