The “Big Tech software era” may be ending, according to James Anderson, a former Baillie Gifford fund manager whose comments have been circulating among investors focused on artificial intelligence. Anderson’s core claim is simple but potentially far-reaching: the biggest rewards from the AI race are shifting away from the software layer—where much of the early excitement and valuation has concentrated—and toward the hardware and infrastructure that make large-scale AI possible.
It’s an argument that sounds like a familiar “picks and shovels” story, but Anderson’s framing suggests something more specific than just a generic rotation into semiconductors. He implies that the economics of AI are changing in a way that favors suppliers of compute capacity, not merely the companies that package AI into apps, platforms, or model interfaces. In other words, the center of gravity may be moving from code and distribution to power, performance, and supply chains.
To understand why this view is gaining traction, it helps to look at what has actually happened since the AI boom accelerated. The early phase of the market was dominated by software narratives: new model releases, platform partnerships, developer ecosystems, and the promise that AI would become a universal layer across industries. But as adoption widened—from research labs to enterprises, from pilots to production—the bottleneck increasingly became physical. Training and inference at scale require enormous amounts of compute, memory, networking bandwidth, and energy. Those requirements don’t scale smoothly with software ingenuity; they scale with engineering, manufacturing, and logistics.
That is where Anderson’s thesis lands. If the “software era” is over, it doesn’t mean software stops mattering. It means software is no longer the only—or even the primary—constraint. When the limiting factor becomes hardware availability and system-level performance, the value chain changes. The companies that can deliver the right chips, the right systems, and the right infrastructure at the right time can capture outsized returns compared with those competing primarily on features, interfaces, or distribution.
A shift from product to capacity
One reason the hardware story is resurfacing is that AI spending is increasingly about capacity rather than experimentation. Many organizations began with proof-of-concept deployments: small models, limited workloads, and constrained user bases. Those phases were often justified by strategic urgency and learning. But production use introduces different metrics—latency, throughput, reliability, cost per query, and uptime. Once those metrics become central, the conversation moves quickly from “Which model is best?” to “What does it cost to run it here, at this scale, with these service-level expectations?”
That change tends to favor suppliers who can provide end-to-end compute solutions. Chips are essential, but they are only one part of the stack. Data centers need power delivery, cooling, rack density, high-speed interconnects, storage throughput, and orchestration software that can keep workloads running efficiently. Even if a company’s software layer is excellent, it still depends on the underlying hardware ecosystem to deliver performance and manage costs.
Anderson’s comments resonate because they align with how AI budgets are being structured. Large buyers are not simply purchasing “AI features.” They are funding buildouts: new data center capacity, expanded GPU fleets, upgraded networking, and the operational tooling required to run inference continuously. In that environment, hardware suppliers and infrastructure providers can become the beneficiaries of long-duration capital expenditure cycles.
The “AI war” as an industrial contest
The phrase “spoils of the AI war” is telling. Wars are not won by ideas alone; they are won by logistics, manufacturing, and the ability to sustain operations. In the AI context, the “war” is often described as a competition for model leadership, but the practical reality is that model leadership requires compute access. Compute access requires supply. Supply requires factories, yields, packaging capacity, and component availability.
This is why the hardware narrative has repeatedly reappeared during each wave of AI enthusiasm. When demand spikes, the market discovers that the limiting factor is rarely the algorithm. It is the ability to produce and deploy enough compute to train and serve models at scale. That dynamic creates a kind of industrial momentum: once a buyer commits to a compute roadmap, it tends to lock in multi-year procurement plans, which can benefit suppliers across the chain.
However, the unique angle in Anderson’s framing is the suggestion that the software layer may not capture the same upside as before. In earlier tech cycles, software companies often captured disproportionate value because they could scale globally with relatively low marginal costs. AI software can scale too, but the marginal cost of serving AI—especially for high-demand workloads—can be heavily tied to hardware utilization and energy costs. If the economics of inference are constrained by compute supply and power availability, then the “profit pool” can shift toward those controlling the scarce inputs.
Where the value may flow next
If the biggest gains shift toward hardware suppliers, investors will likely look beyond the most obvious names. The hardware value chain includes:
1) Semiconductors and accelerators
GPUs and specialized AI accelerators remain central, but the competitive landscape extends to memory bandwidth, interconnect design, and the ability to deliver performance-per-watt. As models grow and inference demands rise, efficiency becomes as important as raw speed.
2) System integration and networking
AI clusters are only as strong as their ability to move data quickly between components. High-speed networking, low-latency interconnects, and scalable cluster architectures can determine whether a system meets training timelines and inference targets. This is a less glamorous area than consumer-facing software, but it can be decisive.
3) Data center infrastructure
Power delivery, cooling, rack design, and storage are not optional. They are the enabling layer that turns chips into usable compute. As AI workloads increase, data centers face constraints around electricity availability and thermal management. Suppliers that help solve those constraints can become critical.
4) Manufacturing and supply chain resilience
Even when demand is clear, supply can lag due to manufacturing capacity, packaging bottlenecks, and component lead times. Companies that can secure supply and deliver reliably can capture value during periods of scarcity.
5) Operational tooling and orchestration
While Anderson’s comments emphasize hardware, the operational layer matters because it determines utilization. Efficient scheduling, workload management, and monitoring can reduce waste and improve throughput. In practice, this can translate into better cost per token or cost per inference, which influences buyer decisions.
The key point is that “hardware suppliers” is not a single category. It’s a network of capabilities that collectively determine whether AI systems can be built, scaled, and operated profitably.
Why software may still win—but differently
Saying the software era is over can sound like a dismissal of software’s role. That’s not necessarily what Anderson is arguing. Software still matters, perhaps even more than before, but its economic position may change.
In many AI deployments, software differentiates through workflow integration, user experience, domain adaptation, and governance. Yet the cost structure of running AI services can be dominated by compute. If buyers can get similar software functionality from multiple vendors, they may choose based on total cost of ownership and performance. That pushes bargaining power toward those who can influence compute costs—either by providing cheaper hardware, improving utilization, or reducing energy consumption.
Software companies may still capture value, but the “easy money” of the early cycle—where software platforms benefited from rapid adoption without immediate pressure on compute economics—could be replaced by a more competitive environment. Buyers will ask harder questions: How much does this cost to run? What is the expected utilization? How does performance scale with additional users? Can we control latency? Can we manage risk and compliance without sacrificing throughput?
Those questions tend to reward vendors that can demonstrate measurable improvements in efficiency and operational outcomes. In that sense, software may evolve into a layer that optimizes hardware usage rather than a layer that stands alone.
The data center as the new battleground
One of the most visible signs of the shift is the data center buildout. AI is not just a software phenomenon; it is a physical infrastructure project. The world is watching for new campuses, expansions, and upgrades—because those are the places where AI capacity becomes real.
Data centers are also where the constraints become tangible. Electricity is often the limiting factor. Cooling is another. Even if chips are available, the facility must be able to support the power draw and heat removal required for dense compute racks. That means the winners may include companies that can accelerate deployment timelines, secure power agreements, and deliver scalable designs.
This is why the hardware/infrastructure thesis can feel compelling to investors: it connects AI demand to capital expenditure cycles that are already underway. Unlike some software trends that can fade quickly, infrastructure investments tend to be sticky. Once a data center is built and configured for AI workloads, it can support multiple generations of models and applications, creating a durable platform for ongoing spending.
A more nuanced “rotation” in investor thinking
Investors often talk about rotations—moving from one sector to another as narratives change. Anderson’s comments suggest a rotation, but not necessarily a clean one. The market may not abandon software; it may reprice it relative to hardware and infrastructure.
In practice, this can mean several things:
– Higher scrutiny on software margins
If software revenue growth depends on expensive compute, investors may demand clearer evidence of margin durability.
– Greater emphasis on unit economics
Cost per inference, utilization rates, and energy efficiency can become key valuation drivers.
– Preference for companies with supply leverage
Hardware suppliers and infrastructure providers may benefit from scarcity dynamics and long-term procurement commitments.
– Attention to “time-to-capacity”
The ability to deliver compute quickly can matter as much as the theoretical performance of a system.
This is not just a thematic bet. It’s a shift in what investors consider the binding constraint in AI adoption.
The global dimension: geopolitics and industrial policy
Hardware supply chains are also deeply political. AI accelerators and related components are subject to export controls, national security concerns, and industrial policy. That adds another layer to Anderson’s thesis: the AI race is not only about technology; it is about sovereignty