Compute & Infrastructure
Top Line
More than 75 data centre projects worth $130 billion have been blocked in the first quarter of 2026 alone — matching the full-year total for 2025 — as bipartisan community opposition over power and water costs emerges as the most immediate structural constraint on US AI infrastructure expansion.
Nvidia has raised the RTX Pro 6000 Blackwell workstation GPU to $13,250, a 55% increase over original MSRP in under a year, while simultaneously raising the price of its DGX Spark personal AI system to $4,699 — signalling that Nvidia is extracting maximum margin from supply-constrained professional markets.
AMD has entered the personal AI compute market with the $3,999 Ryzen AI Halo, undercutting Nvidia's DGX Spark by $700 with 128GB of unified memory, representing the first credible hardware-level challenge to Nvidia's dominance at the workstation inference tier.
The UK government is reportedly moving to directly purchase AI chips from domestic technology firms as a retention mechanism, confirming that sovereign compute acquisition is now an active industrial policy tool in Europe.
Direct-to-chip liquid cooling is transitioning from a niche deployment to a baseline requirement for AI data centres, driven by accelerator thermal densities that exceed the practical limits of air-cooled infrastructure.
Key Developments
US Data Centre Buildout Faces a Structural Blocking Crisis
A research firm cited by Tom's Hardware reports that over 75 data centre projects representing $130 billion in capital have been successfully blocked in Q1 2026 — a rate that has already matched the entirety of 2025 blockages in three months. The opposition is explicitly bipartisan, with local and state governments from across the political spectrum resisting projects on the grounds of grid stress, water consumption, and infrastructure cost socialisation onto existing ratepayers.
This is the most consequential near-term constraint on US AI infrastructure that does not originate from the semiconductor supply chain. Hyperscalers and developers have announced enormous capacity commitments, but the permitting and grid interconnection layer is emerging as the actual rate-limiter. The disconnect between federal-level encouragement of AI development — including the Trump administration's stated push for domestic AI buildout — and local-level resistance reflects a governance gap that has no fast resolution. Projects blocked at the planning stage represent lost capacity that cannot easily be recovered on a 12-to-24-month horizon.
The Brookfield Head of AI Infrastructure indicated awareness of this dynamic in a Bloomberg appearance during London Tech Week, noting that community and grid integration challenges are embedded in infrastructure investment underwriting. This is confirmed positioning, not speculation: major infrastructure investors are already pricing political and community risk into AI data centre deals.
Nvidia's Pricing Power and AMD's Workstation Counter-Move
Nvidia has raised the RTX Pro 6000 Blackwell to $13,250 at official MSRP, with partner-card pricing starting at $11,359 — a 55% increase over where the card launched, according to Tom's Hardware. Simultaneously, the DGX Spark personal AI system has been repriced to $4,699. This is confirmed pricing, not a market secondary-price anomaly. Nvidia is actively exercising pricing power in professional segments where no competitive alternative previously existed.
AMD's response is the Ryzen AI Halo developer kit at $3,999, based on the Ryzen AI Max+ 395 with 128GB of unified memory, running Windows 11 — directly targeting the DGX Spark's use case at a $700 discount, as reported by Tom's Hardware. The unified memory architecture is a genuine architectural differentiator for inference workloads requiring large context windows or model weights resident in memory. This is an announced product with a confirmed price point, though volume availability and developer ecosystem adoption remain to be established. The competitive pressure on Nvidia's personal AI compute segment is real but early-stage.
UK Sovereign Compute Strategy: Chip Procurement as Retention Policy
The UK government is reportedly preparing to directly purchase AI chips from British technology companies as a mechanism to incentivise those firms to maintain UK domicile and operations, according to a Bloomberg report citing the Telegraph. This is an announced policy intention, not a confirmed procurement programme with contracted volumes or named counterparties. However, the strategic logic is clear: faced with outmigration of AI companies to the US — where compute access, talent density, and capital markets are more favourable — the UK is using state purchasing power as an anchor mechanism.
This approach mirrors sovereign compute strategies being pursued across multiple jurisdictions, including UAE state commitments to NVIDIA clusters, France's domestic AI investment programme, and various national AI compute initiatives under the EU AI Act framework. The UK's framing as a chip-purchase-for-retention scheme is somewhat novel but reflects the same underlying recognition: compute access is now a sovereign competitiveness variable, not merely a commercial procurement decision. The effectiveness of this approach depends entirely on whether the volumes and procurement terms are competitive with what firms could access through hyperscaler relationships or US-based investment.
Cooling Infrastructure: Liquid Cooling Becomes a Baseline Requirement
A detailed analysis published by Data Center Dynamics confirms that direct-to-chip liquid cooling is moving from an advanced deployment option to a functional requirement for AI and HPC infrastructure. The driver is straightforward: modern AI accelerator clusters — particularly NVIDIA H100 and B200-class systems — generate thermal densities that exceed the practical removal capacity of air-cooled raised-floor data centre designs. A rack of H100s can exceed 60kW; air cooling at scale tops out well below that threshold under normal data centre operating conditions.
The capital and operational implications are substantial. Retrofitting existing air-cooled data centres for direct-to-chip liquid cooling requires significant civil works and ongoing management of coolant distribution infrastructure. Greenfield AI-native data centres are now being designed with liquid cooling as a default, but the global installed base of available colocation capacity was built for air cooling. This creates a two-tier capacity market: liquid-cooled AI-ready capacity commands premium pricing and is in short supply, while legacy air-cooled capacity is increasingly unsuitable for frontier AI workloads regardless of power availability.
Signals & Trends
AI Hardware Cost Inflation Is Driving Enterprise Substitution Toward Open-Source and Chinese LLMs
Reporting from Tom's Hardware documents a pattern where enterprise customers are hitting subscription cost ceilings for frontier model access and pivoting to Chinese LLMs and open-source alternatives. The analysis notes that at utilisation rates above 5.7%, OpenAI and Anthropic subscriptions can become loss-generating for the providers — a structural tension between usage growth and unit economics. The demand-side response of substituting toward cheaper alternatives has direct hardware implications: open-source model deployment is an on-premises or self-managed cloud workload, increasing enterprise demand for owned or leased GPU capacity rather than API consumption. This is a secondary demand signal for mid-tier inference hardware — the AMD and consumer-grade Nvidia segments — rather than the hyperscale H100-class clusters, and it will likely accelerate as frontier model pricing continues to inflate.
Nvidia's Pricing Escalation Pattern Suggests Sustained Supply Constraint, Not Normalisation
The 55% price increase on the RTX Pro 6000 Blackwell over approximately one year, combined with the DGX Spark repricing, follows a consistent pattern of Nvidia raising prices on professional hardware rather than holding them stable as capacity nominally expands. If supply were genuinely catching up with demand, competitive pressure — including from AMD and custom silicon deployments — would be expected to cap or reduce pricing. The continued escalation suggests that either TSMC packaging capacity constraints remain binding on Blackwell production, or that enterprise and government demand is absorbing new supply at rates that prevent any competitive pricing equilibrium from forming. Analysts and procurement teams should treat Nvidia's professional hardware pricing trajectory as a real-time indicator of the tightness of the underlying semiconductor supply chain, not merely a margin-maximisation behaviour.
The Federal Government as Direct AI Infrastructure Investor Introduces New Market Distortions
Bloomberg's Wall Street Week coverage notes the US federal government is making direct investments in private technology companies, specifically referencing quantum computing as a current target. This follows the pattern of the CHIPS Act and associated semiconductor investment vehicles, and the UK's chip procurement programme discussed above. The emerging signal is that Western governments are transitioning from regulatory and incentive-based AI infrastructure policy toward direct capital deployment — functioning as institutional investors rather than simply as facilitators. For infrastructure professionals, this matters because government capital often comes with domestication requirements, security constraints, and political continuity risk that differs materially from private capital. It also creates procurement channels and preferred-supplier relationships that can distort competitive hardware markets in ways that are difficult to model from public information alone.
Explore Other Categories
Read detailed analysis in other strategic domains