The Grading Sheet, Written Before the Call

In the last piece I described the current price as a verdict deferred. The proposition that has arrived — $3 trillion of off-balance-sheet commitments and their funding cost — is being priced. The one that has not — renewal repricing — has not landed. The first material for that verdict arrives on the 26th.

The calendar adds weight to this call. Nvidia's results on the evening of the 26th, and Warsh's first Jackson Hole keynote as chair on the 28th — the market's biggest earnings event and its biggest Fed event sit roughly 40 hours apart. Whichever moves first, the other becomes an amplifier.

Where the stock sits as I write this is worth recording too. It meets the print after falling for seven straight sessions ◎ — the longest run since 2022. From $225.30 on August 13 to $208.48 on August 24, down 7.5%. Three days before the print, a Bloomberg report landed: Nvidia has told major customers that prices for AI server systems are going up more than 15% (up to 17% on racks) ○, that Vera Rubin and Grace Blackwell systems are among those affected, and that it applies to systems shipped early next year. The driver is the surge in memory prices. So the call opens with expectations already marked down once, and a fresh question on the table.

The grading sheet — a beat is not information

Start with the baseline. What the market already knows: revenue guidance of $91 billion ±2% ◎, gross margin of 74.9% GAAP and 75.0% non-GAAP ◎, and China data center revenue at zero — the company stated it assumed no China data center compute revenue in the outlook at all.

So here is one grading rule. Beating the guide slightly is not information. Nvidia has beaten its own guide as a matter of habit, and the market prices that in. Information appears when the baseline moves. Three baselines are in play on this call.

First, does China move off zero. From mid-August there were reports that H200 chips had begun reaching Chinese customers (ByteDance and Tencent receiving about 10,000 each ○). Yet the company's official baseline is still zero. If any China revenue enters the guidance, that is a turn from zero to positive, which travels far. But Beijing's sales review and the 25% remittance structure remain, so the likelier outcome is "hold at zero, mention the option." The shock comes when they break the expectation and actually put it in.

Second, does the 75% margin hold. This was the only downside candidate with a clean number attached, and the price-hike report three days out has made it the whole market's question. The structure of the pressure is this — conventional DRAM contract prices rose 58–63% quarter over quarter in Q2 ○, and server DRAM is projected to rise another 13–18% in Q3 ○. As for how much of a Vera Rubin system's cost is memory, honestly, the estimates diverge — one puts memory at 62% ○, others at about 29% of system BOM ○ or roughly 26% on a rack basis ○. That divergence is itself a grading point. If the share is in the 60s, a 15–17% hike is not pass-through but partial pass-through, and the rest comes out of Nvidia's margin. If it is in the 20s, the same hike covers nearly all of it. Which means this call's margin guide also works backwards, telling us which estimate was right. And there is a lag. The increase applies to systems shipped early next year while memory costs are rising now, so the quarters in between sell at old prices and pay new costs. This call's next-quarter margin guide sits on that sandwich. If the margin guide comes down into the low 70s, the market will look at margin no matter how much revenue beats. Confirm 75% on top of this cost curve, though, and you have the supplier-side version of the repricing argument — depreciation is fixed, price decides profit.

Third, the quality of the backlog. The company has spoken of bookings visibility on the order of $1 trillion through 2027 ○. But the grading standard for that number is composition, not size — how much is binding contract and how much is a letter of intent. There is precedent worth holding: Nvidia's up-to-$100 billion with OpenAI was a non-binding LOI as of last December, and the Ohio data center that was reported in July as a roughly $250 billion backstop was settled in an August 17 filing as a guarantee of up to $105 billion ○. The gap between the report and the confirmed figure was $145 billion. Whether Q&A produces a concrete answer on composition matters more than restating the size.

Upside and downside are nearly symmetric — and that symmetry is the price

If there is upside: China revenue returning (above), a stated timing for first Rubin revenue recognition (full production ramp is already confirmed, so declaring recognition has begun means the four-architecture cadence is arriving as actual revenue), a raised backlog with the binding portion disclosed.

If there is downside: a lowered margin guide (above), another China inventory write-down (a rerun of last year's H20 charge of $4.5 billion ◎ — the initial expectation was $5.5 billion, and about $1 billion was salvaged through material reuse), Q&A that reveals the backlog as a pile of LOIs, and a defensive tone on circular-financing questions (the $105 billion guarantee on the OpenAI Ohio project, and the structure where places Nvidia invested in turn buy Nvidia GPUs, will come up).

One observation to record here. Most of the downside candidates are not numbers in the prepared remarks but the tone of the Q&A. Margin is the only downside that gets fixed as a number; the rest take their direction after the fact, from how things are said. That upside and downside stand this symmetrically is exactly why the current price is a deferred verdict. The market is not failing to pick a direction — the material for picking one is still inside this call.

And the price-hike report planted a fresh mine in the Q&A — the question "doesn't demand break at that price" will certainly come. The answer cuts both ways. If Jensen Huang backs "demand still overwhelms supply after the increase" with backlog numbers, the bad news of three days ago flips wholesale into evidence of pricing power — the increase raises revenue per unit. If he takes it rhetorically with no numbers, the story hardens into a hike forced by costs. Because the report came before the call, the framing advantage sits with the company — whether they use it is the thing to watch.

This call tests both ledgers at once

Translated into the last piece's frame: if the binding portion of the backlog, the margin, and the Rubin revenue timeline are confirmed — the proposition that has not arrived (revenue) is beginning to. If a China write-down appears or the backlog proves to be a pile of intentions — the proposition that has arrived (cost and risk) gets heavier. Both ledgers go on the stand inside one call.

The worst combination is worth writing down: a lowered margin guide together with damage to backlog quality. That would mean both that revenue-side repricing does not work at the supplier level and that the demand meant to carry $3 trillion of commitments was intention rather than contract — and Jackson Hole two days later becomes the amplifier. The best combination is the reverse.

On the record

This piece is published before the evening of the 26th. On the 27th I will grade the criteria above against what was actually said (filings and transcript), and write down the hits and the misses as they fall. When grading I will separate the types of evidence — what was fixed by company filing, what existed only as reporting, and what was my own judgment — because only then does it show where the error was. Same method as the last two (the Alphabet preview, the other-four observation): the point is not to be right, but to have the direction I read checked against the result.

A note on differentiation. The consequences of this call are booked differently by layer. If margin defence is confirmed, the supply layer's (the three memory makers') pricing power is reconfirmed; if margin is compressed, the cause of that compression — the HBM4 increase — is already sitting on the memory side's ledger. Either way the invoice for scarcity has been issued. If backlog quality is damaged, the hit lands harder on the structures that financed against that backlog (neoclouds, SPV notes) than on Nvidia. This piece is a record of grading criteria, not a recommendation in any direction.

Sources

Baseline (company filings)

• Q2 FY2027 guidance: revenue $91B ±2%, gross margin 74.9% GAAP / 75.0% non-GAAP · Nvidia Q1 FY2027 results (2026-05-20) and CFO commentary ◎ · SEC 8-K
• Outlook assumes no China data center compute revenue · same filing ◎
• H20 inventory charge: $5.5B expected at announcement, $4.5B actually recognised after material reuse · company filing and multiple reports ◎

Share price

• Seven consecutive down sessions, 2026-08-14 to 08-24, $225.30 → $208.48 (−7.5%) · daily closes, internal price database ◎ · the first run of seven or more sessions since 2022-01-01; the prior longest was six (January, April and August 2022)

Price increase reporting

• Customers notified of AI server price increases above 15%; Vera Rubin and Grace Blackwell included; effective on systems shipped early next year · Bloomberg 2026-08-22 ○
• Up to 17% on rack products (Grace Blackwell 300, Vera Rubin 200) · Seoul Economic Daily ○ · Bloomberg's own wording stops at "more than 15%"; the 17% figure comes from Korean follow-up reporting

Memory cost

• Conventional DRAM contract prices projected +58–63% QoQ in 2Q26 · TrendForce 2026-03-31 ○ · no server-DRAM-only 2Q26 figure appears in the public releases, so the conventional DRAM basis is used
• Server DRAM projected +13–18% QoQ in 3Q26 · TrendForce 2026-07-09 ○
• Vera Rubin memory share of cost — estimates diverge: 62% (Digitimes 2026-08-11) / about 29% of system BOM (TrendForce 2026-07-28) / about 26% on an NVL72 rack basis (memory ~$2M of a ~$7.8M total) ○ · the report giving 62% does not state the unit that share is measured against. That is why the body records the divergence rather than picking one

Backlog

• Over $1 trillion of Blackwell plus Vera Rubin visibility through 2027 · Jensen Huang, GTC 2026 keynote (2026-03-16) ○
• OpenAI Ohio data center: reported in July as a ~$250B backstop under discussion → settled in an August 17 filing as a guarantee of up to $105B · WSJ and CNBC reporting, and coverage of the filing ○ · the $105B is a guarantee against lease and power obligations, not a chip purchase contract
• Nvidia–OpenAI up to $100B was a non-binding LOI as of December 2025 (intent to deploy 10GW+) ○

China

• ByteDance and Tencent each received about 10,000 H200 units · FT reporting (2026-08) ○ · the licensed ceiling is 75,000 per company, so the received volume is about 13% of it

Calendar

• Nvidia Q2 FY2027 results 2026-08-26 after the close / Jackson Hole symposium Aug 27–29, Warsh's first keynote the morning of Aug 28 · company calendar; Kansas City Fed ◎

Adjusted from the draft

• The draft's "six consecutive down sessions" was updated to seven using prices as of publication.
• The draft's "server DRAM +53–58% in Q2" could not be confirmed in public sources; it was replaced with the confirmable conventional DRAM +58–63% for Q2 and server DRAM +13–18% for Q3.
• The draft's "reported at $600B in July" could not be confirmed. The confirmable reported figure is $250B, and the August 17 number is a guarantee, not a signed purchase contract.
• The draft's "36 hours apart" was adjusted to roughly 40 hours against the actual schedule.

Related

• The two-ledger frame: "The Reached and the Unreached" (Silicon Nexus, 2026-08-19)
• The on-the-record series: "July 22, Alphabet's Answer" (2026-07-20) → "The Answer of the Other Four" (2026-07-23) → this piece

※ ◎ = filing or primary data; ○ = reporting or estimate. Where estimates diverge, the divergence is recorded as such. Prices and schedules are as of publication and will move.

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