Data-Center Power, Financing and Compute: A Panel with Yixing Xu, Alex Yeh and Bruce Liu

At the Silicon Valley 101 Alignment conference, moderator Christina Xu returns to a theme from the previous year: data centers as the new oil. She asks energy strategist Yixing Xu, investor Bruce Liu and GMI Cloud CEO Alex Yeh what has changed as demand for AI infrastructure has grown.

Yixing Xu says he works on data-center planning and energy strategy at Google, after similar work at Microsoft and ByteDance and earlier work at Breakthrough Energy. Yeh describes GMI Cloud as a provider of GPU infrastructure, a Kubernetes management layer and a token service.

Asked about the electricity bottleneck, Yixing Xu says the US faces a difficult transmission buildout. In his assessment, many developers are responding by placing generation at or near the data-center site rather than waiting for new long-distance lines.

He estimates that 30–50% of data centers planned that year involve behind-the-meter generation, a figure he presents as an estimate rather than a comprehensive survey. He describes interest in gas turbines and fuel-cell suppliers, while warning that supply constraints and technology maturity remain risks.

Some suppliers claim they can bring on-site power online in a matter of months, he says. Faster delivery is attractive, but a proposed schedule is not a guarantee: equipment availability and whether a technology works at the required scale still matter.

Liu agrees that power is a major constraint but says shortages extend across the AI supply chain. Equipment is only part of the challenge: construction labor, permits and support from local communities also determine whether a project can be completed.

Liu argues that projects need to demonstrate benefits to the communities hosting them. He remains optimistic about infrastructure investment, while acknowledging that new construction has both positive and negative local effects.

Asked to pick the scarcest component, Liu points to shortages across GPUs, memory, equipment and smaller electronic components. He says different parts of the supply chain can benefit as demand grows, and highlights the changing power requirements of computing hardware.

Christina Xu notes GMI Cloud’s announced $668 million fundraise and asks what differentiates the company. Yeh identifies dependable delivery, a geographically diverse footprint and financing relationships in Asia as three parts of its strategy. The fundraising and company performance figures are presented by the panel.

Yeh says access to Asian capital can lower his company’s borrowing costs compared with some US peers. He adds that powered facilities remain difficult to secure: in his account, power and a ready building are still central constraints on delivering compute.

Yixing Xu estimates that very large US data centers can take roughly two to two and a half years to turn a power commitment into usable capacity, with smaller projects potentially faster. The timing varies with the project and should not be treated as a guaranteed schedule.

He warns that delays may not yet be fully visible because many projects are planned for future delivery. Power, electrical equipment, regulatory requirements and local opposition can each push a schedule back, with financing costs continuing while revenue is delayed.

He cites restrictions near airports and community opposition as examples of risks. His account of political and media responses reflects his perspective; the practical concern he raises is that permitting and local acceptance can affect construction timelines.

To illustrate financial exposure, Yixing Xu discusses reported rental prices and sketches how delays across multiple gigawatts could defer substantial revenue. His large industry-wide number is a scenario based on assumptions, not a measured loss or forecast.

Yeh says gigawatt-scale projects may combine grid power and on-site generation. He stresses that gas-fired power also requires air permits, and estimates that even a smaller modular project not yet under construction can take well over a year to become usable compute.

He says some customers who once wanted at least 100 megawatts are now asking for 20 or 30 megawatts if it can be delivered quickly. He sees opportunities in parts of Asia where, depending on location, energy availability and government support may allow faster development.

On delays versus cancellations, Yixing Xu estimates that a portion of projects could slip by more than a year, while acknowledging that the overall scale is uncertain. He believes most affected projects are delayed rather than cancelled; this is his estimate, not a verified industry count.

Yeh describes data centers as a regional business: a pause in one market can make capacity elsewhere more attractive. GMI’s approach, he says, includes speculative projects in less concentrated regions so that capacity may be available when demand shifts.

Yeh identifies a financing mismatch: data-center leases often run 12–15 years, while compute contracts may run four to six. Facility owners want certainty across the longer lease, so developers are seeking ways to bridge the difference.

He describes structures such as placing a portion of contracted cash flow into a special-purpose vehicle to pay the facility operator first, or providing a residual-value guarantee. These are examples of financing structures discussed by a market participant, not assurances that they eliminate credit risk.

Liu places the discussion in a broader view of debt and growth. He argues that private-sector borrowing can fund the buildout while AI revenues develop, and that investment-grade credit and other capital sources may provide capacity. These are his macroeconomic judgments, not established outcomes.

He interprets wider credit spreads for companies changing business models as a repricing of risk rather than automatically a collapse. He also cautions that some newer cloud and power builders have much wider spreads than established companies. Leverage can absorb timing gaps, but it also leaves borrowers exposed if expected revenue does not arrive.

Asked whether there is an AI bubble, Liu argues that market valuations should be considered alongside growth expectations and financial conditions. His valuation estimates and optimism reflect a particular point in time and his investment view; markets and individual companies can behave differently.

Yeh contrasts long contracts with speculative capacity. He says long-term agreements support more predictable financing, while uncommitted capacity may capture higher prices when supply is tight. His company uses a mixture, he says, with borrowing separated into project vehicles.

Why sell tokens as well as GPU hours? Yeh compares a cloud to a shared laptop: utilization determines how much value an owner can extract. Long contracts fill the largest blocks of capacity; shorter contracts and smaller units can fill some of the gaps.

He compares long contracts to pebbles in a glass, shorter contracts to sand and token-level sales to water filling the remaining space. Spare nodes held for service obligations and externally purchased compute can feed a token service, he says; utilization and margins still depend on demand and operating costs.

For the closing predictions, Liu says he remains positive on AI infrastructure, while Yixing Xu expects continued demand for power-related equipment and behind-the-meter generation, despite likely ups and downs. Both are forecasts, not guarantees.

Yeh predicts compute shortages could persist through 2030, citing inference, robotics and possible further advances in autonomous AI. His expectations about those technologies and their timing are speculative. The moderator closes by suggesting the audience revisit the predictions the following year.