For investment analysts, corporate finance desks, and fintech product managers, parsing SEC filings is a
critical but notoriously time-consuming workflow. To streamline this process, many teams have
experimented with general-purpose artificial intelligence models like ChatGPT as well as finance-specific
research systems. The useful question is not which label wins universally, but what evidence, adjacent
disclosures, calculations, citations, and limitations each workflow returns for a defined task.
To demonstrate the difference between general-purpose systems and a finance-native AI, we put
ChatGPT and StockUp’s flagship model, Quan, to the test in a
head-to-head comparison. We asked both systems to analyze one of the most complex filings of the year:
**NVIDIA's ($NVDA) Fiscal Year 2026 Form 10-K** (filed February 25, 2026).
The Test Prompt
We submitted the following specific, multi-part prompt to both models, focusing on NVIDIA's massive
inventory commitments and supply chain risks:
"Analyze the latest NVIDIA ($NVDA) 10-K filing. Locate the discussion on 'Purchase Obligations'
(often in the Commitments and Contingencies note) and compare it to the 'Inventory' asset on the
Consolidated Balance Sheets.
1. What are NVIDIA's total outstanding inventory purchase obligations/commitments compared to the
inventory they currently hold?
2. How have these purchase commitments changed year-over-year?
3. What are the specific risk factors NVIDIA outlines regarding these purchase commitments, and what
are the potential cash flow impacts if demand shifts?"
Methodology limitation
This page preserves an illustrative run, not a controlled universal benchmark. The original run did
not record exact model snapshot IDs, equalized tool access, latency, or total cost. Grounding redirect
URLs can also expire. Treat the observations below as a case study to reproduce—not proof that either
system wins every SEC filing task.
Below are the exact, unedited transcripts of the responses returned by ChatGPT and Quan.
Side-by-Side Response Transcripts
ChatGPT (General-Purpose LLM)
ChatGPT
The latest annual filing is the fiscal year ended January 25, 2026 (filed
February 25, 2026).
1. Inventory purchase obligations vs. inventory on hand
The key comparison is:
Item (Jan. 25, 2026) Amount
Inventory (balance sheet asset) $21.4 billion
Outstanding manufacturing, inventory purchase, long-term supply & capacity commitments $95.2 billion
This means:
NVIDIA has $95.2B of outstanding supply/inventory-related commitments.
It currently carries $21.4B of inventory on its balance sheet.
Purchase commitments are therefore approximately 4.45× the inventory currently held (95.2 / 21.4 ≈
4.45x).
The commitments exceed inventory by roughly $73.8B.
Importantly, these commitments are not inventory today. They represent future contractual
obligations to secure manufacturing capacity and components, reflecting NVIDIA's strategy of locking
in scarce AI supply chain capacity.
2. Year-over-year change
The increase has been dramatic.
Fiscal Year Inventory Purchase & Supply Commitments Inventory
FY2025 (Jan. 26, 2025) $30.8B $10.1B
FY2026 (Jan. 25, 2026) $95.2B $21.4B
Year-over-year:
Purchase commitments:
+$64.4B
+209% YoY (roughly tripled)
Inventory:
+$11.3B
+112% YoY
The filing attributes the prior year's increase to commitments supporting new product introductions
(including Blackwell), while the FY2026 filing reframes these as broader manufacturing, supply and
capacity commitments reflecting "datacenter-scale production" and longer ordering horizons.
3. Risk factors NVIDIA identifies
The filing discusses several specific risks associated with these commitments.
A. Demand forecasting risk
NVIDIA states that inventory and purchase commitments are based on forecasts of future customer
demand. If those forecasts prove incorrect, the company could incur: excess inventory, obsolete
inventory, excess purchase commitments, or inventory write-downs.
B. Long manufacturing lead times
The filing emphasizes that manufacturing lead times can exceed 12 months for certain products and
NVIDIA may place non-cancellable inventory orders well before historical lead times.
C. Supply agreements are only partially flexible
NVIDIA notes that some agreements can be cancelled, rescheduled, or adjusted before firm purchase
orders are placed. However, it also warns that changes to these agreements may result in additional
costs.
D. Inventory write-down risk
The accounting policy explicitly states that cost of sales includes provisions for excess inventory,
obsolete inventory, and excess product purchase commitments.
4. Potential cash flow impacts if demand shifts
1. Cash outflows continue despite weaker demand: The company notes that substantially all of the
$95.2B commitments are expected to be paid through fiscal 2027.
2. Inventory and purchase obligation charges: NVIDIA may have to record inventory write-downs or
charges for excess purchase commitments. export restrictions affecting China resulted in a $4.5
billion charge associated with excess H20 inventory and purchase obligations after demand for those
products diminished.
3. Working capital expansion: cash flow could be reduced as capital is locked up in supplier
commitments.
StockUp Quan (Finance-Native AI)
Quan Model
🔬 DEEP RESEARCH DOSSIER
3 verified sources analyzed and cross-referenced
NVIDIA ($NVDA) Fiscal Year 2026 Purchase Obligations vs. Inventory Analysis
An institutional-grade analysis of NVIDIA's Form 10-K for the fiscal year ended January 25, 2026,
reveals an aggressive capital-deployment strategy characterized by unprecedented off-balance-sheet
commitments to secure next-generation semiconductor and packaging capacity.
1. On-Balance-Sheet Inventory vs. Off-Balance-Sheet Commitments
As of January 25, 2026, NVIDIA's inventory assets on the Consolidated Balance Sheets stood at
$21,403 million ($21.403 billion) [1]. In contrast, NVIDIA's outstanding manufacturing, supply, and
capacity commitments (disclosed in Note 12 of the 10-K) reached $95.2 billion [1].
- Commitments-to-Inventory Ratio: NVIDIA's outstanding purchase obligations are approximately 4.45
times (or 445%) its physical on-balance-sheet inventory [1].
- Concentrated Horizon: NVIDIA expects substantially all of the $95.2 billion in purchase
commitments to be paid out during Fiscal Year 2027 [1, 2].
- Other Purchase Obligations: Beyond inventory commitments, NVIDIA also holds $27.0 billion in
multi-year cloud service agreements (primarily hosting DGX Cloud and supporting internal R&D),
bringing its broader commitment footprint even higher [2].
2. Year-over-Year (YoY) Changes
NVIDIA's supply chain exposures and commitments have scaled dramatically to support the rollout of
its Blackwell and upcoming Rubin architectures [1, 2]:
FINANCIAL METRIC | FY2025 (ENDED JAN 26, 2025) | FY2026 (ENDED JAN 25, 2026) | ABSOLUTE CHANGE | YOY
% CHANGE
On-Balance-Sheet Inventory | $10.080 billion [1] | $21.403 billion [1] | +$11.323 billion [2] |
~112.3% [1]
Inventory Purchase Obligations | $30.8 billion [1] | $95.2 billion [1] | +$64.4 billion [2] |
~209.1% [2]
Cloud Service Obligations | $10.9 billion [2] | $27.0 billion [2] | +$16.1 billion [2] | ~147.7% [2]
3. Risk Factors and Demand Shifts
Under Item 1A (Risk Factors), NVIDIA outlines severe operational and financial risks associated with
its massive commitment expansion:
- Asymmetric Cancellation Terms: While NVIDIA's commitments to foundry partners (such as TSMC) are
firm and increasingly structured as Non-Cancellable and Non-Returnable (NCNR), its sales are
primarily based on purchase orders [3]. This means customers can cancel or delay orders on short
notice without penalty, while NVIDIA remains contractually bound to pay for the underlying
components [3].
- Demand Mismatch & Margin Contraction: Overestimating demand can lead to immediate inventory
write-downs, asset impairments, and vendor penalty fees [3].
- The H20 Empirical Precedent: This risk materialized in Q1 of FY2026 due to sudden export license
restrictions on China-bound H20 products [2]. Consequently, NVIDIA recognized a $4.5 billion charge
for excess inventory and outstanding purchase obligations, driving its FY2026 gross margin down to
71.1% (from 75.0% in FY2025) [3]. NVIDIA’s total inventory and purchase obligation provisions
reached $7.2 billion for the fiscal year [2, 3].
- Prepayments and Cash Flow Exposure: Entering into prepaid manufacturing agreements locks up
significant corporate cash reserves [3]. If operational disruptions arise or if hardware
architectures change rapidly, pre-ordered components may become obsolete before delivery, leading to
severe underutilization charges and cash flow depletion [3].
---
Note: NVIDIA's current regular-session price is $204.12, according to Yahoo Finance [SERVER VERIFIED
QUOTE].
VERIFIED RESEARCH SOURCE DIRECTORY
[1] As of January 25, 2026, NVIDIA’s reported on-balance-sheet inventory was $21,403 million
($21.403 billion) compared to $10,080 million at the end of the previous fiscal year.
URL: https://vertexaisearch.cloud.google.com/grounding-api-redirect/...
[2] NVIDIA's manufacturing, supply, and capacity commitments reached $95.2 billion as of January 25,
2026, up from $30.8 billion in FY2025.
URL: https://vertexaisearch.cloud.google.com/grounding-api-redirect/...
[3] NVIDIA had to take a massive $4.5 billion charge in Q1 FY2026 associated with written-down H20
excess inventory and purchase obligations.
URL: https://vertexaisearch.cloud.google.com/grounding-api-redirect/...
Deep-Dive: How the Models Compare
ChatGPT provided a strong direct summary of the figures requested. In this recorded case, Quan also surfaced
adjacent commitments, connected provisions to margin effects, and organized citations differently. Those
observations apply to these saved outputs and should be retested against current model versions.
1. Footnote Completeness (The $27 Billion Cloud Gap)
ChatGPT answered the prompt directly, extracting the $95.2 billion purchase commitments. However, it
completely missed another massive off-balance-sheet commitment located in a adjacent footnote:
$27.0 billion in multi-year cloud service agreements (which grew 147.7% YoY from $10.9
billion).
Because Quan is engineered explicitly for corporate disclosures, it automatically parsed and correlated
all adjacent commitment categories. This gave analysts the full picture of NVIDIA's $122.2 billion
combined off-balance-sheet footprint ($95.2B inventory/capacity + $27B cloud services), rather than just
the subset explicitly named in the prompt.
2. Margin-Impact and Provisions Modeling
Both models identified the $4.5 billion China export license restriction charge on H20 chips. However,
ChatGPT only reported the raw number as a generic risk consequence.
Quan cross-referenced these impairment charges to model their exact impact on the income statement:
pointing out that the China-bound write-down drove NVIDIA's consolidated gross margin down to **71.1%**
(from 75.0% in the prior fiscal year). It also dug deeper into the accounting notes to locate the total
inventory and purchase obligation provisions for the entire fiscal year (**$7.2 billion**).
3. Identifying Asymmetric Contract Risk
An essential aspect of analyzing commitments is understanding the underlying legal agreements. ChatGPT
described NVIDIA's commitments as "only partially flexible."
Quan pinpointed the exact structural mechanism: **Asymmetric Cancellation Terms**. It extracted the
critical detail that NVIDIA's commitments to silicon foundry partners (like TSMC) are structured as
**NCNR (Non-Cancellable, Non-Returnable)**. In contrast, NVIDIA's customer revenue rests on flexible
purchase orders. This contract asymmetry is the exact structural risk that could trigger margin
compression if demand shifts, a vital insight for any hedge fund or credit desk.
4. Auditability and Verifiable RAG Citations
Perhaps the most critical difference for institutional compliance is source traceability. ChatGPT
presented figures as self-contained claims. If an analyst needs to verify the $95.2 billion commitment
figure, they have to manually comb through the 10-K to locate it.
Quan's architecture features a **Verified Research Source Directory**. Every single metric and risk
statement is tagged with grounded links pointing to the exact footnote paragraph in the SEC database.
This creates an immediate, auditable trail that allows analysts to verify data points in one click.
Conclusion: Evaluate the Whole Research Workflow
This case illustrates why finance applications should evaluate more than prose quality. Useful tests also
measure numerical correctness, adjacent-disclosure coverage, source entailment, latency, cost, and the
ability to mark unavailable evidence. StockUp is expanding this case into a versioned benchmark with
recorded model IDs and repeatable scoring.
To learn more about deploying Quan inside your organization's research pipelines, explore the StockUp Enterprise
Page. For details on integrating the API, check out our developer documentation: Quan API Enterprise:
Developer Setup Guide.
