Ep. 039 - [EMERGENCY] Our Numbers Are "Insane." Here's the Proof. (Datacenters) | Jeremie Ontiveros, Reyk Knuhtsen, Jordan Nanos

Summary

• SemiAnalysis is raising, not cutting, its 2027 data center forecast—from the 43 gigawatts Brad Gerstner called too aggressive on All-In/BG2 to roughly 50GW in a single year. Gerstner publicly cut the forecast in half, and the episode discussed a roughly 25GW alternative. The team’s response: “There’s a lot of thought and work that went into it,” and they stand by it.
• The methodology rebuttal is granularity: the model tracks thousands of sites “building by building, permit by permit, power allocation by power allocation,” and uses satellite photos. On moratoriums, Reyk Knuhtsen says they mapped over 400 against data centers in the relevant jurisdictions and found about 10% of the mapped data centers were affected. Behind-the-meter generation can provide an alternative where grid headroom is lacking.
• External haircuts to the number can amount to double counting, because delays are already built into the estimates. The team adds one to three quarters of delay depending on what they observe, treats sub-12-month timelines as a red flag, and assumes conservative ramps—for example, Oracle New Mexico has over a gigawatt under construction but contributes a “fairly low” share to the ’27 number. The Crusoe Childress critique was wrong: the forecast includes under 200 megawatts, not a full gigawatt.
• The 2027 number rests on trackable leading indicators: over $1 trillion in signed data center leases, plus self-build construction starts rising quarter after quarter. Jeremie Ontiveros points to a waterfall of signals: they predicted an order surge at early-stage constructor Sterling, and track private operators (QTS, EdgeConneX, Aligned, Vantage) whose contracts public markets may not see, alongside crypto-miner backlogs (Applied Digital, Core Scientific) now in the tens of billions.
• Modular construction is an important reason the team thinks the build-out can happen, compressing traditional 18-month timelines to roughly 6–9 months in cited examples. A modularized power hall can require much less labor than having “20 electricians” stick-build it. Reyk cites Anthropic reference designs, Meta tents, new Amazon buildings, and Project Houdini, which he thinks has a stated goal of 6GW a year for two or three years. He says developers that don’t follow modular approaches may struggle to get the labor they need.
• The tradeable risk: chips are oversupplied versus data centers today, but the constraint could flip to data center oversupply in 2028. Jeremie’s hedged call is that it’s currently easier to get GPUs than data center megawatts. He says the key watch item is how fast chip supply expands, which would support CapEx such as new semiconductor fabs and equipment.
• The meta-critique: bearish forecasts may rest on “not the most solid foundation of information,” including site trackers built by prompting Claude. Jordan Nanos’s direct grievance: Gerstner cited their data to audiences of 330,000 and 265,000, then dismissed it. Jordan’s response: “Cite our data, provide specific critiques to it, great. We can always learn”—but an offhand “this is the forecast; we think it’s wrong” seems strange.

Deep dive

1. The Gerstner callout — and a three-year track record

• The episode’s trigger: Brad Gerstner showed SemiAnalysis’s 43GW-in-2027 forecast on the All-In Podcast and his BG2 interview with Jensen, calling it too aggressive. The episode discussed an opposing view of roughly 25GW; the team’s answer isn’t a defense of 43—it’s an increase toward 50GW.
• Jeremie’s history lesson: the data center model launched nearly three years ago, with a March 2024 article, “The AI Energy Dilemma: Race for Data Center Space,” putting out forecasts “materially ahead” of McKinsey and peers—“that has largely, of course, been proven right.” In discussing a chart from “Stop Saying Half of U.S. Data Centers Are Delayed,” Jeremie initially referred to 2027, but Jordan corrected him and Jeremie agreed the chart was for 2026; that forecast had “barely moved” across model revisions.
• His reaction to the public haircut: “Dear Brad called us out publicly and cut our forecast in half. We’re like, ‘Damn, this is actually at least some number we made up.’” He then emphasized that “a lot of thought and work” went into it and that they stand by it.

2. The FUD list is “broadly true” — but does not apply to every project

• Reyk’s taxonomy of the bear case: moratoriums, grid connections, labor, equipment supply chains—“all of that is true,” but “being able to measure it accurately is a different story.” He says their granularity means tracking “not even at the campus level, but building by building, permit by permit, power allocation by power allocation,” and buying satellite photos.
• The moratorium answer with numbers: they mapped over 400 moratoriums against data centers in the relevant jurisdictions and found about 10% of the mapped data centers were affected. On energy, grid headroom is one consideration; behind-the-meter generation can provide an alternative for some projects.
• The logic: “When I’m physically looking at all of these data centers under construction, I don’t know how I could haircut it.” Reyk’s conclusion is that these issues are real but “don’t apply to every project”; he says the appropriate projects under construction pass their bar.

3. Delays are pre-baked, so an extra haircut can be double counting

• The model treats timelines shorter than 12 months as a red flag, and Reyk says a developer that “just came out of nowhere,” obfuscated permits or power sources, or no identified tenant raises questions. Possible surprises include a northeastern winter storm during a TeraWulf or Nebius buildout and a transformer blowing up at the Core Scientific site. Reyk preserves the caveat: “It’s not to say that these things don’t catch us off guard. They can… I didn’t predict a winter storm. It’s not my job.”
• Jeremie says the article’s chart covers about 40 sites, still under half of the global data centers shown as under construction. The relevant question is when a site’s first phase arrives and how quickly it ramps; their assumptions vary by project. Oracle New Mexico, which they say they were first to flag as having issues, has over a gigawatt of IT capacity under construction, but they forecast a “fairly low” share of that campus’s megawatts in 2027.
• The Crusoe Childress correction: Twitter accounts said the forecast included a whole gigawatt in 2027. Jeremie says that is wrong: “We have under 200 megawatts as of the latest version published.” Reyk says adding another haircut can be “a bit of double counting,” since Jeremie says they purposely add delays of one to three quarters depending on what they observe.

4. Over $1 trillion in signed leases and a waterfall of indicators

• The two signals emphasized in the article are self-build construction starts—tracking when each building starts, with the count rising quarter after quarter—and pre-leasing, where hyperscalers, neoclouds, and AI labs sign leases with data center operators. Jordan cites over $1 trillion of data center leasing signed; Jeremie says these binding contracts help them understand what is likely to be built.
• Jeremie’s supply-chain waterfall includes early-stage constructor Sterling, whose order surge they predicted through project-level tracking; electrical and cooling equipment downstream; and energy lead times. Gas turbines are multiple years out, while reciprocating engines have shorter lead times. The team uses these signals to check project timelines.
• The public markets’ blind spot: crypto-miner backlogs at Applied Digital and Core Scientific are visible and in the tens of billions, but public markets don’t have a good view of every contract between private operators such as QTS, EdgeConneX, Aligned, and Vantage and hyperscalers or AI labs. Jeremie says tracking those private companies lets them gather data that other forecasts may miss.

5. Modular construction and the 2028 oversupply risk

• Reyk’s underappreciated mechanism is modular deployment. He contrasts traditional 18-month timelines with DayOne timelines of about nine months for hundreds of megawatts and BrightRay timelines of about six months in Southeast Asia. He says Anthropic reference designs will employ modular approaches, and points to Meta tents and new Amazon buildings as examples of buildings that can appear quickly. He also cites Project Houdini, which he thinks has a stated goal of 6GW a year for the next two or three years; the transcript does not identify its owner.
• Reyk’s example of the labor savings: a modularized mechanical/electrical fit-out could use a power hall “like a literal shipping container” slotted into the data center shell, rather than “20 electricians” coordinating to stick-build the hall. He says developers that don’t follow modular design strategies probably won’t be allocated the labor they need. BrightRay is talking to neoclouds about selling entire data center modules, not just one-to-two-megawatt crates.
• Jeremie’s balance call, with the hedge intact: “We’re already in a sort of oversupply of chips today,” and it is easier to get GPUs than data center megawatts. Comparing supply-chain expansion into 2028, he sees a risk of the constraint flipping to data center oversupply. How fast chip supply expands is the key watch item; he calls the resulting CapEx, including new semiconductor fabs and equipment, a strong positive.
• On Jensen taking Gerstner’s framing in stride, Jeremie sees supply-chain aggression: Jensen secured GPUs for crypto-mining build-outs in 2021, which “obviously crashed,” and is writing “massive checks” to memory and logic suppliers. Jordan says Gerstner presented to audiences of 330,000 and 265,000, citing SemiAnalysis’s data and dismissing it. Jordan’s response: “Cite our data, provide specific critiques to it, great. We can always learn”—but an offhand “this is the forecast; we think it’s wrong” seems strange. Jeremie says the article was intended to address a broader topic that clients and others ask about, not to respond to one person.