Compute & Infrastructure
Top Line
Nvidia reported Q2 FY2027 revenue of $96 billion — double year-on-year — and has committed to purchase up to $160 billion in memory, bringing total supply-chain commitments to $279 billion, a figure that signals both extraordinary demand confidence and a deliberate strategy to lock up scarce HBM capacity ahead of competitors.
Vera Rubin, Nvidia's next-generation data center platform, is projected to generate $20 billion in Q3 sales alone — approximately 20% of data center revenue — marking the fastest platform ramp in the company's history and accelerating the obsolescence cycle for Hopper and Blackwell infrastructure.
OpenAI has secured regulatory approval for a 3.2GW power deal in Effingham County, Georgia, with Georgia Power constructing the new generation capacity at OpenAI's expense — a scale of dedicated utility buildout with few modern precedents in the technology sector.
Anthropic has reportedly signed a $45 billion compute capacity agreement with Nscale for 460MW at a West Virginia campus, illustrating how frontier AI labs are now locking in multi-decade infrastructure commitments outside the hyperscaler model.
The Trump administration is weighing expansion of semiconductor tariffs to cover finished products including servers, laptops, and consoles, potentially reversing January's data center exemptions and introducing significant cost uncertainty across the entire AI infrastructure supply chain.
Key Developments
Nvidia's Earnings and Vera Rubin Ramp Reveal the Scale of AI Infrastructure Demand
Nvidia's Q2 FY2027 results — $96 billion in revenue, doubling year-on-year — are more than a financial milestone; they are a demand signal for the entire infrastructure stack. The company's $160 billion in memory purchase commitments, bringing total forward commitments to $279 billion, represents an aggressive move to secure HBM supply at a moment when stack-height scaling is running into serious physical constraints. Reporting from Semiconductor Engineering presented at Hot Chips 2026 details the compounding difficulties: thinner dies increase fragility, greater TSV area reduces usable silicon, thermal dissipation worsens with each additional layer, and manufacturing capacity for high-layer-count HBM remains limited. Nvidia's mass purchasing strategy is partly a hedge against these constraints tightening further.
The Vera Rubin ramp is the other critical signal. A $20 billion Q3 projection for a platform that began shipping this quarter — 20% of data center revenue in its launch quarter — is without historical precedent at this scale, as Tom's Hardware notes. It also confirms that the hyperscalers and large AI labs are pulling forward capacity purchases, accepting supply-chain and integration risk in exchange for early access. Notably, Nvidia reports it is generating more revenue per gigawatt of data center capacity than ever before — a metric that reflects both pricing power and the shift toward denser, more expensive accelerator configurations.
Power Infrastructure Reaches New Scale: OpenAI's 3.2GW Georgia Deal and Anthropic's $45B Compute Contract
OpenAI's 3.2GW power approval in Effingham County, Georgia, confirmed by Data Center Dynamics, represents a fundamental shift in how AI compute infrastructure is procured. Georgia Power will construct the generation capacity, with OpenAI bearing the cost — a model where an AI lab effectively acts as an anchor customer for new utility generation. At 3.2GW, this is not a data center power contract in any conventional sense; it is the equivalent of commissioning multiple large power plants. The precedent matters: it normalises AI companies as direct participants in energy infrastructure financing, bypassing the traditional utility rate-base model.
Separately, Anthropic's reported $45 billion agreement with Nscale for 460MW in West Virginia, covered by Data Center Dynamics, illustrates a parallel strategy — locking in third-party compute capacity at scale rather than owning it outright. Both deals reflect a broader pattern: frontier AI labs are making infrastructure commitments that exceed the capital bases of most industrial companies, raising the question of whether current training and inference demand projections actually justify these positions or whether labs are securing optionality against a competitive worst-case scenario.
Tariff Expansion Threat Introduces Structural Cost Risk Across the Server and Data Center Supply Chain
The Trump administration's reported consideration of extending semiconductor tariffs to finished products — servers, laptops, and gaming consoles — would represent a significant escalation beyond the chip-level duties already in place, according to Tom's Hardware. January's data center exemptions, which shielded hyperscaler procurement from the first round of tariffs, are reportedly under review for removal. This is still at the deliberation stage — not confirmed policy — but the signal is directionally clear: the administration is seeking broader coverage of the semiconductor-embedded product ecosystem.
For infrastructure operators, the exposure is material. AI servers — particularly those containing multiple Nvidia GPUs and HBM memory — represent the highest-value items in the potential tariff expansion scope. Given that a significant proportion of server final assembly occurs in Taiwan and, increasingly, Mexico, tariff structures would either compress integrator margins or force procurement cost increases onto hyperscalers and AI labs already committing hundreds of billions in forward capacity. The timing is particularly acute as Vera Rubin ramp purchases are being finalised.
Custom Silicon Competition Intensifies: OpenAI's Jalapeño and Cerebras' Wafer-Scale Roadmap
Hot Chips 2026 provided substantive technical disclosures on two alternatives to Nvidia's accelerator dominance. OpenAI's Jalapeño ASIC, detailed by Tom's Hardware, does not match Blackwell in raw throughput but delivers superior performance-per-watt and lower latency — characteristics that matter specifically for inference at scale. The chip was notably developed with AI-assisted design tooling. This positions Jalapeño not as a training accelerator replacement but as a purpose-built inference chip that could reduce OpenAI's per-query energy cost and reduce its dependence on Nvidia for inference workloads, which now constitute the majority of deployed AI compute.
Cerebras presented a two-generation roadmap at Hot Chips, including the Nexus rack architecture for CS-4 that triples rack-scale performance, and the forthcoming CS-6 wafer incorporating stacked DRAM — a direct response to the memory bandwidth constraints that limit conventional GPU-based systems, as reported by Tom's Hardware. Wafer-scale integration sidesteps the HBM supply bottleneck by integrating memory differently, but introduces its own manufacturing yield challenges. Simultaneously, the emerging HBF memory standard, presented by Oxmiq Labs at Hot Chips and covered by ServeTheHome, signals that the memory interface layer beneath accelerators is itself under competitive pressure to evolve beyond HBM.
AWS Deepens Nvidia Dependency With 2 Million GPU Expansion; Sovereign and Specialist Operators Accelerate Buildout
AWS has announced deployment of 2 million additional Nvidia GPUs, per Data Center Dynamics, expanding an already dominant partnership at a moment when AWS is also investing in its own Trainium and Inferentia silicon. The scale — 2 million units — dwarfs the GPU fleets of most sovereign AI programs globally and reinforces the concentration risk embedded in the hyperscaler model: AWS is simultaneously AWS's largest customer for custom silicon alternatives and Nvidia's largest cloud reseller. In South Korea, SK Telecom has restructured its infrastructure assets into a dedicated AI data center entity, SK Horizon, with $2.2 billion in committed investment from KKR and IMM, per Data Center Dynamics. This is a confirmed capital commitment, not a proposal, and represents a meaningful sovereign infrastructure play from a major telecom operator seeking to retain AI compute within Korean jurisdiction.
In India, Yotta Data Services is moving toward an IPO to fund expansion of its AI cloud capacity, as reported by Bloomberg. This reflects the broader pattern of non-US operators seeking public markets capital to fund infrastructure that would otherwise require hyperscaler dependency. In the Nordics, startup Fossefall has engaged Armada to deploy five Leviathan modular data centers, per Data Center Dynamics, taking advantage of Nordic power availability and cooling economics. VCI Global's Galatron modular AI data center platform, targeting up to 500MW aggregate capacity per Data Center Dynamics, adds another prefabricated modular entrant to a segment that is growing as speed-to-deployment becomes a competitive differentiator.
Signals & Trends
HBM Supply Is Becoming the Binding Physical Constraint on Accelerator Scaling
Nvidia's $160 billion memory commitment and the technical disclosures at Hot Chips 2026 point to the same bottleneck from opposite directions. On the demand side, Nvidia is pre-purchasing at a scale that suggests it expects HBM supply to remain constrained relative to its accelerator shipment plans through at least 2027. On the supply side, Semiconductor Engineering's Hot Chips coverage documents that each additional HBM layer introduces compounding manufacturing difficulties — thinner dies, increased TSV density, thermal management — that cannot be resolved simply by adding fab capacity. The emergence of HBF as a potential alternative interface, and Cerebras' stacked DRAM integration on CS-6, are early signals that the industry is beginning to route around HBM rather than wait for scaling problems to be solved. Infrastructure operators planning training clusters should treat HBM availability as a near-term capacity ceiling, not a procurement line item.
The Inference Infrastructure Market Is Structurally Different From Training — and Procurement Strategies Have Not Caught Up
OpenAI's Jalapeño disclosure is the clearest signal yet that frontier AI labs are building inference-specific silicon optimised for performance-per-watt and latency rather than peak throughput. This has direct implications for how data center operators should be thinking about rack design, power density planning, and cooling architectures for inference workloads — which are increasingly heterogeneous, latency-sensitive, and continuous rather than batch-oriented. The economics of inference favour lower-power, higher-utilisation configurations that do not match the 10kW+ per GPU rack designs being standardised for training. Operators who build out inference capacity using training-era assumptions — dense GPU racks with maximum HBM — will be left with over-engineered, over-priced infrastructure relative to what purpose-built inference silicon will require within 18 to 24 months.
Tariff and Regulatory Risk Is Becoming a First-Order Variable in Infrastructure Planning
The potential server tariff expansion, OpenAI's customer-funded utility construction model in Georgia, and the proliferation of sovereign compute investments are all symptoms of the same underlying dynamic: the regulatory and geopolitical environment is beginning to reshape infrastructure economics as significantly as technology costs. Infrastructure planners who treat tariff scenarios as tail risks rather than base-case variables are underweighting a policy environment that has already imposed chip export controls, is considering product-level tariffs on servers, and is actively encouraging domestic compute capacity through subsidy and mandate. The practical implication is that procurement and siting decisions made in the next 12 months will be made in a policy environment that is materially different from the one that governed decisions made in 2023 and 2024 — and building in regulatory scenario analysis is no longer optional for capital commitments at this scale.
Explore Other Categories
Read detailed analysis in other strategic domains