US AI and Crypto Czar David Sacks on China and AI, and GPUs
Recent statements reveal much about administration policy confusion
Recent comments from White House and AI and Crypto Czar David Sacks on China, Huawei, and evasion of export controls cast an interesting light on administration thinking on China and AI. What is going on here? Clearly Sacks is listening to a small group of minority voices on these issues, voices with a clear agenda that conveniently ignores many aspects of the broader US China technology competition related to AI.
In addition, in a recent discussion on the All-in Podcast, Sacks highlighted where at least one key official in the Trump administrations stands on a range of AI-related issues: existential risks, Effective Altruism (EA), Biden era AI policy, AI Diffusion, the AI infrastructure deals in the Middle East, AI governance, China and AI, and AI alignment. The conversation reflects a view of these issues coming from a particular part of Silicon Valley—the conservative VC community—and proponents of US dominance of AI at all costs. Let’s stand back and take a fresh look at all of this.
Evolving US policy on AI diffusion, GPU diversion, and export controls
First, the most recent comments: Sacks this week claimed that China had “grown adept” at evading US export controls, and also claimed that China was two years behind the US in terms of semiconductor design capabilities. Sacks was again critical of the Biden administration’s AI Diffusion Rule, and asserted that US allies, including those in the Middle East, would be willing to adhere to certain security policies related to access to advanced compute from US technology companies. This was a reference to recent “hand shake” deals related to AI datacenter buildouts in the Middle East struck during President Trump’s visits to Saudi Arabia, the UAE, and Qatar in May.
Sacks: “….needlessly tight US restrictions on sales of AI chips to American allies could unwittingly create an opening globally for Huawei and other companies by driving countries toward technology originating in China. If we are overly restrictive in terms of US sales to the world, I think that there will be a time where we kick ourselves and say, ‘All of a sudden Huawei is everywhere when we used to have the market to ourselves. Why didn’t we take advantage of that and lock it in?”
Sacks’ contention that easing up on the AI Diffusion rule and allowing unrestricted sales of advanced GPUs to countries that align with the US on security issues will prevent Chinese companies such as Huawei and DeepSeek from making inroads is a strawman, and a view clearly already overtaken by events. Sacks continues to assert that the goal is to make the US AI stack “the global standard.”
In attempting to untangle the complex issue of Chinese company access to advanced AI hardware, and where the AI hardware race stands between the US and China, the game is now being played on multiple fronts:
Domestic Chinese AI company access to existing GPUs, stockpiled Nvidia/AMD/Intel GPUs, in-house designed AI optimized ASICs, and Huawei Ascend based systems. Addressed in detail here.
Diversion of Nvidia GPUs to China-based end users. As I have noted here, this is not a significant source of advanced compute in China, and none of the leading AI developers is using “diverted” or “smuggled” GPUs. This red herring continues to be used as a talking point.
Chinese company access to advanced Nvidia/AMD/Intel GPUs in datacenters outside China, including in Malaysia, Japan, and other countries. While currently a very small portion of the overall Chinese company AI hardware access landscape, this area will become much more important in the coming year. (More on this later.)
The ability of Huawei to ramp production of the Ascend series of processors and export advanced integrated GPU clusters outside of China. This is an evolving and complex question, which looms ever more important in the debate about whether US export controls are “working” or not.
In fact, US export controls have slowed down the ability of Chinese companies to develop the hardware required for training advanced AI models—but even Sacks admits that the gap is now months, and not years, in terms of advanced model development. Specifically on AI hardware and Huawei, the overwhelming focus on US export control policy that I have termed “small gain, high cost,” Sacks asserted that China is supply constrained on GPUs when asked about recent comments from Commerce Secretary Howard Lutnick that Huawei can produce only 200,000 Ascend AI processors—this figure is likely coming from industry “experts” with a clear agenda to show that the controls are working. The real number is almost certainly much higher. Getting at Sacks’ comments on Huawei and what is really happening in China on this score requires understanding what is going on in China in multiple industries, at a time when information around these issue is increasingly sensitive. There is a real need to understand sources and cut through hype. Here are some initial thoughts on this.
First, as I have been stressing since late last year, the more meaningful metrics around the AI hardware issues are now squarely at the systems integration level for large GPU clusters, both for Nvidia and Huawei.
In this regard, the advent earlier this year of the Huawei Ascend CloudMatrix 384 Supernode will likely be a game changer. Huawei’s ability to develop this “supernode” so quickly, after the announcement of Nvidia NV72 cluster in March 2024, has continued to surprise industry observers. Many questions remain around the performance of the Huawei supernode, but it includes a number of features that make it stand out from other similar systems, particularly the all-to-all GPU interconnectivity, the use of optical modules and accompanying power savings, and more. Some industry assessments point to major innovations on heat dissipation that Huawei has made for the CloudMatrix supernode. Most important appears to be the use of optical communications, which lowers latency, increases bandwidth, but comes with challenges in terms of reliability and failure modes.
The United States is trying to lock down Chinese semiconductor technology just like the “Trisolarians”.— Yuan Jinhui, Silicon Flow (美国试图像“三体人”一样，用芯片锁死中国科技) (The Trisolarians: an alien civilisation in Liu Cixin’s science‑fiction trilogy The Three‑Body Problem. After learning that humanity might one day threaten them, they deploy sub‑atomic “sophons” to Earth to block fundamental physics experiments and freeze technological progress for 400 years.)
But again it would be a mistake to see the development of a supernode by Huawei in a vacuum. Huawei Cloud is working closely with cloud provider Silicon Flow, and with DeepSeek on the model development and software environment side. So in addition to being no longer useful to compare GPU to GPU, it is also no longer useful to compare company to company, as we see the emergence of Team China Advanced AI (TCAA). Huawei is at the center of this effort, and the phrase “当历经了鸿蒙蛰伏、海思攻坚、昇腾崛起……凭借华为云CloudMatrix 384超节点，一道坚不可摧的国产算力防线正在构建” (After enduring HarmonyOS’s years of quiet incubation, HiSilicon’s uphill battles, and Ascend’s rise to prominence... a formidable domestic computing power stronghold is now taking shape, powered by Huawei Cloud’s CloudMatrix 384 supernode) captures the spirit of the effort being undertaken here.1 The Cloud Matrix 384 is now deployed at three of Huawei Cloud’s core hubs in Gui’an, Guizhou, Ulanqab, Inner Mongolia, and Wuhu, Anhui. In another example of the importance of systems level thinking, the actual data centers where the Cloud Matrix 384 are also an important part of the overall picture, in terms of power supply and optical communications—again, here is where Huawei’s experience with cooling, interconnectivity, and running data centers will play a pivotal role in how advanced AI clusters perform and are utilized.
Here, we are also seeing publication of papers assessing the performance of the Cloud Matrix 384 that combine innovations from both Huawei and DeepSeek, such as optical communications and all-to-all on the Huawei side, and EP and MoE approaches on the DeepSeek side. This type of synergy going forward was exactly what I was talking about last year when I asserted that: “By the end of 2025 we should be in a better position to determine the near- and medium-term prospects for China’s semiconductor industry to have progressed on a combination of closing some of the technology gaps, overcoming some of the technology “chokepoints,” and folding in domestic innovations of a sufficiently broad and novel nature to allow the industry to reach some level of higher comparability to western systems of a similar size and complexity.”
Comparing these systems now involves assessing an increasing number of fluid variables.2 The second order problem of how “easily” or quickly Huawei can scale up production of systems such as the Cloud Matrix 384 is equally challenging. This depends on factors such as the ability of domestic foundry SMIC to continue to raise yields on its 7 and 5 nm processes, balance demand for advanced semiconductors from Huawei and other Chinese firms, and continue to repurpose existing DUV lithography gear and integrate this equipment with existing US, Japanese, and domestic tools to expand capacity for its most advanced node processes—a tall order. It is very difficult to assess where this effort is, given the sensitivity of the issue and lack of visibility on foundry operations caused by domestic persons controls. For more on this, see my forthcoming detailed update on last November’s paper, likely available later this summer, on how China’s semiconductor industry is reacting to US export controls.
Hitting the Digital Silk Road, AI at the center
In terms of expansion of the Huawei/DeepSeek combination outside of China, Sacks this week also again referred to the Malaysia case, where the Deputy Ministry of Communications announced a project using Huawei and DeepSeek to develop sovereign AI in May, and then the Ministry of Industry, Trade, and Investment issued a clarification which said the project had not been coordinated or endorsed by the Malaysia government. Sacks tweeted about this incident as an example of why the Biden US AI Diffusion Rule was a mistake, and took additional shots at Biden era officials behind US export control policy (see below). Some recent media reporting suggests that Huawei and telecom rival ZTE are both expanding AI offerings along Belt and Road countries, including Malaysia, Uzbekistan, countries in Southeast Asia, and the Middle East. How Huawei and ZTE will source AI specific hardware for expanding or building new AI data centers in these regions remains unclear.
The issue comes down to how fast domestic foundry leader SMIC, working closely with Huawei and Huawei’s design arm HiSilicon, as well as with domestic toolmakers such as SiCarrier, can ramp production of Ascend series processors, and incorporate them into supernodes that can be installed at datacenters outside China. In the case of the Malaysian effort to develop sovereign AI based on Huawei hardware, which would be the first deployment of Huawei Ascend hardware outside of China, the initial project was set to feature a generative AI cloud environment run by Malaysian firm Skyvast Cloud, powered by an AlterMatic DT250 AI Server with eight Huawei Ascend GPUs. During the announcement of the project in early May, Skyvast and Chinese tech firm Leadyo revealed plans to deploy 3,000 advanced GPUs across multiple infrastructure zones by 20263, presumably all Ascend GPUs; it is unclear whether this would be a set of 10 or so Cloud Matrix Supernodes, or other, less advanced Huawei server configurations.
The target for the AlterMatic DT250 is in fact sovereign and private cloud operators that cannot import NVIDIA H100‑class hardware, or who want an Arm‑native, lower‑power alternative. The Skyvast Cloud project is the first announced external customer, positioning DT250 at the core of the country’s “sovereign AI” project. As previously noted, the status of this project remains unclear, particularly in the wake of the Commerce Department’s mid-May guidance that use of the Ascend series anywhere in the world could constitute a violation of US export controls. Chinese trade negotiators were livid at the threat implied by the guidance, and it has become an issue tied up with tense negotiation around additional US export controls and Beijing’s imposition of a strict export licensing regime around rare earths that remains unresolved. For more on this, see here.
Hence, despite almost certainly prioritizing shipment of systems like the Cloud Matrix 384 and other server configuration using the Ascend series for domestic customers, including all the major AI developers in China, Huawei has sufficient stocks to consider beginning to develop export-oriented versions of AI optimized hardware. For SMIC and Huawei, the issue of prioritizing production of advanced node semiconductors, at 7 and 5 nm, between Huawei’s Kirin smartphone ASIC and Ascend series AI processors gets more and more difficult and complex, given high demand for both and the emergence of export markets and what could be high demand from countries that wish to avoid the US AI stack touted by Sacks.
Taking a peak ahead: New US AI Diffusion rule, more export controls?
On the US side, the race to beat the China AI Stack and TCAA to key markets globally remains an administration priority. Faced with pushback in Washington to what appear to be mostly handshake deals in the Middle East during President Trumps’ May swing through Saudi Arabia, the UAE, and Qatar, White House AI Czar Sacks has repeatedly hit back at his critics over the past month. Sack’s view, as expressed on X and in episodes of the All-In podcast, is that in Washington, Effective Altruists have rebranded as “China Hawks” and are opposing providing access to large numbers of advanced GPUs for AI datacenters in the Middle East, using the “China card.” Of course, Sacks is using his own “China card” justification for pursuing the ME deals in the first place, just the flip side of the “China Hawk” argument on China and AI. Confused? You are not alone.
Sacks and his staff are arguing for easing of controls on advanced GPUs in the new version of the AI Diffusion Rule, which is moving steadily through the interagency process in Washington and could come soon. China critics will insist the new rule has sufficient security safeguards to prevent diversion—a red herring when it comes to 2 ton Nvidia superclusters such as the NVL72—and ensure know your customer (KYC) processes are implemented to prevent Chinese companies and researchers from accessing advanced AI hardware. But other major issues remain unresolved, such as how the massive promised investments and shipments of Nvidia and AMD GPUs to data centers in the Saudi Arabia, the UAE, and Qatar will be managed—no details of the handshake agreements made during the Trump Middle East trip have been forthcoming—or how the dollar-for-dollar multi-billion US investments from major Middle East sovereign wealth funds Sacks claims are part of the deal will be handled within the US government. The complexity of these AI infrastructure buildouts will require a serious and long-term industrial policy, something similar to the CHIPS Act effort which saw the establishment of the CHIPS Program Office (CPO) to manage complex investment and supply chain issues. The forthcoming OSTP AI Action Plan could shed new light on how the administration envisages handling this issue and other critical issue around AI infrastructure buildouts such as energy.
At the same time, the issue of continuing to impose export controls on Chinese firms to slow their ability to develop advanced AI models will now need to account for Beijing’s willingness to leverage the rare earths card. A new package of export controls now in the works would see implementation of the 50 percent rule on ownership of subsidiaries for Entity List companies, suspension of all licenses related to semiconductor manufacturing tools, and additions of more Chinese firms to the Entity List. Release of such a package is being pushed by some officials as part of a continued “escalate to de-escalate” policy, but Beijing is not likely to want to play this game. Chinese officials at recent talks in London likely pushed for a major rollback of US export controls, particularly those rolled out late in the Biden administration. More on this next week.
Meanwhile, Huawei and SMIC will continue to make steady progress on capacity, DeepSeek is readying release of its V4 and R2 models, other Chinese AI firms such as Minimax, Alibaba, Tencent, and Bytedance continue to crank out more advanced models and agentic AI platforms, and the race to AGI has become much more complex.
Key to this formulation is the 12 character phrase 鸿蒙蛰伏、海思攻坚、昇腾崛起. Here is more in-depth analysis of this phrase:
1. “鸿蒙蛰伏” – Harmony’s long hibernation
Historic/mythic allusion:
The word “鸿蒙” originates in ancient Chinese cosmology (e.g., 《庄子》 and other Daoist texts), referring to the primordial chaos before the birth of the universe—a time before order, before separation of heaven and earth.By invoking “鸿蒙”, Huawei indirectly casts its HarmonyOS not merely as a software platform, but as something cosmologically foundational, signaling a new beginning for China’s digital ecosystem.
“蛰伏”, or hibernation, evokes the image of a dragon sleeping in the deep, silently gathering strength before a dramatic reemergence. It’s a classic Chinese literary metaphor for latent power.
2. “海思攻坚” – HiSilicon’s grinding battles
Military metaphor:
“攻坚” (gōngjiān) literally refers to storming a fortified position, often used in PLA military lexicon (e.g., 攻坚战, “a battle to breach strongholds”). It evokes the hardship of breaking through an entrenched adversary’s defense—here, symbolizing the U.S. tech blockade after 2019.Stylistic significance:
The phrase frames HiSilicon’s chip development efforts (especially around Kirin SoCs) as a form of technological warfare. There is also a possible subtextual echo of Maoist struggle language, framing Huawei’s engineers as the new vanguard of national resistance and revival.
3. “昇腾崛起” – Ascend’s soaring rise
Rising imagery:
“昇腾” (shēngténg) is a lofty, almost spiritual word meaning to soar upward, often used to describe dragons or phoenixes ascending to the heavens. It lends a sense of grandeur, destiny, and inevitability to the rise of the Ascend AI chip series.“崛起” (juéqǐ) is geopolitically loaded: since the early 2000s, China’s rise on the world stage has been described using this term (“中国的和平崛起” – “China’s peaceful rise”).
By using it here, Huawei situates Ascend not merely as a technological success, but as part of China’s broader national resurgence.
For example, an assessment of the Huawei Cloud Matrix 384 paper yields these types of conclusions, specifically comparing the Ascend based Cloud Matrix system with a system based on Nvidia H100 GPUs:
For deployments of similar capacity (≈ 4 PFLOPS INT8 per 8‑GPU/H100 node vs 8 PFLOPS INT8 per 8‑NPU Ascend node) CloudMatrix delivers equal or better per‑device service efficiency and competitive raw throughput, chiefly through an INT8‑optimised stack and a peer‑to‑peer UB fabric that slashes MoE traffic costs. For organisations outside China that rely today on H100‑class systems, CloudMatrix shows that comparable—or better—LLM serving economics are achievable with Ascend hardware.
CloudMatrix’s inference stack couples a near‑flat, 400 GB s‑class UB fabric with operator‑level tricks (AIV‑Direct, fusion, micro‑batch pipelining) and a dedicated RDMA lane for cold paths. The result is that communication ceases to be the limiter: expert‑parallel MoE decode and long‑context prefill both run close to arithmetic peak while maintaining sub‑50 ms end‑to‑end latency—performance that current H100 + InfiniBand topologies struggle to match once a job crosses a single chassis.
According to information released in early May, Skyvast will use the Huawei and DeepSeek based AI infrastructure to offer AI-as-a-service and GPU-as-a-service (GPUaaS) to government agencies, businesses and universities, enabling them to improve services, boost productivity and drive innovation. Early adopters were to include the Prime Minister's Office, the Ministry of Communications, Permodalan Negeri Selangor, Cyberview and Universiti Teknologi Malaysia.
In addition, the AI infrastructure launch event saw the unveiling of the Malaysia-China Trusted Data Zone, touted as the first bilateral corridor linking Cyberjaya with the China (Shanghai) Pilot Free Trade Zone (FTZ) Lin-gang Special Area.
This data zone – operated by Skyvast Data on the Malaysian side and Shanghai International Data Port on the China side – is intended to enable seamless cross-border AI development through infrastructure, joint innovation platforms and digital capability integration.