Compute & Infrastructure

Colocation operators are moving beyond their traditional role as passive real estate and power providers, now acquiring GPU hardware directly and offering it as a financed asset class to enterprise customers. According to Data Center Dynamics, this shift reflects both the difficulty enterprises face in securing GPU allocations through hyperscaler cloud channels and the willingness of lenders to treat high-end accelerators as collateral-worthy assets with predictable depreciation curves. The implication is that the GPU financing market is maturing: hardware that was once viewed as too specialised for traditional asset finance is now being underwritten by institutional capital.

This structural change has significant competitive implications. Colocation providers that build GPU inventory gain pricing leverage and customer stickiness that pure wholesale power-and-rack operators lack. It also redistributes hardware risk down the value chain — colos now carry GPU obsolescence exposure that previously sat with cloud hyperscalers or end users. As accelerator generations turn over on roughly 18-24 month cycles, the residual value assumptions embedded in these financing structures will be tested. The trend is confirmed by deal activity, not merely announced intent.

A natural gas pipeline project supplying fuel to a planned Oracle AI data centre in New Mexico has been delayed by approximately six months, now expected to come online in 2027, according to Bloomberg. The delay is notable not because gas pipelines are unusual — on-site gas generation is increasingly common for AI campuses seeking power independence from strained grids — but because it reveals that the supply chain constraints limiting data centre buildout now extend into midstream energy infrastructure, not just grid interconnection queues or transformer lead times.

This development reinforces a broader pattern discussed in Bloomberg's commodity market analysis, where AI power demand is reshaping investment flows into uranium, natural gas, and critical minerals. Hyperscalers and their suppliers are increasingly forced to secure energy supply chains years in advance, effectively becoming energy infrastructure developers. The Oracle delay is a confirmed operational setback, not a speculative risk — it will push back capacity availability at a facility that was presumably already factored into Oracle's cloud expansion timeline.

Ukrainian intelligence has claimed that recovered Russian S-71 'Monochrome' cruise missiles contain Nvidia Jetson Orin NX modules, allegedly used for AI-assisted terminal guidance, according to Tom's Hardware. The Jetson Orin NX is an edge AI module commercially available at price points accessible to a wide range of buyers — it is not a data centre accelerator subject to the same export controls as H100 or B200 GPUs. This distinction is critical: the policy focus on restricting large-scale training hardware may be leaving a meaningful gap in dual-use edge AI silicon.

The strategic implications extend beyond this specific weapon system. If confirmed, it establishes that commercially available edge inference hardware — designed for robotics, autonomous vehicles, and industrial applications — is capable of providing militarily significant AI guidance capabilities. This will intensify pressure on the US Commerce Department to expand Entity List controls to cover a broader range of Nvidia's product portfolio, and may accelerate allied governments' demands for more granular end-use certification requirements. Nvidia's exposure is reputational and regulatory, not financial in the immediate term, but a broadening of chip export controls could materially affect its addressable market in commercial sectors that currently face no restrictions.

Organised cargo thieves in California have employed vehicle interdiction tactics — including PIT manoeuvres — to disable private security escorts and seize AI data centre hardware shipments worth millions of dollars, according to Tom's Hardware. The sophistication of the attacks, involving deliberate neutralisation of armed escorts, indicates professional criminal networks with intelligence on shipment schedules and the logistical sophistication to monetise high-value AI hardware through established black market channels.

This is a confirmed operational risk, not a theoretical threat. The existence of a viable black market for AI accelerators has direct implications for supply chain security planning: hardware that can be resold at a premium — whether due to export controls, allocation constraints, or simply scarcity — becomes a target for physical interdiction. For infrastructure operators and procurement teams, this signals a need to harden last-mile logistics with the same rigour applied to cybersecurity. It also underscores the premium that restricted buyers — including potentially sanctioned entities — are willing to pay to acquire this hardware outside official channels.

Two research publications this week illuminate the compounding difficulty of sub-2nm semiconductor manufacturing. Imec's paper on 2D Gate-All-Around CFETs at the A2 node — with a contacted poly pitch of 36nm and gate length of 10nm — identifies contact resistance and parasitic co-optimisation as non-negotiable requirements for viable transistor performance, as reported by Semiconductor Engineering. Separately, Nvidia and Duke University published work on using self-supervised AI models to fix design rule violations in advanced-node place-and-route, addressing a problem that conventional EDA tools increasingly cannot solve at scale, per Semiconductor Engineering.

The strategic relevance for infrastructure planning is that each successive node transition is introducing qualitatively new manufacturing and design challenges — not merely incremental scaling difficulties. The imec CFET work is at the research demonstration stage, not production readiness, and should be treated as a signal about where the industry must invest, not what is coming online imminently. Nvidia's EDA research, by contrast, reflects a direct business need: as Nvidia designs increasingly complex chips at TSMC's leading nodes, it has a strong incentive to develop proprietary AI-assisted tools that reduce design cycle time and manufacturing yield risk. This positions Nvidia not just as a chip consumer but as an active participant in the EDA toolchain — a dependency reduction strategy with long-term supply chain significance.