Reading the Culper Research Short Report on NVIDIA: A Counter-Cyclical Bet on the World's Largest Market Cap

On 13 May 2026, Culper Research published NVIDIA (NVDA): The China Problem, a short report against NVIDIA Corporation (NASDAQ: NVDA), the largest listed company in the world by market capitalisation. Culper's headline contention is uncomfortable in its specificity: NVIDIA's chief executive has repeatedly told investors that the company's compute business in China went to "zero" after the April 2025 U.S. trade restrictions, and Culper believes that in reality more than 20% of NVIDIA's FY2026 compute revenues remained driven by Chinese demand — routed through Southeast Asian intermediaries and outright diversion. NVIDIA shares were up roughly 2.7% on the day of publication, closing at a fresh 52-week high. We read the report against the March 2026 Department of Justice indictment of three individuals tied to Supermicro Computer, contemporaneous reporting from the New York Times, Bloomberg, and the Wall Street Journal, the public corporate filings cited by Culper in Singapore and Malaysia, and NVIDIA's own statements on its earnings calls. This post lays out what Culper alleges, where each layer rests in disclosed evidence, why the position is structured as a counter-cyclical macro bet against a name trading at a 46x P/E and a new high, and what the proximate near-term tests look like.

Short reports are not neutral documents. Culper discloses a short position in NVIDIA and stands to profit if the share price falls. That commercial interest is the right starting frame: the report is directional, but the load-bearing inputs are not the author's. The Supermicro indictment is a federal case. Megaspeed's $2.9 billion balance-sheet expansion sits in its Singapore filings. Speedmatrix's $4.6 billion in 2024–2026 imports sits in trade data — and roughly $4.0 billion of that originated from Aivres Systems, a U.S.-headquartered NVIDIA Elite OEM partner that is one-third owned, through its parent, by an entity-listed Chinese state vehicle. NVIDIA's revenue-recognition policy — that compute revenue is reported by the location of the customer's headquarters, not the chip's end use — is in NVIDIA's own 10-K. We treat the directional framing of the report as a given, and weigh each layer underneath it on its own evidentiary footing.

Overview

1. NVIDIA's "zero China" claim and how revenue is reported
2. The Supermicro indictment, OBON Corp, and the "tip of the iceberg"
3. Megaspeed, Speedmatrix, and the alleged Alibaba financing layer
4. Aivres: a California-listed OEM functioning as a China-dedicated channel
5. Novagate Cloud and the Gigabyte/Giga corridor
6. The macro: a counter-cyclical short on the largest market cap on the planet
7. What the short does not claim, and the bull rebuttals
8. What is yet to be heard
9. Where we sit

NVIDIA's "zero China" claim and how revenue is reported

In April 2025, the U.S. tightened export controls to effectively ban all NVIDIA GPU sales to China, including sales to "companies headquartered or with an ultimate parent therein." NVIDIA took a $5.5 billion write-down on its H20 inventory and disclosed an approximately $8 billion revenue loss for FY Q2 2026. Since then, chief executive Jensen Huang has repeatedly told investors that NVIDIA's China compute share is now zero. At a Citadel Securities conference in October 2025, he stated that "at the moment, we are 100% out of China… we went from 95% market share to 0%." On the recent Dwarkesh Patel podcast, he said that in China NVIDIA had "now dropped to zero." On the May 2025 earnings call, he framed the prior addressable market as a "$50 billion China market" that was "effectively closed to U.S. industry."

Culper's argument turns on two structural features of how NVIDIA reports its revenue. First, NVIDIA's customers of record are OEMs and channel partners — Supermicro, Dell, Aivres, Gigabyte, and a small set of others — not the data-centre operators or end users that ultimately install the chips. Second, NVIDIA's geographic revenue split is reported by the headquarters location of the customer of record, not the location of the eventual end customer. NVIDIA's 10-K is explicit on this: "the location of our customers' headquarters provides a better representation of the geographic profile of our revenue." Both of those design choices are defensible at the operating-policy level. They also mean that a chip sold by NVIDIA to a California-headquartered OEM, then shipped by that OEM to a Malaysian operator that is, on the corporate-filings record, financed by a Chinese parent, is reported as a U.S. sale.

The base-rate question is therefore not "did NVIDIA ship chips into China after April 2025." It is "what fraction of NVIDIA's FY2026 compute revenue, if you traced each shipment to the entity that ultimately controls and uses it, is tied to Chinese demand." Culper estimates more than 20% on a bottom-up basis. Sell-side analyst Gil Luria of DA Davidson estimated in mid-2025 that "at least 25%" of NVIDIA chips and "as much as 40%" of revenues ultimately reflect Chinese demand. Wells Fargo flagged in March 2026 that NVIDIA had previously noted that "China accounted for 20–25% of total Data Center revenue." A former NVIDIA director-level employee Culper cites estimated that "20% to 40%" of NVIDIA's data-centre business remained functionally Chinese after the April 2025 restrictions took effect. None of those figures is a verdict; they are a range, and the range is materially wider than the public "zero" framing.

The Supermicro indictment, OBON Corp, and the "tip of the iceberg"

In March 2026, the Department of Justice charged three individuals tied to Supermicro Computer — including a Supermicro co-founder — with conspiring to smuggle at least $2.5 billion of NVIDIA-powered servers to China between 2024 and 2025. The indictment describes the conspirators as routing the chips through a Southeast Asian pass-through entity identified only as "Company-1" to give "the appearance of legitimate commercial activity and to obscure their China-based end customers." A specific operational detail in the indictment is the use of staged "dummy" servers — physically present at the nominal destination data centre — while the real servers had already been diverted to China. Days before Culper's report, Bloomberg identified Company-1 as Thailand-based OBON Corp (One Belt One Network).

NVIDIA investors have, on the trading evidence, treated the indictment as inconsequential — $2.5 billion smuggled over two years against trailing-twelve-month revenue measured in the high tens of billions. Culper's reading is that the indictment is the corner of a much larger picture. Two former NVIDIA insiders Culper cites describe the alleged scheme in similar terms: a "complex and far-flung operation involving multiple additional NVIDIA OEMs, partners, and intermediaries," and a former director-level employee paraphrasing the position as "Megaspeed is just the tip of the iceberg." Culper reads OBON, Siam AI Corporation, Megaspeed, Speedmatrix, Novagate Cloud, Aivres, and Giga Computing as nodes in the same broad network — and the Supermicro–OBON corridor as one of several concurrent flows.

The OBON–Siam connection itself is documented in corporate filings. Siam AI is an NVIDIA cloud-compute partner — reportedly NVIDIA's first, and currently its only, listed cloud partner in Thailand. Siam AI's Singapore entity, Siam AI Corporation Pte Ltd, was incorporated in October 2024 with overlapping officer disclosures: Luan Jingwei, Chew Kah Leong, Ratanaphon Wongnapachant, and Hing Shun Thomas Tsang. Siam AI's CEO Ratanaphon Wongnapachant told Bloomberg days before the Culper report that he is no longer involved at OBON; Culper notes that OBON's Hong Kong affiliate, One Belt One Network Holdings Limited, names Hing Shun Thomas Tsang as an officer — Tsang is the son of Donald Tsang, the former Chief Executive of Hong Kong. The structural overlap is enough, on Culper's reading, to put the "single corridor" framing of the indictment in question.

Megaspeed, Speedmatrix, and the alleged Alibaba financing layer

The single largest piece of the report, by both dollars implicated and analytical novelty, is the Megaspeed structure. Megaspeed International Pte Ltd — based in Singapore — is, on Culper's reading, NVIDIA's single largest chip buyer in Southeast Asia. It became an NVIDIA partner in November 2023, two weeks after its parent Swiftdata Pte Ltd was incorporated. Megaspeed itself was spun out in March 2023 from 7Road, a Chinese gaming and cloud-computing company that Culper describes as having "documented ties to Chinese state-backed investors." NVIDIA has previously vouched for Megaspeed publicly: that it is "wholly owned and operated by a company based and headquartered outside China, with no China shareholders." Culper accepts the precision of that phrasing — there is no Chinese equity in Megaspeed — and draws attention to what it does not preclude: Chinese financing.

The arithmetic of the alleged financing flow is the simplest part of the report and rests on filings rather than testimony. At year-end 2023, Megaspeed's total liabilities and equity stood at $33 million. By year-end 2024, that figure was $3.0 billion. The expansion was driven almost entirely by $2.9 billion in "refundable deposits received" from an unnamed source on the liabilities side, matched by $2.9 billion in "amounts due from subsidiaries" on the asset side — money Megaspeed had passed through to its operating subsidiaries, the largest of which is Malaysia-incorporated Speedmatrix Sdn Bhd. Speedmatrix's own 2024 annual report, filed in Malaysia, shows it owed 7.9 billion Malaysian ringgit (approximately $1.6 billion) to Megaspeed at year-end and added 8.5 billion ringgit (approximately $1.9 billion) of property, plant and equipment in the year — disclosed almost entirely as server equipment. The accounting is consistent with money flowing from an unnamed depositor into Megaspeed, through to Speedmatrix, and out into NVIDIA-powered server purchases.

Culper's claim about who the depositor is rests on a separate corporate-filings chain, in Singapore. In June 2024, Speedmatrix filed a charge — a debenture over its present and future IT and network equipment, intellectual property, related documents, and insurances — in favour of Apex Enterprise Solutions (Singapore) Private Limited. Culper traces Apex back, through its own filings, to Alibaba and to additional loans from two Alibaba-related companies. Apex's annual report for the year ended 31 March 2025 also discloses that, subsequent to that date, Apex paid more than $4.2 billion "for services to be performed." The implication Culper draws is that the unnamed "refundable deposits" inside Megaspeed — and the bulk of the funding for Speedmatrix's $1.9 billion 2024 server build — were in substance Alibaba money. NVIDIA's earlier representation about Megaspeed's equity ownership remains literally accurate; the financing question is one the company has not, on Culper's reading, addressed.

A further operational data point sits with Bain's Bridge Data Centres (BDC). In April 2026, BDC reportedly dropped Megaspeed as a tenant. BDC declined to publicly cite the reason. One source Culper cites attributed the eviction to BDC having found Megaspeed using the data centre as "a place for Megaspeed to install their dummy servers" and "divert the hardware" — an operational pattern that maps precisely onto the dummy-server detail in the March 2026 Supermicro indictment.

Aivres: a California-listed OEM functioning as a China-dedicated channel

The Speedmatrix import flow is the part of the case where Culper's bottom-up trade-data work expands the public record. The New York Times had previously reported that Speedmatrix imported "nearly $2 billion" in servers between June 2024 and June 2025. Culper's review of trade data puts the figure, between December 2024 and January 2026, at $4.6 billion. Of that, $4.0 billion originated from Aivres Systems, Inc., an NVIDIA Elite OEM compute partner headquartered in California. Aivres is the rebranded successor of Inspur Systems Inc.; the Inspur name was retired in May 2023 after the parent, China-based Inspur Group, was placed on the U.S. Entity List. Aivres today is wholly owned by Inspur Electronic, which is approximately one-third owned by Inspur Group — a relationship that, on Culper's reading, allows Aivres to operate as a U.S.-headquartered counterparty for NVIDIA accounting purposes while in effect functioning as a China-dedicated supply channel just under the Entity List restriction thresholds.

Aivres's documented activity extends beyond Speedmatrix. Culper details further Aivres-supplied flows through Indosat to Shanghai's INF Tech, whose founder is a former Alibaba vice-president. In March 2026, the Wall Street Journal reported that Aivres would supply $2.5 billion in Blackwell chips to a Malaysian operator that would lease the servers to ByteDance. Taken together, the picture Culper draws is one of an OEM whose principal export footprint is into operators with disclosed or suspected Chinese end-customer ties, with chip flow scaled in single billions and concentrated in 2024 and 2025.

The argument is not that Aivres is operating illegally — Culper makes no such claim — nor that NVIDIA's accounting policy is inconsistent with U.S. revenue-recognition rules. It is that the policy itself, applied to a partner whose business is structurally tilted toward Chinese demand, generates a reporting outcome under which more than $4 billion in chip shipments through one OEM into one Malaysian operator are recognised as U.S. revenue. The investor-facing line "we went to zero in China" sits inside that reporting frame.

Novagate Cloud and the Gigabyte/Giga corridor

Novagate Cloud is the second pillar of the alleged Megaspeed network. Novagate Cloud Sdn Bhd was incorporated in Malaysia on 11 April 2025 — two days after the U.S. trade restrictions of April 2025 were announced. A separate Singapore entity, Novagate Cloud Pte Ltd, was registered on 2 October 2025 with $100,000 of capital. One week later, the New York Times and CNBC reported that U.S. and Singapore authorities were investigating Megaspeed for "allegedly helping Chinese companies evade curbs on U.S. chip exports," and that NVIDIA had "engaged with the U.S. government" on the matter. Culper reads the timing of Novagate's two incorporations — bracketing the U.S. restriction announcement and the public disclosure of the Megaspeed investigation — as consistent with a successor entity stood up to keep the diversion pipeline functional under closer scrutiny.

The chip-flow evidence Culper assembles around Novagate is similar in shape to the Aivres–Speedmatrix flow but smaller in dollar terms and earlier in its trajectory. Between December 2024 and the cut-off of the trade data Culper reviewed, Giga Computing — the server subsidiary of Taiwan-listed Gigabyte Technology Co (2376 TW) — exported approximately $511 million of equipment, primarily to two Malaysian recipients: $344 million to Speedmatrix (of which $312 million is Giga-branded servers containing NVIDIA's restricted H200 chips) and $142 million to Novagate Cloud (of which $140 million is Giga's "G894-SD3-AAX7" servers containing NVIDIA's Blackwell Ultra B300 and H200 chips). In January and February 2026, Novagate received a further $120 million in computing equipment and components beyond the $140 million Blackwell shipment.

Reuters had previously reported, in April 2024, that China may have illegally obtained NVIDIA GPUs from Supermicro, Dell, and Gigabyte. Culper's Giga shipment data and Novagate's incorporation timeline are consistent with that earlier reporting and incrementally expand it: the alleged diversion pattern, on this reading, is OEM-portfolio-wide rather than vendor-specific, and survives the formation of new pass-through entities.

The macro: a counter-cyclical short on the largest market cap on the planet

The framing within which all of this lands matters. NVIDIA holds the largest market capitalisation on the planet. Its shares traded at $226.65 on the day of Culper's report — a fresh 52-week high — and the stock continued higher into the close. The trailing P/E is approximately 46x. Insider sales over the prior three months total approximately $164 million. The shareholder base is dominated by index, growth, and AI-thematic flows; sell-side coverage is dominated by Buy ratings; the consensus China narrative is, on Culper's characterisation, that improved trade relations represent upside for the FY2027 numbers rather than a tail risk to FY2026.

Against that framing, the position Culper is taking is deliberately counter-cyclical in two distinct dimensions. The first is the direction of the trade itself. Going short the largest-capitalisation listed company in the world, at a fresh high, with a positive consensus, on a thesis that depends on bottom-up corporate-filings work in two foreign jurisdictions and on trade-data triangulation, is a position that is structurally hard to hold. It requires the analyst to be right both on the underlying claim and on the time path through which that claim becomes consensus. Short campaigns of this kind have historically taken months — sometimes more than a year — to land in the share price of mega-cap names with multiple offsetting positive narratives.

The second dimension is the underlying macro argument. Culper's view is not that NVIDIA's China business will go to zero in the future — it is that NVIDIA's China business already substantially exists, is being financed and routed through Southeast Asian intermediaries that the company has at minimum publicly vouched for, and is now actually contracting at the same time as Beijing is moving aggressively into domestic AI alternatives. The combination is what makes the position counter-cyclical rather than merely contrarian: the bull case requires China to come back as upside; the bear case requires China to be re-rated as a downside that is already partially in the run-rate. Both views can be wrong. The asymmetry is in the gap between them — and the gap, on Culper's bottom-up estimate, is more than $30 billion of FY2027 revenue exposure that the consensus is not pricing.

It is worth being explicit about the historical base rate here. Single-name short reports against mega-cap technology companies have, on average, generated relatively small immediate price reactions — multiples below the moves that comparable reports against small- and mid-cap names produce. The Supermicro charges, individually, did not move NVIDIA shares meaningfully. The Megaspeed and Speedmatrix reporting from the New York Times did not move them. The Singapore investigation reporting did not move them. A short report that aggregates those threads with new corporate-filings and trade-data work may move them; the historical record suggests that the more likely path is a slower one, in which the position pays off only if the underlying claim becomes legally or accounting-positively confirmed — through, for example, a follow-on indictment, a Singapore enforcement action, or NVIDIA's own subsequent disclosure.

What the short does not claim, and the bull rebuttals

Culper is careful about what the report does not claim. It does not claim that NVIDIA itself is criminally liable for the diversion patterns it documents. It does not claim that NVIDIA's revenue-recognition policy violates U.S. accounting rules. It does not claim that the Megaspeed–Apex–Alibaba financing chain is, in itself, illegal. It does not allege that NVIDIA misrepresented its policies on tracking end customers in any specific filing. The argument is structural and probabilistic: that NVIDIA's public framing of "zero China" is materially misleading on a look-through basis, and that a continued Beijing push into domestic alternatives is taking the actual run-rate down rather than leaving it as upside optionality.

The bull rebuttals to this thesis are well-rehearsed. First, NVIDIA's revenue-by-headquarters policy is consistent with the practice of comparable U.S. semiconductor companies and is the reporting standard analysts and regulators have long worked with — a "look-through" methodology is not something the company is required to apply. Second, the Supermicro indictment names individuals, not Supermicro itself, and Supermicro launched an independent board-led probe of its own promptly. Third, Megaspeed has publicly rebutted the Bloomberg coverage and asserted that U.S. authorities have found "no evidence of violations." Fourth, the demand environment for NVIDIA's Blackwell platform remains supply-constrained outside China; even if a non-trivial share of FY2026 compute revenue is functionally Chinese, the marginal chip can be reallocated to non-Chinese demand without a topline hit. Fifth, the Trump administration approved limited NVIDIA H200 exports to China in January 2026 with conditions, reopening at least a partial legitimate channel. None of these rebuttals dispatches the underlying corporate-filings work in Culper's report; they do define the bar over which the thesis has to clear before consensus moves.

What is yet to be heard

Five things are knowable in the near term, and each will materially update the report. The first is whether NVIDIA's own next earnings call addresses, on the record, the disclosure of geographic revenue on a look-through basis — including whether the company will commit to additional disclosure on partner-by-partner end-customer tracking. The second is whether U.S. Commerce, the Department of Justice, or Singaporean authorities take a further enforcement step against Megaspeed, Speedmatrix, OBON, Siam AI, Aivres, or any of the named OEMs — the existing Supermicro indictment was already a significant action, and a second indictment in the Megaspeed corridor would shift the consensus around the dimensions of the diversion problem. The third is whether Beijing's procurement guidance — the post-2025 push toward Huawei Ascend, Cambricon, and other domestic alternatives in state-affiliated buyers — accelerates, plateaus, or reverses through the rest of 2026. The fourth is whether Bain's Bridge Data Centres or any other independent operator publicly explains why Megaspeed was dropped as a tenant. The fifth, and the one the market will price first, is NVIDIA's own response to the Culper report. As of the date of this post, none of these points have been settled in public.

Where we sit

We build Plainview and Bollwerk Frontier to help research, risk, and compliance teams turn the kind of allegation that surfaces in a short report into a structured investigation of their own. The tooling sits over filings, contracts, trade data, regulatory disclosures, and the cross-jurisdictional corporate-record trail the way an analyst does: 10-K geographic-revenue disclosures, foreign annual reports for off-shore counterparties, charge filings in Malaysia and Singapore, indictment dockets in U.S. federal court, and the partner-by-partner trade flows that link a U.S.-listed OEM to its actual end customers. A short report is one input into a larger investigation, not a verdict. If your team is working through the NVIDIA thesis — or any comparable mega-cap, multi-jurisdiction, OEM-channel-exposed name — and would find it useful to compare notes, write to hello@bollwerk.ai.