Compute diversion

Evidence on the redirection of chips to China from outside China primarily draws from public indictments and journalistic investigations into allegations of diversion, though some investigative reporting also provides evidence of resale within China.

A table of cases that are primarily diversion is shown below. Cases that cover both diversion and resale in China are listed in both tables.

Compute resale

This is not based on end-users, but rather based on evidence from the marketplace for chips within China, with evidence coming from investigative reporting, including interviews, observations of sale contracts, and photos of smuggled chips physically present in China

• WSJ, July 2024
• NYT, August 2024
• Information, August 2024
• FT, July 2025

It is also likely that some of the same chips are observed on both the diversion and resale side.¹²

Some journalistic investigations into Chinese chip smuggling do not attempt to enumerate total volumes moved, and instead discuss specific aspects of the distribution network in more detail. A July 2024 WSJ investigation reported that there were more than 70 active distributors on online marketplaces (of which they spoke to 25), but was vague about total volumes (though noting that many sellers had dozens of chips). An August 2024 NYT piece reported finding a similar number of distributors (nearly 100 stores claiming to sell the chips, of which the reporter spoke to 11) and provided some more detail on order volume (one vendor said companies ordered 200 to 300 chips from him at a time, another described a single 2,000 chip order). While the total number of chips explicitly enumerated in each article is low, the overall inference one can make from the reporting is that, in aggregate, there is a substantial quantity of chips being sold, much of it at a small scale.

Much like the FT piece, there is some uncertainty about how much of the supply obtained by the distributors is coming from the large-scale operations versus a network of smaller-scale smuggling that is described in the pieces. We discuss this further under modeled estimates: for now, we turn to aggregating the explicitly reported cases.

Much of the data that we have on the resale market is about the number of vendors in the marketplace and the range of volumes they face, rather than specifically enumerated individual cases.

Compute diversion

The key uncertainties here are:

• What proportion of diverted compute has been detected?
  • We don’t know what proportion of smuggled compute succeeded at not triggering an investigation
  • There may be ongoing investigations that have not yet been publicized or resulted in an indictment¹⁴
• What proportion of allegedly diverted compute made it to China?
  • Attempted diversion doesn’t mean successful diversion
  • Allegations can be false, as many cases are not yet proven in a court of law

The biggest driver of uncertainty on the diversion side is that we don’t know what fraction of diversion has been observed. The large-scale smuggling schemes detected and reported so far could represent the majority of the volume, or they might be just a small fraction of the total flows. We consider a wide range of detection rates (90% CI: 10–80%; median 25%¹⁵). This is higher than some other illicit flows because successfully identifying a perpetrator can provide substantial information about earlier flows. Detection is easier than interception — if a repeat offender is ever caught, this reveals previous transactions.

In addition to uncertainty about detection, there is also some uncertainty around allegations. Even after indictments are made public, there is not always clear information about how many of what kinds of chips were moved when. In the largest indictment to date, the DOJ alleges that $2.5 billion of chips were sent from Supermicro to a front company, which was then sending those chips onwards to China. The indictment only explicitly claims that $510 million worth of servers made it to China (between April and May of 2025), but other quotes from the indictment suggest that the DOJ believes that at least half of the chips made it to China¹⁶. From the indictment, we also know that the scheme was alleged to have grown over time and switched to aggressively shipping Blackwells once they became available (in 2026). Accordingly, we include both case-specific uncertainty¹⁷ and a small overall reporting-accuracy haircut.

Compute resale

Two uncertainties shape our resale estimate: what the observed vendor data implies about total market volume, and what proportion of smuggled compute flows through resellers detectable by investigative reporters. We address each in turn.

To estimate total marketplace volume, we follow CNAS in multiplying the number of vendors by volume per vendor, but add a hierarchical model where vendors share a market-wide shock but also have individual variation instead of using a single draw of average vendor volume¹⁸.

We parameterize the number of vendors as a lognormal with a 90% CI of [38, 150] for 2024 (following CNAS) and a wider [20, 200] for 2025, reflecting uncertainty about whether the marketplace grew or shrank, as there was less reporting that included the number of vendors in 2025. We follow CNAS by generally parameterizing volume per vendor with a 95% CI of [204, 12,000] chips per year. During the H20 ban period(April–July 2025), we use an elevated upper bound of 40,000 chips per year as the 99th percentile to accommodate Gate-of-the-Era-scale operations, which were reported by the FT as moving almost $400 million in chips during that period.

Because we split 2025 into pre-ban, ban, and post-ban sub-periods to accommodate the elevated per-vendor ceiling during the ban quarter, a shared lognormal year-shock (sigma=0.5) is applied to all three sub-periods. This correlation is needed because summing independent sub-period draws artificially narrows the 2025 distribution relative to 2024’s single annual draw. The shock is calibrated so that 2025’s log-scale spread matches 2024’s.

The above estimates of the number of vendors and vendor volumes are based on investigative reporting about sellers and transactions in the Chinese gray market. There are likely some chips that don’t go through the vendors and brokers of this market; instead, their import is arranged by end users. Gray market sellers are to some extent observable because those sellers also want to find clients, whereas it is harder for journalists to obtain information about entirely internal schemes. However, given the robust gray market (selling export-controlled chips is not illegal within China) and since thus far known cases on the diversion side all went through middlemen, we think that the majority of the chips (90% CI 57–96%, median 82%) go through these intermediaries and scale up our estimate of resale accordingly¹⁹.

Combined results

Both the diversion and resale sides are attempts to estimate the same underlying quantity: total smuggled compute reaching China. Diversion approaches this from the export side (what is leaving legitimate channels), while resale approaches it from the China side (what is showing up in the gray market). We combine them as a weighted blend, with more weight on the diversion estimate (which has more case-level anchors grounded in indictments with specific quantities).

The diversion side yields a cumulative median of ~530,000 H100e (90% CI: 180,000 to 1.6 million), and the resale side gives ~700,000 (90% CI: 230,000 to 2.3 million). The resale side runs higher, driven by the marketplace model’s fat right tail (the product of uncertain vendor counts and uncertain per-vendor volumes compounds into a wide distribution). The combined estimate is ~660,000 H100e (90% CI: 290,000 to 1.6 million).

Note that the combined distribution is obtained by averaging draws for diversion and resale: since these are both similar right-skewed distributions with higher means than medians, averaging the draws pushes the combined median up (and narrows the confidence intervals).

Comparison to other estimates

There are two main public attempts at making such an estimate: Grunewald and Fist from the Center for a New American Security (CNAS) estimated 140,000 chips smuggled to China in 2024 (code), with substantial uncertainty around the point estimate (90% CI from 20,000 to 800,000 chips). Their modeling strongly inspired the compute resale portion of our model.

In April 2025, ChinaTalk analyst Aqib Zakaria estimated (code) that cumulatively 452,000 H100e of compute (90% CI: 194,000 to 1.3 million H100e) had been smuggled into China in circumvention of export controls. This estimate built on the CNAS 2024 figure to project 2025 volumes. This estimate used the aforementioned CNAS 2024 smuggling estimate as a starting point for creating a 2025 smuggling estimate.

Our estimates are fairly close, especially on chip count. There is more divergence on the H100e estimate. In 2024, compared to CNAS, we estimated a higher proportion of smuggled chips are H100s/H200s rather than A100s (~83% Hoppers instead of ~67%). In 2025, compared to ChinaTalk, we estimated a higher proportion of smuggled chips are B200s/B300s rather than H100s, with roughly an even split for the year. There is both diversion and resale evidence on substantial quantities of Blackwells making it to China - the Financial Times’ investigation into smuggling during the H20 ban described Blackwells as the most sought after and widely available chip model. The Supermicro indictment confirmed B200 diversions starting in April, and the purportedly missing Megaspeed chips seem to be mostly Blackwells.

Quarterly extrapolation

We produce quarterly estimates of compute smuggled in 2024 and 2025.

For 2024, we allocate the annual estimate across quarters using H100 production volume, lagged by one quarter — this captures the expected vendor ramp-up over the course of the year as chip supply grows and more sellers enter the market. The resulting Q1–Q4 weights are roughly [0.18, 0.23, 0.27, 0.33]. For 2025, we allocate based on the pre-ban / ban / post-ban structure used in the resale model.