There are times when the market shifts not just its leader but the very question at stake. For the past two years, the answer to the AI question has been simple: in order to gain exposure to artificial intelligence, you had to invest in the underlying infrastructure.
GPUs (graphics processing units), HBM (high bandwidth memory), foundries, advanced packaging, optical interconnects, data centres, energy … the market was not yet seeking to identify where AI would create the most end-user value; it was primarily looking to see who controlled the infrastructure.
SpaceX’s listing on 12 June illustrated this dynamic. This multi-trillion-dollar IPO above all confirmed the power of retail investment flows: valued at $1.7tn (£1.27bn) at launch, the company gained around 19% on the first day and a further 19% the following session, turning its IPO into an immediate victory for early buyers. Trading remained very buoyant in the sessions that followed.
What matters is not so much the volatility as the symbolism: investors are willing to value strategic infrastructure – launch capabilities, satellites, connectivity, orbital transport – as a technology mega-cap. In the space of a week, 11 leveraged ETFs/ETPs linked to SpaceX were reported to have accumulated nearly $3.5bn in trading volumes, with assets under management approaching $1bn.
“In the early stages of a technological revolution, the profits go first to those who build the roads. In AI, these ‘roads’ are semiconductors, memory, foundry capacity, server racks, megawatts and networks.
So far, the message has been simple: the centre of gravity of AI lay not in the applications, but in the physical stack. The market was not rewarding ‘AI’ in general, but rather the bottlenecks.”
SpaceX is often regarded as a strategic asset comparable, in the modern economy, to the East India Company during the era of great maritime exploration. The analogy is imperfect but it does remain illuminating. The East India Company’s power did not stem solely from the trade in goods – it was rooted in its control of trade routes, ports and logistics, as well as state support.
SpaceX now occupies a position in space that echoes this logic: launch cadence, reusable rockets, Starlink, Starshield, defence programmes and now the orbital data-centre project – an attempt to extend the logic of space infrastructure to AI compute, even if the economics of this model remain to be demonstrated. This is not merely a growth enterprise but an infrastructure of sovereignty.
The parallel with AI is clear. In the early stages of a technological revolution, the profits go first to those who build the roads. In AI, these ‘roads’ are semiconductors, memory, foundry capacity, server racks, megawatts and networks. The market has therefore rewarded capacity providers before rewarding end-users. This makes sense: without compute, there is no model; without HBM, there are no high-performance GPUs; without energy, there are no data centres; without a network, there is no large-scale inference.
Concentrated pool
The breakdown of the profit pool for AI-related semiconductors makes this logic almost mechanical. Out of an estimated global net profit of more than $600bn in 2026, the US is expected to capture around $314bn, South Korea $223bn, Taiwan $47bn, mainland China $26bn, Japan $14bn and Europe $14 billion. This industrial map closely resembles a stockmarket performance chart: where the profit pool is concentrated, capital flows follow; where scarcity is evident, valuations rise.
The breakdown by company is even more telling. Nvidia alone is estimated to account for around $207bn, representing nearly two-thirds of the US profit pool for AI semiconductors. South Korea has become a two-horse race in HBM, with SK Hynix at around $113bn and Samsung Memory at around $109bn.
Taiwan remains dominated by TSMC, reflecting the foundry’s role as a key bottleneck. In the US, Micron provides exposure to memory, while Broadcom offers exposure to ASICs (application-specific integrated circuits) and networking. In Europe, ASML remains the critical node for lithography.
So far, the message has been simple: the centre of gravity of AI lay not in the applications, but in the physical stack. The market was not rewarding ‘AI’ in general, but rather the bottlenecks. This is what drew index weightings, passive flows and active risk budgets towards the same names. The market was following the profit pool.
In the short term, the market is favouring scarcity: GPUs, data-centres, electricity ... in the longer term, it will judge the ability to convert capital expenditure into cashflow.”
Yet the next stage could be less linear. Agent-based AI is changing the nature of demand. It is not limited to a one-off query in a chat interface. It introduces agents that work, iterate, write code, call tools, verify, correct and execute workflows. The consumption of tokens is becoming more continuous, more voluminous and harder to predict.
For semiconductors, this is a clear boon – more agents mean more compute – but it also marks the start of another debate. If agent-based AI finally enables genuine monetisation – software productivity, process automation, improved diagnostics, accelerated medical research, better utilisation of proprietary data – then the market could begin to re-evaluate other links in the chain. Orchestration software platforms, verticalised software companies with domain-specific data, could regain some of the attention currently captured by semiconductors.
This is particularly true in healthcare. Models become more useful when they encounter rich data, complex workflows and use-cases with high economic value. Diagnostics, drug design, clinical trials, medical documentation, compliance, patient relations, laboratory productivity – in these areas, AI is not merely a support layer. It can become a driver of efficiency integrated into the production system. Companies that possess the data, customer relationships and business expertise could then capture more sustainable value compared to mere distributors of generic tools.
The same logic applies to vertical software. Generating code is becoming less and less of a differentiator – building a reliable, integrated, secure and monetisable product remains crucial. The likes of Veeva in healthcare, Workiva in regulatory documentation, IQVIA in medical data, CCC in insurance – these companies do not merely sell software; they sell a deep understanding of a sector, proprietary data, critical workflows and an established distribution network. Agent-based AI can enhance this value rather than destroy it.
Central question
The central question therefore becomes: will semiconductors remain the centre of gravity, or will they merely become the first stage of a broader cycle? The answer will depend on monetisation. If tokens remain primarily a cost, the scarcity of compute will remain the dominant narrative. If, on the other hand, tokens become revenue, then margins, then free cashflow for end-users, the market will gradually shift some of its focus towards software, platforms and data-owners.
This shift does not mean the semiconductor theme is over. On the contrary – agentic AI can extend the capex cycle. Figures cited by Taiwanese tech giant Hon Hai highlight the industrial scale of the issue: a 1 GW Nvidia Vera Rubin data-centre would cost nearly $47bn, with an annual electricity bill of around $1.3bn. We are not talking about a lightweight software cycle, but a capital-intensive, energy-hungry cycle – often financed by debt.
This is where return on invested capital becomes central once again. In the short term, the market is favouring scarcity: GPUs, HBM, CoWoS (chip-on-wafer-on-substrate), data-centres and electricity. In the longer term, it will judge the ability to convert capital expenditure into cashflow. Some hyperscalers may temporarily accept negative free cashflow to pre-fund future demand but, if credit windows close, rational players will have to return to a sustainable balance sheet before their competitors do.
A distinction must be made between the primary beneficiary of the revolution – capacity – and the secondary beneficiaries – monetisation.”
The right strategy is therefore neither to sell semiconductor stocks too early nor to believe they will remain the sole focus. A distinction must be made between the primary beneficiary of the revolution – capacity – and the secondary beneficiaries – monetisation. Today, scarcity remains on the side of silicon, memory, foundry capacity and energy. In the future, if agent-based AI delivers on its promises, scarcity could also shift towards data, workflows and distribution.
For the time being, the market functions primarily as a voting machine: it votes for the visible constraints. The weighing machine will return, however – and it will weigh tokens converted into revenue, revenue converted into margins and margins converted into free cashflow.
In this transition, semiconductors will likely remain the bedrock of the cycle, but no longer necessarily its sole centre of gravity. The next phase of AI will not involve abandoning the infrastructure, but identifying which players, above that infrastructure, actually manage to transform usage into value.
Xiadong Bao is a portfolio manager, international equities at Edmond de Rothschild Asset Management
