Hedge funds are making a bet that sounds almost too simple: if artificial intelligence is the engine, then semiconductors are the fuel system. The latest rotation in positioning—away from parts of the software stack and toward chipmakers and the broader “AI infrastructure” supply chain—reflects a shift in how many investors think AI value will be captured over the next phase of the cycle.
This isn’t a claim that software is suddenly irrelevant. It’s more precise than that. The market is increasingly treating software as a layer where differentiation is harder to sustain and margins can compress as capabilities become commoditized. Meanwhile, the physical bottlenecks—compute capacity, memory bandwidth, high-speed networking, packaging, power delivery, and manufacturing throughput—remain stubbornly real. When demand for AI accelerates, those constraints don’t disappear because models get smarter; they intensify because training and inference both require more hardware.
In other words, the “software is out, semis are in” framing is less about ideology and more about economics: where the scarcity is, where the capex flows, and where the supply chain can actually scale.
A portfolio rotation with a logic investors can explain
Hedge funds typically don’t move money based on vibes. They move when the risk/reward changes and when they can articulate a thesis that survives contact with earnings reports, guidance updates, and industry data. The current rotation toward semiconductors is being justified through three interlocking arguments.
First, AI spending is increasingly tied to infrastructure build-outs rather than just software subscriptions. Enterprises may still buy platforms, tools, and model access, but the biggest incremental budgets often show up as data center expansion, GPU procurement, networking upgrades, and storage/memory improvements. Those are not abstract line items; they translate into orders, lead times, and utilization rates.
Second, the “picks-and-shovels” narrative has regained credibility. In past technology waves, investors learned that the companies selling the enabling components frequently benefit even when end-user applications evolve quickly. AI is different in its pace, but the underlying pattern—hardware enabling the wave—remains familiar. If AI adoption is the headline, chips are the supply chain that makes adoption possible at scale.
Third, the market is recalibrating what it expects from software. Many software businesses have benefited from AI enthusiasm, but the competitive landscape is changing. Model capabilities are improving rapidly, and open ecosystems are lowering barriers to entry for certain categories of tooling. That doesn’t eliminate software opportunity, but it changes the distribution of returns. Investors are asking: which software layers will remain defensible, and which will become features bundled into platforms?
Semiconductors, by contrast, sit closer to the constraint. Even if software becomes cheaper or more widely available, compute demand still needs to be met with silicon and systems engineering.
Why the “bottleneck” story matters more than ever
The most compelling reason semiconductors are getting attention is that AI is not a single product—it’s a workload. Training workloads are compute-hungry and memory-intensive. Inference workloads are latency-sensitive and throughput-driven. Both require specialized architectures and careful system design. That means the bottleneck isn’t only “can you run AI?” It’s “can you run AI at the required cost, speed, and reliability?”
When investors talk about AI infrastructure, they’re really talking about a stack:
Compute: GPUs and accelerators designed for matrix operations and parallel workloads.
Memory: high-bandwidth memory and large capacity configurations that prevent stalls.
Networking: high-speed interconnects that keep distributed training and inference from becoming communication-bound.
Packaging and integration: advanced packaging techniques that improve performance per watt and enable higher density.
Power and cooling: data center engineering that determines whether hardware can operate continuously at target performance.
Manufacturing and yield: the ability to produce enough units at acceptable yields and timelines.
Software can optimize workflows, but it can’t fully substitute for missing capacity. If demand rises faster than supply, the price of compute capacity tends to reflect that imbalance. That’s why investors are watching not just chip demand, but also supply chain execution—foundry capacity, advanced node availability, packaging throughput, and component lead times.
The rotation is also a response to how markets price risk
There’s another subtle dynamic behind the shift: how investors price uncertainty. Software valuations often embed expectations about user growth, retention, and monetization. When sentiment turns, those expectations can compress quickly because software revenue streams can be perceived as more elastic or more competitive.
Semiconductor valuations, while also sensitive, are often anchored to tangible drivers: order books, backlog conversion, utilization, and guidance tied to production schedules. Even when the sector experiences volatility, the narrative tends to be more grounded in measurable industrial activity.
That doesn’t mean semis are low-risk. They carry their own uncertainties—cyclicality, export controls, customer concentration, and technology transitions. But in the current environment, many hedge funds appear to believe the near-term risk is skewed toward continued AI-related demand rather than a sudden collapse in infrastructure spending.
So the rotation can be read as a bet on the shape of the next few quarters: that AI capex remains resilient and that the “infrastructure build” phase is not finished.
What “semiconductors” really means in this thesis
It’s easy to say “semis,” but the trade is rarely limited to a single category. The AI infrastructure thesis tends to broaden into multiple segments that each play a role in delivering usable compute.
Accelerators and GPUs: The most visible beneficiaries, because they directly execute AI workloads.
Memory and bandwidth suppliers: Often overlooked until shortages or pricing power emerge. AI workloads stress memory subsystems heavily.
Networking and interconnect: As models scale and distributed training becomes more common, networking becomes a performance limiter.
Data center components: Power delivery, cooling, and system-level integration can matter as much as raw chip performance.
Foundry and manufacturing ecosystem: Advanced nodes and packaging capacity determine whether demand can be met.
This is why the rotation can look like a broad “AI infrastructure” tilt rather than a narrow bet. Hedge funds want exposure to the entire chain that converts AI demand into revenue.
A unique take: the software layer is becoming a distribution layer
One reason software may be deprioritized is that the market is increasingly treating software as a distribution layer for models rather than the core scarce resource. In earlier phases of AI adoption, software companies could differentiate through proprietary data, workflow integration, and domain-specific models. Today, foundation models and model APIs have changed the baseline. Many software products now compete on usability, integration, and vertical specialization—areas where differentiation is possible, but where the competitive intensity can be high.
Meanwhile, the “scarce resource” is shifting toward compute availability and system performance. If customers can access models easily, but can’t afford to run them at scale without expensive infrastructure, then the economics favor the providers of that infrastructure.
This doesn’t mean software margins will collapse across the board. It means investors are more cautious about assuming software will capture the majority of incremental value. Instead, they’re looking for where the marginal dollar spent by enterprises goes first—and that often points back to hardware.
The data center is the real battleground
AI is ultimately a data center story. Every improvement in model capability translates into more compute cycles. Every new use case—customer support automation, coding assistance, document processing, robotics, analytics—adds inference demand. Even when individual deployments are small, the cumulative effect can be enormous.
That’s why hedge funds are paying attention to data center capex signals: announcements of expansions, procurement patterns, and the pace at which facilities can be brought online. The semiconductor thesis is strongest when it aligns with evidence that data centers are scaling fast enough to absorb new hardware.
In this context, semiconductors are not just components; they are the “rate-limiting step” for AI deployment. If the industry can’t deliver enough accelerators, the rollout of AI applications slows. If it can deliver, adoption accelerates—and the cycle reinforces itself.
The rotation also reflects a timing strategy
Another practical reason for the shift is timing. Software can rally on announcements and partnerships, but the monetization timeline can be longer and more uneven. Semiconductors often have a clearer path from demand to revenue recognition, especially when orders translate into shipments and when guidance reflects forward visibility.
Hedge funds frequently aim to position ahead of earnings inflections. If they believe AI infrastructure spending is accelerating, they may prefer to be exposed to companies whose results will reflect that acceleration sooner.
This is particularly relevant when the market is transitioning from “AI as a concept” to “AI as a procurement cycle.” Once procurement begins, the hardware supply chain can show momentum quickly.
What could break the thesis?
No rotation is guaranteed. The semiconductor thesis could face headwinds if several scenarios occur:
AI demand could slow if macro conditions tighten and enterprises delay capex.
Supply could improve faster than demand, reducing pricing power.
Export restrictions or geopolitical constraints could limit addressable markets.
Technology transitions could create inventory risk if customers shift architectures faster than expected.
Competition could compress margins if multiple suppliers chase the same orders.
However, hedge funds are not ignoring these risks—they’re managing them through position sizing, diversification across the supply chain, and active monitoring of guidance. The key difference is that, at present, many appear to view these risks as less dominant than the structural demand for AI compute.
Why this feels like a “phase change” rather than a temporary trade
The phrase “AImaxxing” captures the cultural side of the moment, but the underlying market behavior suggests something more durable. The shift toward semiconductors aligns with a broader transition in how AI is deployed:
From experimentation to production.
From isolated pilots to integrated systems.
From single-model demos to multi-model, multi-tenant environments.
From occasional inference to always-on workloads.
As AI moves into production, the cost structure becomes central. Customers care about total cost of ownership: performance per watt, throughput per rack, and the ability to scale without rewriting everything. Those concerns naturally elevate the importance of hardware efficiency and system