NVDA · Forward model · China returns case
The China returns case, 20 quarters out
Each segment is projected from its own operating driver, rolled up into consolidated cash flow, and discounted back to a fair value per share. The assumptions below are editable — change them and every number on this page moves with them.
Quarters are labelled the way NVIDIA labels them: fiscal, running a year ahead of the calendar, so 2027 Q1 is the quarter ended 26 April 2026. NVIDIA reports two segments (Compute & Networking, Graphics) and, until this quarter, five market platforms. This model carries three verticals cut on the platform axis, because that is where the economics differ: Data Center compute, Data Center networking, and everything else. Every number is disclosed. Data Center compute and networking are tabulated exactly in the CFO commentary for 2026 Q1 through 2026 Q4; for the basis quarter NVIDIA gave the split only in prose and only to $0.1B ($60.4B compute, $14.8B networking against a tabulated $75,246M Data Center), so those two points are marked ESTIMATED. Edge Computing is exact throughout: NVIDIA reports it directly for the basis quarter ($6,369M) and it is the arithmetic remainder of total less Data Center for every earlier quarter, which reconciles to the sum of the four legacy platforms (Gaming, Professional Visualization, Automotive, OEM & Other) to the dollar. Nothing here is apportioned. What is NOT modelled: NVIDIA has begun reporting Data Center as Hyperscale and ACIE instead of compute and networking, and has published that split for three quarters only. Compute versus networking has five, and it is the split that carries the thesis, so the model uses it and will have to move when the disclosure runs out.
NVIDIA's own outlook assumes zero Data Center compute revenue from China, and there were no Hopper shipments at all in the basis quarter against $4.6B a year earlier. This case licences it back. Read it as a step, not a slope: the model applies the tilt as compounding growth, so it understates the first year and only lands the size of the prize near the end of the horizon.
Latest: $283.45B (2032Q1E)
| Period | Value |
|---|---|
| 2026Q1 | $44.06B |
| 2026Q2 | $46.74B |
| 2026Q3 | $57.01B |
| 2026Q4 | $68.13B |
| 2027Q1 | $81.61B |
| 2027Q2E | $90.18B |
| 2027Q3E | $99.68B |
| 2027Q4E | $110.21B |
| 2028Q1E | $121.90B |
| 2028Q2E | $134.86B |
| 2028Q3E | $149.25B |
| 2028Q4E | $165.22B |
| 2029Q1E | $182.94B |
| 2029Q2E | $202.62B |
| 2029Q3E | $224.48B |
| 2029Q4E | $248.75B |
| 2030Q1E | $268.62B |
| 2030Q2E | $270.39B |
| 2030Q3E | $272.18B |
| 2030Q4E | $273.99B |
| 2031Q1E | $275.83B |
| 2031Q2E | $277.70B |
| 2031Q3E | $279.59B |
| 2031Q4E | $281.51B |
| 2032Q1E | $283.45B |
What drives each segment
Data Center — Compute
Units × priceThe GPU business: Blackwell 300 today, Rubin next, sold increasingly as whole NVL72-class racks rather than as chips. Roughly three quarters of the company, and the line on which every other number in this model depends.
Latest: $204.47B (2032Q1E)
| Period | Value |
|---|---|
| 2026Q1 | $34.16B |
| 2026Q2 | $33.84B |
| 2026Q3 | $43.03B |
| 2026Q4 | $51.33B |
| 2027Q1 | $60.45B |
| 2027Q2E | $66.96B |
| 2027Q3E | $74.16B |
| 2027Q4E | $82.14B |
| 2028Q1E | $90.98B |
| 2028Q2E | $100.78B |
| 2028Q3E | $111.62B |
| 2028Q4E | $123.64B |
| 2029Q1E | $136.94B |
| 2029Q2E | $151.68B |
| 2029Q3E | $168.00B |
| 2029Q4E | $186.09B |
| 2030Q1E | $200.63B |
| 2030Q2E | $201.11B |
| 2030Q3E | $201.58B |
| 2030Q4E | $202.06B |
| 2031Q1E | $202.54B |
| 2031Q2E | $203.02B |
| 2031Q3E | $203.51B |
| 2031Q4E | $203.99B |
| 2032Q1E | $204.47B |
Assumptions & reasoning
- Modelled as racks shipped times price per rack, not as a revenue growth rate, because NVIDIA's constraint is physical — advanced packaging, HBM supply and power at the customer site — and a single growth number would hide it. The rack count is backed out of revenue at a $3M rack, not disclosed: NVIDIA publishes no unit volumes at all.
- The volume ceiling is the whole argument. 65,000 racks a quarter is roughly 3.2x today's implied output, and the line reaches it in about three years and then stops. A model without a ceiling compounds 10.5% a quarter for five years and arrives at a number larger than the entire semiconductor industry, which is how these models usually go wrong.
- Price per rack drifts DOWN 1% a quarter while the rack gets more capable. That is the historical pattern in accelerated computing — performance per dollar improves — and it is also what competition from custom ASICs would force. If you think NVIDIA holds price, that slider is where you say so.
- The first projected quarter lands about 2% below NVIDIA's own $91.0B guide for 2027 Q2. The model is a smooth ramp and the real thing is lumpy; the gap is stated rather than tuned away.
- EBITDA margin starts at 76% — segment gross margin less direct cost, before the 5% of revenue this model charges centrally for R&D and SG&A — and glides to 66%, which is the assumption that custom silicon and system content eventually cost NVIDIA something.
Data Center — Networking
Units × priceNVLink, InfiniBand and Spectrum-X Ethernet: the fabric that turns racks into one machine. It grew 199% year on year against compute's 77%, because the dollars of networking inside each rack keep rising.
Latest: $67.58B (2032Q1E)
| Period | Value |
|---|---|
| 2026Q1 | $4.96B |
| 2026Q2 | $7.25B |
| 2026Q3 | $8.19B |
| 2026Q4 | $10.98B |
| 2027Q1 | $14.80B |
| 2027Q2E | $16.63B |
| 2027Q3E | $18.70B |
| 2027Q4E | $21.03B |
| 2028Q1E | $23.64B |
| 2028Q2E | $26.58B |
| 2028Q3E | $29.89B |
| 2028Q4E | $33.61B |
| 2029Q1E | $37.79B |
| 2029Q2E | $42.49B |
| 2029Q3E | $47.78B |
| 2029Q4E | $53.72B |
| 2030Q1E | $58.80B |
| 2030Q2E | $59.83B |
| 2030Q3E | $60.88B |
| 2030Q4E | $61.95B |
| 2031Q1E | $63.04B |
| 2031Q2E | $64.14B |
| 2031Q3E | $65.27B |
| 2031Q4E | $66.42B |
| 2032Q1E | $67.58B |
Assumptions & reasoning
- This line is modelled on the SAME rack count as compute, at a different price. Networking is not sold on its own — it is attached to the systems compute ships — so giving it an independent growth rate would let the two drift apart in a way the business cannot.
- The whole claim is therefore in the price: $734k of networking content per rack today, drifting UP 1.25% a quarter. That is the one price in this model that rises, and it encodes scale-up NVLink plus Spectrum-X displacing third-party Ethernet. Networking share of Data Center goes from 20% to about 28% over the horizon.
- The risk this hides: networking is the part of the rack that merchant silicon attacks first, and Broadcom is the reason. If you think content per rack flattens rather than compounds, that slider takes the line back to a fifth of Data Center and takes roughly a tenth off the fair value.
- Margin is set below compute — 68% gliding to 60% — because optics and cabling carry real bill-of-materials cost that a GPU does not.
Edge Computing
Growth pathEverything that is not a data centre: GeForce and consoles, RTX workstations, DRIVE and Jetson, AI-RAN. NVIDIA renamed and merged these four platforms into one line this quarter, which is a fair description of how much it matters to the P&L.
Latest: $11.39B (2032Q1E)
| Period | Value |
|---|---|
| 2026Q1 | $4.95B |
| 2026Q2 | $5.65B |
| 2026Q3 | $5.79B |
| 2026Q4 | $5.81B |
| 2027Q1 | $6.37B |
| 2027Q2E | $6.59B |
| 2027Q3E | $6.82B |
| 2027Q4E | $7.04B |
| 2028Q1E | $7.27B |
| 2028Q2E | $7.50B |
| 2028Q3E | $7.74B |
| 2028Q4E | $7.97B |
| 2029Q1E | $8.21B |
| 2029Q2E | $8.45B |
| 2029Q3E | $8.70B |
| 2029Q4E | $8.95B |
| 2030Q1E | $9.20B |
| 2030Q2E | $9.46B |
| 2030Q3E | $9.72B |
| 2030Q4E | $9.98B |
| 2031Q1E | $10.25B |
| 2031Q2E | $10.53B |
| 2031Q3E | $10.81B |
| 2031Q4E | $11.10B |
| 2032Q1E | $11.39B |
Assumptions & reasoning
- This is the one line carried on a plain growth rate, because there is no operating driver worth modelling: NVIDIA has just stopped reporting the four platforms separately, so the units underneath are no longer visible even in principle.
- The historical points are exact, not apportioned. Total revenue less Data Center gives the same number as adding up Gaming, Professional Visualization, Automotive and OEM & Other for every quarter shown — 2026 Q4 is 3,727 + 1,321 + 604 + 161 = 5,813, which is 68,127 less 62,314.
- 3% a quarter decaying to 1.5% is a deliberately dull assumption for a line NVIDIA has spent a year describing as the home of physical AI and robotics. Those are real and they are small: automotive was $604M in its last separately reported quarter, under 1% of the company. Betting on it here would move the fair value by less than the rounding on Data Center.
- It also holds a real risk that is easy to miss — the CFO commentary blames slower consumer PC demand on elevated memory prices. Micron's supercycle is NVIDIA's cost line, at both ends of the building.
Where each case comes from
China returns case — primary sources
The primary sources this case is built from — filings, calls and posts. Where they are claims by an interested party rather than disclosures, every number in the China returns column is what happens if they are taken at face value.
China is excluded from the company's own numbers
- May 20, 2026 No shipments of Data Center Hopper products to China occurred during the quarter, compared with $4.6 billion in the first quarter of fiscal year 2026.
- May 20, 2026 Revenue is expected to be $91.0 billion, plus or minus 2%. We are not assuming any Data Center compute revenue from China in our outlook.
What the supply behind it looks like
From cash flow to fair value
| Present value of free cash flow, 20 quarters | $1.53T |
| Terminal-year revenue | $1.12T |
| Terminal-year EBITDA | $680.65B |
| Exit multiple, on revenue | 10.0x |
| Terminal value | $11.22T |
| Discounted at 10.0% a year, terminal value becomes | $6.97T |
| Enterprise value | $8.50T |
| Net cash | $41.80B |
| Equity value | $8.54T |
| Diluted shares | 24.39B |
| Fair value per share | $350.31 |
| Against the current price of $216.85 | +62% |
8x terminal revenue against 21x trailing today. The ramp now runs to 2030: volume compounds at about 44% a year through 2030 Q1, decays across the rest of 2030 as it approaches the 65,000-unit-a-quarter ceiling, and from 2031 the line is flat to slightly down. That is the awkward part — the terminal year itself barely grows, at a 59% EBITDA margin, so 8x revenue on it is about 13.5x EBITDA for a business that has stopped compounding. Generous, deliberately, because the alternative is to argue the ceiling is wrong, which is what the Bull case does instead. Note the multiple is a per-case input, not a global one: Bear holds 4x, Bull 12x, China returns 10x. Move the slider: at 4x the fair value is about $151 and at 12x about $339, so this one input swings the answer more than every operating assumption on this page combined. What the model does NOT count: NVIDIA's strategic investments in its own customers, and the $119.0B of supply commitments sitting behind the ramp.
Read the other way round: at $216.85 the market is paying 5.3x terminal-year revenue, holding every other assumption on this page fixed. That is the number to argue about.
The projected path
| Quarter | Data Center — Compute | Data Center — Networking | Edge Computing | Revenue | YoY | EBITDA | Capex | FCF | R40 | PV of FCF |
|---|---|---|---|---|---|---|---|---|---|---|
| 2027 Q2E | $66.96B | $16.63B | $6.59B | $90.18B | +93% | $60.03B | $2.72B | $47.56B | +146 | $46.44B |
| 2027 Q3E | $74.16B | $18.70B | $6.82B | $99.68B | +75% | $65.72B | $3.18B | $51.91B | +127 | $49.49B |
| 2027 Q4E | $82.14B | $21.03B | $7.04B | $110.21B | +62% | $72.03B | $3.68B | $56.73B | +113 | $52.82B |
| 2028 Q1E | $90.98B | $23.64B | $7.27B | $121.90B | +49% | $79.04B | $4.24B | $62.09B | +100 | $56.44B |
| 2028 Q2E | $100.78B | $26.58B | $7.50B | $134.86B | +50% | $86.82B | $4.86B | $68.03B | +100 | $60.39B |
| 2028 Q3E | $111.62B | $29.89B | $7.74B | $149.25B | +50% | $95.47B | $5.54B | $74.64B | +100 | $64.69B |
| 2028 Q4E | $123.64B | $33.61B | $7.97B | $165.22B | +50% | $105.07B | $6.30B | $81.97B | +100 | $69.38B |
| 2029 Q1E | $136.94B | $37.79B | $8.21B | $182.94B | +50% | $115.72B | $7.15B | $90.12B | +99 | $74.48B |
| 2029 Q2E | $151.68B | $42.49B | $8.45B | $202.62B | +50% | $127.56B | $8.08B | $99.17B | +99 | $80.03B |
| 2029 Q3E | $168.00B | $47.78B | $8.70B | $224.48B | +50% | $140.72B | $9.12B | $109.23B | +99 | $86.07B |
| 2029 Q4E | $186.09B | $53.72B | $8.95B | $248.75B | +51% | $155.33B | $10.27B | $120.40B | +99 | $92.64B |
| 2030 Q1E | $200.63B | $58.80B | $9.20B | $268.62B | +47% | $167.07B | $11.25B | $129.33B | +95 | $97.17B |
| 2030 Q2E | $201.11B | $59.83B | $9.46B | $270.39B | +33% | $167.36B | $11.46B | $129.39B | +81 | $94.93B |
| 2030 Q3E | $201.58B | $60.88B | $9.72B | $272.18B | +21% | $167.72B | $11.65B | $129.53B | +69 | $92.79B |
| 2030 Q4E | $202.06B | $61.95B | $9.98B | $273.99B | +10% | $168.15B | $11.84B | $129.74B | +57 | $90.75B |
| 2031 Q1E | $202.54B | $63.04B | $10.25B | $275.83B | +3% | $168.64B | $12.01B | $130.00B | +50 | $88.79B |
| 2031 Q2E | $203.02B | $64.14B | $10.53B | $277.70B | +3% | $169.20B | $12.18B | $130.33B | +50 | $86.92B |
| 2031 Q3E | $203.51B | $65.27B | $10.81B | $279.59B | +3% | $169.81B | $12.34B | $130.70B | +49 | $85.12B |
| 2031 Q4E | $203.99B | $66.42B | $11.10B | $281.51B | +3% | $170.47B | $12.49B | $131.12B | +49 | $83.38B |
| 2032 Q1E | $204.47B | $67.58B | $11.39B | $283.45B | +3% | $171.18B | $12.63B | $131.59B | +49 | $81.71B |
Every row is projected. A year-over-year change is shown only where the quarter it compares against exists — an em dash means there is no comparable quarter, not a flat year.
Model revisions
Assumptions are marked to reality as each quarter prints. Every change is appended here, with the fair value the model produced at the time, so the model's own history stays visible.
| Date | Changed | Fair value then | Note |
|---|---|---|---|
| 2026-08-17 | all | $245.07 | First cut, built on the 2027 Q1 CFO commentary. Three verticals on the market-platform axis; rack count and rack price carry both Data Center lines. |
| 2026-08-18 | dcCompute.unitsCeiling, dcNetworking.unitsCeiling, scenarios.discountRate, scenarios.terminalGrowthDelta | $245.07 | Rack ceiling raised from 48,000 to 65,000 a quarter, which is what moved the fair value; the ramp now runs to 2030 before it binds. Each scenario also discounts at its own rate — Bear 13%, Bull 9% — and carries its own terminal growth delta. |