Compute & Infrastructure
Top Line
Google has raised its 2026 capital expenditure guidance to $195–205 billion — a confirmed increase from prior estimates — while its Cloud backlog hit $514 billion, signalling that hyperscaler infrastructure spending is accelerating beyond prior projections and tightening capacity constraints for years ahead.
AMD has committed up to $5 billion in Anthropic and will supply up to 2 gigawatts of Instinct MI450 GPUs via its new Helios rack-scale systems, with the first gigawatt scheduled for H1 2027 — the most significant challenge to NVIDIA's data centre dominance yet announced.
Samsung has secured a confirmed $200 billion chip supply contract with Broadcom, reshaping the custom silicon landscape and reducing Broadcom's dependency on TSMC for AI ASIC production at scale.
NVIDIA chips at the centre of a US export control enforcement action, with a White House official accusing China's Moonshot of illegally accessing banned NVIDIA hardware to build the Kimi K3 system — raising fresh questions about the effectiveness of the existing controls regime.
OpenAI has announced plans for a 3.2 GW data centre campus in Effingham County, Georgia, contracting power from Georgia Power — a project that, if built, would represent one of the largest single AI infrastructure commitments in history, though it remains in the planning stage.
Key Developments
Google's Capex Escalation and the Hyperscaler Spending Arms Race
Alphabet has confirmed full-year 2026 capital expenditure guidance of $195–205 billion, a material upward revision that underscores how AI infrastructure competition is overriding traditional financial discipline. Bloomberg reports investor unease is growing despite record Cloud revenue, a tension confirmed by Data Centre Dynamics noting the market's discomfort with compounding commitments. The $514 billion Cloud backlog — confirmed contracted work not yet recognised as revenue — provides the demand-side justification, but it also locks Google into a multi-year buildout trajectory with limited flexibility to decelerate.
The scale of Google's commitment, alongside OpenAI's announced 3.2 GW campus in Georgia and the broader Nvidia-SK Group and AMD-Anthropic deals covered below, points to a collective infrastructure bet that is compressing energy, land, and hardware availability simultaneously. A separate report cited by Tom's Hardware estimates the five largest AI tech companies carry $1.65 trillion in off-balance-sheet data centre obligations — 122% of their stated balance sheet debt — a structural financial risk that is not yet priced into conventional leverage analyses.
AMD–Anthropic Deal Signals Credible NVIDIA Alternatives Emerging at Scale
AMD has confirmed an investment of up to $5 billion in Anthropic and a supply agreement covering up to 2 gigawatts of Instinct MI450 GPUs deployed in AMD's new Helios rack-scale systems. The Verge and Data Centre Dynamics both confirm the first gigawatt is scheduled for H1 2027 — meaning this is an announced plan with a near-term delivery commitment, not speculative. Anthropic is already deploying AMD MI355X GPUs, establishing a production reference base before the larger MI450 rollout.
The Helios rack-scale system is AMD's answer to NVIDIA's NVL-format rack designs and represents a significant architectural bet that AI workloads increasingly require full-rack optimisation rather than discrete GPU procurement. The financial structure — AMD investing into a customer — mirrors NVIDIA's own strategic relationships and signals that hardware vendors are competing not just on specs but on locked-in capital relationships. ServeTheHome has begun publishing benchmark normalisation frameworks comparing NVIDIA Vera to AMD EPYC Turin, indicating the performance comparison conversation is maturing as real deployments approach.
Samsung–Broadcom $200 Billion Contract Reshapes Custom Silicon Supply
Samsung has secured a confirmed contract worth more than $200 billion to manufacture chips for Broadcom, according to Bloomberg. This is the most significant custom silicon supply agreement in recent memory and directly relevant to AI infrastructure: Broadcom is the leading designer of custom AI accelerators (XPUs) for Google and Meta, and this contract signals a strategic diversification away from sole-source reliance on TSMC for advanced node production.
The deal strengthens Samsung Foundry's position in AI ASIC manufacturing at a time when TSMC advanced node capacity — particularly CoWoS advanced packaging — remains the primary bottleneck for AI chip production. Whether Samsung can execute at the advanced nodes required for Broadcom's next-generation XPUs is the critical open question; Samsung's foundry yield at 3nm and below has historically trailed TSMC. If execution holds, this contract meaningfully reduces the concentration risk around TSMC that currently defines the AI hardware supply chain.
NVIDIA Export Controls Under Scrutiny as Moonshot Enforcement Case Emerges
A White House official has formally accused China's Moonshot AI of improperly accessing banned NVIDIA chips to develop the Kimi K3 system, which drew significant attention last week for its advanced capabilities. Bloomberg reports the accusation involves both restricted NVIDIA hardware and US AI models. This is a confirmed US government accusation, not a concluded enforcement action — the investigation and any penalties are pending.
The case is strategically significant for two reasons. First, it demonstrates that export controls on NVIDIA chips are being circumvented at a sufficient scale to enable frontier model development, undermining the primary US policy tool for maintaining AI compute advantages. Second, Kimi K3's performance level, reportedly sufficient to stun the industry, suggests the capability gap being protected by export controls may be narrowing faster than US policymakers had modelled. The case will sharpen debate about whether hardware controls remain viable as a containment mechanism or whether software and model controls need to become the primary instrument.
Energy Constraints and the Power Pledge That May Not Resolve Them
Nearly 200 US utility companies and data centre developers have signed President Trump's 'rate payer protection pledge', committing to shield residential consumers from electricity cost increases driven by AI data centre demand, per The Verge. The pledge is a political signal rather than a binding infrastructure commitment — it does not resolve the underlying grid capacity problem, which is driven by the sheer scale of new load being added.
The Nvidia–SK Group $500 billion partnership announced this week, which includes a 2 GW AI factory component alongside next-generation memory supply agreements per Tom's Hardware, adds to a compounding picture: announced data centre power demand now spans multiple gigawatt-scale projects (OpenAI Georgia at 3.2 GW, AMD-Anthropic at 2 GW, Nvidia-SK at 2 GW) that cannot all be satisfied by the current US grid buildout timeline. ADATA's chairman has separately stated that green power and DRAM will be the two scarcest global resources for the next decade, dismissing AI bubble concerns until 2040–2050, per Tom's Hardware — a view that aligns with the supply trajectory visible in these infrastructure announcements.
Signals & Trends
Capital Commitment Structures Are Becoming Hardware Procurement Mechanisms
The AMD–Anthropic deal — where AMD invests $5 billion into a customer and receives a 2 GW GPU deployment commitment in return — and the Nvidia–SK Group $500 billion partnership represent a structural shift in how compute infrastructure is financed. Hardware vendors are no longer purely selling into a market; they are co-investing in capacity that creates captive demand for their own products. This mirrors the dynamic seen in semiconductor foundry prepayment structures but applied at the hyperscaler-to-chip-vendor level. The implication is that access to frontier compute will increasingly depend on equity relationships and strategic capital alignment rather than open procurement — a concentration dynamic that disadvantages smaller AI labs and new entrants who lack the balance sheet to participate in these arrangements.
DRAM Scarcity Is Emerging as an Underappreciated AI Infrastructure Bottleneck
While GPU supply and power availability dominate infrastructure discussions, ADATA's chairman's assertion of a decade-long DRAM shortage, combined with the Nvidia–SK Group partnership's explicit focus on next-generation memory supply, points to memory bandwidth and capacity as a constraint that will increasingly bind AI training and inference throughput. High-Bandwidth Memory (HBM) supply is already tight; the SK deal suggests Nvidia is moving to secure long-term memory supply at the partnership level rather than relying on spot market procurement. As model sizes and inference batch sizes grow, memory capacity per accelerator — not raw compute — may determine effective throughput for large-scale deployments, making HBM supply chain control a first-order strategic variable alongside GPU and power.
Silicon Photonics Entering the AI Interconnect Stack as a Credible Scaling Path
Salience Labs' optical switch development, covered by Next Platform, is a signal that the AI infrastructure community is actively pursuing photonic interconnects as a response to the bandwidth and energy limitations of copper at rack scale and cluster scale. As AI training clusters grow beyond single-rack configurations and into multi-building campuses — as the OpenAI Georgia and AMD-Anthropic deployments imply — electrical interconnect bandwidth becomes a genuine throughput ceiling. Silicon photonics switches offer lower latency and higher bandwidth density, but integration with existing GPU architectures and packaging stacks remains the unresolved engineering challenge. Salience Labs is pre-commercial, but the direction of investment is consistent with what hyperscalers are signalling privately about their post-2027 interconnect requirements.
Explore Other Categories
Read detailed analysis in other strategic domains