Cargo Theft Comes for AI Hardware: Who Owns the Risk Between the Factory and the Rack?

Data Centers & Energy By EnergyInnoCore Editor · Published Oct 27, 2026 · 11 min read

When we talk about the risks facing AI data centers, the conversation almost always starts with power. Can the grid deliver it, can the permits be secured, and who pays if the schedule slips. I have written about most of those questions in recent blogs. This week I want to look at a risk that lives somewhere almost nobody checks: the stretch of highway between the factory and the rack.

Over the past several months, trailers carrying AI servers and networking gear have become targets for organized cargo thieves. NBC News reported that thefts of AI data center equipment have resulted in more than $100 million in losses since June of 2026, a figure that Futurism repeated on October 26. FreightWaves, quoting Scott Cornell of TAPA Americas, an industry security association, put known reported thefts at $110.6 million since June 3, with single losses ranging from $2 million to $38 million.

Those are big numbers, and they deserve a careful read. In this post I will separate what is well supported from what is not, explain how these thefts seem to work, and ask the question that matters most for our industry: when a ten million dollar rack of hardware disappears in transit, who owns that risk, and who eats the delay?

How Much of the $100 Million Can We Trust?

Let me start with some honesty, because this is the kind of story where a big number travels faster than its evidence.

The $100 million plus figure is plausible and shows up consistently across sources. NBC and FreightWaves arrived at roughly $100 to $111 million. But there is no public, itemized tally behind it. The FreightWaves number is described as known reported thefts, with no end date and no list of incidents. Several sources also appear to draw on a small pool of the same people and reports, so they are less independent than they look.

Other numbers need even more caution. The supply chain risk company Overhaul published a press release on October 21 claiming more than $150 million stolen in 2026, thefts up 34 percent since 2024, and deception involved in 92 percent of cases. That is vendor marketing with no published methodology, and it does not define what counts as AI equipment. I would treat it as a signal about direction, not a measurement.

The most solid data point is broader. Verisk CargoNet reported that cargo theft in the United States and Canada reached $304.6 million in the second quarter of 2026, compared with $135.7 million a year earlier, even as the number of incidents fell 26 percent. That covers all cargo, not only AI hardware. But it tells an important story: fewer thefts, and much bigger ones.

Some cases are confirmed by official sources. In June the Cook County Sheriff recovered about $1.3 million of equipment in two trailers near Chicago. The largest cases from this summer, roughly $13 million in Fayetteville, Georgia and roughly $10 million in Texas, remain unrecovered with no public arrests, so their values rest on reports and estimates.

One more correction in the interest of accuracy. An October 1 incident involving Nvidia branded trailers made the rounds as evidence that thieves were targeting AI cargo by name. It turned out the trailers held sand, and the thieves left empty handed. Not every dramatic headline is what it seems.

Fraud, Not Force

Picture a cargo theft and most of us imagine armed men and a hijacked truck. The more common reality is quieter, and it happens on a phone and a laptop.

Modern freight moves through a chain of brokers and carriers. A company that needs a load moved posts it, a broker finds a trucking company, and that company sends a driver. Every link in that chain is a point where identity can be faked. Thieves can pose as a legitimate carrier using stolen or spoofed registration details, or set up a ghost carrier that exists only on paper. They can double broker a load, meaning the original carrier hands the job to someone else without the shipper knowing. Whatever the exact method, the pattern is the same: the load leaves the dock willingly, in the hands of someone who looked legitimate on paper.

The real attack surface is not the truck. It is the booking workflow, the moment when someone in a logistics office accepts a carrier and releases a very valuable load.

The Escort Attack and Why Escorts Alone Fall Short

Some shippers respond to the risk by adding security escorts, a second vehicle that follows the truck. Tom's Hardware, summarizing reporting from WIRED, described escort vehicles being rammed or forced off the road in California, and drivers suspected of complicity.

I want to be careful here. WIRED reportedly could not confirm the incidents with five local law enforcement agencies, so this is secondhand and should be treated as unverified. Even so, an escort protects against a roadside attack. It does not protect against a driver who is part of the scheme, or a carrier that was never real. Cornell's advice, as relayed in coverage, points toward the less glamorous controls: vetting drivers, verifying carriers, and slowing down the booking process so that someone can actually check who is picking up the load.

Why Thieves Want This Cargo

A natural question is where the stolen hardware goes. The honest answer is that nobody has shown a full pipeline, but the incentive is clear.

WIRED, as summarized by Tom's Hardware, reported that an Nvidia DGX B300 system sells for more than $1.1 million in China, compared with about $550,000 in the United States. That is double the price, in a market where export controls restrict legal supply. Cornell has estimated that 60 to 70 percent of stolen US loads leave the country, though that is an estimate about stolen freight in general, not a finding about this particular wave.

I want to draw a firm line between incentive and evidence. There is no public proof that these thefts are tied to China or any state actor. What we can say is that export controls, scarce supply, and a large price gap create a black market premium, and wherever a premium exists someone will try to capture it. As long as legitimate buyers wait in line for allocation, stolen units will have a buyer.

Who Owns the Risk Between the Factory and the Rack?

Now to the question that matters most for developers, owners, and lenders. When a truck leaves the manufacturer and a $10 million rack of equipment vanishes, who bears the loss?

The answer depends on a stack of contracts that many project teams never read closely. The terms of sale determine when title and risk of loss pass from the seller to the buyer. If risk passes at the factory door, the buyer owns the loss on the highway. If it passes on delivery to the site, the seller or its carrier does. Then there is the carrier's own liability, which in domestic trucking is often limited in ways that bear little relation to the value of a trailer full of GPUs. Then there is cargo insurance, which may carry transit limits, exclusions, security requirements, and conditions the shipper must meet before a claim is paid.

Here is where it gets uncomfortable. Some policies treat a load voluntarily handed to a fraudster differently from a load taken by force. Depending on the wording, theft by deception can be excluded or limited. Overhaul argues that every dollar in a claim hides about three more in losses. That is a vendor's estimate, but the logic is sound: the insured value of the hardware is often the smallest part of the problem.

If you read my earlier blog on Oracle's force majeure notice, this pattern should look familiar. In that case the question was who pays when a permit delays a campus. Here the question is who pays when a theft does. The answer is the same in both cases: whoever the contract says, which is often not who the project team assumed.

When a Stolen Rack Delays a Campus

Insurance can replace hardware in theory. It cannot replace time.

A rack worth $10 to $13 million is not sitting on a distributor's shelf. Current generation AI systems can be hard to replace quickly, and a replacement still has to be built, tested, shipped, and installed.

Now connect that to the energy side. A data center campus has an energization schedule, power commitments, interconnection milestones, and often tenant commitments with penalties for late delivery. If servers do not arrive when the power does, the project is paying for electricity capacity it cannot use. If the servers arrive and the power is late, the opposite problem appears. Either way, a missing truck turns into stranded capital.

This is why I think cargo theft belongs in the project risk register, right next to permitting, equipment lead times, and interconnection. Questions worth asking include the following. Is a theft or loss of equipment in transit a force majeure event under our contracts? Does the tenant lease treat it as a delay the landlord must cure, or as an excused event? Who is responsible for liquidated damages? Does the construction financing require insurance that actually covers theft by deception? These questions are much cheaper to answer before shipment than after.

A Practical Checklist for Developers

None of this requires exotic technology. Most of it requires discipline in a few places where speed usually wins.

First, verify the carrier at every handoff. Confirm the identity of the carrier, the broker, and the driver through independent channels, not only through the contact details on the booking. Be suspicious of last minute changes, brokers who appear partway through the chain, and pickups by drivers who differ from those on the paperwork.

Second, slow the booking down for high value loads. Add a cooling period or a second approver for any shipment above an agreed value. As Cornell advises, a slower booking process is itself a security control.

Third, build chain of custody. Photograph and document the load at every transfer, seal containers, record seal numbers, and require signed handoffs. When something goes wrong, documentation often decides whether a claim is paid.

Fourth, use trackers with tamper detection and do not trust a single signal. Place more than one tracker, hide some of them, and monitor for signal loss, unexpected stops, and location jumps that suggest spoofing.

Fifth, read the insurance and the contract. Confirm the transit limits, the theft by deception language, the security conditions attached to coverage, and the point at which risk transfers. If gaps exist, close them with endorsements, contract terms, or a different shipping structure before the first truck rolls.

Sixth, write the delay into the schedule. Identify which deliveries sit on the critical path, hold float where possible, and decide in advance who speaks to tenants, lenders, and the power provider if a shipment goes missing.

How to Read the Next Headline

A final word on reading cargo theft statistics, because more will come. When you see a figure, ask four questions. Does it cover all cargo, or AI equipment specifically? Is it a known reported total, or an audited one? Who produced it, and what do they sell? And is it counting incidents or dollars? CargoNet shows why: incidents fell 26 percent while losses more than doubled, and a reader who sees only one number draws the wrong conclusion.

Closing Thoughts

The AI buildout has taught us that power is the bottleneck, then that permitting is the bottleneck, then that equipment lead times are the bottleneck. Cargo theft adds a quieter lesson: the bottleneck can also be a handoff, or a booking made too quickly on a busy afternoon.

The totals are uncertain, the largest cases are unsolved, and nobody has proven where the stolen hardware ends up. But the exposure is real, and the contracts and insurance that govern it are, in many projects, still unread. The sensible response is not alarm. It is to treat the journey from factory to rack as part of the project, with an owner, a budget, and a line in the risk register.

EnergyInnoCore Editor

Independent analysis of power, grid and policy for AI data centers. Every article lists its sources. About the editor · Editorial policy

This article is for general information only and is not engineering, legal, financial or investment advice. See the disclaimer.