[BidClub_]
硅谷101 · · 42 min

"If You Dare Borrow, They Dare Lend": Discussing Data Center Financing, Power, Computing, and the...

Christina XuYixing XuAlex YehBruce Liu

EquitiesEnergyInvestingCompany Building
YouTube ↗
TL;DR
  • Available power is becoming the scarce asset that determines when AI infrastructure can earn revenue. Yixing Xu estimates roughly 30–50% of this year’s planned data centers may use behind-the-meter generation because transmission expansion has proved so difficult. Bloom Energy offers “speed to power,” potentially within three to six months, but Xu cautions that supply chains, scale and technological maturity remain unresolved.
  • Construction delays could leave substantial capital deployed before the associated revenue arrives. Yixing Xu’s rough estimate is that 10–20% of data centers could slip by more than a year, although “no one really knows” the scale. Bruce Liu uses a roughly 300 MW SpaceX–Anthropic arrangement at nearly $1 billion monthly as a reference and suggests industry-wide revenue deferrals could reach $100 billion if several gigawatts slip a year.
  • Financing availability does not eliminate the mismatch between buildings and compute contracts. Alex Yeh puts data-center leases at 12–15 years versus four to six years for compute. Operators ask who the investment-grade counterparty is because investment-grade backing, cash-flow lockboxes and residual-value guarantees help make that mismatch financeable.
  • GMI Cloud’s expansion combines supply-chain access, geographic flexibility and cheaper Asian capital. The moderator cites a $668 million raise and almost tenfold contracted-ARR growth in a year; Yeh says Taiwanese-dollar financing costs materially less than the roughly 10% or more paid by US peers. Regional permitting disruptions create openings elsewhere, but speculative construction requires being ready before displaced customers arrive.
  • Contract duration creates a trade-off between financing certainty and exposure to rising compute prices. Yeh contrasts CoreWeave’s five- to six-year contracts with Nebius’s greater spot exposure, citing roughly $15 per watt versus $25–30, potentially $50, without specifying a billing period. GMI combines contracted and speculative builds, while turning spare capacity into tokens to improve utilization and capture what Yeh calls “pure margin.”
  • Bruce Liu sees borrowing as a bridge to AI revenue, while acknowledging weak credits within the buildout. His framing is “If you dare to borrow, they dare to lend,” supported by his assessment that US private-sector balance sheets and the banking system can support more leverage. He calls Oracle’s wider spreads repricing, but identifies neoclouds and some power projects as potential problems. He favors broad AI-infrastructure exposure and puts QQQ at 20–21 times next year’s earnings.
  • The panel’s bullish forecasts differ in horizon and dependence on emerging demand. Yixing Xu expects the power sector for AI infrastructure to do well by the end of next year, despite volatility; he also names behind-the-meter generation and related technologies, while electrical equipment is raised as a beneficiary by another panelist. Yeh forecasts compute shortages through at least 2030, arguing that robotics, consumer adoption and recursive self-improvement add demand beyond coding; the six-month-to-two-year RSI timeline comes from researchers he has spoken with.
Digest · the substance, structured for research

1. The power bottleneck moves inside the data-center fence

  • Google data-center planning and energy strategist Yixing Xu says US transmission expansion has proved almost impossible at the required scale. If transmission cannot be built like a highway, companies are looking at generation beside the campus. He estimates roughly 30–50% of this year’s planned centers, potentially opening in one or two years, will use behind-the-meter solutions. Large GE gas turbines are sold out for years, increasing demand for smaller units.

  • Bloom Energy’s appeal, in Xu’s account, is “speed to power”: three to six months, with the company sometimes claiming 60 days. The moderator asks whether hyperscalers use it alongside neoclouds; Bruce Liu agrees. Xu nevertheless says to remain cautious: deployments are still small, supply chains are constrained, and suitability across data-center requirements remains uncertain.

  • Bruce Liu says shortages extend beyond GPUs to memory, substrates, MLCCs and other equipment. He notes that the GPU-to-CPU ratio has shifted from one-to-one to four-to-one, with some bullish on eight-to-one or more, creating potential in CPU compute. He calls this “a growing-pie story,” while noting that technological changes can still create winners and losers. Yixing Xu adds that higher power use and voltages make power semiconductors, once a “boring, cyclical” business, more exciting. Construction capacity, labor and permits also take time to expand.

2. Delays threaten cash-flow timing while shifting demand between regions

  • Bruce Liu puts US delivery at roughly two to two-and-a-half years for large campuses of 500 MW or more, versus perhaps a year-and-a-half for smaller projects. Many announced builds are not due until 2027–28, so problems have yet to surface. Yixing Xu later gives a rough estimate that 10–20% could slip more than a year, while saying “no one really knows” the scale.

  • Permitting adds constraints beyond electricity. In the discussion of delay risks, Xu cites FAA restrictions near airports and says community opposition is influencing politicians ahead of November’s midterms, including a reversal in the Texas governor’s stance. Yeh emphasizes emissions permits for natural-gas generation and estimates roughly 16–17 months from white space ready to token ready if construction has not started: about 14 months, plus time for modular skids and tokens. Customers who previously showed little interest below 100 MW now seek 20–30 MW: “they’ll take everything.”

  • Liu illustrates the financial exposure with a roughly 300 MW SpaceX arrangement with Anthropic for nearly $1 billion monthly; the moderator calls that triple last year’s price. Liu suggests deferred industry revenue could reach $100 billion if several gigawatts slip a year. Construction capital is already committed while receipts move out, potentially delivering “a huge blow to the industry.”

  • Liu asks whether projects are delayed or canceled; he says cancellation would concern him more. Xu says most are delayed, with a small percentage potentially canceled. Yeh describes a regional business: a pause on a Texas site made GMI’s Denver and Atlanta sites more sought-after assets. His response is speculative construction across East Asia, Southeast Asia and less concentrated US locations, ready to absorb displaced demand.

3. Financing must bridge both contract duration and delayed revenue

  • The moderator cites GMI Cloud’s $668 million raise and almost tenfold growth in contracted ARR in a year. Founder Alex Yeh credits reliable, on-time delivery, Asian supply-chain relationships and an international footprint. Financing is the third advantage: access to Taiwanese-dollar capital provides much lower borrowing costs than US peers paying close to 10% or more. Even with capital, he says, ready powered space remains extremely rare globally.

  • Yeh identifies the structural mismatch: data-center operators want 12–15-year leases, while compute customers typically contract for four to six years. Hence the recurring question, “who’s your investment-grade counterparty?” Nvidia, Broadcom and Google can provide investment-grade support. Proposed structures include placing five or six years of cash flows into an SPV lockbox that pays the operator first, or providing residual-value guarantees.

  • Liu’s macro argument is that heavily indebted governments need growth, while comparatively healthy US corporations and households can increase leverage. AI provides the investment anchor, and he considers the US banking system capable of financing it: “If you dare to borrow, they dare to lend.” Oracle’s widening spreads reflect its transition from a cash-generating business into a more capital-intensive model; he calls this “a repricing.”

  • Liu recalls, with a memory caveat, around half a trillion dollars of new issuance year to date and sees substantial remaining financing capacity. He also acknowledges that neoclouds and some power projects have troubling credit spreads. Borrowing bridges upfront CapEx and later receipts: “The buffer is leverage.” If AI revenue grows faster than expected, less borrowing may be necessary; weaker borrowers remain a possible problem.

4. Long contracts finance capacity; tokens monetize the gaps

  • Yeh contrasts CoreWeave’s emphasis on five- to six-year contracts with Nebius’s mix of contracted and spot capacity, while admitting he does not know Nebius’s exact split. Scarcity lets Nebius capture price increases, while CoreWeave cannot take advantage of them to the same extent: Yeh cites $25–30, even $50, per watt against roughly $15 for CoreWeave, without specifying a billing period. GMI similarly mixes speculative and contracted builds, ring-fencing borrowing within individual SPVs.

  • Cloud economics, Yeh argues, ultimately comes down to utilization. His analogy is a glass: long contracts are pebbles, filling perhaps 50% and making the business cash-flow positive; shorter contracts are sand; tokens are water filling the remaining gaps. Five or ten spare nodes kept to fulfill SLAs may attract little rental demand, but tokens divide that capacity into smaller saleable increments.

  • Yeh says GMI’s compute is sold through the end of 2027, but spare capacity and externally purchased compute feed its token business, which he calls “pure margin.” He expects neoclouds and inference platforms to consolidate into a vertical stack because progressively smaller units of GPU time improve utilization and margins. He also says GMI processes more than a trillion tokens daily.

5. The bullish calls depend on resilient growth and additional compute demand

  • The moderator’s provocation—“so there’s no bubble”—draws Liu’s valuation defense. He puts QQQ at 20–21 times next year’s earnings and cites calculated secular-growth premiums of 26% for QQQ, 14% for the S&P and 5% for small caps. His bullish outlook assumes financial conditions have peaked without damaging growth; he favors resilient AI-infrastructure exposure over betting on valuation expansion. He recommends broader exposure rather than picking winners and losers.

  • Xu expects “pains in the process” and fluctuations from delays, while remaining optimistic over the medium to long term. Pressed for a specific forecast, he says the power sector for AI infrastructure should do well “by the end of next year.” He names behind-the-meter generation and related technologies as beneficiaries; another panelist adds that electrical equipment should benefit.

  • Yeh forecasts a “permanent” compute shortage, at least through 2030. Coding demand from Codex and Claude Code is already priced in, he argues, while recursive self-improvement could add substantial demand as AI trains and improves itself without humans in the loop. Researchers he speaks with predict that capability in as little as six months to two years; this is their forecast.

  • Robotics supplies Yeh’s concrete example: Astra’s robot stretches and “wiggles before moving things,” which he sees as evidence of progress beyond narrow task-specific training. He argues that more compute and egocentric data could support generalization. Consumer demand is another prospective layer. His conclusion is that today’s coding-centered market has yet to account fully for robotics, consumers and RSI.

Full transcript