The fallout from the leaders of the biggest AI companies calling for a slowdown in the race to build an AI superintelligence potentially capable of wiping out humanity has been dominating headlines for most of the last week.

Whether the lab chiefs are genuinely worried about an existential threat or protecting their own position is hard to know from the outside (probably a bit of both). We would argue that from an investment perspective the near-term outcome probably looks the same either way. Simply being long semiconductors and short software is well over now. It's down to picking the right companies within each sector.

The race from smarter to cheaper

The AI race has shifted from building the smartest model to making it the cheapest to run. Cheaper open-source models, primarily Chinese, have already closed the gap with frontier models to just six months. This parity created pressure to stay ahead and invest in expensive training. Now, however, frontier labs are slowing down, either because of existential risk or to show near-term profitability ahead of an IPO, which means that the gap will close much sooner.

Training a model means feeding it enormous amounts of data for months using rows of chips running flat out. Inference is what happens every time someone actually asks the model a question. And while costly training only happens once, inference happens billions of times a day, creating a perpetual demand for chips and electricity. High-margin inference fees were supposed to pay back the original, massive training costs. But with open-source models rapidly closing the performance gap, the debate over whether AI can ever deliver a real return on investment will continue to run hot.

Fundamental AI trade still strong

That said, the AI supply chain still looks fundamentally strong. Our analysis shows demand massively exceeds supply, with order visibility out to 2028. One industry argument is that the death of Moore's law, combined with how central this technology has become, has turned semiconductors from a cyclical industry into a structurally growing one. Nvidia CEO Jensen Huang talks of compute as an asset class growing at 100% a year, using cash flow to expand supply, demand and even the electricity grid needed to run it all.

Time will tell if that holds. Industries don't usually move from being cyclical to structurally growing without genuine economic growth behind them, and right now much of that growth is coming from the AI infrastructure investment itself, rather than from any proven return on it.

Many of the industry experts we speak to remain optimistic over the medium term, but near-term prices are being driven by positioning rather than fundamentals. Everyone is simply watching everyone else, which explains this summer’s volatility.

Figure 1: Volatility in chip stocks has swung sharply

Source: Bloomberg, as of 15 September 2026. Equal-weighted average of one-month at-the-money implied volatility (30-day tenor, mid-market) across the top 15 constituents of the PHLX Semiconductor Sector Index (SOX) by index weight.

Problems loading this infographic? - Please click here

The software pockets

In software, it’s all about customer ownership. Most companies believe they can wrap open-source models inside their own systems and keep selling their own products to customers. We see others taking a riskier path and are making their software directly accessible through third-party tools like Anthropic’s Claude, describing it as “meeting the customer where they are". To us, this looks like a defensive move rather than a growth strategy. Hand over the customer interface to someone else’s AI, and you lose the power to set your own prices later.

If model intelligence were to stop improving as quickly, the focus shifts to implementation and productisation, for enterprise customers and labs alike. The broad software trade is over, and what's left is stock selection, with pockets of opportunity for those willing to look closely at individual businesses.

Vibe coding, bugs and cybersecurity

A clear winner out of the concerns over increasing AI potency is cybersecurity, which gets paid regardless of how the safety debate resolves. Budgets are rising and AI coding tools cut both ways here. The same agents that write software faster are also better at finding weaknesses in other people's code, raising the stakes for defenders. Meanwhile, the rush toward AI-generated code, often called vibe coding, means more software written quickly, which tends to mean more bugs and technical debt to fix later.

Figure 2: Cybersecurity has pulled sharply ahead of software in 2026

Source: Bloomberg as of 16 September 2026. Global X Cybersecurity ETF (BUG) vs iShares Expanded Tech-Software Sector ETF (IGV), normalised to 31 December 2025.

Problems loading this infographic? - Please click here

We are also watching consumer AI closely. Historically, we have seen individuals adopting personal assistants far faster than enterprises, picking them up immediately and finding them easy to use. Teaching an assistant your habits, calendar and preferences takes real effort, which most people will only expend once. Whichever assistant earns that initial effort will likely keep it. As with previous technological trends, following consumer adoption helps track cultural shifts, which helps us uncover consumer internet opportunities and offers a lens to understand how enterprises might eventually adopt the technology.

Putting it all together, a slowdown at the frontier would buy time, giving IT services companies, software companies and enterprise customers room to catch their breath rather than chasing whatever model was released last month.

All data Bloomberg, unless otherwise stated.

Author: Sumant Wahi, a portfolio manager at Man Group, focusing on technology equities. With assistance from Mipham Samten, an analyst on the technology equities team at Man Group.