Compute & Infrastructure

Reports emerged this week that Nvidia had told cloud partners participating in its 'AI Cloud Commitments' programme that they could only lease Nvidia GPUs to customers pre-approved by Nvidia — a condition that would give the chipmaker unprecedented influence over downstream compute access. Nvidia has denied pausing the initiative, but Data Center Dynamics and Tom's Hardware report conflicting accounts, with the former indicating some deals were paused amid partner pushback. The contradiction — Nvidia's denial versus partner-sourced reporting — is itself a signal that the programme's terms are under active renegotiation.

The strategic logic is clear: Nvidia wants to ensure its hardware is associated with approved, high-value workloads and to prevent GPUs from flowing to sanctioned or strategically undesirable end-users. But the mechanism — vendor control over a cloud provider's customer list — represents a qualitative escalation in hardware vendor leverage over the cloud stack. Antitrust regulators in the EU and US have already been scrutinising Nvidia's market position; this initiative gives them a concrete distribution-control mechanism to examine.

The Trump administration is preparing a revised AI diffusion rule that would close a significant regulatory gap: the ability of Chinese entities to access restricted US AI compute remotely through servers hosted in third countries or by non-US cloud providers. Tom's Hardware reports the rule could be shared with trade groups as early as September. This follows the Biden-era AI diffusion framework and represents a continuation — not a reversal — of the bipartisan consensus on restricting Chinese access to frontier AI compute.

The practical enforcement challenge is considerable. Remote access to compute is difficult to audit, particularly when routed through intermediary jurisdictions. The rule will likely require cloud providers and colocation operators to implement know-your-customer controls at the workload level — a compliance burden that will fall disproportionately on smaller operators and non-US hyperscalers. It also raises questions for Alibaba Cloud, which this week announced its first Brazilian cloud region with AI services, potentially expanding the network of non-US compute nodes accessible to a wide range of customers.

SK Hynix formally broke ground this week on an HBM assembly and packaging facility in the United States — the first of its kind on American soil. As Tom's Hardware reports, the facility will assemble HBM stacks from DRAM wafers produced in South Korea, with production targeted for 2029. The architecture is a hybrid model: wafer fabrication remains in Korea, but the final packaging and integration — the step that binds DRAM dies into HBM stacks — moves to US soil.

HBM is a chokepoint component for AI accelerators. Every Nvidia H100 and B200 ships with SK Hynix, Samsung, or Micron HBM, and SK Hynix commands the largest share of HBM supply to Nvidia. Bringing assembly onshore reduces geopolitical risk on one critical link in the supply chain, but the 2029 production date means current AI infrastructure buildout — running at maximum pace through 2027 and 2028 — will remain entirely dependent on Korean production. The CHIPS Act framework likely underpins the investment, though the specifics of federal support were not disclosed in this announcement.

Two capital raises this week illustrate the maturing financing architecture for AI infrastructure. Lambda secured approximately $1 billion in short-dated private debt, specifically tied to GPU procurement for its Microsoft collaboration, as reported by Bloomberg. Simultaneously, Andreessen Horowitz closed its dedicated AI infrastructure fund at $1.1 billion, according to Bloomberg. The Lambda deal is particularly instructive: short-dated debt against hardware assets with a hyperscaler offtake agreement is a classic asset-backed structure, suggesting that institutional debt markets are now pricing AI GPU collateral with enough confidence to underwrite at scale.

The a16z fund represents a different dynamic — equity capital targeting the infrastructure layer rather than application software, reflecting the recognition that the value capture opportunity in AI has shifted toward compute provision. Together, these two transactions add over $2 billion in fresh capital to the AI infrastructure ecosystem in a single week, and they represent categorically different risk appetites: lenders are comfortable with GPU-collateralised short-term debt; equity investors are betting on long-duration infrastructure returns.

Two distinct supply-side pressures on the semiconductor ecosystem crystallised this week. First, 80% of Micron's Taiwanese workforce has signalled willingness to strike over bonus pay, as Tom's Hardware reports — a tension driven by the AI-driven memory boom benefiting company financials while workers seek a profit-sharing structure comparable to Samsung and SK Hynix employees. Micron is the only US-headquartered DRAM and NAND manufacturer of scale, making any disruption to its Taiwanese operations directly relevant to US memory supply security. Second, the Trump administration is reportedly considering extending semiconductor tariffs to include data centre servers, per Data Center Dynamics — a move that would materially increase capex costs for every hyperscaler and cloud provider importing server hardware.

These two pressures compound in a specific way: tariffs on servers would increase demand for domestically assembled hardware, but labour unrest at Micron — a key supplier to those servers — would constrain the memory supply needed to populate them. The administration's simultaneous consideration of semiconductor tariffs and the HBM onshoring subsidy for SK Hynix reflects a policy posture that is increasing costs at one end of the supply chain while trying to build resilience at another.