NVDA · NVIDIA Corporation · Stock research

Current Thesis

Prior coverage left one question open: whether July's multiple compression was a summer reset or the front end of a durable de-rating. Two things have since landed on it, and neither was public when the last note went out.

The first is a mechanism for the drawdown. Leopold Aschenbrenner's Situational Awareness fund fell from roughly $45 billion in assets to about $10 billion in a matter of days on margin calls, and was forced to unwind its entire leveraged public-equity book — AI-infrastructure longs paired against software shorts — selling to Citadel at a discount, with Bank of America, Goldman Sachs and JPMorgan working the margin (CNBC, 2026-07-30 and 2026-07-31). Reported leverage ran as high as 400%. That puts a large, dated, non-fundamental seller inside the exact window in which SOXX fell 21.2%. It does not settle the valuation argument. It does mean a measurable share of July's price action was a book being liquidated while the estimate base sat still.

The second is the shape of the reversal. SOXX gained 16.8% over the four sessions through 2026-08-04, its biggest four-day move since March 2020. SMH, SOXX and SOXL topped weekly fund flows for the first time. SOXX itself drew a record $6.9 billion of net inflows during July, the largest monthly inflow since the fund's inception, in the worst month it has had since December 2002 — money arrived while price fell.

NVDA closed $211.94 on 2026-08-04, up 2.56%, above both its 50-day average of $205.69 and its 200-day at $193.41, carrying a $5.13 trillion market cap at 21.22 times forward earnings (stockanalysis.com, 2026-08-04). The record close of $235.74 dates to 2026-05-14 and sits about 10% higher.

The demand headline that drove the session is unusually specific. On SpaceX's first quarterly call as a listed company, Elon Musk said the company has "decided to build exclusively on Nvidia" because the Vera Rubin architecture "is the best architecture," and guided SpaceX past two gigawatts of compute by year-end 2026, several times that by end-2027. SpaceX and NVIDIA separately announced a joint design for the Starmind AI1 satellite compute payload carrying Rubin GPUs and Vera CPUs, after SpaceX raised the satellite's peak power 67% to roughly 250 kW and its average power 33% to 160 kW — enough to run one Rubin NVL72 rack of 72 GPUs. Prototype testing is scheduled for early 2027.

The narrative leg on offer: a supply-constrained accelerator franchise absorbed a levered-fund liquidation, kept its guidance, added a named customer that publicly foreclosed the alternatives, and trades at a multiple set before the hyperscaler capex guides landed. Life-cycle is MATURING. The May record has not been challenged in eleven weeks; the late-July and early-August coverage set ran to Michael Burry's expanded puts (2026-07-31), Jim Chanos calling it "closer to a '99-type moment than a '97" (2026-07-31), debt-tsunami framing (2026-08-03) and a CUDA-moat challenge (2026-08-03) — argument about the trade, alongside genuinely new order flow. Well known, still working on fundamentals, flow running both directions.

Bullish and bearish views on NVIDIA Corporation

The model's bull view on NVIDIA Corporation (NVDA), in brief: 2026-08-04: Musk committed SpaceX to build "exclusively" on NVIDIA and guided the company past 2 GW of compute by year-end 2026, several times higher by end-2027. The bear view: Consensus sits above the guide. Q2 FY27 was guided to $91.0 billion ±2% at 75.0% non-GAAP gross margin; sell-side compilations run near $93.5 billion revenue and roughly $2.08 EPS. A print at the midpoint clears the company's bar and misses the market's. That guide assumes zero… Both cases follow in full.

Bull Case

• 2026-08-04: Musk committed SpaceX to build "exclusively" on NVIDIA and guided the company past 2 GW of compute by year-end 2026, several times higher by end-2027. A named buyer publicly closing the door on merchant and custom-silicon alternatives on a live earnings call is a different datapoint from a sell-side capex estimate.
• 2026-08-04: Starmind AI1 payload announced with Rubin GPUs and Vera CPUs; SpaceX raised peak satellite power 67% to ~250 kW to fit a Rubin NVL72 rack, with prototype testing in early 2027. Orbital compute is a demand surface that did not exist in any 2026 model.
• 2026-08-03: BofA's Vivek Arya put hyperscaler capital expenditure at $1.2 trillion over the next twelve months, carried a $350 objective against a $200.75 reference, and described the binding constraint as supply rather than appetite.
• Morgan Stanley put Amazon, Alphabet and Microsoft at roughly $1.4 trillion of combined capex with returns on invested capital of about 25–50%, and hyperscaler GPU rental at 25–40% ROIC — a return case for the spend that funds the order book.
• Valuation compressed without the estimate base moving: 21.22x forward on 2026-08-04 against a record close set 2026-05-14, and cheaper than the broader market on JPMorgan's framing while it keeps Overweight ratings across NVDA, AVGO and MRVL into late-August prints.
• 2026-08-04: the White House AI oversight framework was finalised as voluntary — labs get up to 30 days to submit models pre-release, with no public release of the framework itself. NVIDIA was in the room. A binding regime on the deployment layer that consumes accelerator capacity was a live overhang; it landed voluntary.
• Flow confirmed the bid rather than chasing it: record $6.9 billion into SOXX during a 21% down month, then semiconductor ETFs at the top of weekly fund flows for the first time (2026-08-04).

Bear Case

• Consensus sits above the guide. Q2 FY27 was guided to $91.0 billion ±2% at 75.0% non-GAAP gross margin; sell-side compilations run near $93.5 billion revenue and roughly $2.08 EPS. A print at the midpoint clears the company's bar and misses the market's.
• That guide assumes zero China Data Center compute revenue. H200 shipments to China have begun but volume has been described as minimal, and Blackwell remains barred from direct export.
• Credit has not confirmed the equity. NVIDIA's five-year CDS printed a record 82 basis points on 2026-07-27, the largest one-day widening since the contract began actively trading in November 2025; a 95bp quote for 2026-07-28 has also been reported and sources differ on the level. No published narrowing accompanies the four-day August rally, and NVIDIA has issued no statement on the reported up-to-$250 billion OpenAI backstop the spread is pricing.
• Vendor-adjacent financing keeps widening. NVIDIA participated in the 2026-08-04 Volta Infra round — $300 million equity plus $5 billion of financing at a $2.4 billion valuation, co-led by Andreessen Horowitz and Altimeter, with Michael Dell also in — and Volta simultaneously disclosed a $10 billion six-year compute contract with an unnamed AI developer. Every such structure adds a buyer the seller helped capitalise.
• High-grade AI debt supply reached $270 billion by early July, more than double all of 2025 (Morgan Stanley data). Apollo's Torsten Slok names hyperscaler issuance among three drivers of 30-year Treasury yields at a 20-year high.
• Insider supply runs one way. Fifteen insiders have filed Form 4 over the past eighteen months with no reported purchases.
• 2026-08-03: a one-year-old startup claimed it rebuilt a CUDA-equivalent software layer in ten hours. Unverified as production-grade, but it puts a number on how fast agent-written tooling could attack the switching cost.

Setup & Price Structure

• Close $211.94 on 2026-08-04, roughly 3.0% above the 50-day at $205.69 and 9.6% above the 200-day at $193.41. The move is four sessions old.
• 52-week range $164.07–$236.54; record close $235.74 on 2026-05-14. Current price is about 10.1% below that record, so the base built through late July has been reclaimed but the prior high is untested.
• The 200-day held throughout the drawdown. The worst close of the July unwind was $196.51 on 2026-07-27, when NVDA fell close to 5% on the backstop reports and shed roughly $250 billion of market value in a session.
• Crowding observables, stated as observed: record $6.9 billion July inflow into SOXX during its worst month since 2002; semiconductor ETFs top of weekly fund flows on 2026-08-04; a confirmed earnings date 21 calendar days out; scheduled 10b5-1 insider dispositions with no offsetting purchases; short interest of 324.05 million shares, 1.34% of 24.22 billion shares outstanding, which supplies no squeeze fuel in either direction.
• The distance above the rising 50-day is small. The extension risk in this setup sits in the four-day sector move and the print, rather than in price stretched from its average.

Catalyst Calendar (next 30 days)

• ~2026-08-10 (est.) — TSMC July 2026 monthly revenue release. June revenue came 2026-07-13 and May on 2026-06-10, so the cadence sits near the 10th. Highest-frequency public read on AI accelerator wafer volumes between NVIDIA prints.
• 2026-08-11 — CoreWeave Q2 2026 results, 5:00 p.m. ET. Company guidance is $2.45–2.6 billion against an LSEG consensus near $2.69 billion; Q1 ended with a $99.4 billion revenue backlog and about 3.5 gigawatts of contracted power. The largest listed GPU tenant, and one of the names liquidated out of the Situational Awareness book.
• 2026-08-14 — Q2 2026 13F filing deadline. Shows whether July's unwind was concentrated in the levered vehicles that failed or broad across institutional holders.
• ~2026-08 (est.) — definitive terms, restructuring or withdrawal of the reported up-to-$250 billion OpenAI backstop. A non-recourse or smaller structure narrows the credit-versus-equity gap; a confirmed guarantee validates the circular-financing critique.
• 2026-08-26 — Q2 FY27 results for the quarter ended 2026-07-26, after the close, call 2:00 p.m. PT. Tests $91.0 billion ±2% and 75.0% gross margin against a consensus above the midpoint.
• 2026-08-26 — Q3 FY27 guidance and its China Data Center compute assumption. The current outlook carries zero; any figure entering the guide is revenue not in the price today.

Elapsed catalysts

• Elapsed: the White House AI framework review took place 2026-08-04 and resolved voluntary, removing that specific overhang. (passed 5d ago)

What Would Change Our Mind

The 200-day average is the line that never broke. NVDA's worst close of the entire July unwind was $196.51 on 2026-07-27, still above the average, which is why the liquidation reading survives at all. Lose it on a closing weekly basis and the July drawdown stops looking like a forced seller and starts looking like a repricing that simply took a pause: a weekly close below $193 ends that reading. Under it, the record CDS print of 2026-07-27 becomes the leading indicator rather than the noisy one.

Second, the print. Revenue landing at or below the $91.0 billion guide midpoint on 2026-08-26, with Q3 guidance again assuming zero China Data Center compute, would mean the multiple compression of July was a forecast rather than an overreaction, regardless of where price sits that week.

Third, the label. A close above the 2026-05-14 record of $235.74 on the back of the August print would argue the cycle has re-accelerated rather than matured, and would require this note to be re-dated. A new high that fails back inside two weeks would argue the opposite — mainstream coverage without a marginal buyer behind it.

Correlation Notes

• NVDA moves as one factor with TSM, MU, AMD, AVGO and MRVL. Holding three or more of these is one concentrated AI-compute exposure wearing five tickers.
• SOXX, SMH and SOXL are the flow vehicles: SOXX -21.2% in July, +16.8% over the four sessions through 2026-08-04, with a record $6.9 billion July net inflow. Sector-ETF flow now sets the marginal price on many sessions.
• CRWV is the levered read-through — a debt-funded tenant whose 2026-08-11 print lands fifteen days before NVIDIA's, and a name that was inside the liquidated Situational Awareness book.
• SPCX became a live correlate on 2026-08-04. SpaceX's compute build guides SpaceX capex and NVIDIA revenue off the same number, and the two now trade against a shared headline.
• MSFT, AMZN, META and GOOGL capex guidance is the demand input; Morgan Stanley's ~$1.4 trillion combined figure and BofA's $1.2 trillion twelve-month hyperscaler number are the estimates in circulation.
• Long-end Treasury yields are a second-order channel: $270 billion of high-grade AI debt supply by early July, more than double all of 2025, is one of three drivers Apollo names behind 30-year yields at a 20-year high. Financing cost feeds back into the capex the order book depends on.
• RSP versus SPY is the breadth check. A $5.13 trillion index weight is bid by passive flows; sustained equal-weight leadership removes that support.