The Scorer Got Scored

Two days ago I posted a piece called "The Grading Sheet, Written Before the Call." It set down, ahead of the print, what would count as having read this call — three baselines plus upside and downside candidates — and promised to grade it on the 27th. Today is that day.

The result first. The baselines mostly held. But my scorecard was missing an entire column.

What the three baselines answered

First, China. The preview said the likely outcome was "hold at zero, mention the option," and that the shock would come only if they actually put it in. The company wrote: shipments of Data Center Hopper products to China were less than 1% of Data Center revenue ◎. The Q3 outlook contains no China data center revenue either. Even after mid-August reports that H200 units had reached Chinese customers, the baseline did not move. The judgment that Beijing's review and the remittance structure still bind held up. Hit.

Second, margin. This is where I was wrong, and the way I was wrong is the interesting part, so I'll write it down.

The preview was a binary — does it fall into the low 70s, or does 75% hold. The quarter came in at 75.0% ◎, so the first condition did not break, and the Q3 guide came down only one point to 74.0% ◎. Stop there and it looks like the answer landed between my two branches, making the binary a miss.

But the call went further. Margins trough at 71–72% in the fourth quarter and settle around 72–73% in FY28 as the price increases take effect ○.

``` Reported 75.0% (GAAP and non-GAAP alike, held) Q3 guide 74.0% (±50bp, GAAP and non-GAAP alike) Q4 71–72% ← trough FY28 72–73% (after the increase lands) ```

71–72% is exactly the "low 70s" I named as the downside condition. The company guided to that range itself. So the margin call was not a partial hit but a hit — it simply arrived a quarter later than I was looking.

The lesson: I asked "does it hold or break," and the company answered with how many points, charged when. In a pass-through phase margin is a dial, not a switch. And the whole dial — 75 to 74 to 71–72 to 72–73 — was disclosed in one sitting. A binary cannot hold that.

Third, backlog quality. The preview said the standard is composition, not size: how much is binding contract and how much is a letter of intent. The question held; the answer did not take the shape I imagined.

Instead of splitting contracts from LOIs, the company answered with what it did with its own money. Supply commitments rose from $119 billion to $279 billion in a single quarter ◎, and the company wrote that this was "primarily related to the procurement of memory." $160 billion committed in ninety days, mostly to memory.

That is stronger evidence than an LOI breakdown, because prepayment is not recoverable. My downside candidate — that the backlog would prove to be a pile of intentions — is rejected in direction by that number.

The fourth baseline, the one the market wrote

This is the body of the piece. The person who wrote down the grading criteria got graded by the criteria the market actually used.

My scorecard had gaps in both directions.

The upside gap — guidance for roughly 70% revenue growth in FY2028 ○. None of my three upside candidates had the shape "guides the whole next-year growth rate." And the CFO framed that number not as a boast but as a constraint: customer forecasts point to growth doubling next year, while the guidance reflects supply limits. The company formally conceded that the ceiling on its own guidance is memory.

The downside gap — free cash flow halved, from $48.55 billion to $21.34 billion ◎. Inventory went from $25.8 billion to $31.6 billion ◎; receivables stood at $63.1 billion with 60 days sales outstanding ◎. The company issued $25.0 billion of senior unsecured notes ◎ and, in the same quarter, returned nearly $26.0 billion through buybacks and dividends ◎.

My scorecard had no cash-flow column at all.

But these are not two separate misses. They are two sides of one thing. The $279 billion commitment is the supply that underwrites 70% FY28 growth (upside) and the reason free cash flow halved (downside), at the same time. What I missed was not two line items but one layer — the balance sheet. I built the scorecard at the income-statement and guidance layer; the company answered at the balance-sheet layer.

And the market treated the two sides oppositely. In the first session it priced the upside gap immediately (down 1.59% in the regular session, then up 4.71% after hours ○) and has not yet priced the downside gap. In the last piece I wrote that there is a lag between the proposition that has arrived and the one that has not. That lag replayed itself in a single day. The bill comes later.

The cause is worth recording too. I framed the timing sandwich as an income-statement problem only. Yet the first-quarter CFO commentary already carried the warning — it discussed DSO, and it listed commitments at $119 billion. It was in my own archive and I did not go back for it.

Entering the lesson: a pass-through scorecard needs columns from all three statements. The cost of pass-through reaches working capital before it reaches margin.

Did "a beat is not information" survive?

The preview's premise was that beating the guide slightly is not information — information appears when a baseline moves.

The premise survived. The list of baselines did not.

Revenue came in at $96.2 billion against consensus near $92.2 billion, a beat of about 4%, with adjusted EPS of $2.22 ◎. But pull apart what actually moved the stock and it is not the beat. It is the $279 billion commitment, the FY28 70% guide, and the 75% held this quarter — all baseline-grade disclosures. The market behaved exactly as the preview's principle said it would. Two of those baselines simply were not on my sheet.

One discipline I will nail down: a single session's pattern is not evidence for a narrative. Warsh's Jackson Hole keynote has not happened yet (the 28th), and the rule I set in the last piece — not to settle an argument with a decline — applies symmetrically to a rally. The evidence for grading is filings and remarks, not the tape.

And one more honest note: the preview's opening did mention a positioning cushion ("expectations already marked down once"). It did not call a rebound. Observing a cushion and calling a direction are different things.

The axis of judgment has changed

The one-line verdict stands. The baselines held, and the bill arrived on the balance sheet.

But one refinement. This call did not close the question of whether the demand is real. $279 billion is Nvidia's conviction, not proof of end demand, and the question of whether OpenAI and Anthropic can pay the higher price is untouched.

The precise statement is this: Nvidia underwrote the demand question with its own balance sheet. The form of underwriting expanded from a guarantee (the $105 billion on the Ohio project) to working capital ($279 billion of prepayment and swelling receivables). I wrote in late July that Nvidia had guaranteed $265 billion of demand in seven days; this quarter's number is the August follow-through on that.

So the next question changes shape. Does the underwriter's cash hold? With prepayment going out above and financing coming in below, does the hub's cash conversion stay intact? The demand question did not disappear — it became a credit question about the underwriter.

For the next scorecard:

• The hub's cash conversion — how fast the $279 billion of commitments moves from inventory into revenue, and the quarterly free-cash-flow path
• FY28 margin — whether 72–73% holds once the new contract prices are fully reflected. That is the endpoint of the trajectory disclosed here, so that is the real test
• Carried over — the renewal-window rate, GPU rental spot prices, and the next funding terms for OpenAI and Anthropic

A note on differentiation. The consequences are booked differently by layer. If $279 billion of commitments is "primarily memory," the destination of that money is the three memory makers' ledgers — what comes out of Nvidia's margin sits down on theirs. A dial that turns as far as 71–72% is itself a back-calculation of how large the transfer is. For the structures financed against the backlog (neoclouds, SPV notes) the disclosure cuts both ways: FY28 70% thickens the collateral, while the admission that supply caps that growth pushes the collateral's timing out.

Grading by evidence type. In this round, what was fixed by filing (commitments, free cash flow, margin, China) and what exists only as spoken remarks (FY28 70% growth, the margin trajectory) do not carry the same weight. The FY28 guide is not in the CFO commentary; it is call commentary. Whether that remark hardens into a filing next quarter is itself a scoring item.

This piece is a grading record, not a recommendation in any direction.

Sources

Filings (NVIDIA Q2 FY2027, 2026-08-26)

• Revenue $96.2B (+106% YoY, +18% QoQ); Data Center $89.0B (+117%) · CFO commentary ◎
• Gross margin 75.0% — GAAP and non-GAAP alike · same document ◎
• Q3 outlook $108.0B ±2%; gross margin 74.0% ±50bp, GAAP and non-GAAP alike · same document ◎
• Supply commitments $119B → $279B, "primarily related to the procurement of memory" · same document ◎ · the commitments span the remainder of FY2027 through FY2029
• Free cash flow $48,554M → $21,341M (−56.0%) · same document ◎
• Inventory $25.8B → $31.6B · receivables $63.1B · DSO 60 days · same document ◎
• $25.0B of senior unsecured notes issued · nearly $26.0B returned via buybacks and dividends · same document ◎
• Net income non-GAAP $53,954M (year-ago $24,763M) · GAAP $59,688M · same document ◎ · the body uses the non-GAAP figure
• Net gains from equity securities $7.8B (prior quarter $15,936M) · same document ◎
• China Hopper shipments under 1% of Data Center revenue · same document ◎

Earnings call remarks (not filed — weighted separately)

• FY2028 revenue growth of about 70%, described as reflecting supply constraints · call remarks ○ · not present in the CFO commentary
• Margin trajectory: 71–72% trough in Q4, settling 72–73% in FY28 · call remarks ○
• Vera CPU in full production; CPU revenue expected to more than double in FY28 · call remarks ○ · this is the Vera CPU, not the Rubin GPU
• Spectrum-X Ethernet revenue up 2.6x year over year · call remarks ○
• "AI has reached its inflection point" · Jensen Huang, call remarks ○

Market

• Regular session close $209.66 (−1.59%) · daily close, internal price database ◎
• After hours +4.71% · reporting ○ · not a close

What this piece corrected from its own skeleton

• The margin verdict was raised from "partial hit" to hit, because the Q4 guide of 71–72% lands exactly on the preview's downside condition ("low 70s"). The skeleton looked only at the Q3 74% and concluded the low 70s had not broken; the call disclosed the rest of the trajectory.
• The skeleton's "Rubin in full production" was corrected to Vera CPU.
• The skeleton's "Data Center networking +138%" could not be confirmed and was replaced with Spectrum-X 2.6x.
• The skeleton's "roughly $99B remaining buyback authorization" could not be confirmed. What is confirmable is a new $80B authorization plus $38.5B remaining, about $118B in total; it is not used in the body.
• Consensus figures differ by provider, so the body says "near $92.2 billion."

Related

• The piece being graded: "The Grading Sheet, Written Before the Call" (2026-08-25)
• The two-ledger frame: "The Reached and the Unreached" (2026-08-19)
• The series: "July 22, Alphabet's Answer" (2026-07-20) → "The Answer of the Other Four" (2026-07-23) → this piece

※ ◎ = filing or primary data; ○ = call remarks or reporting. Filings and spoken remarks are deliberately marked apart — that distinction is one of the items being graded here.

This piece is for information only and is not investment advice. Figures were re-confirmed against primary sources before publication, and remark- or report-based figures are marked as such.