The Foundation of AI Infrastructure: Power at Scale
Modern AI workloads are fundamentally different from traditional compute environments. Training large language models and supporting inference at scale requires exponentially greater power density at the rack level.
This shift has cascading implications:
- Higher rack power (moving from ~30–50kW to 100kW+ and beyond)
- Increased reliance on efficient power conversion
- Greater complexity in voltage regulation across components
At the center of this evolution are PMICs and VRMs, components responsible for converting and regulating power from high-voltage inputs down to the precise levels required by CPUs, GPUs, and memory.
Historically, these components were critical, but not scarce.
That dynamic is changing.
Demand Is Accelerating, Rapidly
According to the Datacenter Power report from Edgewater Research, demand for power semiconductors tied to AI infrastructure is accelerating at a pace that is beginning to strain supply.
The report notes:
- DC power semiconductor demand is increasing significantly, driven by AI datacenter buildouts
- Tightness is emerging across VRMs, PMICs, and the broader power stack
- Demand is expected to remain strong as new architectures, particularly HVDC (high-voltage direct current), gain traction
This demand is not theoretical; it is already being reflected in purchasing behavior.
Hyperscalers are:
- Raising forecasts
- Pulling in demand for next-generation designs
- Competing aggressively for available supply
In some cases, customers are requesting a 16–20-week buffer of inventory to ensure continuity.
This is a clear signal:
The market is beginning to move from awareness to action.
Pricing Has Inflected, And Is Being Accepted
One of the clearest indicators of a supply-demand imbalance is pricing behavior.
Here, the data is unambiguous:
- VRM suppliers are pushing price increases in the range of 20–35%
- In some cases, pricing has increased from ~$0.75–$0.80 to above $1.00 per unit for certain components
- Price increases are being applied broadly, including to major customers such as hyperscalers
Perhaps more importantly, these increases are being absorbed.
This is a critical distinction. In many markets, price increases are met with resistance, substitution, or delayed demand. In this case, customers are accepting higher pricing because:
- Supply is constrained
- Alternatives are limited
- The cost of disruption is far greater than the cost of components
This dynamic mirrors the early stages of memory tightening cycles—but with an important difference: The market is less prepared.
Supply Constraints Are Broad and Systemic
Another key takeaway from this month’s analysis is that supply constraints are not isolated to a single component or supplier. Instead, they are systemic.
Recent reports highlight tightening across:
- Semiconductor devices (MOSFETs, ICs)
- Materials and supporting components (cables, connectors)
- Engineering resources required for design and validation
This last point is particularly important.
In highly customized AI infrastructure environments, engineering capacity becomes a bottleneck in its own right. Suppliers must allocate not only manufacturing capacity but also design and application resources to support complex customer requirements.
As a result, even when physical capacity exists, the ability to deploy it may be constrained.
This is one of the defining characteristics of a supply-driven shortage, and one that is often underestimated until it is too late.
A Structural Shift: The Move to HVDC and 800V Architectures
Beyond near-term supply and demand dynamics, a more fundamental shift is underway in how power is delivered within datacenters.
This is a transition toward:
- High-voltage direct current (HVDC) architectures
- 800V power systems
- Increasingly complex multi-stage conversion processes
This shift is driven by the need to:
- Reduce energy loss
- Improve efficiency at higher power densities
- Support next-generation AI workloads
However, it also introduces new layers of complexity.
For example:
- Power must be stepped down from 800V to 54V, then to 12V or 6V, and ultimately to sub-1V levels required by processors
- Each stage requires specialized components and introduces potential points of constraint
- Architectural differences are emerging between vendors (e.g., Nvidia’s 800V approach vs. hyperscaler preference for ±400V systems)
The timeline for full adoption extends into 2027–2030+, meaning: This is not a short-term adjustment; it is a long-term structural transformation.
The Technology Layer: GaN vs. SiC
As power architecture evolves, so too does the underlying technology.
There is a growing divergence between:
- Silicon carbide (SiC) – dominant in AC/DC conversion
- Gallium nitride (GaN) – emerging as the preferred solution for high-frequency DC/DC conversion
Specifically:
- GaN is expected to play a critical role in 800V-to-low-voltage conversion, where high switching frequency and efficiency are required
- Suppliers such as Infineon, Texas Instruments, and Navitas are actively advancing GaN solutions
At the same time:
- SiC remains essential for high-voltage AC/DC applications
- Capacity, cost, and manufacturing maturity continue to influence adoption
This layered technology landscape adds another dimension of complexity to supply planning.
The Broader Context: Infrastructure Is Under Pressure
The tightening in power semiconductors does not exist in isolation.
Multiple reports are reinforcing this broader theme: AI infrastructure is placing unprecedented strain on the entire ecosystem.
Key observations:
- Rising input costs and extending lead times across global manufacturing
- Significant electrical equipment shortages impacting datacenter development
- A meaningful portion of planned U.S. datacenter capacity facing delays or cancellations due to infrastructure constraints
In addition:
- Power-related equipment (transformers, switchgear) is heavily dependent on global supply chains
- Geopolitical factors and tariffs introduce additional risk
Taken together, these dynamics point to a larger reality:
The challenge is no longer just building compute; it is supporting the infrastructure required to run it.
Why This Risk Is Often Overlooked
If the data is clear, the question becomes: Why isn’t this receiving more attention?
There are several reasons:
1. Memory Has Dominated the Narrative
DRAM and NAND shortages have been highly visible, well-documented, and widely discussed. As a result, they have captured the majority of executive attention.
2. Power Components Are Seen as Secondary
PMICs and VRMs are often viewed as supporting components within the bill of materials, rather than strategic drivers of system performance.
3. Complexity Masks Risk
Power delivery involves multiple layers: devices, materials, architecture, and engineering. This complexity can make it more difficult to identify early warning signs.
4. The Market Is Earlier in the Cycle
Unlike memory, which has already entered a well-defined tightening phase, power components are in the early stages of constraint formation.
This combination creates a classic blind spot:
By the time the issue becomes widely recognized, mitigation options are significantly reduced.
What This Means for OEMs and Supply Chain Leaders
The implications for OEMs, contract manufacturers, and procurement leaders are clear. This is not simply another component category to monitor, it represents a shift in where risk resides within the system. Organizations that respond effectively will do so by focusing on three core capabilities:
1. Deeper Supplier Visibility
Understanding availability is no longer sufficient. Leading organizations are developing visibility into:
- Capacity allocation and prioritization
- Engineering resource constraints
- Technology roadmaps and architectural alignment
This level of insight enables earlier decision-making and reduces reliance on reactive sourcing strategies.
2. Strategic Buffering (Not Blanket Stockpiling)
The move toward buffering is already underway, particularly among hyperscalers. However, effective buffering requires precision:
- Aligning inventory levels with real demand signals
- Focusing on high-risk components
- Balancing working capital with supply continuity
We are seeing requests for 16–20 weeks of buffer inventory becoming more common in many segments.
3. Execution Discipline
Perhaps the most important, and most challenging, factor is execution. Navigating a tightening market requires:
- Cross-functional alignment between procurement, engineering, and operations
- Willingness to act before constraints become obvious
- Confidence in decision-making amid pricing volatility
These are not new principles. But in a supply-driven environment, the speed and consistency of execution become defining advantages.
The Rand Perspective: Seeing the Whole System
At Rand Technology, our approach has always been rooted in a simple principle: Supply chain challenges are rarely isolated, they are interconnected.
What we are seeing in power semiconductors reinforces that view. The same forces driving constraints in memory: AI demand, architectural shifts, limited capacity, and geopolitical complexity, are now extending into adjacent layers of the system.
Our role is not simply to respond to these dynamics, but to help our customers:
- Anticipate where constraints will emerge next
- Understand how those constraints interact across the bill of materials
- Develop strategies that reduce exposure and improve planning confidence
In many cases, that means shifting the conversation from:
“Can we get the part?” to “Do we understand the system well enough to avoid the problem?”
The Constraint Is Moving
The evolution of AI infrastructure is creating unprecedented opportunities but also redefining where risk exists. For the past several years, that risk has been concentrated in compute and memory.
Today, it is beginning to move.
Power delivery, once considered a supporting function, is becoming a critical gating factor in the ability to scale AI systems. The data is clear:
- Demand is accelerating
- Supply is tightening
- Pricing is rising
- Complexity is increasing
And perhaps most importantly:
The market is still early in recognizing the full impact.
The organizations that succeed in this environment will not be those that react fastest to visible shortages. They will be the ones who identify emerging constraints early and act before they become obvious.
