---
ticker: "NVDA"
company: "NVIDIA Corporation"
date: "2026-10-16"
view: "Buy"
fair_value: 251.13
price: 214.914
price_source: "last trade (intraday; prior official close $217.23 on 2026-10-15, SIP)"
price_time: "2026-10-16T19:20:54Z"
price_vs_fair_value: "-14.4%"
bear: 73.64
bull: 544.34
exit_price: 385.28
p_price_justified: "20-50%"
restated: "2026-10-26"
method: "case-spread-2026-10-24"
original_fair_value: 234.3
original_view: "Buy"
page: "https://alpha.dming.si/research/2026-10-16-nvda-2"
---

# NVDA · NVIDIA Corporation · Buy

Research report as of 2026-10-16, partly unblinded. Page: https://alpha.dming.si/research/2026-10-16-nvda-2

We value NVDA at $251 a share. The report's $214.91 price (Oct 16, 2026) was 14% below fair value: Buy.

_Restated Oct 26, 2026 (method case-spread-2026-10-24); the report's original value was $234, Buy._

| Lens | Range | Marks |
| --- | --- | --- |
| Our report | $73.64 to $544 | fair $251; original $234 |
| Our simulation | $99.70 to $451 | median $211 |
| Street targets (25) | $210 to $515 | average $339, median $325 |
| Price | | $214.91 |

Bands run from our bear to bull (the hollow mark is the original $234), our simulation's 10th to 90th percentile and the Street's low to high. Street targets forecast future prices, not value today. Market data, captured Oct 16, 2026.

| Figure | Value |
| --- | --- |
| Fair value | $251 |
| Price | $214.91 (market data, Oct 16, 12:20 PM PT) |
| Five-year return | 15.2% (a year, at the report's price) |
| Bear / bull | $73.64 / $544 |
| Exit price | $385 (on Jan 25, 2032) |
| Confidence | 35% |
| Price against our scenarios | The price is justified in 20-50% of our scenarios (20,000 simulated outcomes, each drawing the uncertain drivers from ranges we chose: our judgement, not a measured chance) |

**What the price needs.** Any one of these, on its own, would justify the price:

- Buyers' cash-flow growth, 2027-31: needs +9.7% a year, we have +13.2% a year
- Buyers' long-run capex per dollar of cash flow: needs 0.62, we have 0.8
- NVIDIA's share of buyers' capex, drift a year: needs -1.4% a year, we have -0.5% a year

_This text argues the original call. The current figures are at the top._

**Original call, Oct 16, 2026 (before the Oct 26, 2026 restatement)**

```
NVDA — NVIDIA — research view (market: Buy at $214.914)
As of 2026-10-16 · partly unblinded (one Street revenue row in the sector note, not used; see assumptions.md) · after one review round

Fair value today      $234 per diluted share, discounting at a 10.59% cost of equity (the yearly
                      return investors require); the blind commit, our model before it saw the
                      price, had $259
Price                 $214.914, last trade, 2026-10-16T19:20:54Z (intraday; official
                      2026-10-15 close $217.23, SIP)  →  fair value 9.0% above the price
Five-year exit price  $337 at the end of FY2032, 2032-01-25: the average of two estimates, a long fade
                      of returns toward the cost of capital ($213) and an outside view from what
                      comparable companies trade at ($462). Fading over 10 or 20 years instead
                      gives $336/$339; value-driver formula 12.5x against the fade's 12.9x
Five-year return      +12.6% a year from today's price on our central case, dividends
                      and payouts included (five_year_return.py); 0.6 point above the 12% hurdle,
                      the yearly return our call rule requires
Purchase price        $220 to earn the 12% hurdle; the price is 2.5% under it
Call rests on         A narrow Buy that depends on one setting of the method: the outside view at
                      the peers' latest P/E, 31.4x. Purchase price (and the view it gives) with one
                      setting changed, the rest central, at $214.914: peers' ten-year median 22.0x
                      $183 (Hold); fade alone, no peers $152 (Avoid); NVIDIA's own beta 1.85 $212
                      (Hold); own beta and 22.0x $175 (Avoid). Two forecast choices also sit near
                      the line: ACIE ratio (AI-cloud, industrial and enterprise revenue over
                      hyperscale revenue) 0.89 instead of 0.95 $214 (Hold); FY28 take (NVIDIA's
                      share of the buyers' capex) at the H2 FY27 rate $200 (Hold)
Bear / bull today     $86 / $405 from coherent operating cases (each a single consistent business
                      outcome); weighting bear, central and bull 25/50/25 gives $240
Confidence            35% (subjective, not calibrated): the chance that value on 2027-10-16, re-estimated
                      on what is then known, lands within ±20% of $234
Moat                  The default AI platform (GPU, networking, software, racks) that buyers fund first;
                      custom silicon caps its price and share, not its access
Market expects        The four largest buyers' (Amazon, Microsoft, Alphabet, Meta) operating cash
                      flow growing 11.4% a year over 2027-31 (our 13.2%), or their long-run capex
                      over cash flow at 0.70 (our 0.80), or NVIDIA's share of their capex falling
                      1.0% a year (our 0.5%); each alone, everything else central
Price needs           One of: buyers' operating cash flow growth +11.4% a year (central +13.2%);
                      long-run capex over operating cash flow 0.699 (central 0.8); change a year in
                      NVIDIA's share of their capex -1.0% (central -0.5%). Three further drivers
                      reach the price alone: AI clouds and enterprise revenue at 0.774 of
                      hyperscale (central 0.95), FY2032-33 gross margin 63.1% (central 69.0%) and
                      the 2027 build-ahead at -0.73 times 2026's rise (central +0.5)
Odds in our scenarios 44% that fair value reaches the price (P(price justified); subjective, not
                      calibrated): simulated median $203, 10th to 90th percentile $109 to $313; 59%
                      in the base world (weight 75%), under 1% in the digestion world (25%) and over
                      99% in the bull-leaning one. A diagnostic of our scenario spread, not the
                      odds of a 12% return
Key driver            What the four largest buyers spend and NVIDIA's share of it: buyers' cash-flow
                      growth at 8%/16% a year moves value -$50/+$34; NVIDIA's share drifting -1.5/+0.5
                      points a year moves it -$40/+$43
Alpha model           Rank 129 of 713, score 82.0 (Strong); growth A, momentum C, quality C,
                      valuation C, expectations C (as of 2026-10-16)
Street                Mean target $339.17, median $325, range $210-$515 (Street consensus, read
                      2026-10-16): 3 analysts in the last month at $261.67, 25 in the last
                      quarter at $332.32, 92 in the last year at $301.45; an expectations
                      reference, not validation
```

**We rate NVIDIA a narrow Buy at $214.914: our central fair value is $234.30
a share, 9.0% above the price, and the central case returns 12.6% a year over
five years against our 12% hurdle.** The Buy is fragile. It exists only
because our method values the end of FY2032 partly at the peers' latest P/E
(31.4x). At their ten-year median of 22.0x the same model says Hold, and on the
fade alone Avoid. Two forecast choices (the take and the ACIE ratio, defined
below) also sit near the line. The case against is the buyers' funding: buyers'
cash flow growing 8% a year instead of the central path gives $184.

This revises [the 2026-10-15 NVDA report](https://alpha.dming.si/research/2026-10-15-nvda), which
stays as it was. Model: model.py, output output.txt,
blind record assumptions.md, drivers drivers.json.
The blind commit is `2e65bc4` on `main`. After the market pass one review round
changed the central from the blind commit's $258.55 to $234.30; the
answer table at the end says what changed and why.
`assumptions.md` is left as the blind record, so where it differs from this
report (the take step, the ACIE ratio, the bull's path), this report is current.

**Terms used below.**
- **Bear, central and bull cases.** The central case is our best estimate and
  sets fair value. The bear is a severe but plausible bad outcome, and the bull
  is the matching good outcome.
- **Blind commit.** The model as committed before we read the price.
- **Fade and outside view.** Two ways to value the end of FY2032. The fade lets
  returns on new capital drift toward the cost of capital. The outside view
  applies what comparable companies trade at, here the peers' trailing P/E. The
  exit price is their mean.
- **The four buyers.** Amazon, Microsoft, Alphabet and Meta, the four largest
  buyers of NVIDIA's data-centre systems. Their capex divided by their
  operating cash flow is their **intensity**.
- **Take.** NVIDIA's hyperscale revenue divided by the four buyers' capex.
- **ACIE.** AI clouds, industrial and enterprise (neoclouds, AI labs,
  sovereigns, Oracle-type clouds, enterprise buyers). The **ACIE ratio** is
  their revenue divided by hyperscale revenue.
- **Blackwell Ultra and Rubin.** NVIDIA's current and next generations of
  data-centre systems.
- **Build-ahead.** How much of 2026's rise in intensity the buyers repeat in
  2027.
- **Fiscal years.** NVIDIA's FY2028 ends in January 2028, so it is mostly
  calendar 2027.
- **Source labels in brackets.** [F] reported fact, [M] management claim, [S]
  outside estimate (the Street or a research firm), [A] our assumption, [I] our
  inference or model result.

## Why the revision: the old central case was not one consistent world

Our chief investment officer's evaluation (2026-10-16) found two faults in the old central case. Its
FY2029-32 hyperscale revenue ($346-382B) was 1.8-2.0 times what its own
buyer cross-check allowed ($181-191B), and the report said so without
choosing. And it held FY2030 revenue flat with operating income down 4.2%,
a digestion year it could not date.

This revision makes the central one world. NVIDIA's data-centre revenue is
no longer chosen; it is what its buyers can spend times NVIDIA's share of
it, in code, for every case:

- **Top-four capex** (Amazon, Microsoft, Alphabet, Meta) = their operating
  cash flow × capital intensity (capex over operating cash flow).
- **Hyperscale revenue** = NVIDIA's take × top-four capex.
- **AI clouds, industrial and enterprise** (ACIE: neoclouds, AI labs,
  sovereigns, Oracle-type clouds, enterprise) = hyperscale × a ratio.
- **Revenue** = hyperscale + ACIE + Edge Computing.

The link is not a theory. NVIDIA's data-centre growth has tracked the four
buyers' capex growth: buyers +65% in 2025 against NVIDIA data centre +68% in
FY2026; buyers +95% in 2026 (their guides) against +100% in FY2027 (our
estimate). NVIDIA's take of their capex was 29.0% in H1 2025 and 31.1% in H1
2026 (10-Q Note 13; market data cash flows) [I].

## Which side moves: FY2028 grows 38.9% against management's 70%, then 3-4% a year

**FY2028 (calendar 2027): NVIDIA's side moves down.** Management calls FY28
"approximately 70%" higher [M, Q2 call 2026-08-26]. At our take and ACIE
ratio that needs the four to spend about $1.26T in 2027, 1.36 times their
cash flow, borrowing about $335B (`output.txt`, cross-checks) [I]. No buyer
has given a 2027 figure. What they have said: Amazon short of capacity
through 2027, Alphabet's 2027 capex to "increase significantly", Microsoft's
FY27 capex up, Meta "we have continued to underestimate our compute needs"
[M, Q2 calls July 2026]. They funded 2026 by borrowing $137B net in six
months, and Amazon, Alphabet and Meta bought back no stock in that time;
Microsoft did [F, market data]. Our central lets intensity rise by half of 2026's
rise (0.95 to 1.10), on cash flow +20%: 2027 capex $1.02T, +39% [A].

The take path, stated plainly. NVIDIA's hyperscale revenue over the four's
capex is 31.1% in H1 2026 (measured). The model carries the Q1-to-Q2 fall in
hyperscale's share of data centre (57.2% to 54.7%) through the second half,
which gives a second-half take of 25.0% on 52 weeks and a full-year FY27 take
of 27.5% [I]. FY28 holds that full-year 27.5%, then drifts down half a point
a year to custom silicon [A]. There is **no Rubin step**: the blind commit
added 2.1 points, the rise from H1 2025 to H1 2026 (the Blackwell Ultra
year), but that is a half-year change added to a full-year average, and no
filing yet shows what Rubin does to the take (management's content per
gigawatt, $40B against $25B, has no capex-per-gigawatt beside it) [M]
(answer 2). ACIE is 0.95 of hyperscale, 48.7% of data centre: FY27E's 0.89
plus the measured half-year rise of 0.06, about the model's second-half 0.96,
and close to management's "roughly half" [I, M] (answer 3). Data centre then
grows 41% on capex +39%, in line with the last two years (growth over capex
growth 1.02 and 1.03; FY28 1.015). FY28 revenue is $582B, **+38.9%**, with
the guidance beat already in FY27.

**FY2029-32: both sides move.** The old cross-check used our in-house models of the four buyers
that cut capex 17% in 2028 by a mechanical fade while their own cloud
revenue grows 30% a year; the buyers' statements contradict them (our
semiconductor sector note). Dell'Oro's forecast of $3T of world data-centre capex by 2030
(2026-08-18) would need the four to keep spending more than their cash flow
to 2030 [S, I]. We use neither. The four's cash flow grows 15%, 12%, 10%,
9% (from 20.8% a year in 2021-25 [F]), and intensity halves its gap to a
long-run 0.80 every two years: above the pre-AI 0.41-0.54 (2021-24: 0.46,
0.54, 0.41, 0.48; about 1.7 times their 0.47 average), because AI capacity
takes more capital, and below 1.0, because borrowing cannot pay for capex
for ever [A]. The four then borrow a net $106B over 2027-31 and start
repaying in 2029. NVIDIA's take drifts down half a point a year as custom
silicon grows [A]. Top-four capex grows about 5% a year; NVIDIA revenue
+4.3%, +3.5%, +3.2%, +3.4% in FY29-32. FY32 revenue is $670B, against $826B
in the old central and $631B in its buyer-consistent variant.

**No digestion year in the central.** Operating income rises every year
(the smallest rise is +1.7%). Two changes made that coherent, both dated:
operating expenses now bend with revenue (the old path grew them 10% a year
on flat revenue), and gross margin keeps about half the memory-cost fall in
our in-house memory model (semiconductor sector note), ending at 69% instead of 67%. A
downturn is possible (NVIDIA has had two profit falls in ten years: FY2020
NOPAT -30%, FY2023 -57%), but nothing dates the next one, so it sits in the
bear and in a 25% digestion world in `drivers.json`.

## How fair value moved: $229 in the 2026-10-15 report to $234.30 here

Central fair value per share, each step valued by the shared method at
2026-10-16 (`bridge()` in `model.py`; the review rows are the last four):

| Step | Fair value | Change |
| --- | ---: | ---: |
| The 2026-10-15 report's published central (its own method) | $229 | |
| Same forecast, exit = mean of fade and outside view (not the median with a judged multiple), at the old cost of equity 11.59% | $266 | +$37 |
| The method's -1.0 point cost-of-equity calibration, a setting not a finding (cost of equity 10.59%) | $287 | +$20 |
| Groq licence consideration accrued, $0.986B (10-Q; in the old assumptions, missing from its code) | $287 | -$0.02 |
| Pre-tax cost of debt at a market yield, 5.79%, not the 4.4% coupon | $286 | -$0.11 |
| Hugging Face's $11.9B paid in FY28, not FY27 (8-K 2026-09-03: close in H1 2027) | $286 | -$0.29 |
| Q3 FY27 at the $108B guide plus the median 4.7% beat of six quarters ($113.1B) | $295 | +$9 |
| FY28 revenue from the buyers' world, +70% to +52.7% | $261 | -$33 |
| FY29-32 revenue from the buyers' world, +4.3/+3.6/+3.3/+3.6% (was +8/0/+4/+6%) | $254 | -$8 |
| FY33, the fade's starting growth, +3.5% (was +6%) | $245 | -$9 |
| Operating expenses bend with revenue: +25% in FY28, then to 8% of revenue by FY32 | $251 | +$6 |
| Gross margin FY30-33 70.0/69.5/69.0/69.0% (was 69/68/67/67%) | $258.55 | +$7 |
| Blind commit's central | **$258.55** | |
| Review: Groq accrual out of working capital (answer 6) | $258.58 | +$0.02 |
| Review: interest paid on the notes at 4.75% (answer 7) | $258.52 | -$0.05 |
| Review: no Rubin step in FY28's take, 27.5% not 29.6% (answer 2) | $239.80 | -$18.72 |
| Review: ACIE ratio 0.95 not 1.00 (answer 3) | $234.30 | -$5.50 |
| **New central** | **$234.30** | **-$24.25 on the blind commit** |

The exit rule is +$37 and the cost-of-equity calibration +$20, together the
old +$58 method change; on the old forecast the calibration is worth $20 (on
the new central it is $234 against $218 without it, $16). The blind commit's
forecast and bridge changes are -$28, and the review round -$24. On one
method, this revision lowers the central from $287 to $234 (-18%).

| | 2026-10-15 report (its method) | Blind commit | This report, after review |
| --- | ---: | ---: | ---: |
| Central fair value | $229 | $258.55 | $234.30 |
| Bear | $80 | $85.95 | $85.93 |
| Bull | $483 | $456.84 | $404.59 |
| Exit price, central | $329 | $372.68 | $337.40 |
| Purchase price, central | $223 | $243.12 | $220.34 |
| Case-weighted | $255 | $265 | $240 |
| View at $214.914 | | Buy | Buy |

Bear and bull were rebuilt as buyers' worlds too (below), so their change
from the old report mixes the method and the case.

## NVIDIA has beaten its own guide six quarters running, and who funds its customers is the largest hidden risk

NVIDIA sells the systems AI data centres are built around: GPUs, networking
and racks (data centre 92.5% of Q2 FY27 revenue). Q2 FY27 revenue $96.2B
(+106%); data centre $89.0B, of which hyperscale $48.7B and ACIE $40.3B
[F, 10-Q]. Q3 guide $108B ±2%, gross margin 74% [M]. The company has beaten
its own revenue guide in each of the last six quarters by 2.5-5.7% (median
4.7%), above the top of its ±2% range in five: it guides low on purpose
[F, CFO commentaries]. We forecast Q3 at $113.1B.

Who funds the customers is the largest hidden risk, unchanged from the
2026-10-15 report: supply commitments of $279B (FY28 $87B, FY29 $88B), $99B of
equity stakes and $25B more committed, $36B of AI-cloud capacity buy-backs,
guarantees up to $105B for an OpenAI site from FY29, and labs NVIDIA backs
at about a quarter of FY28 business [F, M, 10-Q Note 10; Q2 call]. The bear
case charges that through a lower ACIE ratio and a larger expected loss in
the bridge.

The only filing since the 10-Q is the 8-K of 2026-09-03: Hugging Face for
about $11.9B plus up to $1.0B of retention stock, closing in the first half
of 2027 [F]. Its cash now leaves in FY28 as an acquisition.

## Forecast and valuation: the Buy rests on the peers' latest P/E

| Central, $B | FY27E | FY28 | FY29 | FY30 | FY31 | FY32 | FY33 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Top-four capex (calendar year before) | 732 | 1,019 | 1,078 | 1,133 | 1,189 | 1,251 | 1,317 |
| Take | 27.5% | 27.5% | 27.0% | 26.5% | 26.0% | 25.5% | 25.0% |
| Data centre | 387 | 546 | 568 | 586 | 603 | 622 | 642 |
| Revenue | 419 | 582 | 607 | 628 | 648 | 670 | 693 |
| Growth | | +38.9% | +4.3% | +3.5% | +3.2% | +3.4% | +3.4% |
| Gross margin | 73.7% | 72.5% | 71.0% | 70.0% | 69.5% | 69.0% | 69.0% |
| Operating income | 273 | 378 | 384 | 390 | 399 | 409 | 423 |
| NOPAT | 228 | 314 | 319 | 324 | 331 | 339 | 351 |
| EPS ($) | | 13.01 | 13.35 | 13.71 | 14.14 | 14.62 | 15.25 |

Shares are held at 24.242B and stock pay is expensed; 68% of free cash flow
after interest is paid out (the H1 FY27 rate), and the rest builds to $609B
of cash and securities by FY32 [A]. The fade (15 years to 3% terminal
growth) starts at the volume-only return and ends at a 30% mature return
(FY2016-21, 32%), so value hardly moves with it (±$1). The exit is the mean
of the fade ($213) and the outside view ($462).

**The outside view is 117% above the fade, and the call rests on it.** The
input that drives the gap is the peers' multiple: MSFT, AAPL, CSCO and QCOM
at a 31.4x trailing P/E median on FY32 net income of $354B, against a fade
that capitalises year-six NOPAT at 12.9x. About half of the gap is today's
multiples against their own history: at the peers' ten-year median of 22.0x
the outside view is $325, fair value $194 (-$40) and the purchase price $183,
below the price. On the fade alone the central is $161 and the purchase price
$152 (an Avoid). So the Buy exists only because the method averages in the
peers' latest P/E. The method (valuation.md) sets that rule for every report,
and it is not changed for this one. The rest of the gap is the business: the
peers are mature franchises with steadier profits, while NVIDIA's profit has
fallen 30% and 57% in past downturns. We believe the fade more; read the exit
as fade-low, peer-high.

The call under one setting changed at a time, the rest central (price
$214.914; Buy at or below the purchase price, Avoid when fair value is under
0.85 times the price, $182.7; `output.txt`, last block):

| Setting | Fair value | Purchase price | View |
| --- | ---: | ---: | --- |
| The method: peers' latest P/E 31.4x, beta 1.50 | $234.3 | $220.3 | Buy |
| Peers' ten-year median P/E 22.0x | $193.9 | $182.6 | Hold |
| Fade alone, no outside view | $161.1 | $151.8 | Avoid |
| NVIDIA's own beta 1.85, P/E 31.4x | $211.7 | $212.2 | Hold |
| Own beta 1.85, P/E 22.0x | $174.0 | $174.5 | Avoid |
| Own beta 1.85, fade alone | $135.3 | $135.6 | Avoid |
| ACIE ratio 0.89 in every year (answer 3) | $227.7 | $214.1 | Hold |
| FY28 take at the H2 FY27 rate, 25.0% (answer 2) | $212.8 | $200.1 | Hold |
| The blind commit's choices (Rubin step, ACIE 1.00) | $258.5 | $243.1 | Buy |

Bridge at the FY32 exit (central): cash $609B less operating cash $20B and
notes $33.5B; equity investments $99B less $6B tax on gains; non-current
income tax payable $5.6B (all deducted; $5.0B of it is unrecognised tax
benefits); Groq $1.0B; expected loss on guarantees and AI-cloud commitments
$7.5B (15% of $36B plus 2% of $105B).

Cost of equity 10.59% = 5.29% + 1.50 × 4.20% − 1.0 point (the method's
calibration, a setting, not a finding). Beta 1.50 is Damodaran's semiconductor
industry beta. NVIDIA's own beta, from the market pass, is 1.85 (below). Beta
1.2 gives $259, 1.57 gives $229, 1.8 gives $215, 1.85 gives $212; without the
calibration $218.

Sensitivity of central fair value, one driver at a time (`output.txt`):

| Driver | Low / high | Fair value |
| --- | --- | ---: |
| Buyers' cash-flow growth, 2027-31 average | 8% / 16% | $184 / $268 |
| Long-run buyer intensity | 0.60 / 1.00 | $196 / $273 |
| NVIDIA take drift a year | -1.5 / +0.5 points | $194 / $277 |
| ACIE ratio | 0.73 / 1.15 | $210 / $256 |
| FY32 gross margin | 60% / 72% | $205 / $244 |
| 2027 build-ahead (× 2026's intensity rise) | 0 / 1 | $226 / $243 |
| Mature return | 20% / 40% | $233 / $235 |

The 2027 build-ahead matters little because intensity converges to the
same long-run level; how much of their cash flow buyers keep spending, and
NVIDIA's share of it, matter most.

## Bear ($86) and bull ($405): each one consistent world

**Bear ($86; exit $115).** Buyers stop building ahead: 2027 capex at their
cash flow ($864B, +18%), then a 17% cut in 2028 (our in-house buyer models'
cut) and intensity settling at 0.68. Neoclouds and labs lose funding, so
ACIE falls to 0.70 of hyperscale (below H1 FY26's 0.73). Custom chips take
share (take 26% falling to 22%) and price (gross margin 70% to 58%; FY2023's
57% is the ten-year floor). Revenue $450B in FY28 (below Q4 FY27's annual
run rate), $336B in FY29 (-25%), then about $310-318B. Operating income
falls 38% in FY29. Operating expenses are sticky on purpose (answer 4):
$42B in FY28 and $43-49B after, 9.4% of revenue rising to 15.6%.
Expected loss on guarantees and AI-cloud commitments $33.8B.

**Bull ($405; exit $592).** Buyers keep borrowing to build: 2027 capex
$1.25T (1.30 times cash flow, near the $1.3T NVIDIA cites for the top five),
then $1.37T, $1.43T and $1.50T in 2030, Dell'Oro's half of $3T (intensity
easing from 1.30 to 1.03 by 2030 and 0.95 by 2032; answer 5). Rubin lifts the
take to 31%; AI clouds and sovereigns outgrow hyperscalers (ratio 1.15). FY28
+101%, about what management says its customers' forecasts imply. Operating
expenses follow the central's rule, a share of revenue rising from 5.5% to 8%
by FY32 (answer 4). The four borrow $690B net over 2027-31, the bull's
weakest joint; FY31 revenue is $1,050B.

**Why the other side could be right.** For a higher value: NVIDIA guides
low on purpose, its +70% is "preliminary" and "supply-constrained", and its
supply commitments for FY29 ($88B) equal FY28's ($87B), so the company plans
no FY29 fall. If the four spend $1.26T in 2027, FY28 is +70% and the
central is about $52 too low on that year alone. For a lower value: we let
intensity settle at 0.80, well above the pre-AI level; at 0.60 the central
is $196, and a buyer digestion like our in-house models' gives the bear. The
call itself is the lower side's best case: at the peers' ten-year median
multiple the purchase price is $183, under the price.

## Catalysts and falsifiers

| Date | Event | What it settles |
| --- | --- | --- |
| Late October 2026 | Q3 calls of Alphabet, Microsoft, Meta, Amazon | First 2027 capex numbers. The four's 2027 plans summing to $1.26T or more support management's +70% (our central too low); $880B or less points to the bear |
| 2026-11-17 | NVIDIA Q3 FY27 results (company date) | Q3 revenue against our $113.1B; Q4 guide against our $127.8B; FY29 supply commitments; ACIE against hyperscale (our FY27E ratio 0.89, H2 0.96) |
| 2027-02 | NVIDIA FY27 results; buyers' 2026 actuals | Take for the year (ours 27.5%); intensity (ours 0.95) |
| 2028-02 | NVIDIA FY28 results | FY28 revenue against our $582B (±10%); an ACIE ratio below 0.8 means funding of labs and neoclouds is failing (bear) |
| 2029-02 | Buyers' 2028 capex | Down more than 10% is the bear's digestion; flat to +10% is the central |

## Sources

- NVIDIA Form 10-Q for Q2 FY27, filed 2026-08-26, masked (`data/filings/`);
  8-K 2026-09-03 (Hugging Face); CFO commentaries and press releases Q4 FY25
  to Q2 FY27 (8-K exhibits; `data/filings/guidance-record-cfo-commentaries.txt`).
- Q2 FY27 call, 2026-08-26, as recorded in the 2026-10-15 report's assumptions.
- Market data statements: NVIDIA annual and quarterly; AMZN, MSFT, GOOGL, META,
  ORCL, CRWV cash flows (`data/`, fetched 2026-10-16).
- Buyers' Q2 2026 calls (2026-07-22 to 2026-07-30) as recorded in the AMZN,
  MSFT, GOOGL and META-2 reports; Meta's Q2 call 2026-07-29 (S&P Global
  preview 2026-07-23; press summaries).
- Oracle Q1 FY27 results (September 2026) and CoreWeave Q2 2026 results
  (2026-08-11), press summaries.
- Dell'Oro Group, "AI buildout maintains momentum as data center capex
  surpasses $3 trillion by 2030", 2026-08-18; "Data center capex grew 92
  percent in 2Q 2026", September 2026.
- Sector note `docs/research/sectors/semis-ai/note.md` (2026-10-15).
- Damodaran US industry betas (January 2026) and implied premium
  (2026-10-01), through `method.py`.
- Peers' P/E: market data ratios-ttm 2026-10-15 19:19 UTC; market data `ratios` annual
  (2026-10-16).

After the blind commit `model.py` gained: the outside-view line at the
ten-year median P/E, the table of exit methods and betas, the take, ACIE,
finance-lease and coupon sensitivities, the split step bridge, and the history
of operating expenses. They print and change no value; the values that did
change are in the review round below.

## Market pass (2026-10-16): price, odds, own beta, Alpha and the Street

Run by the side that checks the analyst's work, after the blind commit (`2e65bc428`). The numbers in this section were rerun after the review round below; the first version (central $258.55) is in `git log` for this folder.

**Price and view.** NVDA $214.914 (last trade, 2026-10-16T19:20:54Z; the official close on 2026-10-15 was $217.23, SIP). The price is below the purchase price ($220.34), so the view is Buy, by 2.5%. The central case returns 12.6% a year to the exit, from `five_year_return.py` on the saved `forecast.json`.

**What the price requires** (`alpha research odds`, `odds.json`): a reverse underwrite, one driver at a time with the rest central. The price needs the four buyers' operating cash flow to grow 11.4% a year over 2027-31 against our 13.2%, the smallest step. At the low end of the buyers' range (8% a year) fair value is $184, so the price sits inside the range the report itself gives. Long-run capex over operating cash flow of 0.699 against our 0.80 is the next, and a 1.0% yearly fall in NVIDIA's share of their capex against our 0.5%. The ACIE ratio (0.774 against 0.95) and FY2032-33 gross margin (63.1% against 69.0%) also reach the price alone; the mature return cannot (best end gives $228).

**Odds.** 44% that fair value reaches the price, under our scenario spread (20,000 draws, seed 20261016; subjective, not calibrated). The 25% digestion world, our assumption, gives under 1%; the base world gives 59%. Buyers' cash-flow growth explains 69% of the variance alone. The simulated median is $203, 13.5% under the central $234.

**Fair value at the company's own beta.** Not in the analyst's step. Computed here as the OLS slope of NVDA's weekly returns on SPY's over the last 104 weeks, from the stored adjusted closes (`readonly-data-run.sh`, `BetaService`, last close 2026-10-15; the same weekly raw beta the repository uses elsewhere): 1.85 (R-squared 0.48, 104 weeks). Putting it into the same forecast and method, the same code that gives $234.30 at the model's 1.50 gives cost of equity 12.06%, fair value $211.7, purchase price $212.2 and exit $323. Blume-adjusted (0.67 times the raw beta plus 0.33, beta 1.57) fair value is $229 and the purchase price $218.5, still a Buy. **At the raw own beta the price is above the purchase price: the view is Hold, not Buy.** The first version of this paragraph said the Buy did not depend on the beta; that was true at the blind commit's $243 and is not true now.

**Alpha's dated view** (Alpha is Capfolio's stock-ranking model; `alpha rank NVDA --raw --json --allow-stale-model-cache`, read in place, 2026-10-16; cache status partial, integrity verdict CLEAN with a warning count of 2 that was not reviewed). Rank 129 of 713, score 82.0, tier Strong. Category grades: growth A (rank 22), momentum C (260), quality C (237), expectations C (357), valuation C (498). Why we differ is our inference and was not tested: Alpha's only top grade is growth, and its valuation grade is the weakest, which reads today's multiples; this report values the buyers' spending that produces the earnings, so a Buy at 9% under fair value and a middle rank can both hold. The rank does not say whether the price already reflects the growth.

**Street** (market data, read 2026-10-16). Mean target $339.17, median $325, range $210 to $515: 3 analysts in the last month at $261.67, 25 in the last quarter at $332.32 and 92 in the last year at $301.45. The mean is 58% above the price and 45% above our fair value. Consensus revenue is $409.9B for FY2027 (41 analysts), $695.2B for FY2028, $926.4B for FY2029, $1,106.5B for FY2030 and $1,161.0B for FY2031 (42, 38, 30 and 35 analysts), against our $419B, $582B, $607B, $628B and $648B (the vendor's year labels are read as NVIDIA's fiscal years: our inference from the FY2027 match within 2%). We match the Street through FY2027 and sit 16% under in FY2028 and 35%, 43% and 44% under in FY2029-31; consensus EPS of $15.79, $21.40, $25.60 and $26.60 for FY2028-31 is 21%, 60%, 87% and 88% above our $13.01, $13.35, $13.71 and $14.14. The gap is the one the report states: the Street keeps growing after FY2028 where our buyers' world allows 3% to 4%. Market data does not publish the Street's drivers, and targets are an expectations reference, not validation.

## Review round (2026-10-16)

Reviewed in one round by two outside reviewers: a report check (research-eval,
our procedure for checking a report's math and sourcing) and a model auditor
(who reruns the model and checks its inputs). Rules: `references/review.md`. Every finding is answered below as fixed, disclosed
or rejected, and an assumption changed only on dated evidence that does not
depend on the price. Among the blind-commit files, `model.py`, `output.txt`,
`forecast.json` and `valuation.json` changed (the last two are written by
`model.py`); `assumptions.md` and `data/` did not. After the blind commit
`drivers.json`, `record.json`, `odds.json` and this report changed; the stale
`odds.json` was set aside and `alpha research odds` rerun.

| # | Finding | Answer | Before, after | Dollars |
| --- | --- | --- | --- | --- |
| 1 | The Buy rests on the outside view's 31.4x peer P/E: at the ten-year median 22.0x the call is Hold, on the fade alone Avoid; the report shows the $214 sensitivity but not that the call turns on it | **Disclosed.** The decision block now has a "Call rests on" line, and the forecast section has the rows (table there). The method (valuation.md: peers' latest multiples) is the rule for every report and is not changed for one | Block without the line, with it. Central at 22.0x: fair value $193.9, purchase price $182.6 (Hold; the auditor's $214 and $201.5 were on the blind central). Fade alone $161.1 and $151.8 (Avoid). Own beta 1.85 at 31.4x $211.7 and $212.2 (Hold); at 22.0x $174.0 and $174.5 (Avoid) | $0 (-$40 at 22.0x; -$73 on the fade alone) |
| 2 | Take basis: FY28's 29.6% is FY27's 27.5% plus a 2.1-point Rubin step (a half-year change from H1 2025 to H1 2026), while the model's own H2 FY27 take is 25.9%; no filing supports Rubin's effect | **Fixed.** Evidence, all in the blind data: Note 13 shows hyperscale's share of data centre falling (57.2% to 54.7% from Q1 to Q2); the model's H2 take is 25.0% on 52 weeks (25.9% unadjusted); the +2.1 points were measured on half-years in the Blackwell Ultra year and added to a full-year average; Rubin's $40B against $25B per gigawatt is a management claim with no capex-per-gigawatt beside it. The auditor's counter (29.6% is below H1 2026's measured 31.1%) is true, and 31.1% is the high end of the take's range ($265), not a base. FY28 now holds FY27's full-year 27.5%, which is above the H2 rate; Rubin stays in the bull (31%) and the take-drift range. The basis is stated in "Which side moves" | FY28 take 29.6% to 27.5% (FY28 data centre $604B to $560B at ACIE 1.00); central $258.52 to $239.80. At the H2 rate (25.0%) $212.8 and purchase price $200.1 (Hold); at 31.1% $265.3 | -$18.72 |
| 3 | ACIE ratio: hyperscale times a ratio that jumps from the model's FY27E 0.89 (measured 0.73, 0.79, 0.83) to 1.00 on a management claim ("roughly half"); items 2 and 3 together put the purchase price at $214.2 | **Fixed in part, and labelled.** 1.00 had one source, management [M]. The measured half-year rise (0.73 to 0.79, +0.06) added to FY27E's 0.89 gives 0.95, which is also about the model's H2 FY27 ratio (0.96) and 48.7% of data centre, "roughly half". The Q1-to-Q2 jump (0.75 to 0.83) would reach 1.0 by Q4 FY27 if it continued; we do not bank it, and take the lower sourced figure. 0.89 flat has no source for a flat ratio against a measured rise, so it is not used, but it is the line: with it the purchase price is $214.1, a Hold at $214.914 | ACIE ratio 1.00 to 0.95 in every year; central $239.80 to $234.30. Central with both items at their harshest (take 27.5%, ACIE 0.89): fair value $227.7, purchase price $214.1 (the auditor's $214.2) | -$5.50 (-$12.1 at 0.89) |
| 4 | Bear and bull opex are the old cases' dollars on different revenue (bear 9.4% rising to 15.6% of revenue, bull 5.5-6.7%), while the central's bends with revenue | **Bull fixed; bear rejected with evidence.** Bull: opex follows the central's rule, +30% in FY28 then a share of revenue rising linearly to 8% by FY32 (5.5% to 8.0%; the coherence rule in valuation.md). Bear: the central's rule would cut opex from $42B to $27B in FY29, -36%, and NVIDIA has never cut it: in FY2016-26 opex fell in no year; in the two downturns (FY2020 revenue -7%, FY2023 flat) it rose 16% and 50%, to 36% and 41% of revenue, against the bear's peak of 15.6% (`output.txt`, history). The bear's +3% in FY29 on revenue -25% is already kinder than FY2020 and FY2023 | Bull: opex +30%, +20%, +15%, +12%, +12%, +10% (FY32 $79B, 6.5% of revenue) to +30% then to 8% (FY32 $88B). Bear unchanged; bent to 8% it would give $98.3 | Bull -$7.5 on the old path (-$1.9 after the re-anchor in 5). Bear $0 (+$12.4 not taken) |
| 5 | The bull's FY31 revenue ($1,166B) is within 0.5% of the Street figure in the sector note the analyst read; its 2030 top-four capex ($1,671B) is 11% above the Dell'Oro anchor it cites ($1.5T); anchored to $1.5T it is $408 | **Fixed, and the coincidence disclosed.** The report said "about Dell'Oro's half of $3T" for $1.67T; that was 11% off. The bull's intensity now eases to 1.03 in 2030 so 2030 capex is $1,496B, with 2027 unchanged at $1.25T (NVIDIA's $1.3T for five buyers). Whether the sector note's Street row (the partly unblinded item in assumptions.md) shaped the old bull cannot be ruled out; new FY31 revenue is $1,050B, 9.5% under the Street's $1,161B | Intensity 1.30, 1.25, 1.20, 1.15, 1.10, 1.07 to 1.30, 1.20, 1.10, 1.03, 0.98, 0.95; 2030 capex $1,671B to $1,496B; FY31 revenue $1,166B to $1,050B; bull $456.81 to $406.47 (the auditor's $408) | Bull -$50.3 (-$52.2 with row 4); central $0 |
| 6 | Groq accrual ($986M) sits inside the accrued liabilities that set WC_Q2 and again in the bridge | **Fixed (working capital side).** The accrual is a claim, so it stays in the bridge and leaves the working capital base (the 14% ratio rounds the same) | WC_Q2 $52.615B to $53.601B | +$0.02 |
| 7a | Cost of debt coupon 4.4% against new notes at 4.75-5.6% | **Fixed.** Interest paid on the $33.5B is 4.75%: $8.5B of older notes cost about $0.29B a year (market data, Q4 FY26 interest expense $73M a quarter) and the $25B issued in June 2026 about $1.3B at a 5.2% midpoint of the auditor's range (10-Q Note 9, not saved in `data/filings`) [I]. The cost of debt in WACC was already the 5.79% market yield | 4.4% to 4.75% | -$0.05 (the auditor's under $0.1) |
| 7b | UTB $5.0B against $5.602B income tax payable | **Fixed (wording).** The bridge line is the whole non-current income tax payable, $5.602B, deducted in full; unrecognised tax benefits (with interest and penalties) are $5.0B of it. Line renamed in `model.py`, `forecast.json`, `valuation.json` and this report | "Unrecognised tax benefits $5.6B" to "non-current income tax payable $5.6B ($5.0B unrecognised tax benefits)" | $0 |
| 7c | Capex basis: Meta's and Microsoft's guides include finance leases, market data is cash capex (about 1%) | **Disclosed.** The size of the finance leases is not in the saved sources, so the guide sum is kept at $732.5B. With 1% off it, the intensity and the take base move together and the central rises | 2026 guides $732.5B, or $725.2B at -1% | +$1.6 if corrected (purchase price $221.8) |
| 7d | Pre-AI intensity range omits 2022's 0.54 | **Fixed (wording).** 2021-24 intensity was 0.46, 0.54, 0.41, 0.48 (market data annual cash flows; Microsoft on its June year), so the pre-AI range is 0.41-0.54, and 0.80 is about 1.7 times its 0.47 average. `drivers.json` has the four years | "0.41-0.48" and "almost double" to "0.41-0.54" and "about 1.7 times" | $0 |
| 7e | "Stopping buybacks" is not true of MSFT | **Fixed (wording).** Amazon, Alphabet and Meta bought back no stock in H1 2026 (market data quarterly cash flows, as `assumptions.md` already says); Microsoft did. The blind record's phrase "buybacks stopped" is left as written there | "borrowing $137B and stopping buybacks" to "borrowing $137B; Amazon, Alphabet and Meta bought back no stock, Microsoft did" | $0 |
| 8 | The step bridge lumps the exit-method change with the lower cost of equity (+$58): split it (the calibration alone is worth +$18) | **Fixed.** `bridge()` now values the old forecast first with the mean exit at the old 11.59% cost of equity, then with the calibration. On the old forecast the calibration is +$20.48; on the new central it is $16 ($234.30 against $218 without it); the auditor's +$18 was the same quantity on the blind central ($259 against $241) | One +$58 row to +$37.10 and +$20.48 | $0 (no value changes) |

Review flags. The automated checker (`research-check.py`) raises one: flag 1,
the net move from the blind value after the market was handed over, $258.55 to $234.30 (-$24, -9.4%). The
largest change is row 2, the Rubin step (-$18.72). It rests on the 10-Q's own
Note 13 data and the model's own second-half arithmetic, not on a management
claim or a Street figure found after the price, and it was made on a finding
about internal consistency. It moves the value toward the price, so the test
applied is whether it would have been made at any price: it would. The one
choice that decides the view is row 3, ACIE 0.95 against 0.89 ($6.6 of value,
the line between Buy and Hold); we took the sourced middle and say so. The
view did not change from the blind view (Buy), so flag 2 (the view changing from the blind view) does not fire.

Old against new:

| | Blind commit | After review |
| --- | --- | --- |
| Central fair value | $258.55 | $234.30 |
| Bear / bull | $85.95 / $456.84 | $85.93 / $404.59 |
| Case-weighted | $265 | $240 |
| Exit price | $372.68 | $337.40 |
| Purchase price | $243.12 | $220.34 |
| Five-year return at $214.914 | 14.9% | 12.6% |
| P(price justified) | 53% | 44% |
| View | Buy | Buy, narrowly |