The market loves the engine. I’m looking at the electrical system keeping it alive.
When investors talk about artificial intelligence, the conversation usually begins and ends with GPUs. We discuss accelerator road maps, training clusters, model sizes, memory bandwidth, and whichever chief executive has most recently held up a large piece of silicon under theatrical lighting.
I understand the fascination. The GPU performs the glamorous work. It is the muscle behind training and inference, the component that gets photographed, named, back-ordered, and discussed as though it possesses a minor royal title.
But a GPU cannot eat enthusiasm.
It needs electricity—an enormous amount of it—delivered at the correct voltage, at precisely the right moment, with minimal waste, limited heat, and enough reliability that a multimillion-dollar AI rack does not become an extremely sophisticated space heater. Between the utility grid and the processor sits a long chain of power conversion, control, protection, monitoring, and thermal decisions. Every watt has to travel through that chain. Every percentage point of inefficiency becomes heat. Every failure threatens expensive equipment and valuable computing time.
That is why I believe the AI infrastructure story is wider than the accelerator market, and why Monolithic Power Systems, ticker MPWR, deserves more attention than its comparatively unromantic products receive.
Monolithic Power Systems does not manufacture the GPU that writes the poem, generates the image, or confidently invents a historical quotation. It develops high-performance power-management solutions that help the hardware receive and use electricity. That may sound like a supporting role. In AI servers, the supporting role is becoming mission-critical.
The opportunity is exciting. The valuation is demanding. The risks are real. And that combination makes MPWR one of the more interesting “picks and shovels” companies in the AI buildout—provided I resist the investor habit of calling every shovel priceless after the gold rush has already noticed it.
The GPU Is the Celebrity; Power Delivery Does the Work Backstage
An AI server is not merely a box containing accelerators. It includes CPUs, memory, networking, storage, fans or liquid-cooling equipment, controllers, voltage regulators, protection circuits, and multiple layers of power conversion. At rack scale, the electrical architecture becomes even more complicated.
Power arrives at a data center in forms that are unsuitable for direct use by a processor. It must be converted and distributed, often through several stages, before reaching chips that operate at low voltages while drawing extremely high currents. Those voltages must remain stable even as workloads change rapidly. A processor can move from a lighter load to intense computation quickly, creating transient demands that the power system must handle without wandering outside tight operating limits.
The engineering problem is easy to underestimate because electricity is invisible when everything works. Nobody posts a celebratory video because a voltage rail remained stable for another afternoon. Yet the entire system depends on that boring success.
As accelerator power rises and racks become denser, the challenge compounds. Higher current creates more conduction loss. More conversion stages create more opportunities to waste energy. Lost energy appears largely as heat, which then requires additional energy and infrastructure to remove. Poor power density consumes board space that could otherwise support computing, memory, or networking. Weak control can compromise performance and reliability.
This is where power-management companies earn their place.
They design the integrated circuits, modules, and control systems that convert voltage, regulate output, sequence rails, monitor conditions, and protect components. Their products may represent a smaller portion of the bill of materials than the accelerators, but they influence whether those accelerators operate efficiently and reliably.
I think of it this way: buying a world-class GPU while treating power delivery as an afterthought is like installing a jet engine in a lawn chair and then becoming surprised by the handling.
Why AI Changes the Power-Management Opportunity
Traditional servers already needed careful power design. AI accelerators make the requirements more severe.
First, the absolute power per accelerator and per rack has climbed. Second, high-performance systems increasingly pack more accelerators into smaller spaces. Third, workload changes can be abrupt. Fourth, data-center operators care intensely about efficiency because electricity is both an operating cost and a deployment constraint.
The result is more semiconductor content devoted to power management, more value placed on efficiency and density, and more pressure to integrate functions that once occupied separate components.
The industry is also moving toward higher-voltage distribution within server racks, including 48-volt architectures. Carrying the same amount of power at a higher voltage reduces current, which can reduce resistive losses and the amount of copper required for distribution. The voltage still has to be converted down near the processors, however, so the system needs efficient intermediate bus converters, point-of-load regulators, controllers, power stages, and increasingly sophisticated modules.
That last conversion is especially demanding. Advanced processors operate at low core voltages, yet they may need hundreds or even thousands of amps across multiphase power systems. Delivering that current efficiently in limited space is not something an engineer solves with positive thinking and a longer extension cord.
MPWR sells into this complexity. Its portfolio includes power-management integrated circuits and modular solutions for data-center, storage, networking, automotive, industrial, consumer, and communications applications. The company’s stated strengths include system-level expertise, semiconductor process knowledge, packaging, and integration.
That matters because customers do not merely want an isolated chip with an impressive specification. They want a power solution that works inside a specific system, meets transient requirements, fits physical constraints, communicates with other components, manages heat, and arrives on time. The closer a supplier gets to solving the full power problem, the more value it can potentially capture.
MPWR has spent years describing its evolution from a component vendor toward a broader silicon-based solutions provider. The phrase is corporate, but the strategy makes sense. A discrete component competes heavily on price. A validated, integrated power solution can become deeply embedded in a customer’s design.
The Numbers Finally Show the AI Story Clearly
For a while, MPWR’s AI exposure was wrapped in the usual fog of customer confidentiality, product transitions, and uneven quarterly comparisons. The first half of 2026 made the direction much harder to miss.
In the second quarter of 2026, MPWR reported revenue of $980.6 million, up 47.6% from the same quarter a year earlier. Enterprise Data revenue reached $380.6 million, rising 44.8% sequentially and 164.3% year over year. Management attributed the increase primarily to higher sales of power-management solutions for AI and server applications.
That segment represented 38.8% of total quarterly revenue. In other words, the AI and server opportunity is no longer a charming subplot mentioned near the end of the call. It has moved into the kitchen, opened the refrigerator, and started labeling shelves.
Across the first six months of 2026, total company revenue reached $1.785 billion, up 37.1% year over year. Enterprise Data revenue increased 132.4%. Communications revenue also rose 67.1%, helped by power solutions for optical modules and switches—another reminder that AI clusters require networking as well as compute.
The profitability was equally notable. Second-quarter GAAP gross margin was 55.2%, almost unchanged from a year earlier, while GAAP operating income reached $303.9 million. Net income was $257.3 million, compared with $135.0 million in the prior-year quarter.
MPWR also ended June with about $1.414 billion in cash, cash equivalents, and short-term investments. The balance sheet gives the company room to fund development, support customers, return capital, and withstand the inevitable moments when the semiconductor cycle remembers it is a cycle.
Management’s third-quarter outlook called for revenue between $1.14 billion and $1.16 billion. At the midpoint, that would imply another substantial sequential increase from the already record second quarter. The company also expanded its total repurchase authorization to $1 billion.
These are exceptional growth figures. They also create exceptional expectations, which is the part of every investor presentation that tends to arrive in a smaller font.
Why MPWR May Capture More Content per AI System
The simplest bullish argument is that more AI servers mean more power-management chips. I think the better argument is that the power value per server can rise as systems become denser and more complicated.
An accelerator generation that consumes more power may need additional phases, improved current sensing, tighter control, more advanced packaging, and better thermal performance. A transition toward 48-volt distribution creates opportunities in voltage conversion. More optical networking creates power demand around transceivers and switches. More memory and storage expand the supporting electrical system.
The content opportunity therefore extends beyond unit volume. MPWR can potentially benefit from both more systems and more dollar content within each system.
Integration may strengthen that effect. Power modules can combine controllers, power stages, inductors, and other elements into compact packages. Customers may pay more for the module than for a single integrated circuit, but they can receive advantages in board space, design time, performance predictability, and time to market.
This is particularly valuable in AI infrastructure, where customers race to deploy new platforms and design mistakes are expensive. A proven solution can reduce engineering work and qualification risk. Once designed into a platform, a supplier may enjoy revenue across the product cycle.
I do not assume this creates an impenetrable moat. Power semiconductors are competitive. Customers often qualify multiple suppliers, negotiate aggressively, and develop internal expertise. But system knowledge, packaging capability, product breadth, and a record of reliable execution can create meaningful switching friction.
The less glamorous a component appears, the easier it is for investors to overlook how difficult it is to replace. Nobody puts a voltage regulator on a conference T-shirt. That does not mean a server designer casually swaps it on Friday afternoon.
The Opportunity Extends Beyond Compute
I also like that MPWR is not a pure bet on one accelerator socket.
AI data centers need networking to move enormous amounts of data among accelerators. They need storage to feed training and preserve results. They need power conversion across boards, racks, and facilities. They need cooling systems with motors, pumps, controllers, and sensors. Each layer creates adjacent semiconductor demand.
MPWR reports Enterprise Data, Storage and Computing, Communications, Automotive, Industrial, and Consumer end markets. In the second quarter of 2026, Enterprise Data was the growth engine, but Communications reached $139.1 million and rose sharply year over year. Storage and Computing generated $199.8 million. Automotive contributed $157.1 million.
Diversification does not eliminate risk, particularly when one segment is growing quickly enough to dominate the story. It does give MPWR more ways to deploy its engineering platform. Technology developed for one market can inform products in another. High-efficiency conversion, compact packaging, motor control, battery management, and sensing are useful far beyond AI servers.
This broader platform matters to me because today’s fashionable application eventually becomes tomorrow’s comparison problem. AI infrastructure spending will not accelerate at the same pace forever. When growth normalizes, I would rather own a business with multiple strong end markets than one whose corporate strategy amounts to standing near a single customer with an invoice.
Now for the Risks the Excitement Would Prefer I Ignore
The first risk is customer and platform concentration.
MPWR does not always name the end customers behind its growth, but the market has long connected its data-center opportunity to leading accelerator platforms. A major design win can produce tremendous revenue. A lost socket, architecture change, dual-sourcing decision, or customer inventory adjustment can reverse that momentum quickly.
Power-management content is critical, but critical does not mean guaranteed. Large customers want supply security and bargaining power. They may qualify competitors, redesign boards, change voltage architectures, or bring parts of the solution in-house. Competitors will chase a market growing this fast with the enthusiasm of seagulls discovering an unattended sandwich.
Second, MPWR remains dependent on external manufacturing partners. It is a fabless company, which keeps capital intensity lower than owning leading-edge fabrication facilities, but also creates reliance on foundries, assembly providers, and supply chains. Capacity constraints, quality problems, geopolitical disruption, tariffs, and export controls can affect costs or availability.
Third, the product must work. Power failures can damage expensive systems, create warranty costs, and strain customer relationships. The more integrated and system-level the solution becomes, the greater the value—but also the greater the consequence of a design or manufacturing problem.
Fourth, AI capital spending can be volatile. Hyperscalers are spending enormous sums because demand appears large and strategic stakes feel existential. That does not repeal economics. If returns disappoint, deployments pause, architectures become more efficient, or customers digest prior purchases, suppliers can encounter sudden order changes.
Semiconductor demand has a habit of being described as structurally different shortly before inventory becomes structurally visible.
Fifth, MPWR disclosed a material weakness in internal control over financial reporting related to revenue recognition processes. The company stated in its June 2026 quarterly filing that disclosure controls were not effective as of quarter-end, while also saying the financial statements fairly presented its results in accordance with GAAP. Management described ongoing remediation.
I do not treat a material weakness as evidence that the entire business is unsound. I also refuse to wave it away because the revenue chart is attractive. Controls matter, especially when growth is rapid, product arrangements can be complex, and investors are paying a premium for trust in the numbers. I want to see remediation completed and independently tested.
The Valuation Is Where My Enthusiasm Meets Adult Supervision
As of September 4, 2026, MPWR traded around $1,214 per share, with a market capitalization near $60 billion. Based on trailing figures, the price-to-earnings ratio was roughly in the mid-70s.
That is not a valuation for “pretty good.” It is a valuation that expects years of strong growth, durable competitive positioning, successful product execution, and very few unpleasant surprises.
The company may deliver. Revenue growth has been impressive, margins are strong, the balance sheet is healthy, and the AI power opportunity appears to be expanding. But valuation determines how much of that success I am paying for in advance.
At a high multiple, even good news can be insufficient if it is not good enough. A quarterly beat can be greeted with disappointment if the next-quarter guide lacks fireworks. A customer transition can compress the multiple before the long-term opportunity has changed. A modest margin decline can suddenly become the market’s entire personality.
This is why I separate my view of the business from my view of the stock.
I can admire MPWR’s technology, strategy, and execution while remaining disciplined about entry price. “Great company” and “great purchase at any valuation” are different sentences. The market occasionally combines them, usually shortly before providing an educational demonstration.
Rather than pretend I know the exact correct multiple, I watch several variables: Enterprise Data growth, design-win diversification, gross margin, operating expense discipline, cash conversion, inventory, customer concentration, and progress on internal controls. I also compare the market value with plausible earnings several years ahead, not merely next quarter’s momentum.
If the valuation assumes near-perfect execution, I require a margin of safety the stock may not currently offer. If growth continues to outrun expectations and power content expands materially, today’s price could look reasonable in hindsight. Both statements can be true, which is why investing remains more difficult than selecting the most exciting paragraph in an earnings release.
What Would Strengthen My Bull Case
I would become more confident if MPWR demonstrated that its enterprise-data growth is broad across customers, platforms, and product families. One giant design win is lucrative; multiple durable relationships are a business foundation.
I also want evidence that modules and broader solutions are increasing content per system without damaging margins. Revenue growth bought with unfavorable mix is less attractive than growth driven by differentiated value.
Continued communications strength would support the idea that MPWR benefits from the entire AI cluster, not only accelerator boards. Optical networking and high-speed switches are essential as clusters scale. Powering those systems expands the addressable opportunity and reduces dependence on one component category.
Stable gross margin would be encouraging. MPWR maintained a GAAP gross margin slightly above 55% in the first half of 2026 despite rapid growth. Preserving that level while scaling would suggest the company retains pricing power and manufacturing discipline.
Finally, closing the material weakness would remove an unnecessary discount to confidence. When investors are asked to award a premium multiple, the accounting infrastructure should not be arriving late to the premium party.
What Would Break the Thesis
My thesis would weaken if Enterprise Data growth proved to be a short-lived inventory build rather than sustained deployment demand. I would pay close attention to sequential revenue patterns, customer forecasts, inventory levels, and management’s language around visibility.
A meaningful loss of content on a major accelerator platform would also matter, especially if competing suppliers demonstrated better performance, cost, or integration. The AI market can keep growing while an individual vendor loses share. A rising tide is helpful, but it does not inspect every boat for holes.
Persistent margin erosion would raise questions about competitive intensity or product mix. Rapid revenue growth paired with declining economics can still create a weaker business than the headline suggests.
Failure to remediate internal controls promptly would be another warning. One disclosed weakness can be repaired. A pattern deserves a different level of concern.
And, naturally, valuation itself can break the investment case. If the stock rises much faster than underlying earnings power, future returns can deteriorate even as the company performs well.
My Bottom Line on MPWR
The AI boom is often presented as a race to manufacture more GPUs. I think that view is incomplete.
The real buildout is a systems problem. Accelerators need memory, networking, storage, cooling, facilities, and—running through all of it—power. As AI racks consume more electricity and pack more computation into limited space, power conversion becomes harder, more valuable, and less forgiving.
That is the core MPWR opportunity.
The company supplies technology that converts, regulates, controls, and monitors the electricity feeding advanced systems. Its Enterprise Data results in the first half of 2026 show that this is not merely a theoretical adjacency. AI and server demand has become a major revenue driver, while communications growth suggests additional exposure to the infrastructure surrounding compute.
I view MPWR as a high-quality business with a credible path to benefit from rising power content in AI systems. I also view the stock as priced for excellence. At roughly 74 times trailing earnings as of this writing, the market is not asking whether MPWR will participate. It is assuming the company will participate spectacularly.
For me, that makes MPWR a company to own only with valuation discipline. I would not chase it simply because AI revenue accelerated. I would consider building a position during meaningful pullbacks, provided the operational thesis remained intact, or wait for earnings growth to reduce the multiple. Investors with a long horizon and high tolerance for volatility may accept the premium. Those seeking a generous margin of safety may need patience.
The larger lesson is more durable than any one entry price.
AI servers need more than GPUs. They need an electrical architecture capable of feeding those GPUs efficiently, reliably, and at scale. The more powerful the accelerator becomes, the more demanding that hidden architecture becomes. MPWR operates inside that constraint.
The GPU may receive the applause. Power management makes sure the lights stay on long enough to hear it.
Sources
Monolithic Power Systems, Second-Quarter 2026 Results and Third-Quarter Outlook.
Monolithic Power Systems, Form 10-Q for the Quarter Ended June 30, 2026.
Monolithic Power Systems, 2025 Form 10-K.
Monolithic Power Systems, Q2 2026 Investor Overview.
Market data are as of September 4, 2026. This article reflects my analysis and is for informational purposes only. It is not personalized investment advice. Stock prices, valuation multiples, and business conditions can change quickly.
Comments
Post a Comment