What is the “five-layer cake” of AI?

A framework Jensen Huang uses to describe the AI industry stack: (1) energy, (2) chips, (3) infrastructure (land, power, and shell — data centers, cloud and neocloud services), (4) models, and (5) applications. His central point is that the model and application layers are impossible without the energy and infrastructure layers beneath them.

Why does agentic AI need ~1,000× more compute than generative AI?

Agentic systems don't just produce one response — they generate internal reasoning chains, plans, and tool commands, iterating many times before delivering a result. Each reasoning step consumes tokens (and therefore GPU cycles and megawatt-hours), and demand is multiplied further by rapid growth in the number of users.

Are AI data centers really different from traditional data centers as grid loads?

Yes. Traditional retrieval-era data centers were relatively steady loads. AI training campuses are far denser (100–600+ kW per rack, liquid-cooled), connect at transmission voltage, and can exhibit fast, synchronized MW-scale power swings — which is why utilities and RTOs increasingly require dynamic load models, ride-through evaluation, and in some cases EMT-level studies before interconnection.

What did Huang say about the U.S. power grid specifically?

He called the U.S. grid “a little antiquated” and described AI demand as the best opportunity in decades to modernize it — because, for the first time, market forces provide abundant customers willing to pay for new generation, including nuclear and other sustainable resources. He also indicated NVIDIA's near-term investment focus includes helping ensure U.S. “land, power, and shell” projects are sufficiently funded.

What should a developer do differently because of all this?

Treat interconnection as a first-order design input. Enter the queue with credible EMT/PSCAD models, realistic ride-through and reactive capability, complete study data, and a compliance plan for applicable NERC standards. Projects engineered for study-readiness move through the queue faster and avoid costly re-studies — which, in this market, is the difference between capturing the AI load wave and watching it land somewhere else.

AI Data Center Power Infrastructure & Grid Engineering Guide

A Coordinated Electric System Interconnection Review—the utility’s deep-dive on technical and cost impacts of your project.

Challenge: Frequent false tripping using conventional electromechanical relays
Solution: SEL-487E integration with multi-terminal differential protection and dynamic inrush restraint
Result: 90% reduction in false trips, saving over $250,000 in downtime

ERCOT enforces all of the above through simulation, which means your model is your compliance case. The bar is now high:

Whole-facility scope. The model must represent everything the IT load, the UPS and power conversion, the cooling plant, the protection and control systems in formats compatible with ERCOT's study platforms (PSS/E, PSCAD, TSAT).
Real control loops, not approximations. Generic textbook representations are unacceptable. The model must capture the actual inner control behavior of your power electronics.
Hardware-validated converter models. For electronic loads, the PSCAD model must be benchmarked against actual hardware testing including voltage ride-through and subsynchronous response. A model assembled from standard PSCAD library blocks fails by definition, because a generic block has never been tested against your vendor's hardware. The good news: validation is a hardware-type test, so results for a given converter product are reusable across every facility that uses it.
Format migration. Facilities that previously submitted the older composite load model (CMLD) format must transition to EPRI's PERC1 format.
Three checkpoints. Models are reviewed before the stability study begins (no model, no study), before each quarterly stability assessment, and for electronic loads one final time before energization, when you must submit as-built models with a documented comparison against the previously studied data and a sworn attestation that the model matches actual field settings. ERCOT's review takes 10 business days, extendable by 20 put it on your critical path.
A living obligation. Change your technology, controls, or relay settings in a way that affects ride-through including converting a crypto mining site to an AI data center — and you've triggered a new interconnection study, even if your megawatts don't change.

Parameter	Detail
System	230 kV / 138 kV transmission corridors, wind and wet-snow icing exposure
Data basis	15 years of minute-resolution forced-outage records + regional weather observations
Core methods	Event grouping, MVA performance curves, time-to-95%-restore, area outage rate curves, fragility modeling, rerun-history benefits, exceedance and log-domain risk metrics
Headline result	≈85% of maximum resilience benefit at 60% of original capital; worst-event restoration window cut from 11 days to 5 in rerun-history terms
Decision supported	Capital portfolio selection; resilience plan filing; post-investment verification framework

System / Topic	Governing Standard(s)	What It Controls
Overall plant electrical distribution	IEEE 141 (Red Book); IEEE 666	Distribution architecture, voltage selection, design of generating station auxiliary service systems
Power system studies	IEEE 399 (Brown Book); IEEE 551	Load flow, symmetrical/asymmetrical short circuit, motor starting methodologies down to the lowest LV panelboard
Protection & coordination	IEEE 242 (Buff Book); IEEE 3004.5; IEEE C37 series	Generator relaying (21, 59N, 87G), time-current coordination, selective clearing between LV and MV tiers
GSU / UAT / SST transformers	IEEE C57.12.00 and C57 family	Transformer ratings, impedance, testing, loading
HV switchyard breakers	IEEE C37.06	AC high-voltage circuit breaker preferred ratings
MV switchgear (13.8 kV)	IEEE C37.20.2; IEEE C37.20.7	Metal-clad construction, compartmentalization, vacuum breakers; arc-resistant design with plenum venting
MV cable	UL 1072; ICEA S-93-639 (NEMA WC 74)	Type MV-105 shielded cable, 133% insulation level for HRG systems
LV switchgear (480 V)	IEEE C37.13; UL 1558	Metal-enclosed LV power circuit breaker switchgear to 635 V, draw-out ACBs with electronic trip units
Motor control centers	UL 845; NEMA ICS 18	LV-MCC construction, MCCB/MCP protection for motors under ~200 HP
Motors	NEMA MG-1	Motor performance, starting characteristics, service factors
DC & battery systems	IEEE 485; IEEE 946	Lead-acid battery sizing (125/250 VDC), DC auxiliary system design
Grounding	IEEE 80; IEEE 142 (Green Book)	Ground grid step/touch potential limits; system grounding including high-resistance grounding
Lightning protection	IEEE 998	Direct-stroke shielding of switchyard and outdoor generator structures
Arc flash & electrical safety	IEEE 1584; NFPA 70E	Incident energy calculation; worker safety boundaries and PPE
Fire protection	NFPA 850	Fire protection and risk management for combustion turbine generating plants
Installation code	NEC (NFPA 70); NESC	Wiring methods inside the plant fence; overhead/outdoor clearances at the switchyard
Interconnection & compliance	FERC LGIP; NERC MOD-025/026/027, PRC-019/024/029, FAC-008	Interconnection process, model validation, protection/ride-through coordination, facility ratings

IFC / Construction Deliverable	Purpose
Stamped IFC packages	Legal basis for construction; P.E. responsible charge
Final relay settings & TCCs	Protection as-installed matches the coordination study
Calculation archive	Owner records; NERC audit evidence trail
Commissioning procedures	Safe, sequenced energization; MOD field testing
Construction support	RFIs, field changes, FAT/SAT witness
As-builts & model handoff	Operating baseline; future study currency

At the 2026 Global Conference, NVIDIA CEO Jensen Huang sat for a wide-ranging conversation about where artificial intelligence is heading. Most of the coverage focused on models, market caps, and the “boomer versus doomer” debate. But buried in that conversation was a message aimed squarely at our industry: the binding constraint on AI is no longer software or even silicon. It is energy, land, and grid infrastructure — the physical foundation that every megawatt of “intelligence production” sits on.

Huang described AI as a “five-layer cake”: energy at the base, then chips, then infrastructure (land, power, and shell — the data center buildings, cloud services, and neoclouds), then models, and finally applications. His point was blunt: without the underlying layers, no model is useful. For power systems engineers, utilities, transmission planners, and large-load developers, that means the AI build-out is now fundamentally an electrical infrastructure program — arguably the largest sustained load-growth event the U.S. grid has faced since post-war industrialization.

In this article, Keentel Engineering breaks down the four themes from Huang's remarks that matter most to the power sector — energy, chips, infrastructure, and U.S. re-industrialization — and translates each into concrete engineering and interconnection implications.

Huang framed the last two years as a phase change. Generative AI the ChatGPT moment taught machines to generate text, images, and video. But the industry quickly realized that generation enables two deeper capabilities: reasoning (generating internal chains of thought) and tool use (generating commands to control browsers, spreadsheets, and software). The result is agentic AI: systems that understand intent, reason, plan, and act.

The electrical punchline is in the arithmetic. Huang estimated that the computation required for agentic reasoning is on the order of 1,000 times more than simple generative AI and that demand is then multiplied again by rapid growth in the number of users. Older GPUs sold four or five years ago are appreciating in price because there is not enough compute to go around. He also noted that the computing paradigm itself is shifting from retrieval (pre-recorded content fetched from the cloud) to generation (every response computed fresh, in context). Nothing can be pre-computed and cached; everything must be generated on demand.

For grid planners, this is why data center load forecasts keep being revised upward. Retrieval-era data centers were storage-heavy and comparatively power-light. Generation-era AI factories are computation-dense, thermally intense, and power-hungry and their duty cycle looks less like an office building and more like a continuous industrial process with fast, large power swings driven by synchronized GPU training steps.

Keentel Engineering Insight

Agentic AI workloads change the electrical character of the load, not just its magnitude. Training clusters can exhibit coordinated MW-scale power steps in milliseconds as thousands of GPUs synchronize. Interconnection studies for AI campuses should evaluate voltage flicker, ride-through behavior, harmonic injection from high-density power electronics, and dynamic load-shedding schemes not just peak MW and thermal limits.

Layer	What It Includes	Power-Sector Implication
1. Energy	Generation, transmission, fuel supply, grid capacity	The pacing item. New gas, nuclear, renewables + BESS, and transmission must be financed and interconnected years ahead of load.
2. Chips	GPUs, memory, silicon photonics, packaging (7+ chip types per system)	Chip fabs are themselves 100+ MW loads with extreme power-quality requirements (SEMI F47 ride-through, ultra-low harmonics).
3. Infrastructure	Land, power, shell — data centers, cloud and neocloud operators	Site selection is now grid-driven: POI availability, queue position, and study timelines determine where capital lands.
4. Models	Frontier and open-source AI models	Model economics turned profitable in 2025–2026, meaning capacity expansion is now self-funding — load growth accelerates.
5. Applications	Healthcare, transportation, finance, retail, engineering	Every industry becomes a compute consumer — distributed inference load appears across the entire distribution system.

One remark deserves special attention from the finance and utility community: Huang observed that in the last three to six months, the gross margins of leading AI companies turned strongly positive. When a product is profitable, the rational response is to make more of it which is why AI companies are now “racing for capacity.” In power-system terms: the load growth is no longer speculative venture-funded demand. It is margin-funded industrial expansion, which historically is the most durable kind.

Huang gave the audience a physical picture of NVIDIA's current-generation systems: a single rack-scale computer roughly twice the width of a stage, weighing about three tons, containing on the order of 1.5 million parts, costing $4–5 million — with silicon photonics, advanced 3D-packaged memory, and liquid cooling throughout. A single AI factory contains a football field of these racks.

Translate that into electrical engineering terms and the design challenge becomes clear:

Extreme power density. Rack densities of 100–600+ kW push facilities to medium-voltage distribution deep into the white space, direct-to-chip liquid cooling, and in emerging designs, +/-800 VDC or high-voltage DC distribution architectures.
Transmission-level interconnection. Campuses of 300 MW to multi-GW connect at transmission voltage (115–500 kV) and require dedicated substations, often with on-site generation, BESS buffering, and NERC-jurisdictional facilities.
Dynamic behavior. Synchronized training workloads create fast MW-scale load steps; combined with 100% power-electronic interfaces, these loads raise real questions for frequency response, voltage stability, and protection coordination on the bulk system.
Compliance and study requirements. NERC's work on large-load reliability, ERCOT's large-load interconnection framework (including NOGRR282-era ride-through requirements), and evolving RTO/ISO study practices mean large loads are increasingly studied with the same rigor as generation including EMT-level modeling where dynamics matter.

Design Reality Check

An AI factory is best treated as a hybrid facility: part industrial load, part power plant (when co-located generation and BESS are included), and part power-electronics laboratory. The engineering disciplines that utility-scale solar and BESS developers learned over the past decade — EMT/PSCAD modeling, ride-through design, reactive capability planning, protection and control integration at the POI — now apply directly to the load side.

Huang made a forceful economic argument: AI is the United States' best opportunity to re-industrialize. He identified three categories of manufacturing plants driving the build-out — chip plants, computer plants, and AI factories — and estimated the opportunity at several trillion dollars of domestic capital formation, creating hundreds of thousands of jobs over the next four to five years. He credited market forces (anchored by roughly half a trillion dollars of purchase orders steered to suppliers willing to build in the U.S.) with accomplishing what subsidy programs alone had struggled to do.

Every one of those three plant types is an electrical infrastructure project first:

Chip plants. Semiconductor fabs draw 100–500 MW continuously with the tightest power-quality tolerances of any industrial customer. Voltage sags measured in cycles can scrap wafer lots worth millions. Fab interconnections demand redundant transmission supply, static transfer and ride-through design, and rigorous harmonic and flicker studies.
Computer plants. Server and system assembly plants are more conventional industrial loads, but they cluster around fabs and AI factories, compounding regional load growth on distribution and sub-transmission systems that were never planned for it.
AI factories. The AI factories themselves are the anchor loads the gigawatt-class campuses reshaping resource plans in ERCOT, PJM, MISO, the Southeast, and the Desert Southwest.

The strategic corollary Huang drew that America should export energy, chips, infrastructure, models, and applications at every layer reinforces the same conclusion: U.S. competitiveness in AI is now inseparable from U.S. competitiveness in building power infrastructure quickly, safely, and to standard.

Perhaps the most striking statement in the entire conversation, from a grid perspective, was Huang's direct assessment: the U.S. power grid is “a little antiquated,” and AI represents the first opportunity in decades to use market forces — rather than mandates — to fund grid modernization and new generation, including nuclear and other sustainable resources. His logic: for the first time, there are abundant creditworthy customers willing to pay for new power at scale.

He also disclosed where NVIDIA's own capital attention sits: closer to home and nearer on the horizon ensuring that “land, power, and shell” projects in the United States are sufficiently funded, potentially backstopping financings to get “power inserted.” When the world's most valuable chip company says its strategic investment focus is helping energize sites, the message to our industry could not be clearer.

Keentel Engineering sees this playing out across five fronts:

Grid Modernization Front	What the AI Load Wave Is Driving	Engineering Work Required
Generation adequacy	Gas peakers/CCPPs returning, nuclear uprates and SMR interest, utility-scale solar + multi-hour BESS as fastest-to-power options	Interconnection studies, PSCAD/EMT models, PRC-029 ride-through assessments, plant electrical design 30/60/90/IFC
Transmission expansion	New 345/500 kV lines and substations to deliver GW-class loads; network upgrades identified in cluster studies	Transmission line and substation design, POI engineering, short-circuit and stability analysis
Large-load interconnection	RTO/ISO frameworks (ERCOT large loads, PJM/NYISO load interconnection processes) formalizing study requirements for loads	Load ride-through modeling, dynamic load models, FAC/MOD/PRC compliance support
Co-located & bridge power	On-site turbines, fuel cells, and BESS bridging multi-year queue timelines; grid-plus-island hybrid architectures	Islanding studies, protection coordination, grounding design (WinIGS), microgrid controls
Power quality & reliability	Power-electronic loads and IBR generation converging on the same buses; low short-circuit-strength operation	Harmonic and flicker studies, weak-grid EMT screening, arc-flash and protection settings

Huang closed with an observation about national posture: the United States won the last industrial revolution not because it invented the technology, but because it applied it. His stated fear is not foreign competition but domestic hesitation — that fear of AI slows adoption and forfeits the lead. Whatever one's view of the broader AI debate, the infrastructure version of that argument is already settled fact: the countries and regions that can interconnect large loads and new generation fastest will capture the investment.

That puts uncomfortable but useful pressure on our industry's timelines. Multi-year interconnection queues, serial study processes, and late-stage discovery of network upgrade costs are now national-competitiveness issues, not just developer frustrations. The practical response is the one Keentel has advocated across every market we work in: treat grid interconnection — FERC LGIP processes, NERC MOD/PRC/FAC compliance, RTO-specific large-load rules — as a first-order design input from day one, not a late-stage administrative step. Projects that enter the queue with credible EMT models realistic reactive and ride-through capability, and study-ready data packages move faster, re-study less, and reach commercial operation sooner.

Key Takeaways

(1) Agentic AI multiplied compute demand ~1,000× and shifted computing from retrieval to generation — load growth is structural, not cyclical. (2) The bottleneck has moved to the bottom of the five-layer cake: energy, land, and shell. (3) AI factories, chip fabs, and computer plants are electrical infrastructure projects with generation-grade study requirements. (4) Profitable AI economics mean margin-funded, durable load growth. (5) The U.S. grid modernization opportunity is real — but only for projects engineered for interconnection speed from day one.

What is the “five-layer cake” of AI?
A framework Jensen Huang uses to describe the AI industry stack: (1) energy, (2) chips, (3) infrastructure (land, power, and shell — data centers, cloud and neocloud services), (4) models, and (5) applications. His central point is that the model and application layers are impossible without the energy and infrastructure layers beneath them.
Why does agentic AI need ~1,000× more compute than generative AI?
Agentic systems don't just produce one response — they generate internal reasoning chains, plans, and tool commands, iterating many times before delivering a result. Each reasoning step consumes tokens (and therefore GPU cycles and megawatt-hours), and demand is multiplied further by rapid growth in the number of users.
Are AI data centers really different from traditional data centers as grid loads?
Yes. Traditional retrieval-era data centers were relatively steady loads. AI training campuses are far denser (100–600+ kW per rack, liquid-cooled), connect at transmission voltage, and can exhibit fast, synchronized MW-scale power swings — which is why utilities and RTOs increasingly require dynamic load models, ride-through evaluation, and in some cases EMT-level studies before interconnection.
What did Huang say about the U.S. power grid specifically?
He called the U.S. grid “a little antiquated” and described AI demand as the best opportunity in decades to modernize it — because, for the first time, market forces provide abundant customers willing to pay for new generation, including nuclear and other sustainable resources. He also indicated NVIDIA's near-term investment focus includes helping ensure U.S. “land, power, and shell” projects are sufficiently funded.
What should a developer do differently because of all this?
Treat interconnection as a first-order design input. Enter the queue with credible EMT/PSCAD models, realistic ride-through and reactive capability, complete study data, and a compliance plan for applicable NERC standards. Projects engineered for study-readiness move through the queue faster and avoid costly re-studies — which, in this market, is the difference between capturing the AI load wave and watching it land somewhere else.

Keentel Engineering is a power systems and grid interconnection consulting firm serving utilities, developers, data center operators, and industrial customers nationwide. Our services span the exact intersection of the AI build-out and the grid:

Grid interconnection engineering. Load and generation interconnection studies, POI engineering, feasibility through facilities studies, and queue strategy across ERCOT, PJM, MISO, CAISO, NYISO, and the Southeast.
EMT modeling and studies. PSCAD/EMT model development and benchmarking for inverter-based resources, BESS, and dynamic large loads, including weak-grid screening and ride-through (PRC-029) assessments.
Substation and T-line design. Substation and transmission line design from 30% through IFC, protection and control, grounding system analysis (WinIGS), and arc-flash studies for AI campuses, fabs, and co-located generation.
NERC compliance. NERC O&P compliance support across the MOD, PRC, and FAC standard families for both generation and emerging large-load obligations.
Owner's engineer. Independent owner's engineer services for utility-scale solar, BESS, gas, and data center power infrastructure — from due diligence through commissioning.

If your organization is planning an AI factory, a chip fab, co-located generation, or the transmission to serve them, our team can help you engineer the interconnection path before it becomes the critical path. Contact Keentel Engineering at www.keentelengineering.com.

This article is published by Keentel Engineering for general informational and educational purposes only. Keentel Engineering is an independent engineering consulting firm and is not affiliated with, endorsed by, or sponsored by NVIDIA Corporation, the Global Conference or its organizers, or any other company, organization, or individual referenced herein. Quotations and paraphrases from the referenced conversation are provided for commentary and analysis. All trademarks, product names, and company names are the property of their respective owners. The content herein does not constitute engineering advice for any specific project; readers should consult a licensed professional engineer regarding their particular facts and circumstances.

In 1995, Sandip (Sonny) R. Patel earned his Electrical Engineering degree from the University of Illinois, specializing in Electrical Engineering . But degrees don’t build legacies—action does. For three decades, he’s been shaping the future of engineering, not just as a licensed Professional Engineer across multiple states (Florida, California, New York, West Virginia, and Minnesota), but as a doer. A builder. A leader. Not just an engineer. A Licensed Electrical Contractor in Florida with an Unlimited EC license. Not just an executive. The founder and CEO of KEENTEL LLC—where expertise meets execution. Three decades. Multiple states. Endless impact.