Lenovo Expands U.S. AI Server Production as Power Demands Surge

Lenovo has doubled its North Carolina manufacturing footprint to 890,000 square feet, despite persistent doubts about producing advanced AI infrastructure in the United States. The expanded Whitsett campus now assembles and tests dense server systems near Lenovo’s American research operations. A detailed tom hardware tour shows why the location matters, and why local assembly alone cannot secure the entire supply chain.

Lenovo converted an empty warehouse into an operating server line within nine months. The facility can produce 1,400 server nodes daily across 28 assembly stations, according to the factory tour. It currently employs roughly 400 line workers and has room for 1,000.

Those numbers frame the central tension. Lenovo is a Chinese company expanding American production while U.S. customers increasingly want locally assembled infrastructure. Its strategy challenges the assumption that advanced server manufacturing must remain concentrated in established Asian production centers.

Dell, HPE, and Supermicro face the same demand for fast delivery, dense cooling, and access to scarce components. Lenovo’s answer is regional assembly placed close to customers and its Morrisville research hub. The factory still depends on globally produced processors, accelerators, memory, networking equipment, and storage.

That distinction is important. Whitsett represents meaningful domestic production, but not a completely domestic server. The expansion shortens the last major stretch between parts, integration, testing, and deployment.

Tom Hardware Reveals What Changed Inside Whitsett

Lenovo has turned its North Carolina operation from a familiar computer facility into a high-density AI server production center.

The expansion includes a 26,000-square-foot kitting area, where processors, memory, storage, and other parts are organized before assembly. Fifty workers reportedly handle more than 150,000 parts per day across ten kitting lines.

Kitting is the controlled preparation of every component needed for a specific system. Accuracy matters because AI racks contain many expensive, tightly matched parts. One missing adapter or incompatible module can delay an entire customer deployment.

Workers currently prepare about 200 racks weekly, according to the tom hardware account. Lenovo says the kitting area can support 400 racks per week as production rises.

Components then enter a 53,000-square-foot assembly area containing seven lines. Its 28 stations can collectively produce one server every 60 seconds, with stated capacity reaching 1,400 nodes per day.

A server node is an individual computing system installed inside a rack. Multiple nodes can work together with networking, storage, power distribution, and cooling equipment as one larger AI system.

The pace sounds similar to conventional electronics assembly, but the finished products present different challenges. AI servers pack accelerators, high-capacity memory, fast interconnects, and specialized cooling into a dense physical space.

Lenovo therefore needs more than assembly capacity. It must verify that each completed configuration operates under the heat and electrical loads expected at a customer site.

The Whitsett testing area contains ten temperature-controlled pods, each supporting 16 test racks. Large computer-room air handlers provide cooling, with each unit rated for one megawatt of capacity.

The site also has 30 Vertiv Liebert XDU liquid-cooling units, and Lenovo can expand that number to 50. Each unit can support up to 300 kilowatts during testing.

A coolant distribution unit transfers heat between the liquid loop inside a server rack and a facility’s cooling system. It helps technicians test liquid-cooled hardware before shipment.

This infrastructure changes what Lenovo can deliver from North Carolina. The plant is not merely fastening commodity servers into cabinets. It can assemble, connect, power, cool, and test integrated systems before customers receive them.

That matters because deploying an AI rack involves more than unpacking individual servers. A configuration problem discovered inside a data center can consume scarce engineering time and delay the start of expensive computing capacity.

Testing closer to the customer gives Lenovo a chance to find faults before shipment. It also creates a shorter feedback path between buyers, manufacturing teams, and engineers at Lenovo’s Morrisville operation.

Lenovo first established computer manufacturing in Whitsett in 2013. The latest expansion builds on that history, but the workload has changed substantially. Traditional PCs do not create the same power and thermal demands as accelerator-filled racks.

The factory also carries symbolic weight. Lenovo is demonstrating that a company headquartered in China can perform sensitive, high-value server integration inside the United States.

That positioning is especially relevant for public institutions and regulated enterprises. Those buyers often examine supply-chain location, system handling, support access, and delivery certainty alongside raw performance.

The expansion does not resolve every concern about component origin or cybersecurity. It does, however, give buyers a U.S. production and testing point that did not exist at this scale before.

Local Assembly Has Become a Delivery Strategy

The strongest case for Whitsett is speed and accountability, not a claim that every component originates in America.

Lenovo’s Morrisville research hub sits slightly more than an hour from Whitsett. The surrounding region also includes universities, community colleges, logistics networks, and experienced manufacturing workers.

Ben Massey, Lenovo’s vice president of supply chain, described an operating corridor between North Carolina’s Triangle and Triad regions. The company expects that proximity to connect research, supply-chain planning, assembly, and customer delivery.

The claim needs real performance data before buyers should accept it completely. However, the mechanism is straightforward. Shorter physical and organizational distances can reduce handoffs when engineers must diagnose a production problem.

A local line also gives Lenovo more control over final configuration. Enterprise AI buyers rarely order identical systems in unlimited quantities. Their racks can differ by accelerator, network interface, storage layout, cooling design, and software requirements.

Late configuration becomes valuable under those conditions. Lenovo can bring globally sourced parts into Whitsett, assemble the requested system, test it, and ship the completed rack within the region.

That approach reduces some exposure to long-distance transport for finished systems. Fully integrated racks are heavy, sensitive, and complicated to move. Shipping components for local integration can provide greater flexibility.

The strategy also addresses customer unease about where systems are handled. Lenovo executive Ashley Gorakhpurwalla told visitors that customers increasingly want infrastructure produced closer to them.

That demand is not necessarily an outright rejection of overseas manufacturing. It reflects a broader effort to understand vendors, component routes, service responsibility, and recovery options.

For enterprise buyers, locality can simplify factory inspections and acceptance testing. It can also shorten communication loops when a rack arrives with a cooling, firmware, or cabling issue.

Still, “made in America” requires careful interpretation. Whitsett performs substantial assembly and validation, but critical silicon largely comes from specialized fabrication and packaging networks outside North Carolina.

Nvidia accelerators, advanced memory, processors, networking components, and storage devices remain exposed to global constraints. American final assembly cannot manufacture a missing GPU or replace unavailable high-bandwidth memory.

The real comparison is therefore regional integration versus distant integration. Lenovo is betting that bringing the final stages closer to U.S. customers creates enough operational value to justify the investment.

This model pressures Dell, HPE, and Supermicro in different ways. Each competitor already emphasizes supply-chain reach, integrated systems, and services. Lenovo must prove that Whitsett improves delivery rather than simply adding impressive factory capacity.

Supermicro has long promoted U.S. server manufacturing and rack-scale integration. Its own expansion plans demonstrate that proximity has become an industry-wide competitive lever, not a Lenovo-only argument.

Dell brings extensive enterprise relationships and a broad infrastructure portfolio. HPE combines server systems with networking, services, and private-cloud offerings. Neither company can ignore a rival adding tested rack capacity near American customers.

Buyers should ask each vendor the same questions. Where will their system be integrated, how early can its configuration be locked, and where will full-load testing occur?

They should also ask what happens when a component misses its delivery window. A nearby assembly line provides limited comfort if unavailable accelerators leave its stations idle.

Lenovo’s expansion is consequently a promise about execution. It must coordinate global procurement with local labor, electrical capacity, cooling equipment, and customer schedules.

If that coordination works, Whitsett can reduce deployment friction. If it fails, the doubled building will expose the difference between theoretical capacity and completed systems.

The Production Line Is Also a Power Test

Whitsett shows that manufacturing AI servers now requires infrastructure resembling a small data center before any rack reaches its customer.

Lenovo and its contractors secured up to 20 megawatts of electrical capacity for the facility. Duke Energy reportedly moved the building from temporary service to full power within seven months.

That power does not merely operate conveyors, tools, and lights. Lenovo needs substantial electricity to test completed racks under loads that resemble real use.

The tom hardware tour cited Lenovo’s estimate that an Nvidia GB300 rack consumes 140 kilowatts. The company expects a future Vera Rubin rack to reach 230 kilowatts at peak load.

Those figures are Lenovo’s reported expectations, not independent measurements from production systems. Vera Rubin hardware had not yet arrived at Whitsett during the tour.

Lenovo expected its first Vera Rubin equipment by the end of August 2026. That arrival will test whether the facility’s cooling and validation systems match the next generation’s demands.

Rack-level testing explains the large air handlers, liquid-cooling equipment, electrical conduits, and water pipes visible inside the building. The assembly line cannot operate separately from those supporting systems.

Direct liquid cooling moves heat through liquid placed close to processors and accelerators. Liquid carries heat more efficiently than air, which helps tightly packed systems sustain higher electrical loads.

However, liquid cooling adds pumps, heat exchangers, connectors, controls, and leak-management requirements. It also creates another interface between server manufacturers and the buildings receiving their products.

A rack that passes testing in Whitsett still needs compatible cooling at its destination. Customers must have appropriate water temperatures, flow rates, electrical feeds, floor capacity, and monitoring systems.

This makes factory validation increasingly valuable. It also makes each sale more dependent on coordination among server vendors, chip suppliers, cooling specialists, utilities, construction teams, and data-center operators.

The manufacturing mechanism has therefore changed. AI server production now includes a preview of the deployment environment, not simply electronic assembly.

That shift favors companies capable of delivering complete racks and supporting facilities around them. Selling individual servers remains possible, but many buyers want integrated systems that can enter service faster.

The pressure extends beyond competing manufacturers. Utilities must decide how quickly they can connect factories and data centers without reducing reliability for existing customers.

Local governments also face questions about land, water, power generation, transmission capacity, and economic development. A server factory can create skilled employment, but its output enables more resource-intensive computing elsewhere.

North Carolina is already attracting related investments. Jabil announced a planned manufacturing facility supporting cloud and AI data centers in Rowan County. The North Carolina project calls for 1,181 jobs by 2030.

Together, these projects suggest that the state is becoming a manufacturing base for the physical AI supply chain. Universities, transportation routes, available industrial property, and utility partnerships all support that direction.

Yet production growth also compounds demand for trained electricians, manufacturing technicians, cooling specialists, and quality engineers. Companies expanding simultaneously will compete for many of the same skills.

Lenovo says Whitsett workers receive about 1,000 hours of training and rotate through production stages. That investment suggests the company views labor capability as part of its manufacturing advantage.

It also introduces scaling risk. Training enough employees for three shifts takes time, and production quality can decline when hiring outruns supervision.

The plant currently has roughly 380 operators, while Lenovo’s full-capacity scenario includes 1,000. Reaching that target requires much more than filling open positions.

New employees must handle sensitive components, follow detailed configurations, document work, and recognize defects. They must also operate safely near high-current electrical systems and liquid-cooling equipment.

The facility’s real constraint may shift over time. Accelerators can be scarce during one quarter, while trained labor, cooling units, or utility capacity becomes limiting during another.

Whitsett is designed to remove final assembly as one bottleneck. Its success depends on whether Lenovo can keep all the surrounding inputs moving at the same speed.

U.S. Production Does Not Mean a U.S.-Only Supply Chain

The central reversal is simple: Lenovo can localize the server’s final journey without localizing every technology inside it.

Public discussion often divides manufacturing into domestic and foreign categories. Modern AI systems resist that neat classification because their value chain crosses many countries and companies.

A rack assembled in North Carolina can contain accelerators designed by an American company and fabricated in Asia. It can use memory from Korean suppliers and cooling hardware built through another international network.

It can still represent meaningful American manufacturing. Workers perform component preparation, assembly, integration, firmware work, quality control, thermal validation, and shipment from Whitsett.

The problem begins when domestic assembly becomes shorthand for complete supply-chain independence. Lenovo has not established an independent American source for every critical part.

This matters because semiconductor availability determines whether server lines remain productive. Memory constraints and accelerator demand can disrupt delivery even when buildings, workers, and electricity are ready.

Industry reports previously indicated that server and PC vendors faced higher memory costs amid AI demand. Lenovo and Dell were both associated with expected increases in the hardware market.

No price figures are needed to understand the operational effect. When memory supply tightens, vendors must allocate components, revise configurations, or extend delivery schedules.

Advanced accelerators create an even narrower dependency. Nvidia controls much of the high-end AI accelerator market, while a limited group of manufacturers handles fabrication, memory, packaging, and system components.

Lenovo’s advantage lies in working across that network and turning parts into customer-ready infrastructure. It does not control every upstream production decision.

The same is true for its competitors. Dell, HPE, and Supermicro depend on overlapping silicon and component suppliers, even when their integration footprints differ.

This shared dependency changes the competitive contest. Vendors cannot rely only on access to the newest accelerator. They must show they can obtain complete component sets and deliver working racks.

Cooling presents another supply-chain layer. A server manufacturer needs validated relationships with companies producing coolant distribution units, pumps, manifolds, hoses, and monitoring controls.

Networking presents another. Large AI clusters depend on fast connections among accelerators and nodes, so missing switches or adapters can delay an otherwise complete installation.

Software also remains part of the delivered system. Firmware, drivers, cluster management, security settings, and workload tools determine whether physical capacity becomes useful computing capacity.

Local assembly can improve coordination across these layers. It cannot erase their interdependence.

Security-sensitive buyers may still prefer the Whitsett model. A domestic integration point can provide greater visibility into handling, testing, chain of custody, and final configuration.

However, buyers should evaluate evidence rather than geographic labels. Useful evidence includes component records, firmware controls, validation procedures, access restrictions, and documented incident responses.

The original tom hardware visit noted that photography was prohibited around proprietary equipment and customer systems. That restriction reflects legitimate confidentiality, but it also limits public verification of some factory claims.

Lenovo supplied many of the published capacity figures. The reporter observed the facility, yet long-term production output and customer lead-time improvements remain unverified.

That is not unusual for a newly expanded factory. Production capacity describes what a line is designed to support, while output measures what it actually delivers.

The gap can be substantial. Staffing, parts shortages, maintenance, configuration complexity, and customer timing all reduce practical throughput.

Lenovo therefore needs to publish or discuss operational results after the ramp. Useful measures would include completed racks, delivery times, first-pass testing rates, and utilization across the three planned shifts.

Competitors face the same burden. Announcing more floor space is easier than keeping it productive while product generations and cooling requirements change.

Vera Rubin offers an early test. If Lenovo receives, integrates, and validates the new systems on schedule, Whitsett’s flexibility argument will gain support.

If hardware arrives late or requires extended facility changes, the episode will show how quickly theoretical readiness can meet upstream reality.

The Community Tradeoff Starts Before Deployment

Lenovo’s expansion brings manufacturing jobs, but it also ties North Carolina more closely to AI’s growing demands for electricity, water, and industrial land.

The factory itself has secured up to 20 megawatts of power. That amount is smaller than a major data center’s demand, but it is notable for a manufacturing and testing operation.

Testing requires the plant to run dense racks before shipment. The process creates heat that must be removed through air and liquid-cooling systems.

The facility’s freshwater pipes and nearby electrical infrastructure make the resource requirements visible. These are not abstract cloud services. They are physical systems with local inputs.

Lenovo’s plant does not consume resources on the same scale as every data center using its servers. Still, it sits near the beginning of a chain that enables those larger facilities.

The debate should not reduce the expansion to either jobs or environmental pressure. North Carolina receives employment, training, investment, and a larger role in advanced manufacturing.

Communities also inherit planning responsibilities. Utilities need generation and transmission plans, while local governments must assess water, roads, emergency services, and industrial development.

The factory’s employment target deserves scrutiny alongside its capacity claims. Lenovo has indicated that it can reach 1,000 workers while operating three shifts.

Hiring hundreds of employees would create a meaningful local impact. The quality, durability, training requirements, and advancement paths of those positions will matter as much as the headline total.

Lenovo has operated in North Carolina for decades, giving it a longer local history than a newly arrived data-center developer. That experience can support trust, but it does not eliminate the need for transparency.

Residents need information about water use, electrical expansion, traffic, incentives, and emergency planning. Customers need information about delivery performance and component security.

The tom hardware story captured this dual reaction. The factory can inspire regional pride while reminding observers that AI demand increasingly competes for physical resources.

That conflict will intensify as rack power rises. Lenovo’s cited shift from 140 kilowatts for GB300 to 230 kilowatts for Vera Rubin indicates how quickly requirements are changing.

Higher rack density can reduce the floor space needed for a given amount of compute. It can also concentrate heat and electrical demand into harder engineering problems.

Efficiency improvements do not automatically reduce total consumption. If demand for computing grows faster than efficiency, overall power and water requirements can continue rising.

That rebound dynamic matters for communities evaluating new infrastructure. A more efficient rack can encourage organizations to deploy more racks and run larger workloads.

Manufacturing locations therefore become part of a broader public conversation. A factory produces systems, but those systems influence where data centers expand and what utilities must support.

Lenovo can strengthen its position by reporting facility performance with clear boundaries. It should distinguish factory electricity from customer data-center consumption and rated capacity from actual use.

Utilities and local officials should publish equally clear information. Residents need to know which grid upgrades serve one project and which improve the wider regional system.

The company should also explain how water moves through its test systems. Closed-loop cooling, discharge requirements, and operational consumption have different community effects.

No public evidence currently establishes that Whitsett is creating a local water shortage. The responsible skeptical position is to request measurements, not assume either harm or harmlessness.

The same standard applies to grid effects. Securing full power within seven months demonstrates fast coordination, but it does not prove that future expansion will carry no regional cost.

Lenovo’s long-term social license will depend on these distinctions. Manufacturing jobs create visible benefits, while utility burdens can be distributed across a much larger population.

An accountable approach requires regular data after the ribbon-cutting. It also requires community participation before another major expansion reaches the construction stage.

Three Signals Will Show Whether Lenovo’s Bet Works

Whitsett should be judged by completed deployments, next-generation validation, and transparent resource performance, not floor space alone.

The first signal is Lenovo’s initial Vera Rubin testing. The company expected equipment based on Nvidia’s next platform to reach Whitsett by the end of August 2026.

Arrival alone will not settle the question. Observers should look for successful integration with the plant’s electrical, cooling, and validation systems.

A smooth transition would strengthen Lenovo’s claim that Whitsett can adapt quickly across hardware generations. Extensive retrofits or delays would weaken that argument.

The power change makes this test particularly important. Lenovo told the visiting press that a Vera Rubin rack can peak at 230 kilowatts, well above its GB300 estimate.

New accelerators often change more than power consumption. Rack layouts, networking, firmware, coolant flow, service procedures, and validation equipment can also require updates.

Whitsett’s value depends on handling those changes without pausing production for long periods. Adaptability matters because AI hardware cycles move faster than conventional facility planning.

The second signal is sustained output relative to installed capacity. Lenovo says its lines can produce 1,400 nodes per day and eventually support 400 racks per week.

Those are design figures. Buyers should watch for evidence that production rises while quality, testing, and delivery remain consistent.

Staffing offers one visible indicator. Moving from about 380 operators toward 1,000 would suggest demand is translating into more shifts and practical utilization.

Customer delivery information would be even more useful. Shorter quoted lead times, fewer on-site integration problems, and repeat orders would support Lenovo’s regional model.

Lenovo’s financial reporting can provide broader context. Its fiscal third-quarter release said AI server revenue achieved high double-digit annual growth, though that statement covered the wider business.

The company’s quarterly results support the demand narrative. They do not isolate Whitsett’s contribution or prove the factory’s output.

That distinction should remain clear. Corporate revenue growth can coexist with underused capacity at one site, especially during an early production ramp.

The third signal is resource transparency. Lenovo, Duke Energy, and public agencies should clarify how actual power and water use develops as testing volume grows.

The facility’s 20-megawatt electrical capacity is a ceiling, not necessarily continuous consumption. Public reporting should separate peak availability from average operating demand.

Water information needs similar precision. The presence of large pipes does not reveal consumption, recycling rates, discharge, or seasonal effects.

Transparent measurements would strengthen confidence that regional manufacturing can scale responsibly. Missing or vague data would deepen questions about how AI infrastructure shifts costs to local communities.

Competition will provide another cross-check across all three signals. Supermicro, Dell, and HPE will continue promoting their own rack integration, delivery, and cooling capabilities.

Jabil’s North Carolina development also shows that Lenovo is not operating alone. More suppliers and manufacturers can create a stronger regional labor pool and vendor network.

They can also increase competition for electricity, workers, contractors, and industrial sites. The region’s advantage will depend on expanding those foundations alongside its factories.

For enterprise buyers, the immediate lesson is practical. Ask vendors for production location, component traceability, full-load testing details, and delivery evidence.

Ask whether a proposed rack has already run under its expected power and cooling conditions. Determine who owns remediation when the destination facility does not match the tested configuration.

Technical teams should preserve those answers with design documents, meeting records, validation results, and change decisions. A searchable engineering knowledge base can keep that evidence accessible during a long deployment.

Lenovo has made a serious physical commitment in Whitsett. The nine-month conversion, seven assembly lines, testing pods, and liquid-cooling capacity establish more than a ceremonial foothold.

However, the expansion should not be mistaken for a fully domestic supply chain. Its strategic value comes from shortening integration, testing, support, and delivery near American customers.

That model now faces measurable tests. Vera Rubin must fit the facility, installed capacity must become dependable output, and local resource use must remain accountable.

The next tom hardware visit should have more than new factory photographs to pursue. It should be able to compare Lenovo’s promises with operating data from a demanding production ramp.

Will Whitsett consistently ship next-generation racks faster than distant integration centers, or will components, staffing, and utilities remain the decisive constraints? Buyers should track those results before treating local assembly as a complete answer.