The most valuable patch of real estate in an AI data center may be a rectangle smaller than a postage stamp. It sits on the back of a server board, directly beneath the processor, in the place where AmberSemi wants to install a 1,000-amp power module called PowerTile. The company’s argument is wonderfully physical: when an AI chip demands a river of current, stop making that river cross the motherboard. Move conversion under the load, shorten the path, and waste less electricity as heat.
That sounds obvious only after someone draws it. Conventional boards arrange banks of voltage regulators beside the processor. Current moves laterally through copper traces, and resistance turns some of it into heat before useful work begins. The pain gets worse as GPUs and custom accelerators pull more current. AmberSemi says PowerTile attacks the distance itself, converting from a higher-voltage supply close to the processor and delivering power vertically through the board.
Specifications and performance comparisons are AmberSemi claims. Volume customer results have not been publicly disclosed.
01 / The problemThe expensive inches nobody sees
Power delivery has become a performance constraint, not a housekeeping chore. A hotter processor needs more copper, more regulators and more cooling. Those supporting parts compete for board area while the processor’s appetite keeps growing. PowerTile measures 20 by 24 millimeters and, at 1.68 millimeters high, is designed to preserve the option of liquid cooling. Several can be tiled together for systems beyond 10,000 amps, according to the company.
A longer run across the board adds resistance, loss and heat.
Backside placement makes the high-current trip radically shorter.
The architecture also attempts to remove conversion steps. At APEC 2025, AmberSemi described a route from 50 volts to a roughly 0.8-volt processor load. Fewer stages can mean fewer losses and components, although direct conversion makes control, magnetics, packaging and transient response harder. AmberSemi’s stated ingredients are integrated magnetic structures, specialized field-effect transistor design, three-dimensional packaging and digital control. This is not one magic transistor. It is a small stack of difficult disciplines.
“Power delivery is becoming a key performance differentiator.”Thar Casey, founder and CEO
02 / The product trailBefore the AI rack, there was the breaker box
AmberSemi did not begin under a GPU. Founded by Thar Casey in 2017, the company first pursued digital control of mains electricity for outlets, switches, breakers, appliances and industrial equipment. Its AC Direct DC Enabler combined conversion, regulation and supervision without the rectifier bridge, transformer or high-voltage bulk capacitor used in many conventional supplies. A companion switch-control platform promised sub-microsecond fault response, load telemetry and deterministic control of solid-state or mechanical switches.
The work produced more than 50 U.S. patents, according to the company, plus reference designs and exploratory partnerships. Infineon worked with AmberSemi on conversion and switching possibilities. Nordic Semiconductor explored wirelessly connected products powered by AmberSemi devices. STMicroelectronics joined a brushless-motor reference design. Jasco and Deako later agreed to explore integrations in connected-home and lighting products.
Those relationships matter, but they are not the same as a purchase order. AmberSemi’s PowerTile customers remain unnamed. AMD, Nvidia and GlobalFoundries appeared with the company in an APEC 2026 session on vertical power, which signals that the topic matters to the ecosystem. It does not make those companies customers.
03 / The pivotThe first thing to fail was the calendar
Hardware roadmaps have a way of turning confident dates into archaeology. At CES 2023, AmberSemi discussed partner integrations beginning that year and end products arriving as early as 2024. At APEC 2024, it said partner products were expected in 2025. By 2026, Casey said the small company had paused its solid-state circuit-breaker line to focus on AI data-center power. PowerTile customer samples were targeted for the third quarter of 2026, while material production volumes were described for 2027.
That is the most instructive part of the story. The original electrical-products thesis did not explode. It accumulated patents, demonstrations, awards and interested partners, but commercialization took longer than the slides. Meanwhile, AI racks jumped from a specialized market to an infrastructure emergency. The same core skill - moving and controlling electricity with less bulk - suddenly had a customer with a sharper pain and a larger budget.
Series B funding, an Infineon alliance, industry recognition and a semiconductor name change.
The Enabler reached silicon and appeared in demos with semiconductor partners. Public end-product forecasts moved right.
AmberSemi presented 50V-to-load vertical conversion for AI servers and deepened product partnerships.
A new tape-out and $30 million Series C fund samples, qualifications and the attempt to reach production volume.
What did that journey cost? Publicly disclosed rounds include a reported $3.3 million Series A, more than $8.5 million in Series B financing, and a $30 million Series C initial close in March 2026. Revenue and valuation are not public. The Series C is meant to scale development, customer engagement and commercialization. In semiconductors, the invoice is not only tape-out. It includes packaging, test, application engineering, qualification, inventory and the years spent convincing a conservative buyer that a startup will still be there when the server ships.
04 / The businessFabless does not mean frictionless
AmberSemi designs chips and modules but does not own a fabrication plant. That keeps capital needs below those of a new foundry, while shifting execution toward manufacturing partners, packaging houses and supply agreements. The intended business is component sales and design-ins: win a place on a server or electrical-product roadmap, help the customer integrate the part, then supply it across a production run.
Casey has said the company built a multi-partner manufacturing chain because supply disruption is the first fear large customers bring to a small semiconductor vendor. That concern is rational. A server manufacturer may admire an 85 percent loss-reduction claim and still choose a less elegant part from a supplier with proven capacity, second sourcing and a decade of reliability data. AmberSemi’s product is PowerTile. Its second product is confidence.
“We try to run a very, very tight ship.”Casey on cautious hiring as the company scales
The customer set spans hyperscalers, accelerator designers, server makers and semiconductor companies. Casey said in 2025 that roughly 50 alpha or beta customers had engaged across AmberSemi’s product lines. The company now talks of about 30 employees and plans to grow toward 50. Its earlier building-electrical work also addressed appliance makers, electricians, builders and smart-home brands. This is enterprise hardware sold through long technical courtships, not a gadget bought from a web cart.
05 / The fieldA clever diagram enters a crowded market
AmberSemi is not alone in chasing the underside of the processor. Empower Semiconductor markets integrated voltage regulators for under-package and near-package conversion. Ferric works on integrated voltage regulation. Established power names including Vicor, Monolithic Power Systems, Renesas, Infineon, Analog Devices and Delta attack adjacent pieces of the AI power stack. The incumbent alternative is not standing still either: conventional multiphase regulators keep improving, and processor packaging is pulling more functions closer to the die.
AmberSemi’s difference is the proposed jump from a higher distribution voltage directly to the low-voltage load in a thin backside module, paired with a claim of very high current density. If it works at customer scale, the payoff is space, efficiency and a simpler board. If it does not, the likely reasons will be mundane and lethal: thermal behavior, transient performance, manufacturing yield, qualification time, cost, supply assurance or a competing architecture that asks customers to change less.
What another builder can steal
- Find the loss the industry has learned to tolerate. AmberSemi chose voltage drop across the board.
- Make the advantage architectural and drawable. “Beside the chip” versus “under the chip” survives a noisy sales meeting.
- Keep the hard-won capability when changing markets. The beachhead moved from breakers to servers; digital power control stayed.
- Use reference designs to lower imagination cost. Partners need to see the new part inside a familiar system.
- Treat supply resilience as part of product design, especially when the buyer is much larger than the vendor.
06 / The limitsWhen the shorter path is not the better business
PowerTile is most persuasive where current is extreme, board loss is painful, backside space is available and the system roadmap can absorb a new power architecture. The pitch weakens in modest-power systems, products that cannot change cooling or board layouts, and programs where a buyer values established supply more than peak density. AmberSemi’s older Enabler is non-isolated, so it is not a drop-in answer where galvanic isolation is required.
Current density is climbing, wasted board power is expensive, backside placement is possible and the customer can co-design packaging and cooling.
Qualification slips, economics miss the incumbent, thermal integration disappoints, or customers refuse single-source startup risk.
The company has reached the awkward, useful stage where the invention is tangible and the business is not yet proven. A successful tape-out means the design has left the drawing board. It does not mean yield, reliability, customer qualification or repeatable volume are solved. A $30 million round buys a serious attempt at those problems. It does not buy the answer.
Still, AmberSemi chose a constraint that software cannot route around. AI processors need enormous current, and every unnecessary millimeter has a cost. The wager is precise: put conversion in the basement, give the processor a shorter straw, and make enough of the modules reliably that conservative infrastructure buyers accept the new address. The next milestone is not another award. It is a customer shipping a server with PowerTile inside.