Independent technical due diligence for GPU cluster and AI datacenter financings

• GPU-backed loans finance the IT, not just the building. The collateral has a disclosed accounting life of four to six years at the companies that publish one^{[52],[51],[50],[1],[53]}, and its revenue depends on power, cooling and operations the borrower often has not yet proven.
• The most common technical gap in a data room is arithmetic: the node count in the financial model needs more power than the site has at the meter. One NVIDIA DGX B300 system is rated at 14.5 kW^[15]; the average rack in Uptime Institute's 2024 survey drew 8 kW^[26].
• “10 MW” means different things. Utility capacity at the meter, critical IT capacity and contracted-but-not-energized capacity can differ by tens of percent. CoreWeave, for example, reported over 850 MW of active power against about 3.1 GW of contracted power at the end of 2025^[1].
• Revenue starts at acceptance. Lenders should see the acceptance protocol, the pass thresholds and the burn-in plan before close, and should tie draws to independently certified milestones.
• Availability SLAs must be checked against what the hardware can deliver. In Meta's 54-day Llama 3 pre-training snapshot on a 16K H100 cluster there were 419 unexpected interruptions, about 78% of them attributed to confirmed or suspected hardware issues^[49].
• An independent technical advisor turns each of these into a finding, a condition precedent or a covenant, and then monitors them through construction, acceptance and operation.

Lending against GPU clusters has become a recognised asset class. CoreWeave's $2.3 billion facility led by Magnetar and Blackstone in 2023^[2] and its $7.5 billion facility led by Blackstone and Magnetar in 2024^[3] were followed by a $2.6 billion secured facility in 2025 to deliver services under a long-term agreement with OpenAI^[4]. Lambda secured a special purpose GPU financing vehicle of up to $500 million led by Macquarie^[5]. Crusoe closed a $225 million facility led by Upper90 to be used exclusively for NVIDIA GPUs and supporting cloud infrastructure^[6]. At campus scale, a J.P. Morgan-led consortium provided a $7.1 billion construction loan for the second phase of a 1.2 GW AI data center in Abilene, Texas^[7], and TeraWulf's subsidiary priced $3.2 billion of senior secured notes whose collateral includes a lockbox account of its GPU cloud tenant^[9].

The structure is consistent. CoreWeave describes it plainly in its annual report: it finances infrastructure “through asset-level debt supported by take-or-pay customer contracts”, the loans available are constrained by a percentage of the depreciable cost of GPU servers, and the facilities are secured by the equity and substantially all assets of special-purpose subsidiaries^[1]. Committed contracts were over 98% of its 2025 revenue^[1].

That structure changes what technical diligence has to answer. A conventional independent engineer's report on a data center underwrites a building: shell, power and cooling plant designed for decades, leased to tenants who bring and own their IT. In a GPU financing the lender is underwriting the IT itself, plus the facility it depends on, plus the operator's ability to keep it available. Four differences matter most.

• Asset life. Servers are depreciated over four to six years by the public companies that disclose it^{[52],[51],[50],[1],[53]}, and Amazon shortened part of its fleet from six to five years citing “an increased pace of technology development, particularly in the area of artificial intelligence and machine learning”^[50]. The loan has to amortize inside that window.
• Obsolescence. Each NVIDIA generation has raised power per system: 10.2 kW for a DGX H100, about 14.3 kW for a DGX B200 and 14.5 kW for a DGX B300^{[11],[14],[15]}, and up to 142 kW for a GB300 NVL72 rack^[17]. A facility designed for one generation may not host the next without rework, which matters for re-leasing and for any refresh assumption in the model.
• Power density. The average rack in Uptime Institute's 2024 survey was 8 kW (7.1 kW excluding outliers)^[26]. NVIDIA's DGX B300 SuperPOD reference design places four systems per rack at about 56 kW per rack^[21]. Rack-scale systems are liquid-cooled. Most of the facility questions follow from this gap.
• Offtake SLAs. Revenue depends on acceptance and then on availability, measured per GPU or per node, often with service credits. CoreWeave's 10-K discloses that it has experienced “insufficient power to service a customer's project” and has been required to provide service credits as a result^[1]. Technical shortfalls flow straight into the cash flows that service the debt.

None of this replaces the conventional facility review. It adds a second layer: whether this specific hardware, in this specific facility, run by this specific team, will produce the contracted service on the contracted date.

In project finance the independent engineer, also called the lender's engineer, is the engineering representative of the lenders on a large capital project. The role is typically selected by the lender and paid by the borrower, and its value depends on being independent of both sides^[10]. In GPU financings the same role is often called the lender's technical advisor (LTA), the technical due diligence provider or the independent technical consultant.

The mandate usually runs in four stages: pre-close technical due diligence and a report the credit committee can rely on; construction and deployment monitoring, including certification of draw milestones; witnessing of acceptance testing; and post-close monitoring of operating performance against covenants. A lender can buy any subset. The earlier the advisor is involved, the more of the findings can be turned into conditions precedent rather than post-close arguments.

The clients are broader than banks. Private credit funds lending to neoclouds, infrastructure funds and private equity investing in AI campuses, equipment lessors, insurers of completion or residual value, and offtakers committing to multi-year capacity purchases all face the same question from different seats: will the capacity exist, on time, at the contracted performance?

The risk map below is the frame we use to organise a data room review. Each row is a way the credit can be impaired by an engineering fact, the failure it causes, and the documents that answer it.

The rows are not equal. The first two, power availability and power sizing, are the hardest to fix after close. A cooling design can be revised and operations staff can be hired. A missing 2 MW of utility capacity can take years to close, because it waits on the utility, not on the borrower.

Utility interconnection and energization

Grid power is the binding constraint on AI capacity. The 2024 Lawrence Berkeley National Laboratory report for the U.S. Department of Energy found that data centers consumed about 4.4% of U.S. electricity in 2023 and projected 6.7% to 12% by 2028^[38]. Utilities and grid operators are responding with new rules for large loads, and lenders should read them as closely as the borrower's schedule.

Three distinctions matter. First, load interconnection (a data center taking power) follows the utility's large-load process and tariff, which is different from generator interconnection. FERC Order No. 2023 reformed generator interconnection queues, moving them from first-come, first-served to cluster studies with readiness requirements^[39]; it matters to a GPU project when the site relies on new on-site or co-located generation. Lawrence Berkeley National Laboratory found that the typical generation project built in 2023 took nearly five years from interconnection request to commercial operation^[40]. A schedule that assumes new generation in months needs evidence.

Second, rules for large loads are tightening. Texas Senate Bill 6 (2025) directs the Public Utility Commission to set interconnection standards for large loads with a 75 MW default threshold, requires those customers to disclose on-site backup generation, and allows ERCOT, in an energy emergency, to direct large loads with backup generation to deploy it or curtail^[41]. A curtailment right of that kind is a technical input to the availability case, not just a regulatory footnote.

Third, grid operators now watch how data centers behave during faults. NERC's incident review of a July 2024 event in Virginia describes a 230 kV line fault after which about 1,500 MW of data center load disconnected on the customer side, by customer protection and controls, and transferred to backup systems^[42]. Ride-through settings, UPS transfer logic and any obligations the interconnection agreement imposes on them belong in the review.

What we ask for: the executed electric service or interconnection agreement, utility study results and the conditions they impose, the energization schedule by power block with the utility's own milestones, any curtailment, demand response or ramp-rate provisions, and the status of substation and transmission upgrades the schedule depends on. “Contracted” capacity is not energized capacity. CoreWeave, the largest public neocloud, reported over 850 MW of active power and about 3.1 GW of contracted power at the end of 2025^[1]: the difference is the deployment pipeline, and every borrower has one.

MW at the meter, IT load and PUE

A site's megawatts are quoted at several points in the electrical chain, and data rooms rarely say which. Utility capacity at the meter covers everything: IT, cooling, power conversion losses, lighting and offices. Critical IT capacity is what the UPS and distribution can deliver to the racks. The ratio between total facility power and IT power is power usage effectiveness (PUE), defined by The Green Grid^[27]. Uptime Institute's 2024 survey put the industry average at 1.56, flat for years, and notes that PUE does not account for IT energy performance^[26].

For sizing, the PUE that matters is the peak, design-day figure, not the annualised one a sustainability report quotes. Cooling energy is highest on the hottest day, which is also when the utility is most constrained. A liquid-cooled site in a cold climate and an air-cooled site in a hot one can have very different peak PUEs from similar annual ones.

Power per node, by platform

Manufacturer maxima are the right starting point for a lender: a system can draw up to its rated maximum unless a power cap is configured, and a cap trades power for performance^[23]. The table shows published figures and, in the right-hand column, the number of systems that fit in 10 MW at the meter under the assumptions in the next section (peak PUE 1.25, plus 10% for network, storage and management).

Two cautions. First, published GPU power is not system power: eight 700 W H100s are 5.6 kW of a 10.2 kW system, the rest being CPUs, memory, NICs, storage, fans and conversion losses^[12],[11]. A model that multiplies GPU count by GPU TDP understates that system's load by more than 40%. Second, OEM HGX systems built on the same baseboard can draw more than the DGX equivalent, particularly liquid-cooled variants run at higher power limits. The system figure in the borrower's bill of materials, not a generic one, is what should be reconciled. Our DGX B300, HGX B300 and GB300 NVL72 datasheets collect the published figures per platform.

Redundancy topology

Uptime Institute's Tier classification distinguishes redundant capacity components (Tier II), concurrent maintainability with redundant distribution paths so any part can be shut down without affecting IT (Tier III), and fault tolerance with physically isolated systems (Tier IV)^[25]. NVIDIA's GB200 SuperPOD reference architecture says the data center should generally meet or exceed Tier 3, or the equivalent TIA-942-B Rated 3 or EN 50600 Availability Class 3, including concurrent maintainability and no single point of failure^[20].

Redundancy interacts with sizing. If a site is 2N, half the installed UPS and distribution capacity is standby. If it is N+1 or block-redundant, a fraction is. A data room that quotes installed capacity without the topology can overstate usable IT capacity substantially. Some GPU campuses deliberately build reduced redundancy for training halls to save capital; that can be a sound choice, but it has to be reflected in the availability commitments of the offtake contract, and the lender should know it was a choice.

At the branch level, the National Electrical Code (NFPA 70)^[35] requires branch circuits supplying IT equipment to have an ampacity of at least 125% of the connected load^[37], which in practice caps sustained loading of a circuit at 80% of its rating. Rack power distribution has to be checked against the node's maximum, not its typical draw. NFPA 75 covers fire protection of IT equipment rooms^[36]. Outside the United States, the equivalent national codes apply, and the review should name them.

Higher-voltage distribution is coming. NVIDIA has published an 800 V DC architecture for future racks of up to 1 MW, with conversion from medium-voltage AC at the facility perimeter^[24]. For a lender, the relevance is residual: a facility built today on conventional AC distribution may need rework to host the generation after next, which affects re-leasing value.

A borrower presents a 10 MW site and a model with 600 eight-GPU Blackwell Ultra nodes (4,800 GPUs). Does it fit? The arithmetic is short, and every input is visible.

IT load of the plan. 600 nodes x 14.5 kW = 8,700 kW. Adding 10% gives 9,570 kW of IT load.

Case A: the 10 MW is utility capacity at the meter. At a peak PUE of 1.3, 9,570 kW of IT needs 9,570 x 1.3 = 12,441 kW at the meter. The site is about 2.4 MW short. Turned around, 10,000 kW / 1.3 = 7,692 kW of IT, divided by 14.5 x 1.1 = 15.95 kW per node, supports about 482 nodes, 118 fewer than the model.

Case B: the 10 MW is critical IT capacity. Then 10,000 kW / 15.95 kW = about 626 nodes at nameplate with zero headroom, and 600 nodes load the IT capacity to 96%. It fits on paper. Whether it fits in practice depends on three further questions: whether the 10 MW is usable capacity after redundancy or installed capacity before it; whether distribution is sized for nodes at maximum draw on every circuit; and whether the utility service behind the 10 MW of IT is itself large enough, which at a peak PUE of 1.3 means about 12.4 MW at the meter.

Generational shift. The same 10 MW at the meter, at a PUE of 1.3, supports about 685 DGX H100-class nodes at 10.2 kW^[11] against about 482 DGX B300-class nodes. A site that was right-sized for one generation holds roughly 30% fewer nodes of the next. That is a re-leasing and refresh question as much as a sizing one.

The point of the exercise is not that the borrower is wrong. It is that “10 MW” is not a specification. The lender needs the measurement point, the topology and the design-day PUE, and then the node count follows. Our GPU power calculator reproduces this example and runs the arithmetic for other platforms and assumptions. Its B300 preset defaults to the 15 kW system maximum in NVIDIA's power table and a PUE of 1.4, a more conservative case than the 14.5 kW and 1.3 used here.

Air versus direct-to-chip liquid

The 8-GPU DGX systems through B300 are air-cooled. A DGX B300 moves 1,500 CFM at 70% fan speed in its PSU version, rejects up to 49,476 BTU/hr and accepts inlet air from 10 °C to 30 °C^[15]. A DGX B200 is specified at 1,550 CFM and a 10 °C to 35 °C operating range^[14]. Four such systems in a rack means about 6,000 CFM through one cabinet, which requires strict hot- or cold-aisle containment and air handling sized for the density. Halls designed around the single-digit kilowatt racks that Uptime Institute's survey still finds most common^[26] were not built to move that much air through one cabinet, even where the electrical capacity can be found.

Rack-scale systems are liquid-cooled. In GB200 NVL72 the GPUs and CPUs are liquid-cooled and the remaining components air-cooled^[20]; the GB300 NVL72 rack is liquid-cooled, with integrated tray-level and rack-level leak detection^[17]. AMD's MI355X platform brochure specifies PG-25 coolant, a maximum liquid inlet of 43 °C and a recommended flow of 2.1 l/min per module^[19]. Direct-to-chip liquid cooling usually still leaves a fraction of the heat to air, so liquid-cooled halls need both systems commissioned.

CDUs, facility water and ASHRAE classes

A liquid-cooled deployment has two loops. The technology cooling system circulates coolant through the cold plates and is separated by a coolant distribution unit (CDU) from the facility water system that carries heat to chillers, dry coolers or towers. ASHRAE TC 9.9 publishes the thermal guidelines that classify both^[28]. The fifth edition redesignated the facility water classes by their upper supply temperature limit: W17, W27, W32, a new W40, W45 and W+^[30]. ASHRAE S-classes describe the technology cooling loop, and the class chosen should be coordinated with the IT manufacturer, since fluid temperatures that are too high or too low affect equipment performance; the loop must also stay above the dew point to avoid condensation^[31].

Warmer water classes allow more hours of chiller-free cooling and a lower PUE; they also leave less margin before the GPUs throttle. The questions for a lender are concrete: which water class the design assumes, what the IT vendor requires, what the supply temperature is on the design day at this site, and what happens to throughput if it is exceeded.

The Open Compute Project maintains open work on cold plates, CDUs and liquid cooling environments^[32], including a methodology for testing and rating CDU performance^[33]. CDU capacity should be verified at the actual approach temperature and flow, with redundancy (N+1 pumps or units) matched to the availability commitment.

Heat rejection, climate and water

Every watt delivered to the IT ends as heat that the site must reject. A 10 MW site rejects roughly 10 MW of heat continuously. Climate decides how: a Nordic site can reject most of it through dry coolers for most of the year, while a hot, humid site leans on mechanical chilling or evaporative cooling, which raises peak PUE or consumes water. The review should check heat rejection at the design-day dry-bulb and wet-bulb conditions for the site (not an annual average), the water supply and its permits where evaporative cooling is used, and the redundancy of the heat rejection plant.

Commissioning

The cooling system should be proven before GPUs depend on it. The standard sequence runs from factory witness testing of major equipment through installation checks and functional tests to integrated systems testing (IST) of power and cooling together under load banks, including failure scenarios. For liquid systems, coolant chemistry, air purging, passivation and flushing are part of commissioning. ASHRAE TC 9.9 issued a technical bulletin in 2026 warning that “it is too easy to violate direct to chip cooling system operational and warranty requirements”, and listing coolant composition, air management, stainless steel passivation and system pressure ratings as areas where deviations cause damage that is not immediately obvious^[29]. Lenders should see the commissioning scripts and the signed results, and a draw tied to mechanical completion should require them.

GPU racks are heavy and the weight is concentrated. A single DGX B300 in its PSU version weighs 370.38 lb (168 kg)^[15]; four of them in a rack, as in NVIDIA's reference design^[21], are about 1,480 lb (672 kg) of servers before the rack, PDUs, cabling and switches. Rack-scale NVL72 systems are heavier still, and their weight should be taken from the OEM's submittal for the exact configuration, because it varies by vendor and options.

The structural review covers slab or raised-floor rating against the rack's point and distributed loads, the path from the loading dock to the hall (ramps, lifts, door widths and corridor ratings), seismic bracing where required, and the rack itself: depth, width, static and dynamic load rating, and compatibility with the power distribution. Rack-scale systems built on Open Compute Open Rack conventions^[34] and NVIDIA MGX use DC busbars fed by power shelves; the GB200 NVL72 uses eight power shelves of six 5.5 kW supplies in N+N redundancy^[16]. The facility's whips, busway and receptacles have to match.

Fire protection and life safety (detection, suppression choice, and the effect of liquid cooling on both), physical security, and the shipping, receiving and staging areas needed to install hundreds of nodes in weeks also belong in the physical plant review.

A training cluster is only as fast as its fabric. NVIDIA's DGX B300 SuperPOD reference architecture specifies an 800 Gb/s Quantum-X800 InfiniBand compute fabric in a rail-optimized, non-blocking, twin-plane fat tree, with a separate storage fabric^[22]. Its scalable unit is 72 DGX B300 systems; the reference is built from eight units, 576 nodes and 4,608 GPUs, with 4,608 node-to-leaf and 4,608 leaf-to-spine compute fabric cables^[21]. The Ethernet alternative, Spectrum-X, and other RoCE fabrics follow the same rail-optimized logic.

For a lender, the network questions are about completeness and schedule. Is the fabric in the bill of materials sized for the node count in the model, at the oversubscription the offtake contract assumes? Are the optics ordered and confirmed? Optics have been a recurring bottleneck: TrendForce reported in December 2025 that NVIDIA's reservation of key laser capacity had extended lead times for some components beyond 2027^[43]. A cluster with nodes delivered and optics missing earns nothing.

The fabric is also where acceptance most often fails. Miscabled links, marginal transceivers and firmware mismatches show up as collective-communication bandwidth below specification, which is why fabric testing is a separate stage of acceptance (below). Our RoCE fabric acceptance testing runbook covers the Ethernet case in detail.

GPU clusters fail often, and availability is an operations outcome. Meta's Llama 3 paper is the best public data point: during a 54-day snapshot of pre-training on a cluster of 16K H100 GPUs, there were 466 job interruptions, of which 419 were unexpected; about 78% of the unexpected ones were attributed to confirmed or suspected hardware issues, and GPU issues were the largest single category at 58.7%. Meta still achieved more than 90% effective training time, with significant manual intervention required only three times^[49]. That outcome reflects mature automation, spares and staff. A new operator should be assessed on whether it can do the same.

What we assess

• Staffing and coverage. Named roles for facility operations and cluster operations, 24/7 on-site or on-call coverage with response times that match the SLA, shift patterns that survive leave and turnover, and training records for liquid-cooling procedures.
• NOC and monitoring. A network operations center (owned or contracted) with telemetry from the facility (power, cooling, leak detection) and the IT (GPU health, fabric errors, job failures), alerting thresholds, escalation paths and incident records.
• Remote hands and field service. Who physically swaps a GPU tray, a transceiver or a cable at 3 a.m., under what contract, with what response time, and whether they are trained on the specific platform.
• Spares. On-site spares by part (GPU trays or baseboards, PSUs, NICs, optics, fans, CDU pumps and filters) sized from failure rates, and a policy for replenishment.
• RMA and vendor support. Support entitlements with the OEM and NVIDIA or AMD, advance-replacement terms, and the historic turnaround the operator has actually achieved.
• Change management. Firmware and driver baselines, a staging environment, and a process for rolling updates without breaking acceptance-tested performance.
• Security and access. Physical access control, segregation between tenant workloads where the contract requires it, and out-of-band management security.

For a first-time operator, the gap between a staffing plan and a staffed team is itself a finding. A condition precedent requiring named hires, signed support contracts and an operating runbook before the first capacity tranche is accepted is a cheap protection.

In a take-or-pay structure, revenue begins when the customer accepts the capacity. Acceptance testing is therefore the point where engineering risk becomes cash-flow timing, and it is the natural point to tie loan draws to verified capacity. A good protocol is staged, written before testing starts, and has numeric pass thresholds.

Diagnostics are a floor, not a certificate: NVIDIA states that DCGM diagnostics are not designed to provide comprehensive hardware diagnostics or to replace field diagnosis tools^[45], which is why burn-in and cluster-scale tests follow. Three details decide whether the protocol protects the lender. First, thresholds must be numeric and agreed in advance: “satisfactory performance” is not a test. Second, the tests must run on the production configuration, at production scale, with the production firmware. A cluster tested at a quarter of its size has not tested its spine. Third, failures during burn-in must be recorded and the replaced parts tracked: an early failure rate well above the vendor's expectation is information about the batch, the installation or the cooling.

Tying acceptance to draws

A clean structure releases funds against independently certified milestones: energization of each power block; mechanical completion and IST of the cooling plant serving a hall; delivery and installation of a defined tranche of nodes; and acceptance of that tranche against the protocol. The technical advisor issues a certificate for each, with the evidence attached. Funds then follow verified capacity rather than invoices. Our GPU cluster acceptance testing runbook sets out the full test sequence, commands and thresholds.

The offtake contract is where technical performance becomes money. Lenders read it for credit terms; it also contains technical schedules that deserve an engineer's reading.

• Availability definition. Is availability measured per GPU, per node, per rack or per cluster? Over what window? A GPU-level definition with a short window is far harder to meet than a cluster-level monthly one.
• Exclusions and maintenance windows. Whether scheduled maintenance, customer-caused faults, utility curtailment and force majeure are excluded, and how much maintenance time the operator is allowed.
• Service credits and caps. The credit schedule against availability shortfalls, any cap, and whether persistent shortfalls create a termination right. CoreWeave's 10-K notes that most of its customer agreements contain service level commitments and that it has provided service credits or refunds after degradation and interruptions^[1].
• Acceptance mechanics. Who runs the tests, who witnesses, how long the customer has to accept or reject, and what happens on partial acceptance.
• Performance commitments. Any throughput, bandwidth or job-completion commitments, and whether the platform and fabric can meet them on the published specifications.
• Substitution and refresh. Whether the operator may substitute hardware, and whether the customer can require newer hardware during the term.

The engineering test is simple to state: given the failure rates in the public record, the spares and staff in the operating plan, and the redundancy in the facility, can this operator meet this availability definition with reasonable margin? If the answer is no, the base case should include credits.

The lender's collateral is a fleet of servers whose value declines with each new generation. Public disclosures give the range of accounting assumptions:

Accounting life is an estimate of how long an asset is useful to its owner, not a market value. For a lender the questions are the contracted term relative to the amortization profile, and what the fleet would fetch, or earn on re-lease, if the borrower defaulted part-way through. Rental market data shows how quickly that can move: Silicon Data's H100 rental series shows hyperscaler rates stepping down by roughly a third in mid-2025 while marketplace and neocloud rates had already settled far below hyperscaler levels^[54].

Credit support is evolving around this risk. Meta's Hyperion joint venture with Blue Owl, with about $27 billion of total development costs, pairs a four-year initial lease term with a residual value guarantee for the first 16 years of operations^[8]. That instrument covers a campus rather than GPUs, but the logic carries: where the lease term is shorter than the asset's economic life, someone has to underwrite the residual.

The engineering inputs to a residual case

• A serial-level asset register reconciled against delivery records and the installed base, so the collateral exists and is identifiable.
• Configuration that affects remarketing: GPU generation and memory, NVLink domain size, CPU, NIC and storage, and whether the systems are standard OEM SKUs or custom builds.
• Condition evidence: operating hours, thermal history, RMA history and any liquid-cooling incidents.
• Portability: an air-cooled 8-GPU server can be moved to most high-density colocation; a liquid-cooled rack-scale system needs a facility built for it, which narrows the buyer pool.
• Software and support entitlements that travel with the hardware, or do not.

These are engineering facts, and they move a stressed residual case more than any generic depreciation curve. Our own-versus-rent TCO comparison and the AI infrastructure investment articles cover the economics in more depth.

Patterns that, in a data room, warrant a direct question before anything else:

• Site capacity quoted in MW with no statement of where it is measured, or with a utility letter of intent in place of an executed agreement.
• A node count multiplied by GPU TDP instead of system power, or a PUE taken from an annual average rather than design-day conditions.
• An energization date earlier than the utility's own milestones, or dependent on substation or transmission work with no committed schedule.
• A cooling design that assumes a water class or supply temperature the IT vendor does not support, or no design-day check of heat rejection.
• Liquid-cooled hardware ordered for a hall whose mechanical design is still air-only, or a CDU selection not yet made.
• A fabric bill of materials that does not cover the full node count, or optics with no confirmed delivery date.
• An operations plan with no named people, no support contracts and no spares budget.
• An acceptance clause without numeric thresholds, or tests defined only at partial scale.
• Availability commitments at GPU level with short measurement windows, combined with reduced-redundancy facility design.
• An asset register that cannot be reconciled to serial numbers and delivery records.

None of these is fatal on its own. Each is a question that should be answered in writing before close, and most can be turned into a condition precedent or a covenant if the answer is not yet available. The NVIDIA DSX due diligence guide and the NVLink Fusion rack due diligence article apply the same approach to specific platforms.

The deliverable is written for a credit committee and for the lawyers drafting conditions precedent. A typical outline:

1. 01Executive summary. Findings ranked by credit impact, each with a severity, a recommended condition precedent or covenant, and the evidence it rests on.
2. 02Transaction and scope. What was reviewed, what was not, the data room index at the cut-off date, site visits made and the questions left open.
3. 03Site and power. Utility arrangements, energization schedule, capacity at the meter, one-line review, redundancy topology, and a reconciliation of the node count against available power.
4. 04Mechanical and cooling. Cooling architecture, liquid loop and CDU design, heat rejection at design-day conditions, water supply, and commissioning status.
5. 05Physical plant. Structural capacity, rack and busway ratings, fire protection and life safety, physical security.
6. 06IT hardware and fabric. Bill of materials against the contract, platform specifications, fabric topology, optics and delivery schedule, firmware and software baseline.
7. 07Operations. Staffing, NOC, monitoring, spares, vendor support and RMA, change management, incident history where it exists.
8. 08Acceptance and performance. Test protocol review, pass thresholds, witnessed results, and the measured basis for availability and throughput commitments.
9. 09Contract technical review. Availability definitions, service credits, maintenance windows, acceptance mechanics and technical termination triggers in the offtake agreement.
10. 10Budget and schedule. Reasonableness of capital cost categories and timeline against comparable builds, and the contingency held for the risks identified.
11. 11Collateral. Asset register reconciliation, configuration and condition, factors that affect remarketing, and the technical assumptions behind any residual value case.
12. 12Conditions and monitoring plan. Proposed conditions precedent, draw certification criteria, reporting covenants and the post-close monitoring cadence.

Every finding states its evidence and its source. Where a figure comes from a manufacturer specification, the report says so; where it comes from a measurement, it says who measured it and how. Where we could not verify something, the report says that too. Reliance language, limitations and the data room cut-off date are set out explicitly so the report can be relied on by the lenders named in the engagement letter.

GPU Smith is an independent engineering firm for GPU clusters and AI infrastructure. We hold no reseller quota, take no vendor incentive and run no cloud of our own. On a financing we act for one side only.

1. 01Scoping from the data room index. We read the index, not the whole room, and return a written scope: the questions that matter for this transaction, the documents that answer them, what is missing, and the timetable, before any open-ended reading starts.
2. 02Pre-close technical due diligence. Document review, management and engineering calls, a site visit where the site exists, and the independent technical report described above, with draft findings shared early so the term sheet and conditions precedent can absorb them.
3. 03Construction and deployment monitoring. Periodic site and schedule reviews, certification of draw milestones (energization, mechanical completion, delivery and installation of each node tranche), and early warning when the critical path moves.
4. 04Acceptance witnessing. We review the acceptance protocol before testing starts, witness or audit the burn-in and performance tests, and certify each tranche of capacity against the thresholds in the loan and offtake documents.
5. 05Post-close performance monitoring. Periodic review of availability, incidents, spares consumption, SLA credits paid and capacity changes, reported to the lenders against the covenants, with an escalation path when the operating data diverges from the base case.

Confidentiality is standard: we sign the lender's NDA before any data room access, do not reuse one transaction's material on another, and never name clients. To start, send the data room index (or a short description of the site, the hardware and the timetable) through gpusmith.com/contact or to [email protected]. We return a written scope.

Every source cited above, numbered as in the text. Each link was opened and the cited figure or statement checked on 2026-10-14. Secondary sources are marked; where a primary source exists it is preferred.

1. [1]CoreWeave, Inc. Annual Report on Form 10-K for fiscal year 2025. U.S. Securities and Exchange Commission (EDGAR).
2. [2]CoreWeave Secures $2.3 Billion Debt Financing Facility led by Magnetar Capital and Blackstone. Secured Finance Network (press release reprint), August 2023.
3. [3]CoreWeave Secures $7.5 Billion Debt Financing Facility led by Blackstone and Magnetar. CoreWeave investor relations, May 2024.
4. [4]CoreWeave Closes $2.6 Billion Secured Debt Financing Facility. CoreWeave investor relations, July 2025.
5. [5]Latham & Watkins Represents Macquarie in US$500 Million Financing for Lambda. Latham & Watkins, April 2024.
6. [6]Upper90 Closes $225M Credit Facility to Crusoe to Expand AI Cloud. Crusoe newsroom, March 2025.
7. [7]Newmark Facilitates $7.1 Billion Construction Loan to Develop AI Data Center. Newmark press release via Nasdaq, May 2025.
8. [8]Meta Announces Joint Venture With Funds Managed by Blue Owl Capital to Develop Hyperion Data Center. Meta newsroom, October 2025.
9. [9]TeraWulf Inc. Announces Pricing of $3.2 Billion of Senior Secured Notes. TeraWulf investor relations, October 2025.
10. [10]Independent engineer. Wikipedia.secondary
11. [11]NVIDIA DGX H100/H200 User Guide: Introduction to NVIDIA DGX H100/H200 Systems. NVIDIA documentation.
12. [12]NVIDIA H100 Tensor Core GPU. NVIDIA.
13. [13]NVIDIA H200 Tensor Core GPU. NVIDIA.
14. [14]NVIDIA DGX B200 User Guide: Introduction to the NVIDIA DGX B200 System. NVIDIA documentation.
15. [15]NVIDIA DGX B300 User Guide: Introduction to the NVIDIA DGX B300 System. NVIDIA documentation.
16. [16]NVIDIA DGX GB200 User Guide: Hardware. NVIDIA documentation.
17. [17]NVIDIA Enterprise Reference Architecture, NVL72 AI Factory: Components. NVIDIA documentation.
18. [18]AMD Instinct MI300X Accelerators. AMD.
19. [19]AMD Instinct MI355X Platform Brochure. AMD.
20. [20]NVIDIA DGX SuperPOD Reference Architecture with DGX GB200: Architecture. NVIDIA documentation.
21. [21]NVIDIA DGX SuperPOD Reference Architecture with DGX B300 (Quantum-X800): Architecture. NVIDIA documentation.
22. [22]NVIDIA DGX SuperPOD Reference Architecture with DGX B300: Key Components. NVIDIA documentation.
23. [23]NVIDIA DGX H100/H200 User Guide: Managing Power Capping. NVIDIA documentation.
24. [24]NVIDIA 800 V HVDC Architecture Will Power the Next Generation of AI Factories. NVIDIA Technical Blog, May 2025.
25. [25]Tier Classification System. Uptime Institute.
26. [26]Uptime Institute Global Data Center Survey 2024. Uptime Institute.
27. [27]PUE: A Comprehensive Examination of the Metric (White Paper #49). The Green Grid.
28. [28]ASHRAE Datacom Series (Thermal Guidelines for Data Processing Environments and companion volumes). ASHRAE Technical Committee 9.9.
29. [29]Technical Bulletin: TCS Coolant Integrity and System Readiness Best Practices (2026). ASHRAE Technical Committee 9.9.
30. [30]Major Changes to ASHRAE's Fifth Edition of Thermal Guidelines, Part 3: Liquid Cooling Chapter Updates. Upsite Technologies.secondary
31. [31]Designing for AI Workloads. Mechanical Engineering magazine (ASME), May 2026.secondary
32. [32]Cooling Environments Project. Open Compute Project.
33. [33]Liquid Cooling CDU Test Methodology and Performance Rating (white paper). Open Compute Project.
34. [34]Open Rack. Open Compute Project.
35. [35]NFPA 70, National Electrical Code. National Fire Protection Association.
36. [36]NFPA 75, Standard for the Fire Protection of Information Technology Equipment. National Fire Protection Association.
37. [37]NEC Requirements for Information Technology Equipment. EC&M.secondary
38. [38]DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers (2024 LBNL United States Data Center Energy Usage Report). U.S. Department of Energy, December 2024.
39. [39]Explainer on the Interconnection Final Rule (Order No. 2023). Federal Energy Regulatory Commission.
40. [40]Queued Up: 2024 Edition, Characteristics of Power Plants Seeking Transmission Interconnection. Lawrence Berkeley National Laboratory.
41. [41]Texas Senate Bill 6 (89th Legislature), enrolled text. Texas Legislature Online.
42. [42]Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions. North American Electric Reliability Corporation, January 2025.
43. [43]AI Data Centers Ignite a Laser Shortage Wave; Nvidia's Strategic Lock-In Reshapes the Global Laser Supply Chain. TrendForce press center, December 2025.secondary
44. [44]NCCL Tests. NVIDIA (GitHub).
45. [45]DCGM Diagnostics. NVIDIA Data Center GPU Manager documentation.
46. [46]The LINPACK Benchmark (HPL). TOP500.
47. [47]MLPerf Training benchmark. MLCommons.
48. [48]MLPerf Inference: Datacenter benchmark. MLCommons.
49. [49]The Llama 3 Herd of Models (Section 3.3.4, Reliability and Operational Challenges). Meta AI, arXiv:2407.21783.
50. [50]Amazon.com, Inc. Annual Report on Form 10-K for fiscal year 2024. U.S. Securities and Exchange Commission (EDGAR).
51. [51]Meta Platforms, Inc. Annual Report on Form 10-K for fiscal year 2025. U.S. Securities and Exchange Commission (EDGAR).
52. [52]Alphabet Inc. Annual Report on Form 10-K for fiscal year 2025. U.S. Securities and Exchange Commission (EDGAR).
53. [53]Nebius Group N.V. Annual Report on Form 20-F for fiscal year 2025. U.S. Securities and Exchange Commission (EDGAR).
54. [54]H100 Rental Price Over Time. Silicon Data.secondary