AI Infrastructure
Portfolio

August was the month AI demand was validated again, but the equity market began charging more explicitly for memory, power, leases, financing and time-to-revenue. Rather than a repeat of July's broad de-rating, investors drew a sharper line between companies converting AI demand into near-term cash flow and those still asking the market to trust distant capacity promises. Owners of near-term bottlenecks were rewarded; owners of longer-dated, less contracted capacity were not.

Underlying demand was stronger than the headline index moves suggested. AWS quarterly revenue rose 37% to US$42.2 billion, and Amazon guided to approximately US$220 billion of annual capital expenditure. That same spending pushed trailing free cash flow negative, illustrating directly how the capex cycle is now consuming cash even at hyperscalers with the deepest balance sheets, and reinforcing why the market is increasingly discriminating between capex that has visible revenue attached and capex that does not yet.

The accelerator layer confirmed both the strength and the constraints of the current cycle. Nvidia reported US$96.2 billion of quarterly revenue, more than double a year earlier, and indicated continued demand strength while flagging memory constraints and excluding China from guidance. The memory-supply commentary in particular reinforces a theme already visible in the regional sleeves: memory has become a genuine gating factor on the pace of the broader AI buildout, not merely a cyclical semiconductor subplot.

Neocloud providers illustrated the shift from demand validation to delivery and financing scrutiny. CoreWeave reported US$104.2 billion of backlog before adding more than US$25 billion of additional wins during the month, a scale of order book that leaves little doubt that demand exists. The market's focus has moved on to whether that backlog can be financed and delivered on a timeline that satisfies increasingly impatient capital providers, echoing the financing-gap concerns already visible in Oracle's credit spreads through the third quarter.

Networking provided the clearest illustration of the market's new discipline. Arista guided above expectations on AI networking and was rewarded accordingly, while Marvell sold off despite reporting strong results, as investors demanded earlier revenue conversion from its custom-silicon design wins rather than crediting the pipeline alone. The broader market moved from asking whether AI demand was real to asking which companies convert that demand into cash before financing conditions tighten further, a distinction that explains the gap between the Nasdaq-100's 3.89% gain for the month and the semiconductor index's much smaller 1.98% advance after a sharp intramonth reversal.

For the Nexus thesis, August sharpened rather than reversed July's central lesson. The bull case for AI infrastructure was never in doubt on the demand side this month; the risk that is being actively priced is time-to-revenue. The portfolio implication is to continue favouring the constraint layer—power, grid, memory and networking businesses with contracted, near-term revenue—while applying continued valuation discipline to capacity that is real but not yet monetising on a timeline the market is prepared to fund patiently.