Jeremie Eliahou Ontiveros

The “half of 2026 US data-center capacity is canceled” story rests on a denominator SemiAnalysis says fails basic checks. The underlying report counted 12 GW scheduled for 2026 but only 5 GW under construction; Bloomberg’s “half of US data-center capacity is delayed” framing then propagated as cancellation evidence. Amazon alone publicly built 4 GW in 2025 and could add 5 GW-plus in 2026, while CoreWeave has another 1 GW underway. Jeremie’s verdict: “There’s a massive issue in the denominator.”

Cancellations are substantial, but they largely prune speculative early-stage options rather than erase committed capacity. Requested US data-center load exceeds 1 TW against roughly 750 GW of current system peak load, because hyperscalers may explore ten sites for one final investment decision. These “cloud-coded projects” can advertise 10 GW and 500 MW next year while showing no permits beyond a “Contact us” page; SemiAnalysis’s forecast moved less than 5% across the displayed vintages.

Oracle’s Project Jupiter illustrates the real risk: an advanced, financed project can still miss schedule because power infrastructure is unprecedented at gigawatt scale. Its desired New Mexico gas-pipeline route remains unapproved, with only a secondary route approved, and the applicable regulatory process has no precedent for finishing inside two years. Ellie said trucked CNG’s only proven delivery is around 200 MW and may even be optimistic. This is not an early-stage “cloud-coded” project, Jeremie argued, but a roughly $100 billion Oracle/OpenAI commitment confronting execution reality.

SemiAnalysis forecasts more than 40 GW of behind-the-meter data-center additions by 2028 because grid delivery promises can be less dependable than self-generation. Jeremie said 3 GW of deals had been signed in the prior two weeks and offered roughly 18 months from signing to first capacity as a rule of thumb. A utility might turn a handshake for 500 MW in 2027 into 100 MW in 2028 and the full amount in 2032; behind the meter lets buyers “control your destiny,” subject to permitting and supplier execution.

The capacity race remains financed principally by hyperscalers and the AI labs behind them. AWS buildout is associated with Anthropic, Microsoft’s with OpenAI and some Anthropic demand, and Meta’s typically with MSL. Less-capitalized neoclouds struggle to fund both buildings and GPUs unless backed by firms such as Blackstone or KKR. Once a gigawatt lease is signed, financing can follow rapidly—Jeremie cited roughly $25 billion raised for Vantage about a month later—while first capacity generally comes around 18 months after signing and full ramp is expected to be much faster than five years.

Power-equipment scarcity is attracting enough suppliers and substitute technologies that execution speed matters alongside turbine nameplate capacity. Major OEMs are expanding, Bloom Energy is described as unusually “AI-pilled,” and automotive factories producing 100 GW-plus annually at 40%-50% utilization could redirect engines into generation. With labor rates “going to the moon” and revenue per megawatt rising, fast installation and predictable schedules can matter more than paying a premium for equipment or electricity.

Gas is the time-to-power answer for the 2020s, while solar, batteries and nuclear remain longer-duration diversification. Solar’s grid capacity value may be only 10%-20% of nameplate, depending on the region, and a gigawatt-scale behind-the-meter design could require 20,000 acres; nuclear and SMR agreements remain slow and often non-binding. Reyk described a potentially bearish-looking “peak turbine” dynamic in 2026, when over-purchased units may become underused or reach the secondary market, while stressing that this would not invalidate the broader behind-the-meter thesis.

Modular on-site gas is becoming the fastest route to energizing AI data centers, with credible hyperscaler and AI-lab projects targeted for 2026-27. Jeremie distinguishes slow, roughly 500 MW-per-unit turbines from factory-made 3-4 MW reciprocating engines and 15-40 MW aeroderivatives. Elon Musk’s Memphis site had roughly 250-300 MW of on-site turbines in 2024 and may add another roughly 500 MW in 2025, alongside Stargate Abilene’s planned 300 MW and 160 MW of GE Vernova aeroderivatives: “Today it’s all about time to market.”

The durable opportunity is not merely new gas demand but gas equipment taking share from diesel backup generators. A project can use gas as primary power until its grid connection arrives, then retain the same equipment for backup—a requirement that becomes more valuable if training GPUs must later serve high-availability inference during a downturn. Andrew’s pushback was that this bridge spending could fade by 2028; Jeremie countered that gas and diesel equipment costs are roughly comparable, though gas makes pipeline access a decisive site-selection constraint.

SemiAnalysis sees hyperscaler capex continuing far above Street forecasts because forward construction and pre-leasing indicators still point sharply upward. After roughly 50% growth in both 2024 and 2025, its estimates imply approximately 35%-40% growth in 2026 and 25%-30% in 2027, taking five major buyers above $500 billion of annual capex in 2026. Meta’s soft indication of roughly another $40 billion—toward almost $100 billion—supports the thesis that year-on-year dollar growth will be roughly similar.

The principal accounting risk is a four-year GPU life replacing today’s roughly 5.5-to-six-year server assumption. Andrew relayed Jim Chanos’s claim that Meta’s overall useful-life figure was 11-12 years, while GPUs were, in Andrew’s recollection, depreciated at about three years; he also said Chanos’s alternative calculation implied 20 years. Historical high-performance-computing systems support five-to-six-year useful lives, but Jeremie concedes that modern GPUs are being run at maximum utilization and “it’s anyone’s guess” whether they wear out faster. A project can still show an estimated 40% EBIT margin, yet a shorter life could later produce a quarterly write-down of perhaps $10 billion.

CoreWeave’s advantage is execution speed in a bare-metal market with limited technical differentiation, but that same strategy embeds severe cycle risk. It contracted more than 2 GW in roughly two years by embracing crypto-miner brownfields and sometimes delivering in under a year, versus the traditional 18-month-to-four-year build cycle. The mismatch is stark: “Your longest GPU contract is going to be five years; your data center deal is going to be 15 years,” leaving ten years of rent exposure if replacement demand disappears.

Oracle is pursuing “CoreWeave’s path” with an investment-grade balance sheet, accepting long-duration infrastructure liabilities to win enormous AI contracts. Andrew cited leaked figures suggesting a roughly 15-year, $15 billion-$20 billion Crusoe commitment supporting an approximately five-year OpenAI deal, creating both major upside and the risk of roughly $1 billion-plus annual obligations after the customer contract ends. Oracle hopes GPU capacity will also attract OpenAI’s CPU, storage, and front-end workloads onto OCI, but SemiAnalysis sees little evidence yet that GPU colocation is producing that broader cloud upsell.

Remote locations will gain share, but cheap fuel or cold weather cannot substitute for labor, logistics, transmission, and reliable grid power. West Texas, Applied Digital’s Ellendale site, and Crusoe’s announced Wyoming project already show the shift away from established metros. Fully islanded generation can accelerate launch, but Jeremie calls the grid “the best power you can have on site”; absent transmission, locations such as the Permian remain expensive bridge-power propositions. Alaska was raised as a possibility, but Jeremie said he had not evaluated it specifically.

Better chips are unlikely by themselves to collapse electricity demand, while robotics should progress more slowly than language models. Nvidia is improving throughput per watt dramatically, but leading labs spend those gains on more capable frontier models because “the money maker is the frontier model,” leaving power and system price closely linked. In robotics, level-two mobile systems may be an interesting trend within one or two years, but scarce physical-world training data makes rapid progress to strong level-four humanoids unlikely over the next few years.

Ontiveros rejects “unlimited” as an investment premise, but his measured forecast is still a widening AI data-center shortage. Global IT capacity grew roughly 4 GW annually from 2019 through 2023; in 2024, NVIDIA hardware alone added about 5 GW of demand and AI accelerators collectively added 7-8 GW. SemiAnalysis expects construction to lag that demand for several more years, leaving upside across the infrastructure chain.

DeepSeek does not break the compute-spending thesis because its efficiency sits within an already extraordinary trend. Ontiveros says that, in two years, a model with GPT-3 quality should cost about 1,200 times less to run inference on. He argues Gemini 2.0 Flash is both cheaper and better than DeepSeek. DeepSeek itself remained inference-capacity constrained, illustrating why cheaper intelligence can still produce demand for much more compute and data-center capacity.

Running out of easily available text data increases the need for compute rather than ending training scale. Synthetic-data generation, reinforcement learning, stronger validation, and using huge models to improve smaller ones all consume substantial compute; Ontiveros cites GPT-4o and Claude 3.5 Sonnet as products of that broader approach. Today’s largest clusters are roughly 100,000 Hopper GPUs and 130 MW, while leading labs are planning gigawatt—and in some cases 2 GW—sites toward 2027.

NVIDIA’s underappreciated inference advantage may be scale-up networking rather than its familiar training-software moat. The GB200 NVL72 connects 72 GPUs in an all-to-all NVLink configuration at unmatched bandwidth, with 3,000-4,000 copper cables in the rack-scale system. Longer-context reasoning models require more memory and networking, leading SemiAnalysis to think NVIDIA’s moat “might actually be stronger in inference than in training.”

The AI upcycle breaks when adoption and monetization fail to support physical investment, not merely when models become more efficient. Ontiveros expects hyperscaler capex to keep beating estimates by 20-30%, but gigawatt-scale sites eventually require more than user growth. The indicators he would watch are product traffic, available indicators of OpenAI revenue and perhaps Claude revenue, and whether consumers keep paying $20 or $200 monthly. At some point, he says, the investment will be too much for the market.

The shortage makes “time to power” the most valuable infrastructure product. Schneider Electric, Eaton, Vertiv, Taiwanese liquid-cooling companies, and smaller engineering or cooling businesses can all participate; much of the equipment is more commoditized than NVIDIA’s GPUs. Behind-the-meter nuclear solutions such as Talen-style arrangements look much trickier than expected.

Bitcoin miners possess scarce powered sites, but converting them into AI campuses is a team-and-capital problem rather than a simple equipment swap. A mining facility costs roughly $0.5 million per MW, versus more than $10 million per MW for an AI data center, and customers want credible operators. Walker’s metaphor captures the optionality—miners are “clown cars that fell into a gold mine”—but Core Scientific and Applied Digital stand apart because they built data-center expertise and industry relationships early.

Miners’ power is not automatically suitable for AI. Walker flags West Texas curtailment agreements and intermittent wind and solar power as potential problems for always-on AI workloads. Ontiveros says generators, on-site batteries, and the UPS systems already used in data centers can cover outages, with battery reserves potentially extended beyond the usual 5-10 minutes.

The United States remains the default location despite cheaper overseas energy because of speed, labor, supply chains, and experience. Roughly 75-80% of hyperscalers’ large-scale self-built data centers were already American, giving the country experience with 100 MW-scale construction. Ontiveros says that no other country “currently knows better how to build large-scale data centers than the United States.”