The decisive comparison is how each supplier turns demand into usable, economical systems. The calls expose gaps between pipeline, committed supply, physical deployment and demonstrated workload performance.
35 companies · 140 quarterly events · 168 selected Q&A probes
Supply assurances refer to different stages: component allocation, equipment delivery, powered-shell completion, utility agreements and service acceptance. CoreWeave identifies shell execution in one delayed project; Equinix identifies power in dense facilities. Those statements concern different constraints and need not conflict.
Decision implication: Require a dated chain of commitments from wafers and memory through integration, energization and workload acceptance. Record the owner and recovery plan at each stage.
Compute vendors increasingly describe orchestration and tool execution alongside accelerator work. Yet Dell says much conventional server growth still comes from established enterprise demand, and TSMC excludes CPUs from its AI accelerator measure. Revenue labels therefore do not identify a universal agentic CPU/GPU ratio.
Decision implication: Size CPU cores, DRAM, storage I/O and network demand from measured agent traces; separate refresh demand from new AI execution.
HBM requirements are rising, while suppliers discuss conventional DRAM, flash, cache offload and stronger agreements. Additional HBM output competes for manufacturing resources with other products. Statements about samples, booked capacity and shipments represent different evidence levels.
Decision implication: Specify bandwidth, capacity, latency, endurance and power by tier. Ask for qualified delivery quantities and contract terms, not only a claim that annual supply is sold out.
Amazon and Microsoft retain multiple silicon choices while investing internally. Oracle emphasizes hardware execution latency; Cisco emphasizes local real-time use cases. These positions reflect different workloads rather than a settled central-versus-edge outcome.
Decision implication: Compare complete request paths: prefill, decode, tools, retrieval, network transit and service-level targets. Include software migration, utilization and fallback costs.
Networking, cooling and power suppliers describe increasing interdependencies. Vertiv explicitly calls the procurement shift gradual and partial. Eaton describes a broader capability set for DC infrastructure, while Schneider expects staged adoption.
Decision implication: Make interface ownership, qualification, protection, maintainability and acceptance criteria part of the system contract. Do not infer universal adoption from a reference design.
Lenovo distinguishes opportunities from orders. Oracle and Meta describe financing structures that change reported capex without eliminating infrastructure investment. Equipment suppliers distinguish manufacturing volume from process intensity and upgrades.
Decision implication: Compare committed versus prospective demand; units versus price; energized MW versus land runway; and all-party investment versus issuer capex. Avoid adding supplier revenues across the same stack.
These are proposed diligence questions, not claims that every analyst asked them. The ranking is an analytical judgment, not a numerical risk score.
| Priority | Area | Question to resolve | Evidence to request |
|---|---|---|---|
| 1 | Capacity delivery | Which quantities are contractually allocated, installed, energized and accepted in each quarter? | A dated milestone table, cancellation rights, responsible party and recovery alternatives. |
| 2 | Fleet economics | What is cost per successful task at a specified latency and quality target? | Matched workload, utilization, software, power, networking and amortization assumptions. |
| 3 | CPU and memory | What does an agent trace consume outside the accelerator? | CPU time, host memory, cache traffic, storage I/O and concurrency distributions. |
| 4 | HBM and packaging | Which customer-qualified products can ship together, and when? | Qualification, yield, package availability and accepted shipment quantities. |
| 5 | Inference placement | Which latency component justifies central, regional or on-device execution? | Measured network, prefill, decode, retrieval and tool latencies, plus privacy requirements. |
| 6 | Power and cooling | Who owns the complete electrical and thermal design across a generation change? | Protection studies, operating envelopes, redundancy, service and integration acceptance. |
| 7 | Demand quality | How much pipeline becomes non-cancellable orders and operating revenue? | Cohort conversion, concentration, financing and deployment dependencies. |
| 8 | Existing asset value | What workloads remain economic on older hardware? | Utilization, maintenance, impairment and workload reassignment evidence. |
Quarters retain the source’s fiscal labels; they are not synchronized calendar periods. Each link opens the transcript used. Dates refer to the call, not the quarter-end. Schneider’s Q1 and Q3 slots are revenue-update calls; its other slots are half-year and full-year results.
| Company | Stack layer | Four calls, oldest first | Progression across the selected questions |
|---|---|---|---|
| NVIDIA NVDA | Compute | CPU orchestration and inference specialization broaden the platform story; requests to rank supply bottlenecks still lack a clear hierarchy. | |
| AMD AMD | Compute | Agentic CPU demand and accelerator ramps strengthen the narrative, while detailed workload comparisons and secured capacity remain open. | |
| Intel INTC | Compute | Foundry discussions move toward readiness and investment, but customer engagement, qualified production and sustained volume remain distinct milestones. | |
| Arm ARM | Compute | CPU positioning broadens into system-level supply questions; architectural efficiency claims still need workload-level validation. | |
| Broadcom AVGO | Compute | Custom accelerators increasingly connect to system delivery, packaging and site readiness; specialization is customer- and workload-dependent. | |
| Marvell MRVL | Compute | The opportunity expands beyond the main accelerator into optics, interconnect and attached memory functions; content mix remains incompletely disclosed. | |
| Qualcomm QCOM | Compute | Roadmap discussion progresses to reported purchase orders and wafer starts; broader adoption still requires demonstrated silicon and customer qualification. | |
| Micron MU | Memory | The discussion broadens from HBM performance to stronger supply agreements, alternative memory tiers and persistent limits on supply growth. | |
| SK hynix 000660.KS | Memory | Higher HBM requirements, stronger agreements and balanced allocation evolve into a more explicit, workload-specific SSD hierarchy. | |
| Samsung 005930.KS | Memory | Earlier customer-qualification confidentiality gives way to ramp and allocation discussion; commercial acceptance still needs customer-specific confirmation. | |
| TSMC TSM | Foundry | Questions shift from translating AI spending into wafers toward shared leading-node allocation among CPUs, GPUs and custom accelerators. | |
| Microsoft MSFT | Cloud | Fleet flexibility persists alongside internal silicon; the unresolved issue is converting spending and efficiency into usable capacity and cohort returns. | |
| Amazon AMZN | Cloud | Trainium adoption depends on software and volume execution; later questions emphasize procurement, inflation and backlog-to-capacity conversion. | |
| Alphabet GOOGL | Cloud | Integrated models and silicon support the strategy, while persistent constraints make compute allocation and investment lead times central. | |
| Meta META | Cloud | Compute supports both existing recommendations and new agents; financing form and buy-versus-sell capacity choices complicate simple capex comparisons. | |
| Oracle ORCL | Cloud | Inference placement and contract design join financing in the discussion; customer or partner hardware can decouple growth from Oracle-funded capex. | |
| Alibaba BABA | Cloud | Management describes priority for enterprise monetization and CPU procurement, then offers a payback model whose assumptions require validation. | |
| Tencent TCEHY | Cloud | Chip availability constrains spending; later discussion distinguishes CPU/network supply and a gradual hybrid device/cloud inference trajectory. | |
| Baidu BIDU | Cloud | Asset impairments highlight obsolescence risk, while proprietary silicon and utilization support management’s subsequent margin argument. | |
| CoreWeave CRWV | Cloud | Powered-shell execution delays precede continued emphasis on flexible AI infrastructure; CPU/storage requirements for inference remain insufficiently quantified. | |
| Dell DELL | Systems | Broad cost pressures become a ranked component shortage discussion; later server growth still largely reflects established enterprise customers. | |
| HPE HPE | Systems | Supply supporting guidance differs from supply meeting demand; price-led server growth must be separated from actual unit delivery. | |
| Supermicro SMCI | Systems | Rack capacity, cooling and component shortages recur; shifting CPU/GPU categories make simple shipment-mix comparisons difficult. | |
| Lenovo LNVGY | Systems | Enterprise inference and cooling guide portfolio changes; management explicitly clarifies that pipeline is not committed orders. | |
| Arista ANET | Network | AI network boundaries blur, but back-end urgency, shipment timing and alternative-component qualification shape near-term delivery. | |
| Cisco CSCO | Network | Edge use cases and custom networking silicon combine with memory redesign and direct supplier engagement; demand remains lumpy. | |
| Vertiv VRT | Power cooling | Thermal and electrical integration grows more important, and larger projects introduce internal logistics and execution dependencies. | |
| Eaton ETN | Power cooling | Higher-voltage DC and cooling expand the discussion from components to complete capabilities, including protection and service. | |
| Schneider Electric SBGSY | Power cooling | Generally normalized equipment lead times coexist with a gradual DC architecture transition; broad electrification also drives growth. | |
| Equinix EQIX | Colocation | Inference interconnection and rack density increase the importance of site-specific power; contracted and expected supply must be separated. | |
| Digital Realty DLR | Colocation | Inference demand spans a wide range of deployment sizes; development runway is not a dated schedule of energized capacity. | |
| ASML ASML | Equipment | AI announcements cannot be mechanically converted into tool demand; volume, process intensity, reuse and productivity all matter. | |
| Applied Materials AMAT | Equipment | The discussion links packaging and device transitions to existing-fab productivity; segment and capacity-growth decompositions remain incomplete. | |
| Lam Research LRCX | Equipment | Facility timing, NAND upgrades and broader packaging adoption shape demand; planning visibility is not equivalent to binding orders. | |
| KLA KLAC | Equipment | Logic, DRAM and packaging create process-control demand; pricing, technology intensity and physical capacity require separate treatment. |
Universe boundary: this covers leading public compute, memory, foundry, cloud, server, network, power/cooling, colocation and equipment companies. It is not every listed supplier. Private model labs and private infrastructure vendors lack comparable quarterly public Q&A. Specialist optical/interconnect suppliers, additional ODMs, storage-only vendors, OSATs and utilities are outside this edition. Examples outside scope include Credo, Astera Labs, Foxconn, Celestica, ASE, NetApp and Pure Storage. Qualcomm is included for its emerging infrastructure role; some of its memory discussion concerns adjacent device markets.
The inventory contains 35 × 4 = 140 quarterly slots. All 140 have transcript sources and selected Q&A coverage. Retrieval and keyword-assisted candidate identification were followed by editorial selection and answer review. The 168 probes are grouped question-level records, with related subparts retained where the answer addresses them together; they are not 168 exhaustively normalized atomic asks. Compute suppliers receive more selected probes than most other companies. Repeated themes retain separate company-quarter provenance and are synthesized above rather than erased.
Source hierarchy: company-hosted material was used where available, supplemented by full transcript mirrors of the actual management exchange. Media-call and prerecorded ASML entries were replaced with analyst calls; a mislabeled Lam Q4 entry and an inconsistent Applied Q1 entry were replaced. Original audio was not systematically replayed. There is no separate online audience-question corpus to reconcile, and absence from the selected ledger must not be interpreted as silence on the call. Translation, speaker-label and transcription errors remain possible; ambiguous attribution is marked. Numerical performance claims and apparent unit errors were not promoted into independently verified specifications.
| Status | Probes / 168 |
|---|---|
| Answered | 75 |
| Partially answered | 91 |
| Deferred | 1 |
| Deflected | 0 |
| Unanswered | 1 |
These counts describe the selected records only. They are not a vendor transparency ranking, audience-interest distribution, unique-questioner count or investor-sentiment measure.
Answered: directly addresses the selected substantive ask, even if qualitative.
Partially answered: provides relevant information but leaves a requested component unresolved.
Deferred: explicitly moves disclosure to a later occasion.
Deflected: substitutes an adjacent issue without addressing the ask; no selected record was assigned this label.
Unanswered: the requested disclosure was not provided in the reviewed exchange, including an explicit refusal to quantify.
A follow-up can be a missing part of the original answer or a new validation requirement proposed here. It does not automatically make a direct answer “partial.” Status applies at that call date; later quarters do not retroactively change it.
| Primary theme | Probes / 168 |
|---|---|
| Supply | 27 |
| Capacity | 25 |
| Inference | 22 |
| CPU | 16 |
| Memory | 16 |
| Custom silicon | 13 |
| Foundry | 9 |
| Power | 9 |
| Platform | 6 |
| Economics | 6 |
| Network | 5 |
| Cooling | 5 |
| Financing | 4 |
| Packaging | 3 |
| Power economics | 2 |
One primary theme per probe; no double counting. These are selection counts, not relative audience interest. “Supply” concerns component or contractual availability; “Capacity” concerns production, deployment or usable output; “Power” and “Cooling” identify the physical subsystem at issue.
Ledger questions and answers are concise paraphrases. Short search excerpts preserve a small piece of original wording to help find the exchange. Source line numbers refer to the captured text and may change on the live page; use the company, quarter, questioner and search excerpt together. The original transcript remains the authority for exact wording and surrounding context.
HBM: high-bandwidth memory. DRAM: dynamic random-access memory. NAND: flash storage. ASIC: application-specific integrated circuit. XPU: generic accelerator label. LTA: long-term agreement. CDU: coolant distribution unit. TCO: total cost of ownership. CSP: cloud service provider. WFE: wafer-fabrication equipment. OSAT: outsourced semiconductor assembly and test. KV cache: stored attention keys and values used during inference. Prefill processes the input context; decode generates subsequent tokens.
Filter locally by company, theme or answer status. Export includes all 168 records, independent of the active filters. Source links open the original transcript page.
168 of 168 probes
Ask: Which constraint could limit growth most?
Management answer: Management identified upstream supply and downstream sites, power and financing; no bottleneck ranking.
Open issue / proposed follow-up: Relative constraint severity.
Ask: What power and content assumptions underpin growth?
Management answer: Management tied rising platform content to generational efficiency and fixed site power.
Open issue / proposed follow-up: Assumptions behind the aggregate forecast.
Ask: Why sell Vera separately?
Management answer: Management described bandwidth-rich CPUs for preprocessing, tools and post-training.
Open issue / proposed follow-up: Independent workload comparison remains needed.
Ask: Which applications drive the next demand wave?
Management answer: Management emphasized agentic coding, followed by physical AI.
Open issue / proposed follow-up: Application adoption and capacity forecasts remain claims.
Ask: Do agentic CPUs replace GPU demand?
Management answer: Management described CPU harnesses and tools generating additional GPU inference; standalone demand was distinguished.
Open issue / proposed follow-up: Fleet-specific CPU/GPU provisioning ratio.
Ask: Why separate hyperscale from other customers?
Management answer: Management emphasized diverse customers needing integrated infrastructure.
Open issue / proposed follow-up: Comparable segment deployment economics.
Ask: Rank the most acute deployment constraints.
Management answer: Management described broad supply-chain pressure without ranking bottlenecks.
Open issue / proposed follow-up: Constraint ranking and relief dates.
Ask: How do agentic workloads and Groq affect inference position?
Management answer: Management promoted platform flexibility; Groq targets high interactivity with throughput/cost tradeoffs.
Open issue / proposed follow-up: Quantified inference share.
Ask: Can components and power support deployment plans?
Management answer: Management described coordinated two-year planning with customers and suppliers.
Open issue / proposed follow-up: Project-specific delivery assurances.
Ask: Is server demand durable beyond seasonality?
Management answer: Management reported broadening hyperscaler CPU plans into 2026.
Open issue / proposed follow-up: Attribution to agentic versus conventional workloads.
Ask: How does Instinct compete with specialized low-latency silicon?
Management answer: Management pointed to chiplet flexibility and workload optimization across inference stages.
Open issue / proposed follow-up: Measured latency and cost comparison.
Ask: What China assumptions and data-center drivers support growth?
Management answer: Management limited modeled China sales and cited EPYC and accelerator ramps.
Open issue / proposed follow-up: Complete growth bridge.
Ask: Can supply support CPU and GPU ramp cadence?
Management answer: Management acknowledged tightness but asserted support for planned growth and coordinated power visibility.
Open issue / proposed follow-up: Committed deployment capacity beyond guidance.
Ask: Why expand the CPU market outlook despite competition?
Management answer: Management attributed faster demand to agent orchestration and data processing.
Open issue / proposed follow-up: Workload-level basis for market and share forecasts.
Ask: When will accelerator revenue overtake CPUs?
Management answer: Management emphasized growth in both and the larger accelerator opportunity.
Open issue / proposed follow-up: Crossover timing.
Ask: What drives the CPU/GPU growth cadence?
Management answer: Management linked added supply to CPU growth and a larger fourth-quarter Helios ramp.
Open issue / proposed follow-up: Exact CPU/GPU contribution split.
Ask: How does NVIDIA collaboration fit inference plans?
Management answer: Management discussed NVLink connectivity and future inference products.
Open issue / proposed follow-up: Detailed product integration.
Ask: How can customer commitments precede capacity investment?
Management answer: Management described engagement and reuse of existing footprint and equipment.
Open issue / proposed follow-up: Binding commitments and qualified output.
Ask: Does unmet server demand migrate to competitors?
Management answer: Management framed demand as an x86 refresh and expected supply improvement.
Open issue / proposed follow-up: Market sizing and share migration.
Ask: When will Intel loosen capacity spending?
Management answer: Management emphasized yield/throughput gains and customer-backed 14A capacity investment.
Open issue / proposed follow-up: Contract details and achieved yields.
Ask: How should agentic CPU demand be modeled?
Management answer: Management offered illustrative GPU/CPU ratios declining from training to agentic workloads.
Open issue / proposed follow-up: Ratios are not a universal fleet sizing rule.
Ask: How does Intel compete with AMD and Arm?
Management answer: Management cited architecture optimization and foundry opportunities without a quantified share answer.
Open issue / proposed follow-up: Comparable performance and market share.
Ask: Does higher capex imply external foundry orders?
Management answer: Management described broad investment, packaging and long-term agreements.
Open issue / proposed follow-up: Customer-specific commitments and allocation.
Ask: What supports external-foundry confidence and spending?
Management answer: Management cited process readiness and PDK milestones; engagement remained distinct from production.
Open issue / proposed follow-up: Binding volume and capex split.
Ask: How does inference expansion affect Arm?
Management answer: Management described diverse inference architectures extending from cloud to edge.
Open issue / proposed follow-up: Deployment mix remains unquantified.
Ask: How is Arm positioned for new AI capacity?
Management answer: Management used power constraints to argue for Arm efficiency and adoption.
Open issue / proposed follow-up: Efficiency assertions need workload normalization.
Ask: What do new memory technologies mean for Arm?
Management answer: Management described ongoing memory and power-efficiency research without a product schedule.
Open issue / proposed follow-up: Specific technology and timing.
Ask: How does inference change the CPU role?
Management answer: Management described more CPU deployment for efficient, low-latency work.
Open issue / proposed follow-up: Workload-specific CPU demand.
Ask: Can supply catch up with AGI CPU demand?
Management answer: Management distinguished demand from secured wafer, memory, packaging and test capacity.
Open issue / proposed follow-up: Incremental supply commitment.
Ask: When might CPU/GPU ratios pass parity?
Management answer: Management emphasized separate CPU capacity beyond GPU head nodes.
Open issue / proposed follow-up: Comparable counting units and timing.
Ask: How will core counts and silicon reporting evolve?
Management answer: Management favored more cores but withheld detailed roadmap commitments.
Open issue / proposed follow-up: Specific roadmap and reporting details.
Ask: What improved confidence in supplying AGI CPUs?
Management answer: Management cited progress across wafers, substrates, memory and testing.
Open issue / proposed follow-up: Supplier-specific allocations remain undisclosed.
Ask: Are key materials secured for the backlog?
Management answer: Management described additional capacity work and advanced packaging investment.
Open issue / proposed follow-up: Component-by-component supply coverage.
Ask: Are new customer deliveries chips or systems?
Management answer: Management explicitly described responsibility for system sales.
Open issue / proposed follow-up: Customer acceptance criteria.
Ask: Could customer-owned designs displace Broadcom?
Management answer: Management defended integrated silicon, SerDes, packaging and networking expertise.
Open issue / proposed follow-up: Displacement magnitude.
Ask: Can Broadcom obtain more wafers and HBM?
Management answer: Management said requirements for 2026–27 were secured and later years under planning.
Open issue / proposed follow-up: Incremental upside capacity.
Ask: Does agentic AI change infrastructure-software demand?
Management answer: Management linked growing CPU core deployment to VMware demand.
Open issue / proposed follow-up: Workload attribution versus licensing effects.
Ask: Why do frontier labs seek direct compute capacity?
Management answer: Management described lab economics and preference for first-party optimized infrastructure.
Open issue / proposed follow-up: Comparable economics and customer mix.
Ask: Which constraints govern XPU deployments?
Management answer: Management identified sites, power, HBM and customer-procured system memory.
Open issue / proposed follow-up: Project-level bottleneck ranking.
Ask: Does Celestial's outlook include memory products?
Management answer: Management included the whole business but expected photonic chiplets to lead.
Open issue / proposed follow-up: Revenue timing remains a forecast.
Ask: Why use cloud capex to model optics?
Management answer: Management called capex a broad proxy; AI optics could grow faster.
Open issue / proposed follow-up: Customer-level optical content.
Ask: What drives custom compute and attached content?
Management answer: Management described participation beyond the main XPU and new customer programs.
Open issue / proposed follow-up: Program-specific revenue allocation.
Ask: Can optics outgrow hyperscaler capex?
Management answer: Management linked growth to accelerator generations and connectivity demand.
Open issue / proposed follow-up: Port and bandwidth mix.
Ask: Are SRAM offload accelerators a design opportunity?
Management answer: Management emphasized SRAM IP and custom-design capabilities.
Open issue / proposed follow-up: Specific wins and performance.
Ask: What does the custom-silicon outlook include?
Management answer: Management distinguished the overall custom market from XPU-only content.
Open issue / proposed follow-up: Customer conversion of forecasts.
Ask: How do memory constraints change custom designs?
Management answer: Management cited CXL expansion and inference-specific acceleration.
Open issue / proposed follow-up: Measured system benefit.
Ask: What products sit within the expanded engagement?
Management answer: Management named inference, NIC, storage, memory-interface and near-memory compute opportunities.
Open issue / proposed follow-up: Contribution by product.
Ask: Why depart from conventional GPU/HBM architectures?
Management answer: Management prioritized inference density, cost and energy through alternative designs.
Open issue / proposed follow-up: Validated performance and software readiness.
Ask: What are deployment economics and ramp timing?
Management answer: Management discussed an earlier revenue ramp but limited comparative economics.
Open issue / proposed follow-up: Normalized system economics.
Ask: How can Qualcomm manage memory scarcity?
Management answer: Management cited broad qualification while noting customers procure most handset memory.
Open issue / proposed follow-up: Handset answer; indirect infrastructure evidence.
Ask: How are data-center programs progressing?
Management answer: Management described disaggregated decode, CPU options and a new memory architecture.
Open issue / proposed follow-up: Customer delivery milestones.
Ask: What is the hyperscaler engagement's scope?
Management answer: Management withheld details until the June investor event.
Open issue / proposed follow-up: Scope and product mix at this call.
Ask: Are wafer and assembly constraints limiting plans?
Management answer: Management reported broad tightness but asserted supply support for its plans.
Open issue / proposed follow-up: Headroom beyond current plans.
Ask: How does prefill/decode specialization change custom silicon competition?
Management answer: Management described specialization by customer, model and workload across its customer roadmaps.
Open issue / proposed follow-up: Compare end-to-end performance and cost; initial question 4 was clarified in question 5.
Ask: How will Qualcomm compete in data-center inference?
Management answer: Management outlined disaggregated inference, CPUs and custom silicon; further roadmap disclosure was scheduled later.
Open issue / proposed follow-up: Product-level comparison and delivery milestones.
Ask: What supports the hyperscaler revenue ramp?
Management answer: Management cited two engagements, purchase orders and wafer starts, with revenue beginning in the December quarter.
Open issue / proposed follow-up: Track shipment acceptance and subsequent ramp.
Ask: How will HBM meet customer bandwidth requirements above standard specifications?
Management answer: Management attributed its capabilities to DRAM architecture, circuit design and the base die.
Open issue / proposed follow-up: Validate customer-qualified bandwidth and power; transcript units are unreliable.
Ask: How firm and broad are memory supply agreements?
Management answer: Management described multiyear DRAM and NAND negotiations with stronger commitments, withholding detailed terms.
Open issue / proposed follow-up: Binding quantities, price mechanisms and cancellation rights.
Ask: What is the opportunity and limitation of high-bandwidth flash?
Management answer: Management acknowledged capacity benefits but identified write-speed, power and retention limitations.
Open issue / proposed follow-up: Benchmark workload suitability against other memory tiers.
Ask: How much worse will memory undersupply be next year?
Management answer: Management expected persistent tightness and explained slow capacity additions without sizing the future gap.
Open issue / proposed follow-up: Demand-supply gap by memory type and delivery quarter.
Ask: How much wafer demand corresponds to a gigawatt of AI capacity?
Management answer: Management declined a conversion because project configurations differ.
Open issue / proposed follow-up: Project-specific chip mix and wafer content; no universal GW conversion.
Ask: Will new capital spending relieve near-term shortages, and is engineering labor limiting?
Management answer: Management placed major new-fab supply later, emphasizing near-term productivity; labor shortage was not quantified.
Open issue / proposed follow-up: Separate productivity gains, new capacity and staffing limits.
Ask: Will CPUs enter TSMC's AI revenue definition?
Management answer: Management said CPUs remain excluded because it cannot reliably identify their destination; historical AI figures were not restated.
Open issue / proposed follow-up: Preserve the narrower accelerator definition in comparisons.
Ask: What is the relative growth and capacity allocation for CPUs and accelerators?
Management answer: Management said they share leading-edge technologies and allocation is coordinated with customers; no specific mix was provided.
Open issue / proposed follow-up: Node-level CPU/GPU/XPU allocations.
Ask: What must Trainium overcome to broaden adoption?
Management answer: Management identified volume delivery, software ecosystem improvement and customer proof points, alongside continued NVIDIA purchases.
Open issue / proposed follow-up: Independent workload and migration-cost comparison.
Ask: How is Rainier progressing and how should its chip count be understood?
Management answer: Management described Anthropic training and additional Trainium use outside Rainier; a complete count reconciliation was not supplied.
Open issue / proposed follow-up: Separate cluster count, other deployments and accepted capacity.
Ask: How will memory and storage inflation affect investment?
Management answer: Management cited early supplier procurement and confidence in supply, but did not isolate the capital-spending effect.
Open issue / proposed follow-up: Price-versus-volume bridge and contract protection.
Ask: How does larger backlog change future capacity and input-cost exposure?
Management answer: Management said known backlog enters planning and discussed contract versus on-demand demand; later capacity was not quantified.
Open issue / proposed follow-up: Backlog-to-commissioned-capacity schedule and cost pass-through.
Ask: How does custom silicon improve efficiency and external monetization?
Management answer: Management linked its integrated stack and TPU/GPU offerings to demand and margin opportunities.
Open issue / proposed follow-up: Incremental economics attributable specifically to silicon.
Ask: When will investment close the compute supply-demand gap?
Management answer: Management expected constraints to persist, emphasizing supply-chain lead times and prior investment.
Open issue / proposed follow-up: Capacity delivery schedule and gap closure date.
Ask: Where will scarce next-generation compute be applied?
Management answer: Management emphasized Search, Gemini and agentic experiences, plus cloud and selected TPU hardware sales.
Open issue / proposed follow-up: Allocation weights and return thresholds.
Ask: How are returns and future spending assessed under supply constraints?
Management answer: Management described broad opportunities and continuing demand above capacity without a quantified return bridge.
Open issue / proposed follow-up: Incremental usable capacity and realized return metrics.
Ask: How does the data-center joint venture affect reported investment?
Management answer: Management distinguished construction spending outside reported capex from Meta's ownership contributions and described flexibility.
Open issue / proposed follow-up: Compare economic commitments rather than capex alone.
Ask: How does more compute translate into advertising performance?
Management answer: Management described transferring knowledge from larger models to smaller runtime models and continuing capacity constraints.
Open issue / proposed follow-up: Marginal performance and inference cost versus compute.
Ask: What is the roadmap for consumer and business agents?
Management answer: Management distinguished nearer-term engagement and advertising work from longer-term interacting agents.
Open issue / proposed follow-up: Deployment milestones and infrastructure requirements.
Ask: Why consider selling compute while buying external capacity?
Management answer: Management described a portfolio of near-term monetization and longer-term internal intelligence investments.
Open issue / proposed follow-up: Timing, chip and workload differences were not disentangled.
Ask: How does fleet strategy support durable infrastructure economics?
Management answer: Management emphasized geographically fungible capacity across research, internal products and external customers.
Open issue / proposed follow-up: Measure actual portability and stranded-capacity exposure.
Ask: How does internal silicon affect inference economics?
Management answer: Management described model-system co-design while retaining external suppliers for fleet-level cost flexibility.
Open issue / proposed follow-up: Comparable workload costs and margin contribution.
Ask: Can physical constraints support the spending ramp and allocation needs?
Management answer: Management expressed procurement confidence but expected tension between internal and Azure demand to persist.
Open issue / proposed follow-up: Accepted capacity by component and customer category.
Ask: How have infrastructure returns and efficiency opportunities changed?
Management answer: Management cited silicon, model choice, token efficiency and fleet operations as levers without a quantitative return comparison.
Open issue / proposed follow-up: Realized returns on successive capacity cohorts.
Ask: How is scarce compute allocated across models, services and GPU rental?
Management answer: Management prioritized higher-value model and inference services over simple GPU rental and described broad component shortages.
Open issue / proposed follow-up: Quantify allocation and realized returns by service.
Ask: How does Token Hub change coordination and priorities?
Management answer: Management emphasized tighter application-model integration for agents rather than a numerical priority hierarchy.
Open issue / proposed follow-up: Operational allocation rules and measurable integration benefits.
Ask: How are resources split between enterprise and consumer AI?
Management answer: Management said most infrastructure currently supports business customers because willingness to pay is clearer.
Open issue / proposed follow-up: Track allocation as consumer demand develops.
Ask: What drove spending and what return supports it?
Management answer: Management cited delivery timing, CPU demand and component inflation, offering a payback framework based on current margins.
Open issue / proposed follow-up: Auditable service-level allocation and payback assumptions; management's framework is not validated return.
Ask: Why impair infrastructure while continuing investment?
Management answer: Management said some existing assets no longer met efficiency needs and described portfolio renewal and utilization improvements.
Open issue / proposed follow-up: Asset-level obsolescence and residual-value assumptions.
Ask: How will AI investment be funded and earn returns?
Management answer: Management cited continuing investment, improving financial metrics and leasing or borrowing alternatives.
Open issue / proposed follow-up: Capital cohort returns and investment schedule.
Ask: How will agents be monetized?
Management answer: Management described application-led value and multiple pricing approaches, expecting value beyond raw tokens.
Open issue / proposed follow-up: Translate application demand into infrastructure requirements.
Ask: How can infrastructure margins improve during investment?
Management answer: Management cited GPU-cloud mix, proprietary chips, utilization and operating scale.
Open issue / proposed follow-up: Isolate each contribution and validate full-stack cost.
Ask: What specifically caused infrastructure delivery delays?
Management answer: Management identified powered-shell delivery rather than power availability for this situation, and described provider diversification.
Open issue / proposed follow-up: Site-specific commissioning milestones; do not generalize its power claim worldwide.
Ask: Will inference economics lead to custom ASIC adoption?
Management answer: Management emphasized customer-led NVIDIA demand without committing to a custom-ASIC roadmap.
Open issue / proposed follow-up: Alternative silicon qualification and economics.
Ask: How do component inflation and capacity expansion affect spending?
Management answer: Management emphasized supply-chain execution and spending preceding revenue; the answer did not fully bridge inflation to capacity.
Open issue / proposed follow-up: Price, volume and commissioning bridge.
Ask: Can the fleet support inference's additional CPU and storage needs?
Management answer: Management said its infrastructure was built for the full AI cycle and could support inference.
Open issue / proposed follow-up: CPU, storage, network and site retrofit requirements.
Ask: How much funding is required for AI capacity?
Management answer: Management described different hardware ownership and supplier rental models rather than a single funding requirement.
Open issue / proposed follow-up: Total economic commitment across financing models.
Ask: Will inference require data centers closer to users?
Management answer: Management argued that hardware execution currently dominates latency and emphasized specialized inference architectures.
Open issue / proposed follow-up: Measure network versus compute latency for each workload and geography.
Ask: How are long-term contracts protected from component inflation?
Management answer: Management described fixed prices when costs are secured and pass-through mechanisms when they are not.
Open issue / proposed follow-up: Confirm coverage and exceptions in actual contracts.
Ask: Can spending slow without limiting growth, and when does cash flow turn positive?
Management answer: Management separated Oracle-funded capex from partner/customer capital and declined a specific cash-flow date.
Open issue / proposed follow-up: Total capital deployed and cash-flow timing.
Ask: Does reduced spending imply lower AI ambition?
Management answer: Management attributed the change to chip availability rather than lower expected token demand.
Open issue / proposed follow-up: Secured supply and delivered capacity.
Ask: How does agentic demand affect cloud pricing amid memory inflation?
Management answer: Management described execution on CPUs, associated memory demand and supply scarcity.
Open issue / proposed follow-up: Specific pricing and cost pass-through.
Ask: Are bottlenecks shifting from accelerators to CPUs and networking?
Management answer: Management differentiated China's accelerator constraints and described securing CPU/network supply through established relationships.
Open issue / proposed follow-up: Supplier commitments and continuing accelerator availability.
Ask: What role will on-device inference play?
Management answer: Management expected a gradual hybrid device/cloud approach, shifting more compute to devices as capabilities improve.
Open issue / proposed follow-up: Workload split, device requirements and transition timetable.
Ask: How will customers respond to component-driven pricing?
Management answer: Management prioritized supply and product mix amid broad shortages, without quantifying demand elasticity.
Open issue / proposed follow-up: Order elasticity and realized price versus cost.
Ask: How can AI server economics hold as demand and memory costs rise?
Management answer: Management reaffirmed its operating-margin framework and emphasized broad demand, without isolating memory cost pass-through.
Open issue / proposed follow-up: Quantify component-cost recovery and margin sensitivities.
Ask: Which constraints matter beyond memory, and in what order?
Management answer: Management identified DRAM, NAND and CPUs first, then hard drives and other components.
Open issue / proposed follow-up: Compare ranked constraints with contracted delivery schedules.
Ask: Is traditional server growth mainly enterprise refresh or AI CPU racks?
Management answer: Management said historical enterprise customers dominate, with emerging neocloud and advanced AI CPU demand.
Open issue / proposed follow-up: Avoid assigning all conventional server growth to agentic AI.
Ask: How is the memory cycle reflected in guidance?
Management answer: Management emphasized supply allocation, procurement capability and business mix.
Open issue / proposed follow-up: Memory-price assumptions and sensitivity.
Ask: Is supply sufficient for guidance, and what memory inflation is assumed?
Management answer: Management distinguished supply supporting its outlook from supply insufficient for all customer demand.
Open issue / proposed follow-up: Quantified memory-price assumptions and incremental allocations.
Ask: Why does storage growth lag servers?
Management answer: Management pointed to migration toward Alletra MP and deferred software revenue within the storage total.
Open issue / proposed follow-up: Normalize platform mix and recognition before comparing growth.
Ask: How much server growth is price versus units, and will supply ease?
Management answer: Management said supply constrained units, with substantial price contribution and stronger units expected while constraints persist.
Open issue / proposed follow-up: Track unit delivery separately from revenue growth.
Ask: What revenue can manufacturing capacity support?
Management answer: Management gave physical rack capacity and discussed Malaysia, but did not convert it into revenue.
Open issue / proposed follow-up: Rack mix, utilization and accepted production.
Ask: Can cooling and system innovations keep arriving ahead of reference designs?
Management answer: Management expressed confidence in earlier integrated delivery based on engineering capability.
Open issue / proposed follow-up: Validated thermals, qualification and actual time-to-operation.
Ask: How much upside is blocked by CPU and other shortages?
Management answer: Management confirmed CPU, memory, storage and some GPU constraints without sizing foregone revenue.
Open issue / proposed follow-up: Component-specific unmet demand and timing.
Ask: How does inference change CPU/GPU shipment mix?
Management answer: Management offered an approximate mix while noting blurred categories for agentic and edge GPUs.
Open issue / proposed follow-up: Consistent unit/revenue definitions and forward mix.
Ask: How does AI back-end traffic affect front-end upgrades?
Management answer: Management said accelerator networks increasingly pressure general cloud networking, making revenue boundaries harder to separate.
Open issue / proposed follow-up: Measure traffic and upgrade needs rather than treating front/back-end categories as fixed.
Ask: Why does network growth not track hyperscaler capex immediately?
Management answer: Management cited deployment sequence, product qualification and shipment timing across years.
Open issue / proposed follow-up: Network acceptance and delivery schedule.
Ask: How does agentic AI change front-end versus back-end demand?
Management answer: Management described strong immediate back-end demand and developing distributed inference and enterprise use.
Open issue / proposed follow-up: A stable traffic ratio and timing of front-end refresh.
Ask: How are upstream decommitments and shortages being addressed?
Management answer: Management described supplier engagement and qualifying alternatives in memory, boards and optics while expecting persistent industry constraints.
Open issue / proposed follow-up: Validate alternate-component qualification and firm commits.
Ask: Where and how will enterprise edge inference deploy?
Management answer: Management emphasized latency-sensitive local use cases and multiple deployment models.
Open issue / proposed follow-up: Workload-specific latency budget and economics versus centralized execution.
Ask: Why do AI orders not rise directly with customer capex?
Management answer: Management pointed to lumpy orders concentrated in a small customer set and incremental use-case wins.
Open issue / proposed follow-up: Compare firm orders and delivery dates instead of capex proxies.
Ask: How is Cisco securing components beyond memory?
Management answer: Management cited silicon commitments, memory-reduction redesigns, migration to newer memory and supplier agreements.
Open issue / proposed follow-up: Distinguish portfolio redesign benefits from AI networking specifics.
Ask: How does Silicon One change control of supply and lead times?
Management answer: Management reported no significant lead-time issues and cited direct TSMC coordination and reduced dependence on merchant-silicon allocations.
Open issue / proposed follow-up: Compare actual deliveries and supplier commitments across product families.
Ask: How does higher-voltage DC change content per megawatt?
Management answer: Management explained fewer conversions and its DC capabilities, but the future architecture's content remained conditional.
Open issue / proposed follow-up: Architecture-specific bill of materials and conversion losses.
Ask: How do cooling changes affect Eaton's opportunity?
Management answer: Management emphasized growth and power/cooling integration and cited evolving accessible content.
Open issue / proposed follow-up: Separate market-content estimates from actual deployed revenue.
Ask: Can Boyd integrate and expand capacity without disrupting delivery?
Management answer: Management described deliberate integration and historically modest capital intensity.
Open issue / proposed follow-up: Specific capacity additions, dates and execution risks.
Ask: How does the DC transition affect Eaton's content and position?
Management answer: Management confirmed its current content estimate and identified transformers, DC breakers, power electronics, cooling and service as required capabilities.
Open issue / proposed follow-up: Future architecture content and efficiency still require design-level validation.
Ask: Can capacity expansion keep up with backlog?
Management answer: Management emphasized existing-site expansion and productivity, declining specific next-year capex.
Open issue / proposed follow-up: Capacity commissioning and throughput by product.
Ask: How will thermal architectures change mix and content?
Management answer: Management expected hybrid thermal chains and continued CDU relevance, dependent on workload and resiliency.
Open issue / proposed follow-up: Validate full-site thermal design rather than extrapolating one GPU configuration.
Ask: Is procurement moving from individual products to integrated systems?
Management answer: Management described a gradual, partial shift toward modules and converged systems with varying customer adoption.
Open issue / proposed follow-up: Define integration responsibility and interface ownership.
Ask: What caused delivery timing shifts and congestion?
Management answer: Management cited project complexity and internal as well as external supply interdependencies.
Open issue / proposed follow-up: Specific delayed milestones and recurrence risk.
Ask: What do different lease sizes reveal about AI demand?
Management answer: Management cited inference-related demand across very large and sub-megawatt deployments.
Open issue / proposed follow-up: Do not map all inference to small edge sites.
Ask: What power is committed and when does development capacity arrive?
Management answer: Management discussed development and preleasing without a complete year-by-year energized-power schedule.
Open issue / proposed follow-up: Utility commitments and energization dates by site.
Ask: How secure is power for new projects?
Management answer: Management described contracted or advanced-stage power arrangements and utility-related investment requirements.
Open issue / proposed follow-up: Separate executed agreements from ongoing discussions.
Ask: When is enterprise inference arriving and why does interconnection matter?
Management answer: Management cited emerging enterprise demand and private connectivity as a deployment enabler.
Open issue / proposed follow-up: Workload-level demand attribution and connection economics.
Ask: Does high-density inference face space, thermal or power limits?
Management answer: Management identified power as the largest constraint and said densification can require reserving adjacent space.
Open issue / proposed follow-up: Usable power and service-level compliance by facility.
Ask: How much expansion has an executed utility agreement?
Management answer: Management grouped contracted power with high-confidence future contracting instead of giving the requested proportion.
Open issue / proposed follow-up: Executed-versus-expected power split and energization gates.
Ask: Are memory customers constrained by fab-shell space?
Management answer: Management said aggregate space appeared available but customer-specific installation and qualification schedules differed.
Open issue / proposed follow-up: Check each customer rather than treating macro capacity as immediately usable.
Ask: Is improving existing-fab output and yield an incremental driver?
Management answer: Management confirmed customer emphasis on productivity and yield while new space takes time.
Open issue / proposed follow-up: Quantify good-wafer output improvements.
Ask: Which segment leads incremental equipment strength?
Management answer: Management described broad AI-related demand across leading logic, DRAM and packaging without ranking them.
Open issue / proposed follow-up: Segment-specific incremental capacity and growth.
Ask: Can AI capacity announcements be translated into EUV demand?
Management answer: Management declined a mechanical conversion and emphasized broader customer participation.
Open issue / proposed follow-up: Firm wafer plans, layer counts and tool productivity.
Ask: Will DRAM architecture changes reduce EUV demand through tool reuse?
Management answer: Management described multi-node planning and process simplification benefits from EUV.
Open issue / proposed follow-up: Net tool additions after reuse and productivity.
Ask: Will EUV equipment become the expansion bottleneck?
Management answer: Management pointed to manufacturing, cycle-time and productivity levers and said it was not currently the bottleneck.
Open issue / proposed follow-up: Compare promised throughput with customer delivery windows.
Ask: How much memory equipment growth is volume versus lithography intensity?
Management answer: Management identified both HBM/DDR volumes and additional layers without quantifying the split.
Open issue / proposed follow-up: Volume/intensity bridge; equipment revenue does not equal bit-capacity growth.
Ask: Does packaging growth include HBM as well as logic?
Management answer: Management expressed confidence in both markets without separating contributions.
Open issue / proposed follow-up: HBM-specific inspection demand and share.
Ask: How does equipment growth differ between DRAM and logic?
Management answer: Management expected faster DRAM growth with both HBM and conventional-memory demand.
Open issue / proposed follow-up: Track actual equipment deliveries and qualified output.
Ask: Is incremental demand from new fabs, migration or yield work?
Management answer: Management described urgent near-term deliveries and broad future greenfield activity, without a precise decomposition.
Open issue / proposed follow-up: Separate capacity additions from technology transitions.
Ask: Are tool price changes cost recovery or value pricing?
Management answer: Management described both costs and delivered value, with new-product and existing-product mechanisms differing.
Open issue / proposed follow-up: Like-for-like pricing, timing and margin bridge.
Ask: What limits acceleration of semiconductor supply?
Management answer: Management said physical facility timing can exceed equipment lead times.
Open issue / proposed follow-up: Customer-specific cleanroom availability and tool acceptance.
Ask: Are NAND migrations slowing, and when does new capacity arrive?
Management answer: Management distinguished DRAM prioritization from continuing NAND upgrades, with capacity additions following.
Open issue / proposed follow-up: Separate technology upgrades from net wafer capacity.
Ask: Do customer memory agreements create equipment commitments?
Management answer: Management described longer planning visibility and operational benefits without confirming equivalent down-payment obligations.
Open issue / proposed follow-up: Binding equipment orders versus planning discussions.
Ask: How is Schneider positioned for higher-voltage DC, and what is the data-center backlog?
Management answer: Management discussed evolving infrastructure requirements but did not provide the requested backlog breakdown.
Open issue / proposed follow-up: Qualified system readiness and comparable backlog disclosure.
Ask: Is strong demand mainly data centers and semiconductors?
Management answer: Management described broader electrification, grid and industrial demand alongside data centers.
Open issue / proposed follow-up: Avoid assigning all electrical growth to AI.
Ask: What is the status of advanced nodes and turnkey memory/packaging?
Management answer: Management described process targets, customer collaboration and integrated capabilities.
Open issue / proposed follow-up: Customer qualification, yields and production volumes.
Ask: How is HBM4 ramping?
Management answer: Management described commercial shipments, booked capacity and a larger second-half contribution.
Open issue / proposed follow-up: Independent customer acceptance and realized ramp.
Ask: How will Samsung balance HBM and conventional DRAM?
Management answer: Management maintained a balanced portfolio objective while expanding HBM and citing future customer agreements.
Open issue / proposed follow-up: Wafer allocation and opportunity cost by product.
Ask: Can HBM meet above-standard customer requirements?
Management answer: Management said enhanced samples and production address higher bandwidth needs and expected requirements to keep rising.
Open issue / proposed follow-up: Verify customer qualification and bandwidth/power together.
Ask: How do new memory supply agreements differ from older ones?
Management answer: Management described stronger mutual commitments than historically flexible volume agreements, withholding contract specifics.
Open issue / proposed follow-up: Binding term, volume, price and cancellation provisions.
Ask: What is HBM4 qualification status, shipment timing and share?
Management answer: Management emphasized customer-aligned ramp readiness and overall competitiveness without the requested detailed schedule or share.
Open issue / proposed follow-up: Customer acceptance, mass shipments and product mix.
Ask: How will SSD products address inference and cache offload?
Management answer: Management described a workload-specific hierarchy of high-capacity QLC, performance TLC and SLC-mode products.
Open issue / proposed follow-up: Measured cache-offload latency, endurance and total cost.
Ask: Has customer HBM qualification completed and what is the roadmap?
Management answer: Management withheld customer-specific qualification under confidentiality, while describing HBM3E sales and HBM4 samples.
Open issue / proposed follow-up: Customer-specific acceptance is not established by shipment growth alone.
Ask: How does liquid cooling support the infrastructure opportunity?
Management answer: Management linked Neptune to energy efficiency and air-to-liquid conversion services.
Open issue / proposed follow-up: Customer-validated energy and deployment results.
Ask: How does restructuring position Lenovo for inference?
Management answer: Management described portfolio simplification, stronger inference offerings and changes to enterprise delivery and sales.
Open issue / proposed follow-up: Verify production adoption and operational savings.
Ask: What customer and training/inference mix sits in the pipeline?
Management answer: Management described broad geographic/customer participation and expected a shift toward inference without a numerical split.
Open issue / proposed follow-up: Pipeline mix and conversion by workload.
Ask: Is the AI pipeline committed orders, and when will it convert?
Management answer: Management explicitly distinguished uncommitted or unqualified opportunities from orders and did not give a conversion forecast.
Open issue / proposed follow-up: Do not add pipeline to backlog; obtain conversion and delivery schedules.
Ask: Is data-center equipment capacity limiting growth?
Management answer: Management said lead times had generally normalized and capacity investment would continue.
Open issue / proposed follow-up: Utilization was not quantified; portfolio conditions differ.
Ask: What is the DC-power architecture roadmap?
Management answer: Management described a gradual rollout, early sidecar solutions followed by more integrated systems and selected partners.
Open issue / proposed follow-up: Qualification dates and integration responsibility.
Ask: Is latency-sensitive AI demand appearing in major metros?
Management answer: Management described broad AI/cloud dialogue and forthcoming capacity, without specific latency-sensitive application examples.
Open issue / proposed follow-up: Named workload requirements and actual latency budgets.
Ask: Are larger enterprise blocks a sign of inference demand?
Management answer: Management described emerging conversations for larger contiguous blocks as inference develops.
Open issue / proposed follow-up: Distinguish discussions from signed, operating inference deployments.
Ask: How fast will advanced packaging grow relative to equipment overall?
Management answer: Management highlighted HBM and 3D stacking plus larger future substrates without quantifying relative growth.
Open issue / proposed follow-up: Packaging growth decomposition and customer qualification.
Ask: What drives the stronger advanced-packaging outlook?
Management answer: Management cited broad HBM and logic packaging adoption, with panel packaging still early.
Open issue / proposed follow-up: Separate technology adoption, pull-ins and sustained volume.