Anyone who has touched a phone charger after heavy use has felt the problem. That warmth is electricity lost while the brick converts wall power into something a battery can use. It’s the same reason a laptop turns into a hotplate during heavy use while charging.
AI data centers lose power the same way at every step between the grid and the processor. As racks grow hungrier, those leaks have become a business problem.
Navitas Semiconductor (NVTS) built its name shrinking chargers with gallium nitride chips. On Monday, Oct. 5, it teamed up with Microchip Technology (MCHP) to apply that idea to AI server racks, according to a press release.
Their reference design converts 800 volts straight down to 6 volts for GPU boards. It folds two conversion stages into one converter, Investing.com reported, cutting wasted heat and freeing space around the chips.
Read Now: Two chipmakers, one demo, and a signal for AI’s next phase
Power is joining chips as AI’s tightest constraint
I think power is becoming AI’s next bottleneck after chips. Microsoft CEO Satya Nadella said as much last November, describing “chips sitting in inventory that I can’t plug in,” TechCrunch reported.
Global data center electricity demand jumped 17% in 2025, according to the International Energy Agency. That is more than five times the growth rate of overall electricity demand.
In the U.S., data centers drove about half of electricity demand growth last year, Fortune reported, citing IEA figures.
The IEA’s latest AI energy report describes a scramble for electricity, grid connections, and factory capacity, not just chips. It expects AI server power density to quadruple again by 2027, testing supply chains for power electronics.
One caveat: Chips are not fully solved. The IEA expects a high-bandwidth memory shortage to last through at least 2027, so power is joining chips as a constraint, rather than replacing them.
Cutting one conversion step matters more than it sounds
Every voltage change leaks some power as heat. The release says racks are moving to 800 volts partly to curb those losses. Removing a whole stage pushes that logic further.
The board targets up to 96% peak efficiency at full load, according to the release. With grid connections fixed, every point saved can be used for computing instead of cooling.
Security is the overlooked detail. The design adds Microchip’s TA100 authentication chip and controllers that support post-quantum cryptography, the release said. That treats power hardware as a hacking target, a risk the IEA flags.
Still, a reference design is a blueprint, not a purchase order. The companies will show it at Microchip’s OCP event in San Jose on Oct. 12-15.
What investors are paying for
Navitas sells gallium nitride and silicon carbide power chips, and AI data centers are now its central pitch. It is worth about $3.2 billion, according to Stock Analysis.
That price leans on the future. Trailing revenue fell 46% to about $37 million as Navitas pulled back from mobile chargers, leaving the stock above 80 times sales, Stock Analysis data shows. Management expects AI infrastructure to top a third of sales by year-end, its second-quarter call revealed.
The stock set an all-time high of $34.17 in June after an all-time low of $1.52 in April 2025, while its 52-week low is $6.85.
Barron’s noted the shares slumped from that peak before a U.S. Army contract sparked a rebound. Eight analysts rate Navitas a Hold, with an average target of $14.08, Stock Analysis noted.
Microchip is the steadier half. It sells microcontrollers and analog chips into cars, factories and data centers, and is worth about $44 billion, Stock Analysis data show.
Its trailing P/E near 120 looks extreme compared with a forward P/E near 21. That gap signals expected earnings recovery.
Microchip’s 52-week range runs from $48.52 to $105.91, Stock Analysis indicated, and that high is its all-time record. TradingView puts its split-adjusted all-time low near 31 cents in 1993.
Twenty-six analysts rate Microchip a Buy, with an average target of $107.84, according to Stock Analysis. For Microchip, AI rack power is one growth lane. For Navitas, it is the thesis.
More AI:
- Morgan Stanley revisits Micron stock price target amid stellar growth
- 27-year-old chip stock gets a rare win as rival bidder steps in
- Nvidia’s AI boom revives a Warren Buffett warning for investors
The fix for AI’s power problem has its own chokepoint
Gallium nitride chips need gallium, and China produced 94% of the world’s supply, according to a 2024 European Union report cited by AFP.
Beijing banned gallium exports to the U.S. in December 2024, then suspended the ban until Nov. 27, 2026, Reuters reported. Shipments still need Chinese export licenses.
Navitas makes its newest GaN chips with GlobalFoundries in the U.S., according to a press release. That secures the factory step, not the raw metal.
For most of this AI boom, investors asked who could make enough chips. The next phase asks who can deliver the power to run them, and who controls the materials underneath.
The Nov. 27 deadline, six weeks after the San Jose demo, is the next real test.
