- Flying petabytes of AI knowledge is China’s newest workaround for strict U.S. chip controls
- Bodily smuggling exhausting drives now bypasses surveillance and digital firewalls throughout a number of jurisdictions
- GPU-rich Malaysian knowledge facilities have gotten floor zero for offshore Chinese language AI coaching
As america continues to tighten export restrictions on superior AI chips, akin to these produced by Nvidia, Chinese language AI firms are turning to a workaround that feels nearly analog in at present’s digital world.
Quite than counting on on-line transfers or sanctioned {hardware}, some companies are bodily transporting huge datasets on exhausting drives throughout borders.
A report from the Wall Road Journal claims 4 Chinese language tech staff not too long ago flew into Malaysia, every carrying 15 high-capacity exhausting drives, totaling an astonishing 4.8 petabytes of knowledge, supposed for coaching massive language fashions.
Huge knowledge continues to enter China regardless of restrictions
US restrictions have made it more and more tough to amass high-end AI GPUs by means of authorized channels.
Though Nvidia maintains, “there isn’t any proof of chip diversion,” on-the-ground studies recommend in any other case, with a black marketplace for smuggled Nvidia GPUs thriving in China.
A few of these chips are reportedly getting into the nation by means of subsidiaries and companions in neighboring nations.
Nevertheless, that route is turning into dearer and riskier attributable to heightened scrutiny and diplomatic strain from Washington on these middleman international locations.
In consequence, firms are shifting techniques: slightly than importing restricted chips, they’re exporting huge volumes of coaching knowledge.
This can be a advanced and resource-intensive course of. Firms rigorously plan the bodily transportation of knowledge, distributing drives to keep away from detection by customs.
In addition they hire GPU-rich servers in third-party international locations akin to Malaysia to course of the information.
One instance entails a Chinese language agency that used its Singapore-registered subsidiary to signal a knowledge heart contract. Nevertheless, the Malaysian companion later insisted on native registration to keep away from regulatory strain, as Singapore started tightening its personal controls.
Regardless of rising efforts by US companies, enforcement gaps and logistical loopholes proceed to be exploited. Whereas transport petabytes of knowledge on exhausting drives could seem outdated, it sidesteps bandwidth limitations and digital surveillance.
The usage of exhausting drives, starting from massive SSDs arrays to high-capacity exterior HDDs, is central to those covert transfers.
Nonetheless, it raises a query: why not use magnetic tape, particularly on condition that fashionable LTO-10 codecs can retailer as much as 30TB uncompressed and 75TB compressed?
The reply possible lies in practicality. Tape options require specialised learn/write {hardware} and lack the plug-and-play comfort of high-end HDDs generally used at present.
