Overhaul and Navium launched the Helix Consortium through Lloyd's of London on June 1, 2026, the first dedicated cargo insurance product for AI infrastructure: GPU clusters, AI chips, liquid-cooled servers and networking equipment in transit. Transit limits run to $75 million under a single agreement party, storage to $25 million, across all transport modes. It is a product priced for an asset class with almost no loss history.

Key Takeaways

  • $75 million transit and $25 million storage limits under one agreement party, against a standard marine cargo policy that typically caps at $5 million to $25 million and splits high-value shipments across multiple underwriters.
  • $700 billion of combined 2026 capital expenditure across Amazon, Microsoft, Alphabet and Meta, up 77%, roughly 75% of it directed at AI infrastructure.
  • $2.5 billion of Supermicro servers containing restricted NVIDIA GPUs were allegedly smuggled to Chinese buyers in an indictment unsealed in March 2026, with at least $510 million reportedly reaching its destination.
  • 98% disruption prevention is the rate Overhaul reports across monitored shipments, and it is doing the work a loss triangle would normally do in the pricing.

What the Consortium Actually Puts on Risk

Three parties contribute distinct capabilities. Overhaul supplies the cargo risk management platform, which the company reports protects $1.4 trillion of cargo value across more than 150 countries. Navium, a Lloyd's coverholder, leads the underwriting. The Fidelis Partnership participates through Lloyd's Syndicates 3123 and 2126, with ten further syndicates on the consortium.

Feature Helix Consortium Standard Cargo Policy
Transit limit Up to $75 million, single agreement Typically $5M-$25M, multi-party
Storage limit Up to $25 million Often sublimited or excluded
Transport modes Road, rail, ocean, air (all modes) Mode-specific endorsements common
Monitoring IoT label tracker per pallet, 24/7 SOC GPS on container, periodic check-in
Onboarding White-glove setup in 15 days 30-60 day underwriting cycle
Incident response 24/7 Global Security Operations Center Carrier claims team during business hours

The single-agreement structure is the part that changes the buyer's position rather than the price. High-value cargo of this size is normally covered by separate policies from multiple underwriters, each taking a slice of the total value, and the seams between those policies are where coverage gaps live. Consolidating to one agreement party removes the seams along with the documentation.

The underlying demand is not speculative. Gartner puts worldwide AI spending at $2.59 trillion for 2026, up 47%, with AI infrastructure alone above $401 billion, and IDC forecasts data center semiconductor revenue at $477.1 billion in 2026 rising to $843.2 billion by 2030. Every dollar of that eventually becomes hardware that moves from a fab in Taiwan to a data center somewhere else.

Pricing a Class With No Triangle

The actuarial problem is stated plainly by the asset. GPU clusters at current scale barely existed five years ago, so loss triangles, development factors and credibility-weighted experience have nothing to operate on, the same bind that shaped AI liability pricing and the standalone market that followed the CGL exclusions.

What replaces them starts with exposure rating off adjacent classes: high-value electronics, pharmaceutical shipments, semiconductor wafer transport. Each shares something with AI hardware and none is a clean analogue, so the credibility weight assigned to them is a judgment with no empirical support underneath it.

Value density is why the analogues do not transfer. A single pallet of H100 or Blackwell GPUs can exceed $10 million, a full container load $50 million or more, against $2 million to $5 million for a standard container of consumer electronics. Liquid-cooled racks add a temperature and vibration tolerance that would be irrelevant for a container of smartphones and destroys thermal interface materials here. And the same density sets the theft incentive: the Supermicro indictment describes conspirators transferring serial numbers between real and dummy servers with hair dryers.

Frequency has a substitute. Overhaul's per-pallet IoT telemetry and its reported 98% disruption prevention rate let the consortium price off a technology-enabled prevention framework rather than historical experience, which is roughly the path commercial fleet telematics took when monitoring data displaced experience rating.

Severity has no substitute, and that is where the parameter risk concentrates. A total loss on a $50 million shipment dominates the book's experience for years, and with no credible tail data the choice among a lognormal, a Pareto, or a mixed distribution with a mass point at total loss is unconstrained. The 98th-percentile loss estimate can move by a factor of three on that choice alone, which means the capital held behind this book is set by a distributional assumption rather than by data.

Lloyd's has run this sequence before. Cyber was a "newer or less well understood sector" in its own annual reports before 2016 and now has 77 cyber risk insurers in the market; satellite launch cover was priced on engineering assessment for decades before experience accumulated. The Fidelis Partnership already runs a Data Center Construction Consortium that committed $1.6 billion of capacity in 2025 across 12 lines of business, so Helix extends an existing arc from construction into transit.

Concentration Breaks the Assumption the Book Rests On

Cargo portfolios are built on independence across shipments. The AI hardware supply chain does not supply it.

NVIDIA holds roughly 90% of the market for AI training semiconductors. TSMC fabricates nearly all advanced AI chips on one island. Assembly sits in a handful of facilities across Taiwan, Malaysia and Vietnam, and the logistics funnel through a small number of ports and airports.

Work the scenario. A typhoon closes a major Taiwanese port for two weeks. Every AI server shipment routed through it takes delay, improvised storage, and elevated theft exposure while hardware sits in unsecured holding. That is not one claim with a large limit; it is dozens of claims arriving together, which is an accumulation profile closer to property catastrophe than to marine cargo.

Per-pallet tracking, route deviation alerts and a 24/7 security operations centre are real mitigations against the losses that happen one at a time. They do not decorrelate a node failure. Monitoring tells the consortium that fifty shipments are stranded at the same port; it does not stop them being stranded.

The export control dimension compounds it in a way the policy language has to answer rather than the technology. Hardware that crosses a border illegally may be uninsurable, which turns a theft or diversion loss into a coverage dispute over whether the shipment was ever an insured interest. Standard cargo wordings were not drafted to adjudicate that, and the $510 million that reportedly reached its destination in the Supermicro case is the shape of the exposure that produces it.

Further Reading

Sources