Compute & Infrastructure

Two large compute contracts announced today collectively represent over $10bn in committed GPU infrastructure spend, both denominated in Nvidia GB300 hardware. Argentum AI's $4.1bn deal with an unnamed customer for 27,000 GB300s, reported by Data Centre Dynamics, and SpaceX's $6.3bn agreement with Reflection AI for GB300 access at Colossus 2, also reported by Data Centre Dynamics, together underscore that the GB300 Blackwell Ultra architecture is now the de facto unit of large-scale AI capacity procurement. The anonymity of Argentum's customer is a material unknown — whether this is a sovereign buyer, a large enterprise, or a new AI lab carries significantly different implications for market structure.

Both deals deepen the ecosystem's dependence on Nvidia as the sole credible supplier for contracts of this scale. No AMD, Intel, or custom silicon is referenced in either deal. This is not merely a market preference — it reflects supply chain reality, where TSMC's CoWoS advanced packaging for Nvidia's HBM-integrated GPUs remains the critical bottleneck. The concentration of two multi-billion-dollar contracts on a single GPU architecture, procured through non-hyperscale channels, also signals that the AI cloud market is fragmenting into a larger set of specialised infrastructure operators competing on GPU access rather than software differentiation.

Nvidia has announced a liquid cooling system that raises coolant inlet temperature to 45°C — above conventional direct liquid cooling thresholds — and claims to eliminate facility water consumption entirely by enabling dry cooling at the facility level, according to Tom's Hardware. The higher coolant temperature is strategically significant: it allows the rejected heat to be dissipated via air-cooled dry coolers rather than evaporative cooling towers, which is the primary mechanism by which hyperscale data centres consume millions of gallons of water annually. Reduced electricity consumption is a secondary claim, likely from eliminating chiller plant energy draw.

Tom's Hardware flags that sustainability challenges remain — a critical caveat. The claim of 100% water reduction applies to the cooling loop itself, but manufacturing the hardware, the embodied carbon of high-density GPU infrastructure, and the source of electricity powering the facility are unaddressed. This announcement should be read as a credible engineering advancement on operational water use, not a holistic sustainability solution. For infrastructure operators facing water-use restrictions in Arizona, Texas, and parts of Europe, this is a meaningful operational unlock — but regulatory and environmental scrutiny of data centre footprints is broadening, not narrowing.

Groq has closed $650m in new growth capital led by Disruptive and Infinitum for AI cloud expansion, per Data Centre Dynamics, while on the same day Cerebras dropped approximately 10% after issuing a 2026 revenue outlook that failed to demonstrate it is capturing a material share of the AI data centre market, per Bloomberg. These two datapoints together suggest the alternative inference silicon market is bifurcating: Groq, with its LPU architecture optimised for low-latency inference, retains investor conviction, while Cerebras — whose wafer-scale engine targets large model training — is struggling to translate architectural differentiation into commercial scale.

The Groq raise is a confirmed capital event; the Cerebras guidance disappointment is a confirmed market signal. Neither constitutes evidence that alternative silicon is displacing Nvidia at scale — the GB300 contract announcements elsewhere today make that clear. What it does indicate is that a viable inference-layer market exists for non-Nvidia architectures where latency and cost-per-token economics favour specialised designs. The strategic question is whether these players can secure long-term contracts before Nvidia's own inference-optimised products commoditise that window.

BlackRock-backed AIP, KKR, and Brookfield Asset Management are reportedly among bidders for Stack Infrastructure's Asia-Pacific data centre portfolio, according to Bloomberg. This is an announced process with named institutional participants — it is not speculative that a sale process is underway, though deal terms and ultimate buyer remain unconfirmed. The involvement of AIP — the BlackRock, Microsoft, and Global Infrastructure Partners joint vehicle explicitly created to fund AI infrastructure — alongside traditional infrastructure funds like KKR and Brookfield signals that Asia-Pacific AI compute capacity is now commanding the same institutional attention as North American and European buildout.

The strategic logic for buyers is straightforward: Asia-Pacific AI demand is growing rapidly, greenfield permitting and grid connection timelines in Singapore, Japan, and Australia are extending, and acquiring operational capacity with established power agreements is a faster path to AI-ready infrastructure than ground-up development. For Stack, this represents a potential exit or recapitalisation at a premium to pre-AI valuations. The outcome will set a pricing benchmark for Asia-Pacific AI data centre assets.