Huawei's 750,000 Ascend chips add up to under a twenty-fifth of Nvidia's 2026 compute

Multiplying the shipment plan by the per-chip rating gives roughly 390,000 H100-equivalents of Huawei silicon for the year ^[1], and Epoch AI's framing of that as under 4% of Nvidia's output ^[10] implies Nvidia itself ships more than 9.7 million H100-equivalents in 2026 ^[2]. Treat that as a bound rather than a measurement, since one term is a company plan reported by Reuters ^[12] and the other two are outside estimates ^[11][10].

The binding input in that estimate is memory. Epoch has high-bandwidth memory as the bottleneck and reckons a Huawei confined to Chinese-made HBM stays near 1% of Nvidia's output through 2028, with its chips three to four years behind through at least 2030 ^[15]. Moving from under 4% to about 1% is a factor of four, which means roughly three quarters of Huawei's projected 2026 standing rests on memory it does not fabricate ^[4].

Nvidia's China AI-semiconductor share is 55% ^[16], leaving 45% to everyone else ^[5], and that is a share of one national market while the sub-4% figure measures Huawei against Nvidia's entire output ^[10]. Subtracting them tells you nothing. Read together they say buyers are switching faster than fabs are scaling: Reuters reported demand for the Ascend 950 surging after DeepSeek released V4, with ByteDance, Tencent and Alibaba interested ^[14], and the 950PR comfortably beats the H20 Nvidia is allowed to sell into China while trailing the H200 ^[13].

The conduct actually on trial is older and more physical than any of this: Cisco router source code and T-Mobile phone-testing technology, across five American companies ^[4], plus Skycom moving prohibited US goods into Iran ^[6]. Reuters put one Skycom shipment of Hewlett-Packard equipment at 1.3 million euros, about 1.5 million dollars ^[7], against the more than 100 million dollars of Skycom-linked transactions that Cryptopolitan reported HSBC approved ^[8], a ratio near 67 to one ^[6]. That ratio is where the case lives, which is to say the lever has always been the correspondent bank rather than the wafer.

Which is the weakness in reading this trial as the fight over China's chip independence. A conviction does not deliver HBM to Huawei, and an acquittal does not withhold it. Or rather, the more interesting version of the proxy argument runs through the banking channel: if RICO penalties make intermediaries price Chinese chip-supply exposure the way HSBC has had to price Skycom ^[8], the grey routes for memory and tooling get dearer, and dearer routes are exactly Epoch's 1%-through-2028 path ^[15]. CSIS argues export controls have limited China's access while accelerating domestic substitution ^[19], and both of those can hold at once.

Enflame shows how the equity side handles the ambiguity. Sales up 1,475% in a quarter and 900 million dollars raised in Shanghai ^[18] are not the same kind of number, since the raise has a denominator and the growth rate does not, and the four firms Reuters Breakingviews calls China's "little dragons" ^[17] are being priced on the one without. For a chip investor the Brooklyn jury prices two decades of past conduct. The stack question gets settled on a memory line.