Nvidia insiders sold $410M five days after a record quarter — and Burry already filed the AI short

The Form 4 (SEC accession0001199039-26-000012) covers a single transaction date of Aug. 31, 2026, with the filing accepted on Sep. 2 at 6:10 p.m. ET. The reporting person is Mark A. Stevens, a director and one of NVIDIA's earliest outside backers. The 1,848,501 shares moved through trusts at $220–$226 per share, leaving him with 29,921,132 shares still under direct and indirect ownership.

The Aug. 31 trade was not an isolated event. Across calendar 2026, Stevens has now filed four sales totaling more than 4M shares — a cumulative ~$856M of insider supply at progressively higher prices. The Aug. 31 filing is the largest by share count and by proceeds, and the first one executed after NVIDIA's record Q2 FY27 print.

What changed is not direction but pace. The June tranche averaged ~$215 across two filings; the August tranche cleared at $222. The acceleration coincides with the period in which NVIDIA stock traded within roughly 4% of its 52-week high and the data-center segment became more than 92% of revenue. Selling into demand is a routine insider move. Selling the equivalent of nine months of volume in three days, into a quarter that beat by $4.2B, is a different gesture.

Michael Burry's last public 13F (Q3 2025, period ending Sept. 30, 2025, filed mid-November 2025) placed the AI basket at the center of his fund. The filing showed put options on 5,000,000 shares of Palantir at a notional $912.1M and put options on 1,000,000 shares of NVIDIA at $186.6M — together roughly 80% of reported portfolio assets. Burry described the position as 'the most concentrated way to express a bearish view on the artificial intelligence trade.'

The point of pulling Burry's filing into an article about an NVIDIA insider sale is not to call the top. It is to note that the AI bubble thesis was already a public, sized position ten months before Stevens filed Aug. 31. The insider flow is now arriving at the same moment the consensus short was originally written against — which is what makes the timing unusual.

On Aug. 22, 2026, Bloomberg reported that NVIDIA has notified customers of price increases 'above 15% in many cases' on systems containing its AI chips, with the hikes effective on systems shipped in early 2027. The increases span the flagship Vera Rubin and Grace Blackwell platforms, and they depend on chip generation and memory configuration — meaning HBM-heavy configurations face the largest hikes.

Memory is now the bottleneck in the AI server stack, not silicon. TrendForce estimates memory has climbed to ~25% of the cost of a high-end AI rack, with GPU silicon's share falling from >80% in prior generations to ~50% in next-generation builds. NVIDIA can pass through that inflation for now — Vera Rubin is sold out, hyperscalers are placing forward orders, and alternative architectures do not yet exist at scale. That pricing power, however, is paired with something the company does not advertise: a gross-margin guide that steps down before recovering.

NVIDIA's 15%+ hike is not an NVIDIA profit event — it is a pass-through. The cost of the rack is rising in HBM, DRAM and NAND, and the price NVIDIA charges is rising to reflect that. The dollar flows therefore land in different places than the headline suggests.

The split explains why NVIDIA's gross margin is guided down even as revenue is guided up. If HBM is now ~25% of a rack's cost and rising double-digit percent Q/Q, the cost-of-revenue line moves faster than the price NVIDIA can charge on systems already under contract. The hike applies to early-2027 shipments; Vera Rubin systems shipping through the rest of 2026 carry the existing cost structure but face the forward guide. The result is a one-to-two-quarter margin trough, not a structural reset — but the trough is the period Burry's puts were written against.

The other half of the bear case is not supply but demand. Hyperscaler capex — the spending that flows into NVIDIA, TSMC, SK Hynix, Micron and the rest of the chain — is now estimated at $600B–$860B for calendar 2026. Goldman Sachs noted in late 2025 that AI capex would need to reach $700B in 2026 to match the peak of late-1990s telecom investment — the analog Burry's generation uses for cycle tops.

Even at the lower end of the estimate range, 2026 capex is more than double the 2024 baseline. The price hike is therefore not opportunistic — it is the visible sign that demand is colliding with a memory supply curve that does not get easier until 2029. Whether the demand persists at $700B+ into 2027 is the open question. NVIDIA's contract-manufacturer order book through2027 is reportedly full, which makes the 2027 number about how much hyperscaler revenue can support that level of spend — not about willingness.

Read individually, none of these data points is unusual. Stevens has trimmed every quarter this year. Burry's AI short is old news. NVIDIA's gross-margin step-down was telegraphed at the print. The 15% price hike is consistent with a memory-tight market.

Read together, they describe the same moment from three directions:

1. Insiders are accelerating sales into a quarter that beat by $4.2B, with a single filing (~$410.8M) larger than the previous four quarters of NVIDIA insider activity combined outside of routine 10b5-1 plans.
2. The company is signaling it can no longer fully absorb HBM cost inflation — pricing is up more than 15% on next-generation systems, and gross margin is guided down to ~71% before recovering on Vera Rubin mix.
3. A public bear thesis sized at ~$1.1B was filed against the AI basket ten months ago, written to mature into this peak.

The alignment is not proof of an imminent correction. The forward guide was still above consensus, the backlog is full, and NVIDIA's data-center segment grew 117% YoY. It is, however, a clear signal that the phase of the cycle has changed — and that the people closest to the order book are now positioning for the back half of the build-out rather than the next leg up.