At the GPU Technology Conference (GTC) 2026 in San Jose, NVIDIA CEO Jensen Huang compressed the entire structural thesis of the modern artificial intelligence boom into five stark words: "Compute is your revenue now."
Huang was not describing a new architectural blueprint for real estate or a variation on traditional commercial construction. Instead, he was describing a fundamental rewrite of the corporate income statement. Across the technology sector, the terminology of the "AI factory" has become so ubiquitous on vendor product pages and marketing decks that it is easy for industry observers to dismiss it as mere corporate branding. That, however, would be a mistake. Beneath the polished veneer of marketing terminology lies a straightforward, unyielding economic question about how infrastructure operators generate revenue—a reality that now applies to any entity that controls industrial-scale megawatts of electrical power.
What an AI factory is, and what it changes
An AI factory is a specialized industrial facility whose primary measured output is token throughput rather than traditional metrics like uptime availability or stored data capacity. NVIDIA’s definition frames this infrastructure as a comprehensive system managing the full lifecycle of artificial intelligence workloads—stretching seamlessly from initial data ingestion through intensive model training to high-volume inference, with intelligence itself treated as the primary commercial product. While NVIDIA has utilized the nomenclature since 2024, it was at GTC 2026 that the concept crystallized into the central organizing thesis of Huang’s keynote address.
In parallel, "AI factory" also functions as a product category name, popularized largely by hardware giants such as Dell and NVIDIA. When hardware vendors deploy the term in commercial discussions, they are typically referring to a validated, pre-packaged technology stack available for purchase. Within the broader infrastructure market, however, the term defines an entire category of physical facilities.
The critical distinction separating these facilities from conventional enterprise data centers is not merely a linguistic rebrand designed to evade regulatory pushback or local opposition from communities wary of power consumption. The divergence reduces to a single mechanical fact. In a traditional colocation model, an idle rack of servers continues to bill the tenant. In an AI factory model, idle compute earns zero revenue while it steadily depreciates in value. This underlying financial pressure directly dictates every operational and utilization decision made further down the line.
Your supply is energy
In this new paradigm, operators do not simply purchase compute capacity; they purchase raw electrical power and convert it into intelligence. This reality turns every subsequent operational phase into a question of conversion efficiency. The defining metric has shifted to tokens per watt, or at broader facility scale, tokens per megawatt—a benchmark that NVIDIA now prominently features when marketing its latest hardware platforms.
Electrical power remains severely constrained, though perhaps not entirely in the manner typically portrayed by mainstream coverage. Speaking at GTC 2026, ERCOT CTO Venkat Tirupati revealed that approximately 230 gigawatts of large electrical loads currently sit within the regional interconnection queue, with roughly 70 percent of those requests originating from data centers, all weighed against a local system peak of approximately 85 gigawatts. The core operational problem is not merely a shortage of power, but a severe mismatch in timing. While a new electrical load can potentially connect to the grid within six to 18 months, constructing new electrical generation capacity routinely requires one to two years, and building out new transmission lines demands anywhere from three to six years. Consequently, massive demand for power arrives years before the utility supply built to serve it can be completed.
This timeline gap explains why operational flexibility has rapidly evolved into a critical interconnection strategy rather than a mere concession. A 500-megawatt grid request structured as 100 megawatts of firm power alongside 400 megawatts of flexible, interruptible power connects significantly faster while simultaneously earning valuable demand-response compensation. For modern AI workloads, this strategy is entirely practical. Firms like Luxor Energy and Bentaus successfully demonstrated a live deployment where an active GPU cluster reduced its power draw by 75 percent in under 500 milliseconds in response to an ERCOT Four-Coincident Peak signal, accomplishing the adjustment without disrupting the active inference workload running on the hardware.
Your product is tokens
Within these facilities, a token serves as the fundamental unit of model input or output, roughly equivalent to a word fragment. The crucial change lies in the accounting: token throughput represents the first data center output metric that maps directly to incoming revenue rather than passive infrastructure capacity.
Successive generations of hardware aggressively accelerate this metric. NVIDIA has cited benchmarks indicating that its GB300 NVL72 systems are capable of delivering roughly 50 times more tokens per megawatt than older Hopper-generation systems. Industry analysts note, however, that these manufacturer-run benchmarks utilize specific, optimal conditions and are best interpreted as performance ceilings rather than conservative planning figures.
The temporal nature of the work has shifted as well. Traditional model training historically operated in distinct bursts: operators booked a cluster, trained a model, and halted operations. In contrast, the autonomous, agentic workloads described by OpenAI’s Sachin Katti during GTC do not stop. A facility that generates revenue around the clock represents a vastly superior business model compared to one that earns revenue in sporadic bursts, even though it introduces significantly higher operational complexity.
Your demand is the compute rental market
This operational framework exposes infrastructure operators to a more challenging economic reality. Model providers establish the baseline price per million tokens, which subsequently dictates what neocloud providers can charge per GPU-hour, which ultimately determines what an underlying facility operator earns per megawatt. Just a few steps down this commercial chain, operators function essentially as price takers, selling GPU-hours at market rates they can observe but cannot actively influence—a dynamic tracked closely by industry benchmarks such as the AI Hardware Price Index.
Operators frequently build two fundamental assumptions into their financial models, both of which face severe market pressure. The first assumption involves the projected useful life of the hardware. Speaking at GTC, CoreWeave CEO Michael Intrator argued that ongoing advancements in inference disaggregation could successfully extend the operational useful life of high-performance GPUs from the traditional assumption of four to five years out toward eight to ten years. This thesis relies heavily on software reliability capable of supporting heterogeneous hardware over extended periods. Should this assumption hold true, lifetime token output would roughly double, making capital expenditure recovery the primary financial concern.
AI factory digital twins and the retrofit question
Within this industrial ecosystem, a digital twin refers to a physics-accurate simulation of a physical facility rather than a virtual avatar. These sophisticated models integrate power, cooling, networking, and structural data into a unified simulation, allowing engineering teams to rigorously test design modifications and failure scenarios before physical construction begins. NVIDIA’s Omniverse blueprint for AI factory design serves as the primary reference implementation for this approach, and its application in greenfield developments is well established.
The retrofit case, however, represents the more pressing challenge facing the majority of infrastructure operators. When dealing with an existing building burdened by fixed physical constraints, operators must determine whether the site can realistically achieve a competitive tokens-per-megawatt efficiency figure before committing substantial capital. Constructing a digital simulation allows teams to evaluate the physical facility systematically, starting with the structural floor: can the concrete slab safely support the immense weight of high-density server racks? Next comes cooling infrastructure: can a closed-loop liquid cooling system be installed, and where will the massive thermal output be exhausted? Finally, operators must evaluate internal power distribution at the rack level. Interestingly, structural floor loading constraints halt more facility conversions than cooling limitations, and it remains the most cost-effective item to verify on the entire checklist.
Do you sell the megawatt or the token?
The entire conceptual framework of the AI factory alters strategic decision-making only if an organization’s primary unit of revenue is the token. An operator’s specific position along the broader value chain dictates whether metrics like tokens per watt represent a critical line item on their own P&L statement or simply concern their tenant.
A traditional real estate landlord supplies electrical power, the physical building shell, and cooling infrastructure, earning predictable revenue through contracted kilowatts while absorbing tenant credit risk rather than direct utilization risk—representing the lowest capital expenditure alongside the lowest revenue ceiling. Conversely, an infrastructure operator that owns the compute hardware and sells GPU-hours carries the highest capital expenditure and the highest revenue ceiling, while bearing both utilization risk and residual asset value risk. Meanwhile, an operator operating under a signed offtake contract receives consistent compensation regardless of whether the underlying compute hardware is actively utilized, as the buyer has absorbed that operational risk. In such cases, growth is constrained primarily by financing availability and the credit quality of the buyer.
Even a traditional landlord must monitor tenant token efficiency, because a tenant operating an inefficient facility will struggle to sustain long-term lease payments. For digital infrastructure operators with roots in cryptocurrency mining, power procurement, grid curtailment economics, and site operations map directly to these new demands, with the emerging demand-response structures closely mirroring how miners have historically interacted with regional power grids.
What the AI factory frame is actually for
The terminology of the AI factory earns its place in industry lexicon because it finally establishes a quantifiable unit of output that correlates directly to revenue—a correlation that traditional uptime metrics and Power Usage Effectiveness (PUE) ratings never achieved. The framework does not dictate that every infrastructure operator must become a direct token producer. Instead, it forces operators to identify precisely which unit they are selling and to price their services accordingly.
Ultimately, adopting the AI factory perspective will not automatically generate profit for an operator. Rather, it clarifies precisely where money is accounted for across the value chain, revealing whether an enterprise actually owns that crucial step or if it belongs to someone else entirely.
Cryptocurrency Mining
Cryptocurrency Mining
Cryptocurrency Mining
