Why Tencent's Lease of 100,000 Nvidia H100/H200 Chips Challenges US Export Controls

Tencent reportedly leased 100,000 H100/H200 accelerators through arrangements described as testing the limits of US export controls, a development that coincides with a $300 million chip-smuggling prosecution and Nvidia's Q1 FY2027 revenue of $81.6 billion, up 85% year-on-year.

The disclosure this month that Tencent had leased 100,000 H100/H200-series chips — the most powerful Nvidia accelerators barred from direct sale to Chinese entities — crystallises a fault line that regulators have struggled to contain since export controls were tightened. According to Table.Briefings, the arrangement reportedly challenges the US export control framework, though the precise structure was not detailed in available reporting. Whatever the mechanism, the news lands against a backdrop of active enforcement: federal prosecutors this month charged California resident Greg Lui, alleging he orchestrated a scheme to smuggle approximately $300 million worth of Nvidia AI servers to China, with four specific chip models named in the indictment. Together, the two cases illustrate the range of pressure being applied to Nvidia's supply chain — from sophisticated financial structures at one end to alleged criminal networks at the other.

That pressure exists precisely because demand has outrun any legal channel available to Chinese buyers. Nvidia's Q1 fiscal year 2027 results — $81.6 billion in revenue, 85% year-on-year growth, with the data centre segment up 92% — leave little ambiguity about the scale of global appetite for the company's accelerators. The magnitude of those numbers makes almost any workaround economically rational, which is partly why US enforcement agencies face an inherently asymmetric problem: the incentive to circumvent controls scales with the performance gap between what is permitted and what is restricted. That gap is, if anything, widening: Nvidia's Vera Rubin architecture has entered full mass production — 336 billion transistors, up to 288GB of HBM4 per GPU — with the company promising up to a tenfold reduction in inference token costs relative to prior generations.

The demand dynamic is also reshaping how Nvidia's hardware is being financed. Amazon is reportedly seeking to move $8 billion in Nvidia chips off its balance sheet through a leaseback arrangement, a structure that Reuters and Wall Street analysts are treating as a potential inflection in how hyperscalers fund AI infrastructure rather than capitalise it outright. Simultaneously, Nvidia is stepping in as a capital provider: a $668 million investment in GMI Cloud, combined with confirmation that Nvidia is the mystery tenant behind a $5.02 billion, 704-megawatt lease at Hut 8's Beacon Point campus in Texas, places the company in the business of operating AI factories rather than merely supplying them. Nvidia's own capital expenditure reached $6.04 billion in fiscal year 2026, up 86.7% from $3.24 billion in fiscal year 2025, against revenue of $215.94 billion — a 3% intensity that reflects deliberate asset-lightness even as its strategic footprint expands. Investor Michael Burry has publicly likened the current AI infrastructure buildout to the 1960s computer leasing bubble, and Reuters has reported growing Wall Street scepticism about Nvidia's capacity to sustain GPU financing at scale.

Beyond the financing debate, the competitive and geopolitical perimeter around Nvidia is being tested from multiple directions simultaneously. On the demand side, xAI is planning to bring approximately 420,000 Nvidia GPUs online at its Colossus facility in November, and OpenAI's GPT-6 Astra Ultrafast reportedly runs on Nvidia Blackwell hardware — evidence of continued dependence at the frontier. Yet OpenAI's custom Jalapeño ASIC is being deployed alongside AMD EPYC Turin processors rather than Nvidia's Vera architecture, suggesting that even the most committed customers are building parallel alternatives. DeepSeek and Huawei this month released open-source programming libraries targeting Huawei's Ascend 950 accelerator, explicitly aimed at reducing lock-in on Nvidia's software stack. In China, Biren Technology is readying its BR20X GPU. None of these challengers is close to displacing Nvidia's market position today — Cerebras shares reportedly fell 20% after OpenAI was said to be shifting compute back toward Nvidia, underscoring the incumbency premium — but the directional bet against concentration is now being placed across hardware, software and capital simultaneously.

Three developments merit close attention over the coming months. First, the regulatory response to lease-based access structures: whether the US Treasury or Commerce Department will interpret arrangements such as the reported Tencent lease as substantively equivalent to a direct sale, and whether that triggers new guidance around third-party chip access. Second, whether Amazon's leaseback structure is replicated by other hyperscalers — which would shift Nvidia's revenue profile from lumpy capex-funded sales toward an annuity-like stream carrying different credit, duration and counterparty risks, the concern that Wall Street's scepticism already flags. Third, whether Vera Rubin's tenfold inference cost improvement materialises at production scale; if it does, the per-token economics shift enough to revive demand from customers currently experimenting with Ascend, TPU v7 or custom silicon. The underlying question is no longer whether Nvidia's chips can be contained — this month's events make plain that the existing framework is under serious stress — but whether a company whose hardware is now integral to the global AI buildout can sustain a model that increasingly requires it to finance the infrastructure it sells into.