Summary

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.

Deep dive

1. Modular gas is compressing the time required to energize AI clusters

Jeremie’s opening call is a surge in on-site natural gas—not primarily the large, approximately 500 MW-per-unit turbines whose lead times stretch multiple years, but smaller modular systems that can be factory-produced and deployed far faster.
Two architectures are gaining traction: 3-4 MW reciprocating internal-combustion engines, essentially “giant car engines,” supplied by companies including VoltaGrid and Caterpillar, and 15-40 MW aeroderivative turbines supplied mainly by GE Vernova and Caterpillar. Jeremie sees credible hyperscaler and AI-lab projects using both in 2026-27, perhaps with slightly greater momentum behind turbines.
The scale is already changing. Memphis represented roughly 250-300 MW in 2024; Elon Musk may add another roughly 500 MW during 2025, while Stargate Abilene is expected to deploy 300 MW and 160 MW of GE Vernova aeroderivatives. “Today it’s all about time to market.”
These installations can take budget from diesel generators rather than represent wholly incremental equipment demand: gas supplies primary power during the grid wait, then becomes backup once the substation arrives. That is the pattern Jeremie sees at Memphis and expects at Abilene.

2. Backup power remains valuable after the grid connection arrives

Andrew’s challenge: spending hundreds of millions on gas equipment that eventually serves only as backup sounds like an upcycle expedient, not a sustainable 2028 design choice—particularly if batteries improve and cheap diesel can cover the residual outage risk.
Jeremie’s rebuttal is that backup has always been a stranded-looking but necessary expense; normalized over its few operating hours, diesel power might cost thousands of dollars per MWh. The Spanish blackout illustrated why data-center operators still prize uninterrupted operation even when the equipment rarely runs.
Gas also preserves flexibility across the cycle. A training cluster may initially tolerate outages, but during a downturn its owner could stop renting fresh capacity and repurpose existing GPUs for inference, where high availability matters. “You just shift from diesel to gas.”
He disputes the assumed cost gulf: comparable Caterpillar diesel and natural-gas engines have roughly similar pricing and cost structures. The real distinction is geography—diesel needs a storage tank, while gas requires proximity to a pipeline.

3. Forward indicators say hyperscaler capex is still accelerating in dollars

Sell-side models generally assume high-single-digit or low-double-digit capex growth after 2025. Jeremie argues that semiconductor cycles rarely settle into smooth plateaus: history more often delivers a strong, persistent upswing followed eventually by a downcycle.
SemiAnalysis instead sees self-built construction starts surging and hyperscaler pre-leasing remaining exceptionally high. Since pre-leased capacity becomes operational later, those commitments are a direct forward indicator for next year’s spending, and they point toward high-double-digit growth rather than stabilization.
The progression is extraordinary: roughly 50% capex growth from 2023 to 2024, another 50% in 2025, then an estimated 35%-40% in 2026 and 25%-30% in 2027. Five major buyers could exceed $500 billion in 2026 alone; the signals still read “up, up, up.”
Meta supplied the clearest early indication after SemiAnalysis published its work: from roughly $65 billion-$70 billion in 2025, its soft guidance implied almost $100 billion in 2026. That is deceleration by percentage, but roughly another $40 billion in absolute spending.

4. GPU depreciation is a real uncertainty, but the bear case is often misstated

Andrew relays Jim Chanos’s bearish framing: Meta’s overall useful-life figure was allegedly 11-12 years, while Andrew recalls GPUs being depreciated at about three years; Andrew also says Chanos’s alternative calculation implied 20 years. Jeremie’s figures are different: about six years for servers, Meta at 5.5 years by his recollection, and CoreWeave at 6.4 years for servers.
Public high-performance-computing systems and industry experience support approximately five-to-six years of useful operation. The honest caveat is workload intensity: operators now try to “sweat every single watt” from expensive GPUs, so historical HPC longevity may not transfer perfectly. “It’s honestly anyone’s guess at this point”; moving from six years to five or four remains a legitimate risk.
SemiAnalysis estimates that Oracle’s largest OpenAI-type projects can produce about 40% EBIT margins at the project level, before corporate overhead. Yet if hardware lasts four years rather than six, the relevant company could eventually record a concentrated write-down—perhaps a quarter containing approximately $10 billion of losses.

5. Contracted project economics matter more than blended corporate ratios

To avoid the fast-growing-bank illusion Andrew raises—where enormous new deployment conceals poor returns on earlier vintages—SemiAnalysis models GPU clouds project by project, itemizing the infrastructure capex and operating expenses rather than inferring profitability from consolidated growth.
Outcomes vary sharply with data-center cost, operating discipline, financing, and network architecture. A year earlier, some neoclouds faced 15%-20% debt costs and 20%-plus equity costs; Oracle’s efficient large-cluster networking is one example of capex optimization.
Taken together, those variables can move the modeled result from a random neocloud at roughly 5%-10% ARR to a 25%-30% AR business, in Jeremie’s wording. The transcript does not define the distinction between those ARR and AR labels.
The hyperscalers appear to underwrite similarly, securing four-to-five-year contracts that provide largely guaranteed ARR unless something goes wrong or AI fails. The long contracts matter because so much capex is incurred upfront.

6. CoreWeave wins a commodity market by delivering capacity first

Jeremie characterizes very large bare-metal GPU contracts for OpenAI, Anthropic, and hyperscalers as closer to a commodity than traditional cloud: the provider builds the facility, installs machines, and networks them, with comparatively little differentiated software layered above.
Long term, the lowest cost structure should win; during the current shortage, speed dominates. CoreWeave contracted more than 2 GW in roughly two years by accepting financial risks others avoided and partnering early with crypto miners whose power and substations already existed.
That brownfield strategy bypassed greenfield timelines. Established data-center operators often take 18 months to four years from construction start to revenue, whereas some Core Scientific-related capacity moved in under a year. “Speed is really what enabled them to gain those contracts.”
CoreWeave’s costs may not have been best-in-class, but scale created trust and room to integrate vertically, including through its Core Scientific acquisition. The company took infrastructure approaches specifically optimized for AI while incumbents were still applying a conventional cloud playbook.

7. CoreWeave and Oracle both carry a dangerous duration mismatch

Andrew’s pushback—worth keeping—is that CoreWeave optimized for precisely what works in an upcycle: maximum speed, aggressive commitments, and risk absorption. A six-month demand pause could expose obligations that looked intelligent only while every available GPU had a buyer.
Jeremie agrees completely. “Your longest GPU contract is going to be five years; your data center deal is going to be 15 years.” Failure to replace the initial GPU contract can leave ten years of rent with no associated revenue.
The underwriting is therefore a long-term bet that OpenAI and the broader AI industry will keep consuming GPUs and power, allowing successive hardware generations and customer contracts to refill the same facilities. It is not a claim that the contractual mismatch has somehow disappeared.

8. Oracle used its balance sheet to make a CoreWeave-style wager

Oracle’s strategy is “CoreWeave’s path” backed by an investment-grade signature. That standing provides access to experienced operators such as Digital Realty, yet Oracle also made a nontraditional bet on Crusoe when it still looked, on paper, like a crypto miner without conventional uptime credentials.
Andrew cited leaked figures suggesting approximately five years of OpenAI revenue supported by a Crusoe facility commitment around 15 years and reportedly $15 billion-$20 billion. If OpenAI does not renew, Oracle could owe $1 billion or more annually for another decade.
Strategic upside extends beyond GPUs. After entering cloud late in 2016-17 and remaining far smaller than the rival hyperscalers, Oracle hopes GPU capacity will attract OpenAI’s CPU, storage, and traditional front-end workloads; Jeremie cited the roughly 700 million weekly users as an example of those needs.
SemiAnalysis has not yet found much evidence that this upsell is occurring, and GPU infrastructure can connect readily to another cloud provider. Still, Andrew’s broader reading survives Jeremie’s scrutiny: Larry Ellison made a “big bold bet” after missing the first cloud wave, and it is paying off so far.

9. Power access and infrastructure, not cold weather alone, determine where the next clusters can scale

Jeremie rejects the premise that AI facilities remain concentrated near major cities. West Texas is expanding; Applied Digital’s Ellendale, North Dakota site is “really far away from everything”; and Crusoe has announced a massive Wyoming development near Cheyenne.
More remote deployment should continue, but labor and logistics can dominate climate or land advantages. Developers must bring thousands of workers to the site, move equipment reliably, and secure fiber, gas, and electrical infrastructure. Alaska was raised as a possibility, but Jeremie said he had not evaluated it specifically.
The Permian Basin demonstrates the missing ingredient: abundant gas without substantial transmission infrastructure. Fully islanded generation works for rapid time to market but is expensive; Jeremie calls the grid “the best power you can have on site,” with gas ideally transitioning to backup. Grid power also offers lower electricity costs and higher uptime.

10. Efficiency gains are being reinvested into intelligence, not power savings

Andrew’s bear scenario is physical bottlenecks provoking a step-change in efficiency: if future GPUs need 50% less electricity, gigawatt data centers could become overbuilt even while compute demand continues growing. Jeremie accepts oversupply as “completely possible,” not an absurd tail risk.
Cloud history nevertheless points the other way. Azure’s reported electricity consumption has tracked revenue growth fairly closely: more efficient CPUs lowered unit costs, but customers consumed enough additional compute that aggregate power demand continued rising.
Nvidia’s roadmap already improves throughput per watt dramatically, even as system price and total power rise together—Blackwell’s two compute dies are the clear example. Hardware efficiency is real; the unresolved choice is whether customers bank the savings or use them to produce more tokens and capability.
Leading labs choose capability because the best frontier model earns the strongest economics. Anthropic’s coding leadership attracts usage, while cheaper mid-tier models face competition from distillation, open-source firms, Chinese labs, and the big labs. Hence “the money maker is the frontier model,” sustaining the incentive to convert efficiency into intelligence.

11. Robotics offers nearer-term level-two adoption, not an LLM-speed leap

SemiAnalysis maps robotics from level zero rigid arms, through flexible pick-and-place systems and level-two mobile quadrupeds, to weak level-three and strong level-four humanoids. The taxonomy is deliberately simplified but makes present capabilities and investable milestones easier to distinguish.
Level-two systems such as mobile robot dogs are already entering early production and could become an interesting trend within one or two years. Strong level-four humanoids remain in the research phase; Jeremie sees little evidence they will arrive within the next few years.
Physical-world data is the bottleneck. Language models inherited trillions of internet tokens, while robotics lacks comparable volumes of high-quality action data, making an LLM-like acceleration difficult even as progress improves.
The labor conclusion stays cautious. ChatGPT can synthesize advanced research yet confidently insist “there are three Bs in blueberry,” and robotics adds further physical-world challenges. Jeremie prefers the spreadsheet analogy: automation changed accountants’ work rather than eliminating it, perhaps shifting human value toward judgment, relationships, and trust.

Industry Watch 32: Embodied Robotics—Bubbles, Challenges, Value Lovable CEO, Anton Osika: The State of Foundation Models, Grok vs OpenAI, and Replit vs Bolt