From Grid to Token: NVIDIA's AI Factory Buildout

Why NVIDIA Stopped Selling Chips and Started Selling Factories

NVIDIA now calls its customers' data centres "AI factories" because the unit of production has changed — from servers housed to tokens manufactured per watt of electricity delivered.

Take the metaphor literally. A conventional plant converts raw material into finished goods; an AI factory converts energy into tokens through accelerated computing, and NVIDIA measures the output the way an industrialist measures a line — tokens per second, tokens per watt, cost per token, utilisation and uptime. Power usage effectiveness describes the building. Tokens per watt describes the business.

Underneath sit five co-designed layers: energy, silicon (the Blackwell and Rubin platforms), physical infrastructure, models such as Nemotron, and the applications that consume them. The workloads are specific — always-on agentic systems, physical-AI reasoning and high-volume inference — and they run hot and continuously rather than in bursts. A plant that runs near full utilisation around the clock earns its capital back through throughput, so idle capacity is not slack; it is lost revenue against a fixed, largely debt-funded cost base. A stranded megawatt becomes intolerable, and a merely wasted one becomes expensive.

Once compute is revenue, everything that gates it — a grid connection, a permit, a coolant loop — becomes a claim on that revenue. NVIDIA states the premise without hedging: in the AI economy, compute is revenue. The corollary is what reorganises the industry: control a gating input, and you control a slice of the revenue behind it.

What Do "Land, Power and Shell" Actually Mean — and Why Is Power the Binding Constraint?

LPS is NVIDIA's shorthand for the site-level scarcities — land, utility power and the physical shell — that now gate GPU deployment more tightly than GPU supply itself.

Land means a site large enough for a multi-building campus. Shell means the buildings and their mechanical and electrical guts. Power means the grid connection, generation and substations — and it has become the tightest link. A company can hold ten billion dollars of GPUs and have nowhere to energise them, because the missing input is not silicon but interconnection: grid capacity, then substations, then buildings, then cooling.

The shift is what makes this strategic rather than logistical. For years the division of labour was simple — NVIDIA designed GPUs, the customer bought them, and the customer found the land, the power and the building. When a single campus needs gigawatts and a multi-year interconnection queue, GPU availability is no longer the only thing standing between capital and compute, and the party with the most to lose from an idle order book is NVIDIA itself. So it moved downstream, from selling accelerators to shaping the power, cooling and financing that decide whether those accelerators ever switch on.

The efficiency arithmetic is where the story turns commercial. In one representative power budget NVIDIA examined, only about 60% of the electricity delivered to a site actually reaches AI compute; the other 40% is drawn by power conversion, cooling, networking, storage, backup and facility overhead. Static rack provisioning makes it worse, reserving each rack's worst-case peak even when real workloads never draw it. NVIDIA's answer is DSX MaxLPS — Maximum Land Power Shell — software and facility planning aimed at pushing as much of a site's fixed electricity as possible into useful computation. The company says the approach can fit up to 40% more Rubin GPUs inside the same power envelope: roughly 40,000 Rubin GPUs in a 100-megawatt factory.

The Grid-to-Token Stack Is One Co-Designed Machine

NVIDIA's DSX platform reframes the data centre as a single system spanning capital, power, cooling, compute, networking and operations software, all tuned to one number: tokens per megawatt.

DSX now reaches across compute, networking, storage, facilities, simulation and operations — the layers a data centre used to procure separately, from vendors who never spoke to one another. The commercial objective is not to fit more servers into a hall; it is to extract more saleable output from each megawatt entering the fence line. Two design choices make the machine coherent. DSX Flex lets workloads respond to grid and price signals, so a factory can shift or throttle load when power is scarce or dear — turning a fixed liability into a grid-interactive asset the utility can plan around. A construction-and-operations digital twin then ties facility controls to GPU behaviour in real time, because at gigawatt scale the cooling plant and the power train can no longer be tuned in ignorance of what the chips are doing.

Reclaiming that power is engineering, not accounting. Static provisioning locks away the gap between a rack's worst-case peak and its real draw as stranded headroom; live telemetry and active power management let a site run closer to its true draw and return the recovered megawatts to compute. That is where the 40%-more-capacity claim comes from — not new silicon, but overhead clawed back. The table below groups the buildout by where value, and with it bargaining power, is accruing.

Where Will the Capital Actually Land?

The money is migrating from balance-sheet IT budgets into project-financed infrastructure, and NVIDIA has assembled institutional partners to mobilise more than US$500 billion to keep it moving.

In August 2026 NVIDIA signed memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build AI-compute financing platforms aimed at mobilising over half a trillion dollars of third-party capital. The framing is deliberate: an AI factory is treated as long-lived infrastructure — closer to a toll road or a power plant than to a rack of depreciating servers. That reclassification changes who lends, at what tenor, and against what collateral. Financing capacity, not chip allocation, increasingly decides whether a project reaches construction.

The Ohio deal shows the model in the concrete. On 17 August 2026 NVIDIA guaranteed SB Energy's PORTS-Pike Technology Campus, an 8-gigawatt site where OpenAI is the tenant on a 20-year lease and SB Energy builds, owns and operates. NVIDIA's role is the tell: it is not buying the campus or leasing it, but guaranteeing the money behind it — providing credit support to secure an initial 4.25 IT-gigawatts, with an option over the remaining 3.75, a guarantee reported in securities filings at up to roughly US$105 billion — and investing $1.5 billion directly into SB Energy. Separately, CoreWeave and NVIDIA said they intend to accelerate more than 5 gigawatts of AI-factory buildout by 2030, with NVIDIA holding an equity stake in the operator. It sits alongside Singapore's own 200 MW DC-CFA2 award, where the binding constraint is likewise power, not floor space.

The operating line is almost entirely electrons. Once built, a factory's largest recurring cost is power, which is why offtake now hinges on power-purchase agreements, on-site generation and behind-the-meter arrangements negotiated years ahead of first load. Financing follows the same logic — special-purpose vehicles, lease guarantees and asset-backed structures that price a twenty-year contract against hardware that turns over every two or three. That mismatch is the quiet fault line: the shell depreciates over decades, the GPUs over a single product cadence, and a Rubin-to-next-generation transition can strand a balance sheet that assumed the two aged together.

Scale is the other half of the lock-in. The ecosystem sells speed through repeatability: NVIDIA-certified systems and enterprise-validated reference architectures from Dell, HPE, Lenovo and Supermicro let an operator stamp out the next fifty-megawatt block without re-engineering it, and the ODM base — Foxconn, Wistron, Wiwynn, QCT, Pegatron, GIGABYTE — turns that design into racks at volume. Repeatability is what makes an eight-gigawatt campus financeable in phases instead of as a single wager. It also means each new phase is easiest to fill with more of the same architecture, which is the point.

The Five Strongest Incremental Beneficiaries

Set NVIDIA and the obvious semiconductor supply chain aside, and bargaining power accrues to whoever controls the scarcest link — which today is powered land, not silicon.

Powered land sits in front of the entire buildout. The asset that commands a premium is not an industrial parcel; it is a site with a credible utility connection, planning and environmental approval, fibre, a construction pathway, a creditworthy offtaker, and headroom for several equipment generations. Unpowered land near a city is not an AI asset. The Ohio arrangement is the template: the developer owns the campus, the hyperscaler signs the lease, and NVIDIA guarantees the power and shell so the silicon has somewhere to go.

Grid and electrical suppliers rank second, because every incremental megawatt needs physical iron — transformers, switchgear, protection, storage. Direct-to-chip liquid cooling ranks third and is the clearest equipment-level winner. Specialist EPC, MEP and prefabrication firms rank fourth, rewarded for commissioning IT and facility systems together and for shortening the interval between first power and production. Networking, optics and connectors rank fifth, with content that scales as clusters grow denser: NVLink for scale-up inside the rack, Spectrum-X Ethernet or Quantum InfiniBand for scale-out across racks, and the optics, connectors and cable that Amphenol and its peers ship by the reel. A Vera Rubin deployment is a fabric of rack-scale systems, not a shelf of standalone servers.

The Operators Carry the Risk Everyone Else Offloads

The highest revenue does not sit with the equipment vendors. It sits with the AI-cloud operators — CoreWeave, Crusoe, Firmus, IREN, Lambda, Nebius, Nscale, Yotta — that monetise the finished factory by renting compute. They also absorb what everyone upstream sheds: debt and financing cost, customer concentration, equipment depreciation, GPU-generation transitions, capacity utilisation, energy-price volatility, construction delay and the enforceability of long contracts. More revenue, and more ways to lose it.

Infrastructure software is the mirror image — a smaller pool, better margins, recurring. Thermal-simulation, power-orchestration and digital-twin vendors such as Cadence, Dassault Systemes, Schneider's ETAP, PTC, Phaidra and Emerald AI sell into every factory regardless of who builds it, and their strategic weight is rising because facility controls can no longer be divorced from workload behaviour. NVIDIA's DSX Exchange, MaxLPS and Flex exist to connect compute, network, cooling, electrical systems and grid signals into one loop. Selling the software that closes that loop is a better annuity than pouring the concrete around it.

How Does 800 VDC Threaten the UPS, PDU and Transformer Incumbents?

NVIDIA's shift to centralised 800-volt DC distribution collapses the conversion stages between grid and GPU, and in doing so it quietly redraws which electrical products gain content and which get designed out.

Announced with Microsoft, Google and more than 80 Open Compute Project partners, the 800 VDC architecture converts medium-voltage AC to direct current centrally, then runs an overhead busway to a row power center supporting up to 2 megawatts per row, with availability expected in 2027. The intent is fewer transformers, less AC switchgear, fewer power-distribution units and no rack-level AC-to-DC conversion. Each stage removed is a stage of loss avoided — and a product line disturbed. "All power equipment wins" is the wrong read.

Timing is the trap. NVIDIA puts 800 VDC availability at 2027, with one-megawatt-plus racks arriving the same year, so today's projects still specify conventional AC and will keep ordering it through the decade. An incumbent that reads that near-term order book as reassurance is measuring the wrong horizon. The exposed position is a franchise optimised for a distribution architecture that a co-design partner has already published a plan to remove — defended only by the fact that the removal has not happened yet. The best-positioned electrical companies are those able to migrate toward centralised rectification, DC distribution, solid-state protection, storage integration and control software, regardless of how large their AC UPS franchise is today.

Why Liquid Cooling Stops Being Optional at Rack Scale

At 100 to 200-plus kilowatts per rack, air runs out of physics — and NVIDIA's Rubin generation is engineered for 100% liquid cooling, coolant included on the networking, with fluid entering at up to 45 degrees Celsius.

The counterintuitive part is the temperature. Warmer coolant, not colder, is the design goal, because 45-degree fluid can reject heat through dry coolers and ambient rejection instead of energy-hungry chillers where climate allows. The chain shortens from GPU to chilled water to chillers to cooling towers, down to GPU to a warm liquid loop to a dry cooler. Electricity saved on chilling is electricity redirected to more GPUs — the MaxLPS logic, expressed in plumbing.

The mechanics are unglamorous and decisive. Heat leaves each processor through a cold plate, into a technology cooling loop, across a coolant distribution unit and out to the facility water system — a closed path with no server fans anywhere in a Rubin system. Get the fluid chemistry, flow rates or leak containment wrong and the most expensive silicon on the planet throttles or trips. CDUs matter most, because they are the interface between the building's water and the technology loop, and NVIDIA positions DSX as the common framework through which cooling suppliers plug in without re-engineering every facility from scratch. That is how an ecosystem hardens into a standard, and a standard into recurring service revenue.

The winners run well past the CDU vendor:

• Cold plates and CDUs — the primary heat path and the facility-to-technology interface, from Schneider's Motivair, Vertiv, Delta and Asia Vital Components.
• Pumps, manifolds, hoses and quick-disconnects — the plumbing that has to seal under load, from CPC and specialist suppliers.
• Dry coolers and heat exchangers — what makes 45-degree, chiller-light rejection possible where climate allows, from Trane and Danfoss.
• Coolant treatment, leak detection and thermal commissioning — the services that keep the loop alive and turn a one-off sale into recurring revenue.

Which Suppliers Get Stranded?

Not every data-centre incumbent wins; the exposed are those whose product still assumes yesterday's low-density, air-cooled, AC-distributed hall.

Traditional AC UPS, PDU and rack-level power-supply vendors keep strong near-term demand — most projects still build conventional AC — but face substitution as 800 VDC advances. Air-cooling-only suppliers stay relevant for general-purpose and hybrid halls, yet lose the highest-density AI footprints outright. Chiller-heavy designs shrink where a warm-water loop and a favourable climate let dry coolers do the work, though they do not vanish. Landowners without secured power discover that acreage is not an interconnection. Generic contractors find that high-density electrical distribution, liquid loops, contamination control and staged commissioning are a different trade from pouring a warehouse.

The pattern is consistent. The stranded assets are not old companies; they are old assumptions — that cooling is a building accessory, that power is someone else's problem, and that a data centre can be assembled from independently procured layers rather than engineered as one interdependent whole.

Why Singapore's SS 726 Turns a Global Thesis Into a Local One

On 27 August 2026 Singapore issued SS 726:2026, the world's first national standard for liquid cooling in tropical data centres — at the moment NVIDIA is standardising the global AI-factory architecture around the same subsystem.

Developed by IMDA and Enterprise Singapore through the Singapore Standards Council, SS 726:2026 codifies what NVIDIA's hardware assumes: cooling-loop and coolant-distribution-unit design, integration with existing air-cooled facilities, floor loading, piping-material selection, fluid-quality requirements, monitoring, commissioning, leak risk, and energy- and water-efficiency measurement. Singapore's release puts a number on the prize — liquid cooling can cut data-centre-level energy consumption by more than 30% against conventional air — while naming the tropical tax: heat and humidity drive corrosion, and the cooling fluid itself invites microbial growth and biofilm.

That tax rewrites the local playbook. NVIDIA's 45-degree dry-cooling logic is climate-dependent, and Singapore is the wrong climate for the easy version. High ambient temperature, humidity and water constraints mean hybrid plants, high-efficiency chillers, dense monitoring and water-treatment expertise stay more valuable here than in temperate markets — an inference the standard's own risk list supports. For a Singapore vendor, the standard reads less like regulation than like a scope of work: coolant-quality testing and treatment, corrosion-resistant piping specified to the standard, leak detection rated for occupied halls, integrators who can retrofit liquid into halls poured for air, and controls engineers who can make a CDU, a chiller plant and a rack of Rubin silicon agree on a single setpoint. None of that ships from Santa Clara. It is bought, installed and maintained within a few kilometres of the raised floor.

The New Scoreboard

The question has changed from "how many GPUs can I buy?" to "how many tokens can I extract from each megawatt I can actually energise?"

That inversion moves the bargaining power. It travels backward down the chain — away from the buyer of chips and toward the holder of the interconnection, the guarantor of the lease, the builder of the coolant loop and the writer of the control software that squeezes overhead out of the power budget. NVIDIA has read the map and is buying position on every square that gates a token: guaranteeing the power, financing the developer, specifying the cooling, and setting the standard the ecosystem plugs into.

In NVIDIA's own representative budget, about 40% of every delivered watt never reaches a processor — lost to conversion, cooling and overhead before it can produce a single token. Whoever closes that gap, or owns the substation feeding it, will price the token. Everyone else will rent.