Compute & Infrastructure
Top Line
Big Tech's combined 2026 capex has reached a projected $725 billion — up 77% from last year's record — with Microsoft alone attributing $25 billion of its AI budget to surging memory and chip costs, signalling that component inflation is now a primary driver of infrastructure spend, not just deployment ambition.
OpenAI claims to have secured 10GW of AI infrastructure capacity ahead of its 2029 target, with 3GW added in the last 90 days alone — a pace of commitment that, if confirmed at the build level, would represent one of the fastest infrastructure ramp-ups in data centre history.
Huawei is forecasting $12 billion in AI chip revenue for 2026 — a 60% increase — as Nvidia's H200 shipments remain stalled in regulatory limbo, accelerating Beijing's push for domestic AI hardware dominance in a market analysts project will reach $67 billion by 2030.
A tightening 2/3nm capacity crunch at TSMC, widening memory shortages projected through 2027, and Apple raising Mac Mini prices due to AI-driven processor supply constraints collectively confirm that semiconductor supply chain stress is broadening beyond GPUs into adjacent components.
Fermi's failure to sign a single data centre client despite promising nuclear-powered AI campuses in Texas underscores the gap between the narrative around nuclear energy for AI infrastructure and the commercial reality of securing power purchase agreements at scale.
Key Developments
Big Tech Capex Reaches $725 Billion — Component Inflation Becomes a Structural Cost Driver
Google, Amazon, Microsoft, and Meta have committed a combined $725 billion in capital expenditure for 2026, a 77% year-on-year increase from what was already a record $410 billion in 2025. The scale alone is significant, but the more analytically important signal is where the cost pressure is originating. Microsoft has explicitly attributed $25 billion of its AI infrastructure budget to increased memory and chip costs — meaning component inflation, not just the number of units being deployed, is inflating the total. Tom's Hardware
Meta's trajectory illustrates the human capital trade-off this spending requires: the company spent $72.2 billion on capex across all of 2025, and its 2026 midpoint guidance would nearly double that in a single year. Mark Zuckerberg has confirmed 8,000 job cuts to fund this infrastructure push, with the company explicitly leaving open the possibility of further headcount reductions. Tom's Hardware This is a structural reallocation — not a cost-cutting exercise — with labour being liquidated to fund physical infrastructure, a pattern that will likely intensify as training and inference workloads grow.
Huawei's AI Chip Ascent and the Bifurcation of the Global Semiconductor Market
Huawei is forecasting $12 billion in AI chip revenue for 2026, a 60% increase, and is on track to become China's dominant AI chip supplier as Nvidia's H200 shipments remain stuck in export control and customs delays. Data Centre Dynamics Tom's Hardware Analyst projections place China's domestic AI chip market at $67 billion by 2030, a figure that — if Huawei consolidates its position — would represent a major shift in global semiconductor revenue distribution away from the US-led supply chain.
The structural risk here is not simply competitive: US export controls are functioning as a forcing mechanism for Chinese domestic chip development, effectively accelerating the timeline at which China builds indigenous capability. Huawei's Ascend chip line remains constrained by access to leading-edge process nodes — it cannot access TSMC's 3nm or Samsung's equivalent — but it is demonstrating that competitive inference performance is achievable at less advanced nodes when software optimisation and system-level integration are prioritised. Separately, Semiconductor Engineering reports that IC tool sales to China have now been stopped, tightening the equipment layer of the supply chain further. Semiconductor Engineering
OpenAI's 10GW Infrastructure Claim and the Invenergy-Nvidia Partnership Signal a New Scale of AI Compute Commitment
OpenAI has announced it has secured 10GW of AI infrastructure capacity ahead of its previously stated 2029 target, with 3GW added in the last 90 days. Data Centre Dynamics This figure requires careful parsing: 'secured' in data centre parlance typically means power purchase agreements, land acquisitions, or signed LOIs — not shovels in the ground. The distinction between contracted capacity and operational capacity is material, and OpenAI has not published a breakdown of how much is under construction versus committed on paper.
Separately, Invenergy, Nvidia, and Emerald AI have announced a partnership on 'flexible AI factories' ranging from edge deployments to multi-gigawatt campuses. Data Centre Dynamics Invenergy's involvement — as one of North America's largest independent power developers — is significant: it suggests Nvidia is moving beyond chip supply into broader infrastructure stack partnerships, positioning itself as a systems integrator for AI compute campuses rather than purely a hardware vendor. SoftBank's parallel plan to launch a US-based AI and robotics firm targeting a $100 billion valuation and a 2026 IPO, with data centre construction as a core activity, adds further capital to the buildout pipeline — though this remains an announced plan with no confirmed construction timelines. Tom's Hardware
Power Architecture and Cooling Innovation as Infrastructure Bottlenecks
Open Compute Project members are advancing a low-voltage direct current (DC) power distribution architecture for AI data centres, targeting efficiency gains by eliminating AC-to-DC conversion losses at the server level. Data Centre Dynamics This is not a new concept, but the OCP coalition backing is meaningful — it suggests hyperscalers are moving toward standardising DC power infrastructure as a design requirement rather than an experiment, which would have significant implications for power distribution unit vendors and facility retrofit costs.
On the cooling side, KAIST has published research on liquid cooling for advanced packages, and the UK data centre sector is being warned that water availability — not just power — will shape the next phase of AI data centre growth. Semiconductor Engineering Data Centre Dynamics Water constraints are a less-discussed but equally binding limit on high-density AI cluster deployments: direct liquid cooling systems reduce air cooling water consumption but require significant facility plumbing investment, and planning authorities in water-stressed regions are beginning to treat data centre water usage as a permitting consideration comparable to power draw.
Semiconductor Supply Chain Stress Broadens: 2/3nm Crunch, Memory Shortages, and Consumer Hardware Price Inflation
Semiconductor Engineering's weekly review flags a tightening capacity crunch at the 2nm and 3nm process nodes, a memory shortage projected to widen through 2027, and the cessation of IC tool sales to China. Semiconductor Engineering These three signals in combination indicate that supply chain stress is no longer isolated to AI accelerator GPUs: it is spreading to the process node capacity that underpins all advanced logic, to memory — a component critical for both training and inference — and to the equipment layer that determines future fab output.
The consumer hardware market is already reflecting this pressure. Apple has raised the Mac Mini starting price from $599 to $799, explicitly citing AI-driven demand draining processor supply as the cause. Bloomberg Qualcomm, meanwhile, has disclosed it has quietly entered the custom hyperscale silicon market, developing a dedicated CPU for an unnamed hyperscaler alongside a new 'agentic' CPU product line. The Register Google's TPU line is now commercially available for external sale, per Semiconductor Engineering, adding another non-Nvidia option to the accelerator market — though TPU availability outside Google's own infrastructure has historically been constrained. These developments collectively indicate the market is responding to NVIDIA's dominance with a broader diversification push, though none of the alternatives yet match H100/H200 ecosystem maturity.
Signals & Trends
Nuclear Power for AI Data Centres Is Failing Its First Commercial Test
Fermi's inability to sign a single client despite a credible pitch around nuclear-powered data centres in Texas is an important early signal that the nuclear-AI narrative is ahead of the commercial infrastructure. The gap between the story — reliable, carbon-free baseload power for always-on AI clusters — and the reality — long development timelines, regulatory uncertainty, and offtake risk that most hyperscalers are unwilling to assume on a speculative basis — is proving wider than the capital markets valued in 2024-2025. This does not invalidate nuclear as a long-term power source for AI infrastructure, but it suggests that the near-term buildout will continue to rely on gas, grid interconnection, and renewable-plus-storage combinations, with nuclear relevant only on a 7-10 year horizon. Infrastructure professionals should treat any hyperscaler 'nuclear partnership' announcement as a long-dated option, not a near-term power procurement solution.
Nvidia Is Transitioning from Hardware Vendor to AI Infrastructure Systems Integrator
The Invenergy-Nvidia-Emerald AI partnership on flexible AI factories — spanning edge to multi-gigawatt campuses — is a structural signal that Nvidia is repositioning itself within the infrastructure stack. Partnering with an independent power developer (Invenergy) and a colocation operator (Emerald AI) gives Nvidia influence over site selection, power procurement, and facility design in ways that a chip vendor historically would not have. Combined with CUDA's ecosystem lock-in and now Nvidia's involvement in campus-level infrastructure design, the company is building leverage at every layer of the AI compute value chain. This creates concentration risk for hyperscalers and sovereign infrastructure programmes alike: the more Nvidia's reference architectures define what an 'AI factory' looks like, the harder it becomes to design around its hardware without accepting significant performance or ecosystem penalties.
The Capex-to-Capacity Gap Is the Central Risk to AI Infrastructure Timelines
The aggregation of announced commitments — $725 billion in hyperscaler capex, OpenAI's 10GW claim, SoftBank's planned IPO vehicle, Invenergy-Nvidia campuses — creates a surface appearance of abundant future capacity. The operational risk is the gap between financial commitment and physical delivery. Grid interconnection queues in the US, UK, and Europe remain multi-year backlogs. Water and power permitting is tightening. Semiconductor supply chain constraints are broadening into memory and advanced logic nodes. And nuclear power, the hoped-for baseload solution, is not commercially validated at the required scale. Senior infrastructure professionals should track the ratio of announced capacity to commissioned capacity across major markets as the leading indicator of whether AI training and inference demand will be met on the timelines that AI model developers are publicly projecting.
Explore Other Categories
Read detailed analysis in other strategic domains