Lambda Prices $926M GPU Loan as IBM Commits $240M to Open-Source Inference

Lambda finances GPU capacity with asset-backed debt, changing cloud economics

Lambda priced a $926 million senior secured term loan to fund GPU cloud infrastructure, marking the first time a non-hyperscaler has used structured debt at this scale for committed customer deployments. The facility is priced at SOFR + 3.00%, matures December 31, 2030, and is secured by the GPU servers and contracted customer cash flows. Lambda described the structure as "fully amortizing" and aligned with the useful life of the hardware, and expects to replicate it for future builds.

This matters because it makes dedicated GPU capacity cheaper for Lambda's customers than traditional cloud pricing. Asset-backed financing shifts capital costs from equity (expensive) to debt (cheaper), allowing Lambda to offer lower per-GPU rates while still covering amortization and interest. For enterprise buyers, this creates a credible alternative to hyperscaler spot and reserved instances, particularly for multi-year commitments where Lambda can lock in capacity and pass through savings.

The financing structure also constrains Lambda. Fully amortizing debt means Lambda must maintain high utilization to service the loan, which reduces flexibility to offer steep discounts or absorb unused capacity. Enterprises negotiating contracts should ask whether their reserved capacity sits in the SPV (subject to lender covenants) or Lambda's balance sheet, and what happens if Lambda cannot meet debt service.

IBM deploys $240M in NVIDIA B300 GPUs for open-source models

IBM and Together AI signed a multi-year, $240 million agreement to deploy NVIDIA HGX B300 clusters on IBM Cloud, with availability expected in Q1 2027. The infrastructure targets open-source model inference, positioning IBM as a managed alternative to AWS, Azure, and Google Cloud for enterprises that want to avoid proprietary model lock-in.

IBM Cloud has historically lagged the hyperscalers in GPU capacity and AI workload share. This build-out narrows that gap, but only for a specific use case: hosting and serving open-source models like Llama, Mistral, and Stable Diffusion. For enterprises already using IBM for mainframe, database, or regulated workloads, this creates a plausible path to consolidate AI inference under a single vendor relationship. For net-new buyers, the Q1 2027 availability date means IBM is not an option for near-term production deployments.

The competitive dynamic is clear. AWS offers open models through Bedrock and SageMaker, but bundles them with proprietary orchestration. Azure is tightly coupled to OpenAI. Google prioritizes TPUs and Vertex. IBM's partnership with Together AI—a platform built specifically for open model hosting—signals a bet that enterprises will pay for neutrality and interoperability. Whether that bet pays off depends on IBM's ability to match hyperscaler pricing and latency, which it has not yet disclosed.

NVIDIA releases AI Enterprise Infrastructure 8.2 and 7.8, clarifying support timelines

NVIDIA released AI Enterprise Infrastructure 8.2 (Production Branch) and Infrastructure 7.8 (Long-Term Support Branch) with explicit end-of-life dates: April 2027 for 8.x and July 2028 for 7.x. These software stacks—driver branch R595 for 8.x, R580 for 7.x—are what OEMs and cloud providers certify for NVIDIA GPU servers, and what enterprises must align to for NVIDIA support.

The 7.8 LTSB release with support into mid-2028 gives conservative IT organizations a multi-year stability horizon. For regulated industries that lock in validated stacks and cannot update drivers every quarter, this is the branch to standardize on. Infrastructure 8.2 targets newer GPUs and bleeding-edge features, but with a shorter lifecycle, forcing more frequent change management and regression testing.

Procurement teams can now bake EOL dates into multiyear contracts and hardware refresh plans, avoiding the risk of deploying new clusters on a branch that expires mid-project. The clearer timeline also narrows one of the cloud's traditional advantages—managed infrastructure with abstracted update cycles—by making on-prem GPU clusters more operationally predictable. For enterprises comparing build versus buy, this is one fewer variable in the total cost of ownership calculation.

What to watch

Lambda's debt structure will be copied if it works. If the company can maintain high utilization and low default risk, other GPU cloud providers will follow with their own asset-backed facilities, potentially lowering the cost of dedicated capacity across the market. If Lambda cannot service the debt, it creates a cautionary tale about the limits of financing GPU infrastructure outside hyperscaler balance sheets.

IBM's Q1 2027 availability date is the forcing function. If Together AI and IBM can publish competitive cost-per-token and latency benchmarks before launch, this becomes a credible multi-cloud option. If they miss the window or fail to match hyperscaler economics, the $240 million build-out becomes another IBM Cloud also-ran. Enterprises planning 2027-2029 AI roadmaps should ask for early access terms and SLAs now, while IBM still has capacity to negotiate.

NVIDIA's branch strategy clarifies the tradeoff between stability and performance. Enterprises must now choose explicitly between long-term support and cutting-edge features, rather than hoping a single stack delivers both. That choice has budget implications—LTSB deployments amortize validation and testing costs over a longer period, while production branch deployments require recurring investment in change management. Factor both timelines into your three-year infrastructure planning.