Skeleton Technologies reaches a billion-dollar valuation on AI data centre power

Skeleton Technologies, the Tallinn-based energy storage company, has raised a nine-figure strategic investment led by Siemens that values it at more than 1 billion dollars. The company did not disclose the amount. It said the money will fund more than 200 million dollars of investment in American operations and research, including cell manufacturing in the United States, as it chases demand for power systems in AI data centres.

The valuation is the headline and the least interesting part of the deal. What makes it worth reading is what Skeleton actually sells, which is neither generation nor bulk storage. It sells the equipment that stops a rack of GPUs from browning out its own building.

What a supercapacitor does that a battery cannot

A supercapacitor stores energy in an electric field rather than through a chemical reaction, which is the difference that matters here. It holds far less energy than a lithium-ion battery of the same size, and it can release what it holds far faster, over a long life, across a wide temperature range, without the fire risk of a cell that can be pierced or overheated. That combination is wrong for a car and right for a power system that has to absorb a spike lasting milliseconds.

• A dynamic response time of 10 microseconds, which the company says is faster than the 20 microseconds needed for peak shaving on GPUs.
• Energy savings of up to 40 per cent in a data centre, according to the funding announcement, and up to 45 per cent more energy efficiency on the company's own site.
• Double-digit improvements in GPU performance, another company figure that will depend on the deployment and its operating conditions.

Why a GPU cluster needs microseconds

AI workloads are not steady. A GPU cluster can change its power draw within milliseconds as training or inference load shifts, and those swings ripple through the electrical plant. The efficiency claim rests on what Skeleton calls artificial compute loads, work run to keep the electrical system stable when real demand moves. Remove the fluctuations, and the padding goes with them, in the company's account.

The macro argument is on the company's side. The International Energy Agency has projected that global data-centre electricity consumption could nearly double between 2025 and 2030, which puts operators under pressure to extract more compute from each megawatt rather than to find more megawatts.

The product is a UPS, and the customer is a hyperscaler

• GrapheneUPS, a double-conversion uninterruptible power supply built for AI data centres, quoted at 242 kilowatts per square metre and up to 50 per cent less floor space than competing units.
• GrapheneGPU, for peak shaving next to the compute, and GrapheneBBU, a battery backup unit, both sold as boxes that sit beside the rack.
• Sidecars and power support for solid-state transformers, the direction the next generation of AI halls is heading, including work on 800 volt direct-current designs.

The Siemens relationship predates the round. The company's Digital Industries division designed and automated the SuperFactory in Leipzig, and Siemens backed an earlier expansion in 2023. Siemens now gets a fast-response component for its data-centre electrical offering, and Skeleton gets a partner with the infrastructure business that a power system has to plug into if it is to be sold as a system rather than a component.

The geography is the part Estonia should notice

Skeleton has been the country's most-cited deeptech company for years, and this round makes it a multinational with an American supply-chain ambition. The chief executive frames the goal as being the only company in the segment with a full United States value chain, from cells to complete systems. The headquarters and the engineering remain in Tallinn. The factories have gone where the customers are, on Skeleton's own explanation that it wants to manufacture closer to the hyperscalers it intends to sell to.

That is not a criticism. A company that makes physical objects has to build them near the buyer, and proximity is part of the product when the buyer is commissioning a data hall. It does mean that the Estonian claim on the value is the intellectual property and the high-value engineering rather than the production line, and that the country's gains from a billion-dollar company will look different from the gains it would book from a factory.

What to watch

• Whether the round is ever quantified. A nine-figure sum spans a wide range, and the difference between 100 million and 900 million dollars matters to anyone valuing the company independently.
• Whether the hyperscaler contracts are named. Skeleton sells to operators whose procurement is visible enough to check, and a named deployment would settle more than a valuation.
• Whether the 40 per cent saving survives a third-party test. Data-centre efficiency is usually measured at the facility level, where many things change at once.
• Whether the American cell plant is built. Domestic manufacturing is the expensive half of the plan and the part most exposed to a change in American policy.
• Whether the European plants keep their role. A company with factories in Germany and a new American supply chain has choices about where the next line goes.
• Whether Skeleton lists, and where. A billion-dollar valuation and a strategic industrial investor are the profile that tends to precede a listing conversation.

The deal is a good day for Estonian deeptech and a better one for the argument that power, not compute, is the binding constraint on AI. Skeleton's supercapacitors do one small thing extremely fast, and the market has decided that this is worth a billion dollars because everything else in the hall depends on it. The next test is not the valuation. It is whether a hyperscaler puts the boxes in a building and lets somebody measure the saving.