How to Underwrite AI Infrastructure Investments and Why GPU Financing Fails

Are you a borrower? Here’s GPU Financing 101, our guide for borrowers, instead.

Evaluating AI infrastructure projects purely based on offtake, the customer contracts that guarantee revenue, is a mistake. Power and hardware procurement often make or break a project. For both full data center builds and cluster (GPU servers) deployments, financiers should look beyond offtake agreements and contracted revenue.

Compute offtake is "easy"

Offtake matters for profitability. But as of early 2026, customer demand far exceeds current supply of compute.

Anthropic's annualized revenue run rate grew from ~$1B in early 2025 to nearly $20B by March 2026. ^{[1]Bloomberg, "Anthropic's Revenue Run Rate Tops $9 Billion as VCs Pile In" (January 2026)https://www.bloomberg.com/news/articles/2026-01-21/anthropic-s-revenue-run-rate-tops-9-billion-as-vcs-pile-in [2]Reuters, "Anthropic aims to nearly triple annualized revenue in 2026, sources say" (October 2025)https://www.reuters.com/business/retail-consumer/anthropic-aims-nearly-triple-annualized-revenue-2026-sources-say-2025-10-15/ [3]Bloomberg, "Anthropic Nears $20 Billion Revenue Run Rate Amid Pentagon Feud" (March 2026)https://www.bloomberg.com/news/articles/2026-03-03/anthropic-nears-20-billion-revenue-run-rate-amid-pentagon-feud} The revenue growth has translated into compute demand.

Operators tell us they have billions of dollars in compute demand. But a single customer will sign letters of intent with three different operators, then go with whoever delivers on time. Only the operators who hit milestones capture the revenue. ^{[4]Janus Henderson, "Data Center Power: Why the Key Risk Is Under Delivery, Not Overbuild" (2026)https://www.janushenderson.com/social/article/data-center-power-why-the-key-risk-is-under-delivery-not-overbuild/}

Schedule risk drives loss

Delays translate to lost customers and failed projects.

• Contract: customers are lost due to missed milestones.
• Cost: interest, development overhead, and fixed commitments accrue while GPUs are not earning revenue. Based on our modeling, for a typical 576-GPU B200 cluster financed at 70% LTV and 15% interest, the operator pays ~$232K/month in interest alone during the ramp period, plus facility costs, all before a single GPU-hour generates revenue.
• Technology timing: a data center built for today's hardware may not be ready for tomorrow's hardware. ^{[5]NVIDIA, "H100 Tensor Core GPU" product page (accessed March 2026)https://www.nvidia.com/en-us/data-center/h100/ [6]NVIDIA, "Blackwell Platform" product pages (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/blackwell-platform/}

What causes delays depends on whether you're building a data center or deploying GPU servers into an existing one.

1. The facility: site, power.
2. The cluster: procuring GPU servers and space.

Real delays are compounding

Of 110 data center projects slated to come online in 2025, 26% were delayed and 10% pushed back their commissioning dates as power, permitting, and construction constraints stalled timelines. ^{[7]Sightline Climate, "Q1 2026 Data Center Outlook" (February 2026)https://www.sightlineclimate.com}

About 16 GW of data center capacity is scheduled for 2026 across 140 projects, but only 5 GW is under construction as of February 2026. Sightline Climate expects 30-50% of these projects to be delayed. 25% of them have not disclosed how they will source power. ^{[7]Sightline Climate, "Q1 2026 Data Center Outlook" (February 2026)https://www.sightlineclimate.com}

xAI shows what aggressive execution looks like: Colossus 1 in Memphis went from empty warehouse to 100,000 operational GPUs in 122 days, using on-site gas turbines to bypass the grid entirely. ^{[8]Bisnow, "Musk's xAI Building Massive Memphis Data Center Hub At Record Speed" (2025)https://www.bisnow.com/national/news/data-center/musk-xai-building-memphis-data-center-bigger-faster-than-ai-rivals-131026} That speed came with regulatory consequences. xAI installed 35 gas turbines without Clean Air Act permits, and within a year the facility became one of Shelby County's largest NOx emitters. The EPA closed the loophole xAI relied on in January 2026, and environmental groups filed intent-to-sue notices. ^{[9]Politico, "'How come I can't breathe?': Musk's data company draws a backlash in Memphis" (May 2025)https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582 [10]WinBuzzer, "EPA Rules xAI Illegally Operated Gas Turbines at Memphis Data Centers" (January 2026)https://winbuzzer.com/2026/01/20/epa-rules-xai-illegally-operated-gas-turbines-at-memphis-data-centers-closing-loophole-that-allowed-year-long-pollution-without-permits-xcxwbn/} Most operators cannot replicate xAI's approach, and those who try face growing regulatory exposure.

OpenAI and Oracle hit the opposite outcome at Abilene, Texas. Their Stargate data center was capped at 1.2 GW instead of the planned 2 GW after power grid availability slipped more than a year beyond projections. Winter 2026 weather knocked portions of the campus offline for days when liquid cooling infrastructure failed. ^{[11]Reuters, "Oracle and OpenAI drop Texas data center expansion plan" (March 2026)https://www.reuters.com/business/oracle-openai-end-plans-expand-texas-data-center-site-bloomberg-news-reports-2026-03-06/ [12]Unite.AI, "OpenAI and Oracle Scrap Stargate Expansion in Texas" (March 2026)https://www.unite.ai/openai-and-oracle-scrap-stargate-expansion-in-texas/} OpenAI redirected future capacity to alternative sites rather than wait.

GPU procurement adds its own timeline

Even when a facility is ready, the hardware may not be. NVIDIA Blackwell B200 and GB200 GPUs are sold out through mid-2026, with a backlog of 3.6 million units from cloud providers alone. ^{[13]TokenRing/Times Online, "Nvidia's Blackwell Dynasty: B200 and GB200 Sold Out Through Mid-2026 as Backlog Hits 3.6 Million Units" (December 2025)http://business.times-online.com/times-online/article/tokenring-2025-12-29-nvidias-blackwell-dynasty-b200-and-gb200-sold-out-through-mid-2026-as-backlog-hits-36-million-units} Smaller orders (tens to hundreds of GPUs) face 6-9 month lead times. Large training clusters (thousands of GPUs for frontier AI labs) can wait over 18 months.

H100 availability has improved: lead times dropped to 2-4 weeks by early 2026, down from months-long waits in 2024-2025. ^{[14]Wecent, "NVIDIA H100 Stock Update Q1 2026: Availability, Lead Times, Global Shipping Trends" (2026)https://www.szwecent.com/nvidia-h100-stock-update-q1-2026-availability-lead-times-global-shipping-trends/} But H100s are previous-generation hardware. Operators building for the new Blackwell generation face a much longer timeline.

Power is the biggest hurdle today

Power projects show that being in line for grid access does not mean you will get power. As of end-2024, about 10,300 projects were waiting to connect to the U.S. power grid. Of those that applied between 2000 and 2019, only 13% ever became operational. 77% gave up and dropped out. ^{[15]Lawrence Berkeley National Laboratory, "Queued Up: 2024 Edition"https://emp.lbl.gov/queues}

U.S. offshore wind shows the same pattern. DOE reported a 53% pipeline increase to 80,523 MW by May 2024, while only three U.S. projects were fully operational at that time. ^{[16]U.S. Department of Energy, "Offshore Wind Market Report: 2024 Edition" (2024)https://www.energy.gov/eere/wind/offshore-wind-market-report-2024-edition}

Georgia as an anecdotal example

Georgia is one of the fastest-growing data center markets in the U.S., with 285 new facilities planned on top of 162 existing ones. Georgia Power's data center pipeline has reached about 35 GW of requested load. That's nearly equal to the state's entire existing grid capacity of roughly 38 GW. ^{[17]Data Center Dynamics, "Georgia Power: Load growth to treble in next decade due to data centers" (2025)https://www.datacenterdynamics.com/en/news/georgia-power-load-growth-to-treble-in-next-decade-due-to-data-centers/}

The state will need to handle grid approvals, substation and transmission upgrades, and power equipment orders, all on tight timelines. Then comes the workforce: trained electricians, lineworkers, and power engineers already in short supply nationally. And community opposition, environmental review, and local permitting add further delay. ^{[17]Data Center Dynamics, "Georgia Power: Load growth to treble in next decade due to data centers" (2025)https://www.datacenterdynamics.com/en/news/georgia-power-load-growth-to-treble-in-next-decade-due-to-data-centers/}

For power projects in the U.S., the median wait time to get connected has nearly doubled, from under 2 years in the early 2000s to over 4 years for recent applicants. ^{[15]Lawrence Berkeley National Laboratory, "Queued Up: 2024 Edition"https://emp.lbl.gov/queues}Georgia will need to beat these wait times to hit its targets.

How to evaluate

Larger projects look safer to underwrite. Well-funded buyers like OpenAI and Anthropic, plus demand that hyperscalers like Microsoft and Google can't serve internally, exclusively want big clusters. But large projects are also where execution risk compounds. Delays accumulate across power, site, procurement, and more.

In practice, smaller projects are safer to underwrite. If one operator fails, the portfolio survives, and shorter timelines reduce exposure to delay. Offtake from well-funded customers is harder to lock in at smaller scale, but demand as of early 2026 makes it manageable.

Utilization is the single biggest variable: keep the cluster at 80% utilization or higher. Based on our modeling, below ~60%, most clusters cannot cover monthly costs. But the lender in a 3-year loan is still safe even when utilization and rates are in the 60-80% range. The operator, rather than the debt holder, absorbs the slow payback.

Inference, running trained AI models for end users, drives demand for smaller clusters. Training contracts from large frontier models drive demand for larger ones. ^{[18]Futurum Group, "AI Workload Priorities Diversify as Enterprises Push Compute Beyond Training" (2026)https://futurumgroup.com/press-release/ai-workload-priorities-diversify-as-enterprises-push-compute-beyond-training/}

Evaluate based on the operator's deployment track record and supply chain relationships. The best projects can clearly deploy in the near term, on timelines backed by evidence.

Evaluating facility buildouts and cluster rollouts

Power and grid access are the bottleneck for facility buildouts as of early 2026. This includes substation upgrades, power equipment orders, and utility coordination timelines. ^{[19]Uptime Institute, "Global Data Center Survey 2024"https://intelligence.uptimeinstitute.com/resource/uptime-institute-global-data-center-survey-2024 [15]Lawrence Berkeley National Laboratory, "Queued Up: 2024 Edition"https://emp.lbl.gov/queues [20]FERC, "Order No. 2023: Improvements to Generator Interconnection Procedures and Agreements" (2023)https://www.ferc.gov/order-no-2023 [21]International Energy Agency (IEA), "Electricity 2024"https://www.iea.org/reports/electricity-2024}

1. Has the operator secured grid access and power before?
2. Has the operator completed similar projects before?

For GPU deployments, evaluate reliable uptime and procurement timelines. Hardware manufacturers and NVIDIA have long order backlogs as of early 2026. ^{[5]NVIDIA, "H100 Tensor Core GPU" product page (accessed March 2026)https://www.nvidia.com/en-us/data-center/h100/ [6]NVIDIA, "Blackwell Platform" product pages (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/blackwell-platform/}

1. Has the operator commissioned clusters of similar scale before?
2. Has the operator secured space with sufficient power and cooling?