The First $10 Trillion Company

Document(s)

RegentQuant NVDA Valuation Model (.xlsx)

Summary

Building a forecast for the most-modeled stock on earth does not require a thousand-row spreadsheet. We assemble one for NVIDIA from five inputs: today's share price ($220.13), today's market cap ($5.32T), the company's average earnings beat over the last five reported quarters (+4.30%), and three exit P/E multiples (30× base, 40× bull, 20× bear). That is the entire model. The result is a clean Base / Bull / Bear arc — through a 270-day horizon at FY2027, a 635-day horizon at FY2028, and a 1,000-day horizon at FY2029 — that any reader can disagree with in a single number. The Base case prints a 30%+ annualized CAGR across all three horizons; the Bull case stretches to a 98% one-year and 45% three-year CAGR; the Bear case carries a 23% drawdown risk into FY2027 before turning positive. The provided bull research adds an illustrative overlay for the segments not yet in the P&L — Physical AI, Sovereign AI, Vera CPU, Automotive, AI-RAN — that can add anywhere from +14% to +50% to today's share price by FY2029 with no multiple re-rating required.

Why Simple Wins

A good forecast isn't the most elaborate one. It is the most honest one — the one whose every assumption you can defend in a sentence, and whose every output you can trace back to a number you actually own. Every additional layer in a model is another place to be wrong silently. We built this one the other way around: starting from the question (what return does NVIDIA owe its current shareholders over the next three fiscal years?), working backward to the smallest set of inputs that could plausibly answer it, and refusing to add anything that didn't earn its slot.

Five inputs went in. Three scenarios came out. The methodology is transparent enough that a disagreement becomes a disagreement about a specific number — Is the historical 4.3% beat still the right forward assumption? Is 30× the right base exit multiple? — rather than a disagreement about which of forty cells got wrong. That is the entire reason for the design.

"The best forecasts aren't the most elaborate. They're the most honest — every assumption defensible in a sentence, every output traceable to a number you actually own."

Methodology

Two equations do all the work.

1. Surprise-adjusted EPS. For each future quarter where consensus has not yet reported, we multiply the consensus estimate by (1 + average historical beat):

• \(\displaystyle\text{Adjusted EPS}_{Q} = \text{Consensus EPS}_{Q} \times \left(1 + \overline{\text{Beat}}\right)\)
• Reported quarters (Q1 FY26 → Q1 FY27) are carried at their actual values, not re-adjusted. Adjustment applies only to estimate cells.

2. Exit-multiple return. At each FY-end, total return is the chosen exit P/E divided by today's surprise-adjusted forward P/E, minus one:

• \(\displaystyle\text{Total Return} = \frac{\text{Exit P/E}}{\text{Adj. Implied Fwd P/E}} - 1\)
• \(\displaystyle\text{CAGR} = \left(1 + \text{Total Return}\right)^{365 / \text{Horizon (days)}} - 1\)
• \(\displaystyle\text{Implied Market Cap} = \text{Current Cap} \times \left(1 + \text{Total Return}\right)\)

That is the entire model. There is no DCF, no WACC, no terminal-growth slider, no segment-by-segment revenue build, no working-capital schedule. There are five inputs (price, market cap, average beat, three exit multiples) and a five-quarter beat history. Everything else is mechanical.

FY-end anchors. Valuation dates are anchored to the as-of date (≈270 / 635 / 1,000-day horizons): 2027-02-15, 2028-02-15, 2029-02-14. The horizons do not drift with the calendar — re-running the model on a later date will mechanically shorten them, not silently slide them forward.

Findings

Key inputs & market data

EPS surprise history — trailing five quarters

The 4.30% blended beat is not a curve-fit. It is the simple average of five consecutive quarters of consensus-versus-actual outcomes, every one of which was a beat, with a remarkably tight range (+3.26% to +5.77%). The discipline of the assumption is in its stability — not in its size.

Forward EPS — consensus vs surprise-adjusted ($ / share)

The right-hand block applies the 4.30% lift only to unreported quarters. Q1 FY27 already reported at $1.87, so it appears identically in both panels. The compounding of a 4.30% beat across the remaining unreported quarters lifts FY2027 EPS from $8.81 to $9.11 (+3.4%), FY2028 from $12.39 to $12.92 (+4.3%), and FY2029 from $14.59 to $15.22 (+4.3%). Today's $220.13 share price translates into a surprise-adjusted forward P/E of 24.2× on FY27, 17.0× on FY28, and 14.5× on FY29 — the three denominators that drive every scenario below.

Valuation scenarios — surprise-adjusted EPS, exit-multiple driven

The same three FY-end horizons (≈270, 635, 1,000 days) are taken through three exit multiples. Read each row left-to-right as the trajectory of the same scenario through time, and each column top-to-bottom as the spread of outcomes at a given horizon. The Base case has roughly the same CAGR (≈30–38%) across all three horizons — the lower terminal multiple at longer dates is offset by the larger absolute EPS base by FY2029. The Bull case earns its biggest one-year number (+98%) and decays naturally as the high exit multiple compounds against a longer horizon. The Bear case is the one scenario that goes through a drawdown before recovering — a useful reminder that even with EPS marching higher, multiple compression alone can deliver a negative one-year outcome.

The headline read. Base (30× exit) compounds at roughly the same rate (≈30–38% CAGR) regardless of horizon — a property of the model, not a coincidence, because the multiple is held constant while EPS grows. Bull (40× exit) is the scenario that requires the AI compute cycle to not normalize over a one-year window; the +98% CAGR figure is what NVIDIA could earn if 2026 looks like 2024 again. Bear (20× exit) is the textbook compression scenario: even with surprise-adjusted EPS at $9.11 in FY27, a re-rate to 20× delivers a −17% one-year return before EPS growth has time to repair it.

Bull Thesis — The Embedded Optionality Not Yet in the P&L

The model above is built entirely on the Data Center earnings engine that already exists. The bull case for the next leg of growth rests on segments that are either pre-revenue or contributing rounding errors today, but where NVIDIA has built full-stack incumbency before the demand curve has bent. We frame these below.

NVIDIA's Q1 FY2027 print, released after the close on May 20, was a reminder of how concentrated the story still is. Revenue hit a record $81.6 billion, up 85% year-over-year, with the Data Center segment alone contributing $75.2 billion — roughly 92% of the total. The newly disclosed Edge Computing segment (consolidating gaming, pro-vis, automotive, robotics, and AI-RAN) generated $6.4 billion, growing a comparatively pedestrian 29%. In other words, the bull case for the next leg of growth does not rest on what's already in the model. It rests on a handful of markets that are either pre-revenue or contributing rounding errors today, but where NVIDIA has built the full-stack incumbency before the demand curve has bent.

1. Physical AI & Humanoid Robotics — the largest TAM nobody is yet paying for

This is the marquee asymmetry. Management sizes physical AI — factories, logistics, transportation, and humanoids — at a $50 trillion market opportunity across factories, transportation, and humanoid robots. Against that backdrop, robotics revenue today is immaterial, buried inside the $6.4 billion Edge bucket. The setup mirrors the data-center story circa 2020: NVIDIA owns the entire enablement stack — Jetson Thor compute at the edge, Isaac GR00T foundation models, Omniverse/Cosmos for synthetic training data, and DGX for cloud-side training — before unit volumes have meaningfully ramped.

The strategic tell is the breadth of the partner roster now building on the platform: Agility, FANUC, Figure, Hexagon Robotics, KUKA, Skild AI, Universal Robots, World Labs and YASKAWA. Huang's framing — "every industrial company will become a robotics company" — captures the bull thesis precisely: if humanoid and industrial robot deployments inflect (the so-called "ChatGPT moment for robotics"), NVIDIA monetizes both the training compute and a per-unit edge silicon attach, a dual revenue stream not present in the cloud GPU model. Even a low-single-digit-percent capture of the addressable physical-AI compute layer over a decade implies a business larger than today's entire company. Near-term, expect robotics to stay a sub-$5B contributor; the option value is in the FY2028–2030 ramp.

2. Sovereign AI — a structurally new buyer class, only just starting to spend

Sovereign AI is the segment closest to the inflection. It tripled to over $30 billion in fiscal 2026, and it now sits inside the new "ACIE" sub-segment (AI Clouds, Industrial, Enterprise), which recovered to $37.4 billion in Q1 FY27 after the China-related trough a year earlier. What makes this a distinct new engine rather than an extension of cloud is the buyer: nation-states (UAE, Saudi, Europe, India, Japan) building domestic AI capacity for strategic-autonomy reasons, with budget cycles and price-insensitivity unlike commercial hyperscalers. This is government capex, not corporate — a demand pool that is largely uncorrelated with the hyperscaler capex cycle that bears worry is maturing. The bull view: sovereign demand provides a multi-year, non-cyclical floor under Data Center growth that the market is not yet underwriting because it has only one or two quarters of clean disclosure to work from.

3. Vera CPU — a ~$200B adjacency NVIDIA didn't previously sell into

The most concrete "new line item" disclosed yesterday is the Vera CPU. Management called it a "major new growth driver," opening up a $200 billion revenue opportunity, with Vera CPU revenue this fiscal year — standalone plus Rubin-bundle configurations — expected to reach approximately $20 billion. This matters because it expands NVIDIA's silicon content per rack into the CPU socket it historically ceded to Intel and AMD. Vera is described as the world's first processor purpose-built for agentic AI workloads. The bull math is straightforward: every Rubin rack now carries incremental NVIDIA-captured CPU dollars on top of GPU dollars, raising dollar content per AI factory without requiring a new customer. A $20B run-rate scaling toward a $200B TAM is the rare growth engine that is already showing first revenue yet remains under-modeled.

4. Automotive & Autonomy — the slow-burn that's reaching design-win maturity

Automotive remains tiny — roughly $5 billion this year, with a trillion-dollar potential as all cars become fully autonomous — but the leading indicator is the design-win pipeline, not current revenue. The Alpamayo end-to-end AV model, the full-stack Mercedes-Benz CLA win, the Uber robotaxi commitment across 28 cities across four continents by 2028, and added programs with BYD, Hyundai, Nissan, and Geely convert into revenue with a multi-year lag as vehicles enter production. With only 1% of the cars on the road today at L2+ capability, the penetration runway is long. Auto won't move the needle in FY2027, but it is the textbook case of pre-revenue design wins seeding a decade of high-margin, recurring software-plus-silicon attach.

5. Enterprise / Industrial AI & AI-RAN — the platform-extension wedges

Two further wedges round out the optionality. First, on-prem enterprise and industrial AI, now explicitly carved into ACIE — the bet that AI factories migrate beyond hyperscalers into the Fortune 2000's own data centers and factory floors. Second, AI-RAN/6G, where NVIDIA has expanded telecom partnerships with T-Mobile and Nokia to integrate physical AI applications into AI-RAN infrastructure and future 6G wireless networks. Both are negligible today; both represent NVIDIA inserting accelerated compute into infrastructure layers (telecom base stations, on-prem enterprise) it has historically not addressed.

Synthesis

The bull case is not that NVIDIA grows Data Center forever — it's that the company has pre-positioned the full stack across at least five adjacencies (physical AI/robotics, sovereign AI, the Vera CPU, automotive autonomy, and enterprise/AI-RAN) whose combined addressable markets dwarf the current ~$320B annualized revenue base, yet which contribute little to nothing today. Q2 guidance of ~$91 billion already implies continued acceleration off a pure Data Center engine. The optionality embedded in the segments not yet in the numbers — where NVIDIA's CUDA-and-stack moat is being replicated market-by-market — is what underwrites the multi-year compounding thesis. Huang's closing line is the entire bull argument in one sentence: "The world is rebuilding computing for agentic AI and robotic physical AI," and NVIDIA intends to sell the silicon, the models, and the software for every layer of that rebuild.

Risks to the thesis (for completeness): each of these TAMs is a forecast, not a backlog — humanoid and AV timelines have repeatedly slipped industry-wide; sovereign demand carries geopolitical and export-control exposure (China remains excluded from guidance); and custom-silicon efforts by hyperscalers pressure the very Data Center base funding these bets. A bull case built on un-booked revenue is, by definition, the least certain part of the model.

Optionality Uplift — RegentQuant Illustrative Estimate

Translating the thesis into the model. The five segments above are essentially absent from the surprise-adjusted EPS in the valuation block (FY27 $9.11 / FY28 $12.92 / FY29 $15.22, all Data-Center-driven). To size their potential contribution we hold NVIDIA's incremental economics constant — diluted shares implied by the model (≈24.17B) and an assumed non-GAAP net margin of 55% — so each $10B of new high-margin revenue adds roughly $0.23 to annual EPS. The same 30× base exit multiple already in the model then converts that EPS into price and market-cap impact. These are illustrative scenario figures for a ~FY2029–30 horizon, not booked revenue.

Read against the model. The conservative path lifts FY2029–30 adjusted EPS by ~$1.0 (about +7% over the $15.22 base) and, at the 30× base exit, adds roughly 14% to the share price — enough to nudge the Base-case CAGR from ~30% into the low-30s. The aggressive path adds ~$3.6 of EPS (about +24%) and ~50% to price, which alone closes most of the gap between the Base (30×) and Bull (40×) scenarios with no multiple re-rating required. Caveat, echoing the thesis: none of this is booked — most lands in FY2029+, and any one segment slipping (robotics/AV timelines, China/export limits, hyperscaler custom silicon) removes a visible slice of the uplift. Treat it as a probability-weighted overlay on the Data-Center base case, not a straight addition.

"Five inputs in. Three scenarios out. Every disagreement is now a disagreement about a number — not a disagreement about which of forty cells got wrong."

Conclusion — A Defense of Simplicity

The temptation, when modeling a name as consequential as NVIDIA, is to keep adding. Add a working-capital schedule. Add segment forecasts. Add a stochastic terminal multiple. Each addition feels like rigor, but most of them are theater: they widen the surface area where the model can be wrong without widening the surface area where the model can be checked. The trade is bad. We chose the opposite trade — fewer assumptions, each visible, each defensible — and we believe that is the version of the model a reader can actually argue with.

Practical takeaways.

• Today's print is not a bubble. At $220.13 and a surprise-adjusted FY27 P/E of 24.2×, the Base case (30× exit) compounds at roughly +30% CAGR through FY2029 without needing a single additional dollar of optionality revenue to land.
• The Bear case is real, but short-dated. Multiple compression to 20× delivers a −17% one-year outcome. By FY2028–29, even at 20×, the EPS base has grown enough to absorb the compression. Drawdown risk is concentrated in the next ~270 days.
• The optionality overlay is where the Bull case actually lives. A 40× exit multiple is doing a lot of work in the Bull scenario. The Conservative-to-Aggressive optionality table shows you can get most of the way from Base to Bull through EPS uplift alone (Physical AI, Sovereign, Vera CPU, Auto, AI-RAN) — without re-rating the multiple at all. That is a more durable construction of the bull thesis.
• The model is meant to be argued with. Disagree with the 4.30% beat — drop it, the FY29 EPS becomes $14.59 and Base CAGR moves from +30.5% to ≈+28%. Disagree with the 30× base exit — drop to 25×, and FY29 Base CAGR drops to ≈+22%. The point is that the disagreement is now a one-line disagreement, not a forty-cell rebuild.

The methodological point, restated. Sophistication and insight are not the same thing. The largest single-name positions we have seen put on against this stock over the last three cycles were not sized off the most elaborate decks. They were sized off the two or three numbers the trade actually depended on. The five-input model is built in that spirit — not because simplicity is virtuous in the abstract, but because every assumption you cannot defend is another place where the position is shorter your conviction than you think it is.

Source model: RegentQuant_NVDA_Valuation_Model.xlsx · As-of May 21, 2026 · Proprietary research · Illustrative — not investment advice.