Corvex Signs $103 Million Take-or-Pay NVIDIA Blackwell Contract to Double Mid-Atlantic ARR
$38 million non-recourse USD.AI debt and $26 million prepayment fund 82% of the $78 million HGX B200 buildout

Corvex, Inc. (Nasdaq: MOVE) disclosed on Monday a three-year, $103 million take-or-pay agreement to supply NVIDIA HGX B200 GPU capacity to an unnamed scaled AI inference platform — the small Virginia-based neocloud's largest contract to date and its third NVIDIA Blackwell deployment at its Mid-Atlantic 1 data center. The deal more than doubles the facility's contracted annual recurring revenue from $22 million to approximately $56 million and validates an increasingly visible playbook for mid-market GPU operators: combine customer prepayment with non-recourse asset-backed debt to fund Blackwell buildouts without issuing new equity.
The hardware commitment totals more than 1,000 NVIDIA HGX B200 GPU servers connected by NVIDIA Quantum-2 InfiniBand, drawing 1.6 megawatts of IT capacity. Total deployment cost is approximately $78 million. Rather than raising equity, Corvex structured the buildout so roughly 82 percent of that cost is covered by a $26 million customer prepayment and a $38 million non-recourse debt facility arranged through compute-finance platform USD.AI — leaving Corvex's equity outlay at approximately $14 million. Full service is expected by the end of the first quarter of 2027.
What 1,000 HGX B200 Nodes Actually Deliver for Inference
The NVIDIA HGX B200 is the flagship Blackwell server configuration: eight Blackwell GPUs per node connected through fifth-generation NVLink, with each GPU delivering 8 terabytes per second of HBM3e memory bandwidth and enabling GPU-to-GPU communication at memory-like speeds across the interconnect fabric. Each GPU carries 192 gigabytes of HBM3e, and the platform's second-generation Transformer Engine adds native FP4 precision support — halving the memory footprint versus FP8, so a 400-billion-parameter model that requires 16 B200s at FP8 can potentially run on eight at FP4, directly doubling concurrent inference capacity on the same hardware.
NVIDIA's internal benchmarks claim up to 15 times higher inference throughput for Blackwell versus prior Hopper H100 systems. NVIDIA's larger GB200 NVL72 rack-scale system achieves 25 times lower energy per inference token versus H100 air-cooled hardware on specific generative AI workloads; the HGX B200 node configuration achieves 12 times better energy efficiency versus the HGX H100. Independent reproduction of these figures at comparable conditions has not been published at scale, so they remain manufacturer claims. The NVIDIA Quantum-2 InfiniBand interconnect provides deterministic, lossless bandwidth across multi-node configurations — the preferred fabric for distributed inference of very large models where GPU-to-GPU communication is the throughput bottleneck. Corvex has previously deployed B200 hardware using Lenovo Neptune liquid-to-air cooling, a rear-door heat exchanger approach that achieves commissioning times of roughly two weeks rather than the seven to twelve months typically required to retrofit a full liquid-cooled data hall.
The USD.AI Non-Recourse Structure Powering the Capital Stack
The $38 million non-recourse facility comes from USD.AI, a compute-finance platform developed by Permian Labs that originates GPU-backed loans and distributes them to institutional buyers. Its standard structure places a first-priority lien on the GPU hardware, routes customer contract cashflows through a bankruptcy-remote special purpose vehicle, and amortizes over 36 months on a straight-line schedule. The loan's 36-month term matches the contract, keeping refinancing risk low but creating a simultaneous renewal decision in 2029 on both the debt and the revenue agreement.
USD.AI originated more than $500 million in GPU financing across 2026, including comparable facilities for NexGen Cloud in Scandinavia and Duos Edge AI in the United States, and launched a GPU Loan Exchange in October 2026 to connect standardized GPU-backed loans with credit funds and institutional buyers. The proliferation of these facilities signals that GPU-backed asset-backed lending has moved from an exotic instrument to a recognizable asset class in under two years.
Corvex's Differentiation in the Northern Virginia Corridor
Corvex completed its Nasdaq listing in March 2026 through an all-stock merger with Movano; its market capitalization at announcement was approximately $189 million — a fraction of CoreWeave's or Nebius's scale. Rather than competing for hyperscale GPU workloads, Corvex targets regulated enterprises and government customers who need dedicated, single-tenant deployments with verifiable security guarantees.
Its Assured AI platform combines NVIDIA GPU Confidential Computing, Intel Trust Domain Extensions, and remote hardware attestation, so workloads can be cryptographically verified before sensitive model weights are released to the system. The Secure Model Weights product holds cryptographic keys that are released only after the hardware environment passes attestation — protecting proprietary model parameters from the infrastructure operator itself. That positioning is niche but defensible in the Northern Virginia corridor, where federal and regulated-enterprise customers have compliance requirements that hyperscale shared infrastructure does not easily satisfy.
Corvex reported $3.8 million in revenue for the second quarter of 2026 against a net loss of $12.8 million — a company still in early ramp, not generating sustainable cash flow. The $103 million contract, once in service, represents an annual run rate of approximately $34 million from a single agreement. The disclosed $21 million in annual contract revenue per megawatt gives the mid-market Blackwell infrastructure segment a rare publicly verified pricing benchmark.
A Repeatable Equity-Light Formula — and Its Remaining Risk
Corvex raised a $33 million PIPE in August 2026 to fund its 5× capacity expansion. For the current $78 million deployment it is contributing roughly $14 million of equity while two external financing sources cover the remainder. Co-CEO Jay Crystal described the approach as "a repeatable formula for efficient growth — adding multi-year take-or-pay revenue backed by non-recourse financing and customer prepayments rather than issuing equity." Management has described requiring that capital, contracted offtake, and physical data center capacity be assembled simultaneously before a deployment is approved — a discipline designed to prevent speculative buildouts that cannot service their debt.
The deal's remaining uncertainty is the customer. A take-or-pay contract eliminates utilization risk for Corvex regardless of whether the customer's own traffic materializes. But a customer unable to pay, or one that renegotiates terms, would stress the matched-tenor debt at a point when three-year-old Blackwell hardware may trade at a significant discount to today's replacement cost. The contract counterparty is described only as a "scaled AI inference platform" — a phrase that leaves investors and analysts unable to independently assess the creditworthiness underpinning both the revenue forecast and the debt service. As inference demand continues to push contract buildouts down-market to operators like Corvex, the opacity of customer identity in take-or-pay disclosures will increasingly be the residual risk that neocloud investors are actually pricing.



