Compute & Infrastructure

Alphabet's Q2 2026 infrastructure spend of $44.9 billion — which pushed the company into negative free cash flow for the first time — is not an anomaly but a declared trajectory. CFO Anat Ashkenazi confirmed that capex will increase further in 2027, with the full-year 2026 envelope now guided at $195-205 billion. The company is explicitly betting on its custom TPU programme as a strategic differentiator, a point made sharper by Nvidia's concurrent $5 billion investment in Safe Superintelligence, which had previously relied on Google TPUs. As reported by Tom's Hardware, that transition represents both a commercial loss for Google Cloud and a validation of Nvidia's strategy to embed itself in frontier AI labs before they scale.

The Nvidia-SSI deal, reported by Data Centre Dynamics, structures the relationship as an investment rather than a simple procurement agreement, suggesting Nvidia is acquiring strategic alignment with a high-profile lab rather than just a customer. This mirrors Nvidia's pattern of ecosystem capture — the risk for Google is that SSI's TPU dependency was a potential showcase for non-Nvidia infrastructure at the frontier, and that case study is now closed.

SK Hynix reported a 557% year-on-year profit increase for the quarter — a figure that would ordinarily draw celebration — but the miss against analyst expectations has crystallised concerns that the AI memory supercycle may be losing momentum. The company simultaneously announced at least $31 billion in capital spending for 2026, a record outlay that reflects its dominant position in HBM supply for Nvidia's H- and B-series accelerators. As Bloomberg reports, the market reaction reflects growing anxiety about overinvestment — a concern that capacity being built today could arrive into a softer demand environment.

The structural issue is that HBM supply chains require 12-18 months of lead time from investment to production, meaning SK Hynix's $31 billion is essentially a bet on demand conditions in 2027-2028. If hyperscaler capex — led by Google's enormous commitments — sustains or grows, that bet looks rational. If even one major cloud provider pulls back, the resulting oversupply in HBM would be severe, given the limited number of customers capable of absorbing advanced memory at this scale.

AMD is pushing rackscale AI system architectures — integrating compute, networking, and memory at the rack level rather than the card level — as its primary mechanism to compete for hyperscaler AI infrastructure contracts. As detailed by Next Platform, the strategy reflects an understanding that Nvidia's moat is not simply the GPU die but the full-stack NVLink and NVSwitch fabric that makes large-scale training clusters coherent. AMD's Infinity Fabric and ROCm software ecosystem are the counter-play, but the software gap remains the most durable competitive barrier.

The significance here is timing: AMD is advancing these roadmaps at exactly the moment when hyperscalers are under pressure to demonstrate capex discipline. Any cloud provider that can credibly argue AMD rackscale infrastructure delivers comparable training performance at lower cost-per-FLOP has a strong internal justification to diversify. The question is whether AMD can close the software ecosystem gap fast enough to make that argument credible before customers lock their 2027 infrastructure procurement.

A sponsored analysis from Data Centre Dynamics articulates a shift already visible in procurement requirements: data centre operators are moving from optimising water usage effectiveness (WUE) to eliminating evaporative cooling entirely. The driver is a combination of municipal water restrictions in key data centre markets — Northern Virginia, Phoenix, Dublin — and the expectation that AI-scale rack densities above 100kW will make air cooling physically insufficient regardless of water availability.

The technology transition centres on direct liquid cooling (DLC) and rear-door heat exchangers that transfer heat to building cooling loops without evaporation. The infrastructure challenge is that legacy facilities were not built for liquid-cooled racks, and retrofitting requires significant civil and mechanical engineering investment. New AI factory builds — the greenfield deployments being planned by hyperscalers and large colocation operators — are increasingly specifying waterless designs from the foundation, which will become a competitive differentiator as environmental permitting tightens.

Nvidia has deployed AI agents internally to accelerate ASIC engineering workflows, with Next Platform reporting on the programme's scope. Separately, Semiconductor Engineering details how agent orchestration frameworks targeting 10x productivity improvements are being applied to verification and design closure tasks. The strategic implication is compounding: if Nvidia can shorten its chip design cycle using AI, it can iterate faster on GPU architectures, widening the lead over competitors who do not have the same in-house AI capability or the internal dataset of proprietary chip design history to train on.

This represents a feedback loop that is structurally difficult for competitors to replicate quickly. Nvidia's GPU revenue funds AI research, which funds AI-assisted chip design, which accelerates the next GPU generation. AMD and Intel are pursuing similar programmes but lack Nvidia's scale of internal AI deployment and proprietary training data derived from decades of GPU design.