Analyzing Consensus Forecast Achievability with AI: The case of $NVDA
Contrasting model performance with Chain of Thought prompting vs a single-pass prompt.
As AI tools rapidly become embedded into financial workflows, one of the most pressing questions is: Can they meaningfully assess the achievability of financial forecasts? When applied to a real-world case like Nvidia’s consensus 2026 estimates, the results are telling and surprisingly effective when used in the right way.
Here’s the summary:
ChatGPT-5 Delivers Solid Baseline Analysis
When prompted with a single question, ChatGPT 5 in Thinking mode can produce detailed and generally accurate assessments. It parses recent financials, extrapolates plausible growth scenarios, and outlines risks and assumptions all within seconds.
Chain-of-Thought Prompting Enhances Depth and Accuracy
Using Chain-of-Thought prompting significantly boosts performance by engaging the model’s reasoning step-by-step. It allows analysts to guide the AI through logic trees, conditionals, and sequential decisions, resulting in more nuanced scenario modeling and justification of forecasts.
Single Prompt Limitations: Conservative and Nuance-Blind
However, when asked to extrapolate growth or margins in a single-pass prompt, the model may miss key nuances in company-specific drivers or industry context. Its baseline estimates often err on the conservative side, limiting its value for high-conviction forecasting unless guided with more structured logic.
Human-AI Collaboration Unlocks the Best of Both Worlds
Even with chain of thought prompting, human judgement is still essential, especially for tailoring the style of analysis to the sector, scenario type, or risk regime. However, once primed, the AI can accelerate the analysis workload exponentially. Earnings call interpretation, scenario analysis, examination of margin trends and growth extrapolation - all can be automated with impressive speed and consistency.
AI as a Tireless, Bias-Free Analyst
Think of your AI model not as a replacement, but as another analyst on your team. It doesn’t tire, doesn't bring cognitive or institutional bias, and can test more permutations in less time than any junior hire. When guided properly, its output can match or even exceed human judgment as shown in several independent research studies comparing LLM outputs with professional analysts.
In an era where speed, accuracy, and adaptability matter more than ever, AI isn’t here to replace human analysts, it’s here to amplify them. Whether you're modeling upside for Nvidia or pressure-testing guidance from another high-growth stock, ChatGPT is no longer just a chatbot. It’s your new research assistant - fast, logical, and ready for its next question.
The Prompt Tests
At the outset of the tests, note that this post has been split into two parts due to its length. The focus is to demonstrate the benefits of Chain of Thought (CoT) prompting when conducting stock analysis with AI, using the world’s most followed stock, Nvidia, as the test case. I present both the prompts and the AI’s responses to clearly illustrate the differences in approach and output.
The core task mirrors what every analyst routinely does: assess the achievability of consensus forecasts and identify potential upside or downside risks. Since earnings, both growth and surprise, are key drivers of excess returns, this kind of analysis remains a cornerstone of any rigorous, fundamental investment process.
Nvidia serves as an ideal example not only because of its global prominence, but also due to a real-world complication: China’s export controls are expected to disrupt its sequential growth trajectory in FY2026. This introduces a layer of analytical complexity that tests how well AI can adapt to nuanced, scenario-driven thinking.
Given the length of the model’s responses, I’ve included the full transcripts in an appendix. The main body of the post features selected excerpts and analysis to highlight the most relevant insights. For those seeking a deeper comparison, the Appendix provides additional detail and context.
Background:
Year-to-date, Nvidia’s share price has risen 27%, and an impressive 60% over the past twelve months. This surge has been fueled by a rapid acceleration in data center capital expenditure, driven by the explosive growth of artificial intelligence. As a result, Nvidia remains the market’s most closely watched AI infrastructure play.
However, 2025 has also brought significant headwinds. U.S. export controls on advanced semiconductors to China have created disruptions. Sales of Nvidia’s most advanced AI chips to Chinese customers have halted pending export licences, and any future sales will now be subject to U.S. taxes, pressuring margins on them. Inventory originally allocated for China was written down in both Q1 and Q2, contributing to a slowdown in growth, though still strong by most standards, but notably softer compared to the prior year. Compounding this, the Chinese government is actively discouraging domestic firms from purchasing U.S. chips, despite their current technological edge.
This has created a state of flux that has interrupted the “beat and raise” rhythm investors had come to expect earlier in the year. The critical question now becomes: with expectations for FY2026 still high and the shareholder base increasingly crowded, can Nvidia exceed consensus forecasts and continue to outperform?
Single Prompt Test:
ChatGPT 5 is presented with the following prompt. No other data is provided to the model other than that specified in the user prompt:
Prompt A:
You are an experienced equity analyst and are to perform analysis of Nvidia's FY2026 consensus revenue and non-GAAP EPS forecasts to assess achievability and any upside or downside potential to those targets.
Nvidia's average consensus forecasts for FY26 are:
Revenue of $206bn
Non GAAP EPS of $4.50
Steps to undertake:
1. Analyze recent business trends and insights from Nvidias most recent earnings call that are relevant to assessment of growth rates and forecasts.
2. Perform an analysis and present insights of Nvidias most recent income statement and balance sheet trends including margins and ratios. Show calculations for any ratios used.
3. In light of this information and insights make an assessment of the achievability of Nvidia's FY2026 consensus forecasts including any upside or downside potential and the factors that drive that conclusion.
4. Indicate your confidence on a scale of 0 to 100 (100 being perfectly confident) in your assessment and show calculations you have made to arrive at any assessments of upside or downside.
While this instruction is presented to the model as a single prompt, it is structured into steps an analyst might undertake themselves to assess forecast achievability. In this way, the prompt calls upon GPT-5’s “Thinking” mode which does deeper research on the subject matter to present greater detail in the response. Role and task are clearly delineated in the system prompt at the top as this mimics the way models are trained during fine tuning and reinforcement learning stages.
The prompt presents the model with the FY26 consensus targets to ensure the correct targets are assessed. Large language models are still hit and miss if you allow them too much latitude to gather their own data as they have a tendency to conflate data from multiple sources which reduces reliability. The models do not have access to proprietary financial databases, so I always prefer to provide the model with the data I know is correct.
ChatGPT-5 Single Prompt Response Extracts:
The model’s response to questions 1 and 2 are adequate but could be improved. It summarizes the key points of management commentary relevant to the growth of the business including product segment trends. The model extracts the key financial metrics including recent quarterly revenue, operating income, net income and EPS and calculates margins and balance sheet ratios.
However the AI model doesn’t specifically examine recent growth trends or link balance sheet line item trends to growth. The analysis is static and the type of summary one would expect in a brief sales trader flash note immediately following lodgement of a company’s quarterly financials with the market. The model summarizes the highlights from the financials such as margins, liquidity and the key swing factor being resumption of China exports however it is notable that its does not address growth rates at this stage of the analysis despite instruction to analyze trends.
Section 3 of the response addresses forecast achievability. This is the first time the model brings growth calculations into consideration. While the model adequately calculates required Q3 and Q4 revenues and EPS to hit consensus targets, there are two aspects that are missing. See if you can spot them in this extract:
3) Can FY26 consensus be met? (Your base: $206bn revenue, $4.50 EPS)
Revenue bridge
What’s locked in/indicated:
H1 actual: $44.1bn (Q1) + $46.7bn (Q2) = $90.8bn. NVIDIA NewsroomNVIDIA Investor Relations
Q3 guide (mid): $54.0bn. Shacknews
Implied need for Q4 to hit $206bn:
Required Q4 = $206 − ($90.8 + $54.0) = $61.2bn.
That is +13.3% q/q vs guided Q3 ( = 61.2 / 54.0 − 1 ).
China sensitivity (if $2–$5bn resumes in H2):
Add $3bn H2 from China → FY total pre-Q4 need falls to $206 − ($90.8 + 54.0 + 3.0) = $58.2bn → Q4 +7.8% q/q.
Add $5bn → Q4 need $56.2bn → +4.1% q/q. InvestopediaFinancial Times
Read-through:
Reaching $206bn is achievable, but most cleanly with either (a) China contribution in H2 or (b) a strong Q4 seasonal step-up on Blackwell/Rubin supply. A flat Q4 at $54bn would land NVDA near $198.8bn, below consensus. (Base math: 44.1 + 46.7 + 54 + 54.) NVIDIA NewsroomNVIDIA Investor RelationsShacknews
EPS bridge
Starting point: H1 non-GAAP EPS ≈ $1.86 (0.81 + ~1.05). NVIDIA Investor RelationsInvestopedia
Revenue-scaled proxy for H2: Using Q2 as a rough scaler (EPS/revenue ≈ 1.05 / 46.743 ≈ 0.0225 per $1bn):
Q3 at $54bn → EPS ≈ 54 × 0.0225 ≈ $1.21.
To reach FY $4.50, needed Q4 EPS ≈ $4.50 − (1.86 + 1.21) = $1.43.
If Q4 revenue = $61.2bn (the revenue needed for $206bn with no China), simple scaling gives $61.2 × 0.0225 ≈ $1.38, leaving a $0.05 gap that could be bridged by mix (networking/software), margin normalization from H20 cleanup, or buyback accretion. If China adds $3–$5bn in H2, the gap narrows further. Investopedia
Bottom line:
Revenue $206bn: Base-to-likely with a strong Q4 and/or China licenses in time.
Non-GAAP EPS $4.50: Doable but tighter — base math suggests $4.3–$4.5 depending on H2 margin/mix and share count.
Problems with this response:
The model immediately defaults to a QoQ growth extrapolation as its basis for assessing achievability of revenue forecasts. This is fraught for Nvidia at this time due to the temporary interruption to sequential growth through Q2 from China export curbs. If these curbs stay in place (and management EPS guidance and consensus do not take resumption into account), Q2 revenue growth (QoQ), being the first full quarter most affected by the China sales ban, is biased downwards. Q3 revenue growth will more naturally restart to reflect the underlying base rate of growth (QoQ) as its comparing apples with apples (2 quarters, both with a full impact of China sales ban) and hence an investor would expect the QoQ revenue growth in Q3 to re-accelerate from that shown in Q2.
It would be preferable to examine Q2 - Q4 revenue and ex-charge EPS growth rates on a YoY basis and extrapolate from there given concerns with QoQ growth rates at the current time. Assuming, as guidance does, that China sales remain subject to curbs, the Q2 YoY growth rate is indicative of underlying growth until we cycle a full twelve months from the introduction of the curbs. Alternatively, the model could have recognized the QoQ revenue growth resumption implicit in the $54bn Q3 revenue guidance (vs $46.7bn in Q2) and work from there. Triangulation across both methods would be considered to be a strong response.
No examination of Nvidia’s sales and margin seasonality was undertaken to strengthen the forecast. The methodology used is simplistic. Given Nvidia’s Data Centre division is very high margin relative to its Gaming and other segments, this could be a factor to consider given Gaming segment revenue is likely seasonal, has lower margins and will impact the margin mix in peak quarters. A complete analysis of EPS achievability needs to take this into account.
Given these problems, the model’s conclusion of forecast achievability appears weak despite its stated confidence of 72% and leaves the investor with a feeling that some small downside actually exists. The model does not identify and present any quantitative basis for supporting a strong Q4 sequential growth rate which is required to hit the $206bn full year revenue target (even though that may well be likely). Notably here, the lack of comprehensive narrative insights from the financial statement analysis appears to hamper the model from presenting a strong case for Q4 to meet consensus. Similarly with respect to EPS, the model actually calculates a likely EPS range below consensus but notes achievement is possible.
I’d grade this response a C. Its a low passing grade but is of only moderate use to an investor.
The Appendix below outlines the full response from the single prompt test. Tomorrow, I will present a structured Chain of Thought prompt set to strengthen the analysis and conclusions made.
Feel free to share this around. It is my desire to bring together a tight community of forward thinking investors who are dedicated to enhancing their own research and analysis using the AI revolution…
Inference never stops. Neither should you.
Andy West, The Inferential Investor.
10 September 2025.
Appendix - Single Prompt Response (Complete):
Got it — using your stated FY26 (year ending Jan-2026) consensus of $206bn revenue and $4.50 non-GAAP EPS, here’s a focused analyst read using the latest call and 10-Q.
1) What the latest call says (Q2 FY26, reported Aug 27, 2025) — growth-relevant takeaways
Run-rate still accelerating. Q2 revenue $46.7bn (+6% q/q, +56% y/y); Data Center $41.1bn (+5% q/q, +56% y/y). Company guided Q3 to $54bn ±2% — notably excluding any China H20 shipments. NVIDIA Investor RelationsNVIDIA NewsroomShacknews
Blackwell ramp: management cited +17% q/q Blackwell Data Center revenue; demand visibility reiterated into Q3. NVIDIA Investor RelationsNVIDIA Newsroom
China dynamics: no H20 sales to China in Q2; press/call commentary frames potential $2–$5bn upside in H2 if licenses clear (not in guidance). NVIDIA Investor RelationsInvestopediaFinancial Times
Margins and returns: non-GAAP GM ~72.7%; Board expanded buyback authorization by $60bn. Yahoo FinanceSEC
Customer concentration (context for risk): reporting around the filing/call highlights outsized orders among a handful of hyperscale/AI buyers. Useful to keep in mind for scenario risk. TechRadar
2) Latest income statement & balance sheet (Q2 FY26 10-Q) — key ratios with math
P&L (GAAP unless stated):
Revenue: $46.743bn; COGS: $12.890bn → Gross margin = 1 − 12.890/46.743 = 72.4% (also shown in 10-Q). Operating income: $28.440bn → Op margin = 28.440/46.743 = 60.8% (per 10-Q table). Net income: $26.422bn → Net margin = 26.422/46.743 ≈ 56.5%. SEC
Liquidity & leverage (Balance Sheet, 7/27/25):
Current ratio = Current assets / Current liabilities = 102,219 / 24,257 = 4.21×.
Quick ratio ≈ (Cash + Mkt secs + AR) / Current liabilities = (11,639 + 45,152 + 27,808) / 24,257 = 84,599 / 24,257 ≈ 3.49×.
Net cash = Cash & marketable securities − LT debt = (11,639 + 45,152) − 8,466 = $48.3bn.
Working capital = 102,219 − 24,257 = $78.0bn. SEC
Working-capital turns (quarterly):
Inventory days ≈ (Inventory / COGS) × 90 = (14,962 / 12,890) × 90 ≈ 104.5 days.
DSO (AR days) ≈ (AR / Revenue) × 90 = (27,808 / 46,743) × 90 ≈ 53.6 days.
Inventory mix: Finished goods $8.7bn, WIP $4.4bn, raw $1.8bn — consistent with a large new-platform transition. SEC
Non-GAAP EPS cadence (for context):
Q1 FY26 non-GAAP EPS = $0.81 (with disclosure of China-related charges and an ex-charge view); Q2 FY26 non-GAAP EPS ≈ $1.05–$1.08 (press/recaps). H1 non-GAAP EPS sum ≈ $1.86. NVIDIA Investor RelationsInvestopedia
Takeaway: Margins are extraordinary (72–73% GM; 61% operating), liquidity is ample, and working capital is building to support the Blackwell (and upcoming Rubin) ramps; China is the key swing factor.
3) Can FY26 consensus be met? (Your base: $206bn revenue, $4.50 EPS)
Revenue bridge
What’s locked in/indicated:
H1 actual: $44.1bn (Q1) + $46.7bn (Q2) = $90.8bn. NVIDIA NewsroomNVIDIA Investor Relations
Q3 guide (mid): $54.0bn. Shacknews
Implied need for Q4 to hit $206bn:
Required Q4 = $206 − ($90.8 + $54.0) = $61.2bn.
That is +13.3% q/q vs guided Q3 ( = 61.2 / 54.0 − 1 ).
China sensitivity (if $2–$5bn resumes in H2):
Add $3bn H2 from China → FY total pre-Q4 need falls to $206 − ($90.8 + 54.0 + 3.0) = $58.2bn → Q4 +7.8% q/q.
Add $5bn → Q4 need $56.2bn → +4.1% q/q. InvestopediaFinancial Times
Read-through:
Reaching $206bn is achievable, but most cleanly with either (a) China contribution in H2 or (b) a strong Q4 seasonal step-up on Blackwell/Rubin supply. A flat Q4 at $54bn would land NVDA near $198.8bn, below consensus. (Base math: 44.1 + 46.7 + 54 + 54.) NVIDIA NewsroomNVIDIA Investor RelationsShacknews
EPS bridge
Starting point: H1 non-GAAP EPS ≈ $1.86 (0.81 + ~1.05). NVIDIA Investor RelationsInvestopedia
Revenue-scaled proxy for H2: Using Q2 as a rough scaler (EPS/revenue ≈ 1.05 / 46.743 ≈ 0.0225 per $1bn):
Q3 at $54bn → EPS ≈ 54 × 0.0225 ≈ $1.21.
To reach FY $4.50, needed Q4 EPS ≈ $4.50 − (1.86 + 1.21) = $1.43.
If Q4 revenue = $61.2bn (the revenue needed for $206bn with no China), simple scaling gives $61.2 × 0.0225 ≈ $1.38, leaving a $0.05 gap that could be bridged by mix (networking/software), margin normalization from H20 cleanup, or buyback accretion. If China adds $3–$5bn in H2, the gap narrows further. Investopedia
Bottom line:
Revenue $206bn: Base-to-likely with a strong Q4 and/or China licenses in time.
Non-GAAP EPS $4.50: Doable but tighter — base math suggests $4.3–$4.5 depending on H2 margin/mix and share count.
4) Upside/Downside factors (what would move the needle)
Upside drivers
China licenses allowing H20/Blackwell shipments (estimated $2–$5bn H2 tailwind). Investopedia
Supply/ASP mix: higher Blackwell mix, NVLink/RoCE networking, and software/services uplift can support >73% non-GAAP GM. Yahoo Finance
Buyback accretion (new $60bn authorization) lowers diluted share count and cushions EPS math. SEC
Downside risks
Capex moderation narrative at hyperscalers into CY26 (order smoothing or digestion). External commentary around this risk grew post-print. InvestorsFinancial Times
Customer concentration risk and any pivot to custom silicon. TechRadar
Export/regulatory timing: slippage in China permissions or tighter controls. Investopedia
5) Assessment & confidence
Revenue ($206bn): Achievable; path A = Q4 +13% q/q; path B = smaller Q4 lift (+4–8%) if $3–$5bn China resumes in H2.
EPS ($4.50): Achievable with less slack; baseline scaling + buyback/mix gets close; cleanly above $4.50 likely needs either (i) Q4 revenue ≥$60bn or (ii) non-GAAP GM drift back toward mid-73% plus repurchases.
Confidence score: 72 / 100.
Why 72? High visibility from H1 actual + Q3 guide and clear product cadence (supportive), offset by genuine binary China timing and Q4 elasticity needed to fully clear both revenue and EPS bars.