AI has been the loudest story in tech markets for more than a year, but the market’s volume is starting to change pitch. For a long time, investors treated artificial intelligence as a software narrative: models, platforms, and applications that would convert compute into revenue. Yet the latest signals suggest that the “picks-and-shovels” layer—semiconductors and memory—may be catching up to the pace of AI infrastructure spending, even as software adoption and monetisation appear to lag behind earlier expectations.
This shift matters because it changes what “progress” looks like. In the early phase of an AI boom, the most visible activity is often procurement: data centres ordering accelerators, systems integrators building racks, and memory suppliers scaling capacity. Software, by contrast, has to prove it can turn that infrastructure into sustained usage, pricing power, and measurable business outcomes. When software moves slower than investors expect, the market doesn’t stop believing in AI—but it starts looking for proof elsewhere. And right now, the proof is increasingly showing up in the hardware layer.
The result is a kind of time-lag trade. Hardware demand ramps first, because training and inference require physical capacity immediately. Software adoption can take longer, because enterprises need integration work, governance, workflow redesign, and—crucially—confidence that the outputs are reliable enough to justify operational risk. That gap between “compute bought” and “value realised” is where semiconductor and memory stocks have found renewed attention.
What’s driving the renewed focus on chips and memory?
Start with the simple reality of AI infrastructure: it is not just about having powerful processors. It is about feeding them. Training workloads are bandwidth-hungry, and inference workloads are latency-sensitive. Both depend on memory capacity and memory speed, as well as the ability of systems to move data efficiently between storage, memory, and accelerators.
In practice, this means that AI build-outs are constrained by more than one bottleneck. Even when accelerator performance improves, overall throughput can be limited by memory subsystems, interconnects, and the efficiency of data movement. That’s why memory—often overlooked in headline discussions—has become a central part of the investment conversation. If the industry is buying more accelerators, it also needs more high-performance memory to keep those accelerators busy rather than waiting.
There’s also a second-order effect: once data centres commit to large-scale deployments, they tend to standardise around proven configurations. That creates a feedback loop. As procurement volumes rise, suppliers gain confidence to expand capacity, customers gain confidence to place repeat orders, and the supply chain becomes more predictable. In that environment, hardware vendors can look less like speculative beneficiaries and more like companies with near-term visibility.
Meanwhile, software companies face a different set of constraints. Many AI products are still in the process of moving from “demo-ready” to “workflow-ready.” The market may have expected faster monetisation, but enterprise adoption is rarely instantaneous. Even when users want AI, they need to integrate it into existing systems, ensure compliance, manage data privacy, and control costs. Those steps take time, and they can slow revenue recognition even when usage grows.
So the market’s attention is shifting from the question “Who will sell the AI app?” to “Who supplies the capacity that makes the AI app possible at scale?” That’s a subtle but important change in investor psychology.
Why the “catch-up” dynamic is showing up now
The phrase “catching up” can sound vague, but in market terms it usually means one of two things: either hardware demand is arriving sooner than expected, or hardware earnings are becoming more visible while software earnings remain delayed.
In AI infrastructure, procurement cycles can be surprisingly front-loaded. Data centre operators and cloud providers often plan capacity months in advance, especially when they need to secure power, cooling, rack space, and delivery timelines. Accelerators and memory components are ordered as part of a broader build-out schedule. That schedule can make hardware demand appear to lead the software story.
Then there’s the reporting lag. Semiconductor and memory companies typically report results based on shipments and inventory movements, which can reflect demand earlier than software companies’ revenue lines. Software monetisation depends on contracts, deployments, and user conversion—processes that can take longer to translate into financial statements.
When you combine these two realities—hardware procurement leading and financial translation lagging—you get a market pattern that looks like momentum shifting. Investors who initially chased software narratives start to rotate toward the supply chain, not because they’ve lost faith in AI apps, but because they’re trying to align expectations with where the measurable demand is showing up.
This is also why the “time lag” framing resonates. It’s not that software is failing; it’s that the path from infrastructure to revenue is uneven. Hardware can be installed quickly relative to the time it takes to redesign workflows, train staff, and establish ROI. In other words, the market is watching the conversion funnel—and noticing that the top of the funnel is filling faster than the bottom.
Chips: the obvious beneficiary, but not the whole story
Chips are the most visible part of the AI supply chain, and for good reason. Training and inference require specialised compute, and the industry has been racing to improve performance per watt, throughput, and scalability. But the chip story is not just about raw compute. It’s about system-level performance: how accelerators interact with memory, how efficiently they handle different workload shapes, and how quickly they can scale across clusters.
That’s why investors increasingly pay attention to platform ecosystems rather than single-chip performance. A chip might be fast, but if the surrounding system cannot feed it efficiently, the real-world performance advantage shrinks. Memory capacity and bandwidth become part of the “effective compute” equation.
Another factor is the pace of iteration. AI infrastructure is evolving rapidly, and data centres often upgrade in waves. That can create periods where demand for certain generations of chips spikes, followed by transitions. Semiconductor companies that can navigate these transitions—maintaining supply, meeting performance targets, and supporting customers through integration—tend to be rewarded. Those that stumble can see demand shift away even if the broader AI trend remains intact.
Memory: the less glamorous bottleneck that keeps getting louder
Memory is where the AI infrastructure story becomes more technical—and more interesting. In many AI deployments, memory is not merely a supporting component; it is a limiting factor. Large models require substantial memory footprints, and even when models fit, the system must manage memory efficiently to avoid performance degradation.
As workloads grow, memory demand tends to rise in multiple ways:
1) More capacity is needed to store model weights and intermediate activations.
2) Higher bandwidth is needed to reduce stalls and improve throughput.
3) Better reliability and endurance matter because data centre operations run continuously and at scale.
Memory also intersects with the economics of inference. Inference is often where companies try to monetise AI, and cost per query is a key metric. If memory constraints force inefficient batching or increase latency, the unit economics can worsen. Conversely, improved memory performance can enable better utilisation and lower cost per token.
That’s why memory makers can experience demand that tracks AI infrastructure build-outs closely. When data centres expand, they don’t just buy more compute—they buy more memory to keep that compute productive. Over time, as deployments mature, memory requirements can intensify further due to model size growth, context length increases, and the expansion of multi-model or ensemble approaches.
The market’s “catch-up” narrative fits this: if software adoption is slower, but hardware procurement continues and memory capacity expands, then hardware-related earnings can start to reflect the AI spend more clearly before software revenue catches up.
What about the software lag? It’s not necessarily a broken thesis
A common mistake in interpreting a lag is to assume it means the AI thesis is wrong. More often, it means the timeline is different.
Software adoption can be slower for reasons that are not about interest, but about implementation:
– Integration complexity: AI tools must connect to existing data sources, identity systems, and application workflows.
– Governance and compliance: enterprises need auditability, access controls, and safety measures.
– Reliability and evaluation: teams need to measure performance against real tasks, not benchmarks.
– Cost management: AI can be expensive at scale, so companies need optimisation strategies.
– Change management: employees must trust and use the tools consistently.
Even when usage grows, revenue recognition can lag because contracts may be structured around milestones, seat adoption, or measured outcomes. Some companies also delay monetisation until they can demonstrate stable performance and reduced risk.
So the software lag can coexist with strong underlying demand. It simply means that the market’s expectation for “fast monetisation” was too optimistic. When that happens, investors often reprice the timeline rather than abandon the theme.
A unique angle: the market is learning to price “infrastructure-to-revenue conversion”
The most interesting part of this story is not that chips and memory are benefiting. It’s that investors are increasingly focused on conversion mechanics—how quickly infrastructure spending turns into revenue.
In earlier phases of AI investing, the market sometimes treated AI as a direct line from model capability to business value. But the infrastructure layer complicates that line. Compute must be purchased, deployed, and integrated. Then the software must be adopted and monetised. Each step has its own friction and timing.
As a result, the market is developing a more nuanced framework:
– Hardware demand can be a leading indicator of AI deployment intensity.
– Software revenue can be a lagging indicator of monetisation success.
– Memory and system-level constraints can determine whether deployments scale efficiently.
This framework helps explain why hardware stocks can rally even when software stocks appear stuck. It’s not simply “rotation.” It’s a recalibration of what counts as evidence.
If you want a practical way to think about it, consider this: chips and memory are closer to the physical reality of AI. Software is closer to the organisational reality of AI. Physical reality tends to move faster than organisational reality.
Where does this leave investors?
For investors, the implication is not that software