AI Chip Investment Outlook

Not investment advice. This page is for informational and educational purposes only. Sterling is not a registered investment advisor. Nothing here is a recommendation to buy, sell, or hold any security. Markets carry risk and you may lose money. Do your own research and consult a licensed professional before making any investment, securities purchase, or trading decision.

Thesis

Why this sector matters to investors right now. Structural, not market timing.

AI infrastructure has become the single largest capex line item in the global economy. Microsoft, Alphabet, Meta, and Amazon guide to a combined 730 billion dollars of capex in 2026 (company guidance midpoints; Microsoft's and Meta's include finance leases), against 376 billion of cash purchases of property and equipment by the same four in 2025 (424.3 billion with Oracle, on its four fiscal quarters to February 2026), so the two totals are not like for like. Management at all four describes AI infrastructure as the majority or dominant share of capex but gives no dollar split (Microsoft FY2027-Q1 call October 28, 2026, Alphabet Q2 2026 call July 22, 2026, Meta Q3 2026 release October 21, 2026, Amazon Q3 2026 release October 29, 2026; company 10-Ks and 10-Qs for 2025). NVIDIA Q4 fiscal 2026 (ending January 25, 2026) reported 68.1 billion in revenue, with data center revenue alone of 62.3 billion (up 75 percent year over year) and full-year fiscal 2026 data center revenue of 193.7 billion versus 115.2 billion the prior year (NVIDIA Form 10-K for fiscal 2026). All four hyperscalers report being supply-constrained, not demand-constrained.

The investible structure splits into five layers: the merchant GPU duopoly (NVIDIA dominant, AMD ramping), the foundry oligopoly (TSMC dominant, Samsung struggling at advanced nodes, Intel Foundry re-emerging), the HBM memory triopoly (SK Hynix leading, Micron rising, Samsung lagging), the lithography monopoly (ASML), and the custom ASIC design layer (Broadcom plus Marvell) that builds the hyperscalers' internal alternatives to NVIDIA. Custom silicon shipments are projected to grow 45 percent in 2026 versus 16 percent for merchant GPUs (Broadcom and Marvell coverage via heygotrade and trefis), meaning the ASIC layer is taking share inside the larger pie. The compute supercycle is the structural story. The competitive question is which layer captures the value as the spend keeps expanding.

Structural drivers

Forces that shape long-run demand and economics. Each driver is sourced.

  • •Hyperscaler capex is unprecedented and rising. 730 billion dollars of combined 2026 capex guidance across Microsoft (about 170 billion), Amazon (about 220 billion), Alphabet (195 to 205 billion) and Meta (135 to 145 billion), plotted at midpoints; Microsoft's and Meta's include finance leases, so the total is not like for like with the 376 billion of cash capex the same four spent in 2025, or 424.3 billion with Oracle (Microsoft FY2027-Q1 call October 28, 2026, Alphabet Q2 2026 call July 22, 2026, Meta Q3 2026 release October 21, 2026, Amazon Q3 2026 release October 29, 2026; company 10-Ks and 10-Qs for 2025). All four report supply-constrained, not demand-constrained.
  • •NVIDIA data center revenue confirms the demand. Q4 FY26 data center revenue of 62.3 billion, up 75 percent year over year, with networking up 263 percent year over year. Full-year FY26 data center 193.7 billion versus 115.2 billion prior year. Source: NVIDIA Q4 FY26 press release, February 2026; NVIDIA Form 10-K for fiscal 2026 (full-year figures).
  • •Memory is the next-tightest bottleneck. The global HBM market is forecast to grow from 38 billion in 2025 to 58 billion in 2026 (Introl Blog). Microsoft attributed roughly 25 billion of its 190 billion 2026 capex specifically to rising memory and component costs (CNBC February 6, 2026).
  • •TSMC N2 (2 nanometer) entered high-volume manufacturing in Q4 2025 at both Hsinchu and Kaohsiung, and TSMC expects a faster ramp in 2026; it gives no monthly wafer capacity for the node. Q4 2025 advanced nodes (7 nanometer and below) were 77 percent of wafer revenue. Source: TSMC Q4 2025 earnings call (January 15, 2026).
  • •ASML EUV shipments expand. ASML recognized 48 EUV systems in 2025 versus 44 in 2024: 44 standard NXE systems for 10,445.8 million euros and 4 High-NA EXE systems for 1,156.9 million euros, an average of about 237 million and 289 million euros of revenue per system (a calculation from reported system sales), and it expects 2026 growth largely driven by a significant increase in EUV sales. Intel's 10-K says its Intel 14A node is being designed to potentially incorporate the first use of high-NA EUV in high-volume logic manufacturing. Source: ASML Form 20-F for 2025, filed February 25, 2026; ASML Q4 2025 results release, January 28, 2026; Intel Form 10-K for fiscal 2025, filed January 23, 2026.
  • •Custom AI ASICs are growing faster than merchant GPUs. Projected 45 percent growth in 2026 versus 16 percent for GPU shipments. Broadcom AI revenue ~12 billion in FY25 with Q3 25 AI revenue 5.2 billion (up 63 percent year over year). Marvell data center revenue 6.1 billion in FY ending February 2026. Source: heygotrade; Broadcom Q3 FY25 earnings.
  • •Compute platform consolidation reduces customer optionality at the silicon layer. Anthropic alone has committed to up to 1 million TPUs via Google Cloud and is also a major Trainium2 customer at AWS (CNBC November 21, 2025). Hyperscaler workloads are becoming the primary buyers of advanced silicon.

Structural risks

Forces that could compress demand, change economics, or break the thesis.

  • •Hyperscaler capex digestion risk. 730 billion of guided 2026 capex (four companies, midpoints, mixed bases) implies revenue scaling that has to materialize. Fortune (April 30, 2026) and CNBC (February 6, 2026) reported the cash impact is starting to show in free cash flow profiles. A multi-quarter air pocket in AI infrastructure spending would compress every layer simultaneously.
  • •China export controls remain volatile. US restrictions on the H20 were tightened in 2025 (requiring licenses) and then partly relaxed (licenses granted). Huawei has plans to double Ascend output in 2026, with US Commerce estimating 200,000 Chinese AI chips produced in 2025 versus Huawei sources at 300,000 to 400,000 (CFR; Council on Foreign Relations; SemiAnalysis). National security policy can swing demand on short notice.
  • •TSMC concentration and Taiwan exposure. TSMC sits at the structural choke point of advanced logic. Any disruption (cross-strait, infrastructure, equipment-export licensing) would cascade through the entire stack. Geopolitical risk is unhedged in the current valuation of the sector's compute layer.
  • •HBM supply is concentrated and qualification-dependent. SK Hynix held 57 to 62 percent HBM market share through 2025; Samsung fell from 41 percent in Q2 2024 to 17 percent in Q2 2025 as it struggled to pass NVIDIA's HBM3E qualification, partly recovering to 22 percent in Q3 2025. Single-supplier exposure to NVIDIA's roadmap is real (Astute Group; Counterpoint Research).
  • •Memory cost inflation is the hidden line item. Microsoft's 2026 capex outlook explicitly attributed 25 billion to memory and component costs. If HBM and DRAM prices keep rising, even at constant unit demand, hyperscaler returns compress (CNBC February 6, 2026).
  • •Custom silicon erodes merchant GPU TAM. Broadcom holds roughly 70 percent of the custom AI accelerator design market (Google TPU, OpenAI's custom chip, Meta MTIA). Marvell ASIC revenue projected at 9 to 11 billion in 2026 (AWS Trainium, Microsoft Maia, Meta DPU). AWS has deployed 500,000-plus Trainium2 chips; Google's Ironwood (TPU v7) shipped November 2025. If hyperscalers internalize more inference and training, NVIDIA growth slows even if total compute keeps rising.
  • •Samsung Foundry has not won meaningful leading-edge AI silicon customers, leaving TSMC as a single source for most advanced AI chips. Intel Foundry has Microsoft Maia 3 on 18A (Tom's Hardware) and possible discussions with Google and Apple, but the timeline is multi-year.
  • •ASML China revenue concentration is unwinding. China fell to 19 percent of ASML system sales in Q1 2026 from 36 percent in Q4 2025 (ASML Q1 2026 results). ASML still guides total 2026 sales of 36 to 40 billion, flat to modestly higher than 2025, supported by non-China EUV demand. A faster-than-expected China step-down would change the equipment-cycle thesis.

Competitive landscape

How to think about the players. Framing along axes (pure play vs diversified, incumbent vs challenger, etc). Not stock picking.

The investible sector sorts into six layers, each with different economics.

1. Merchant accelerators. NVIDIA is the runaway leader: Q4 FY26 data center 62.3 billion (+75 percent), Blackwell in volume production since Q3 FY26, Rubin platform upcoming. AMD Instinct revenue is ramping (MI350 launched 2025; MI400 in 2026 with 432 GB HBM4); AMD targets data center 60-plus percent annual growth and tens of billions in AI revenue by FY27. AMD is projected to take 12 to 15 percent of AI accelerator share by Q4 2026 (Tech Insider).

2. Custom AI silicon designers. Broadcom is the dominant design partner (Google TPU, OpenAI custom chip, Meta MTIA), with ~70 percent of custom ASIC design and a guided 60 to 90 billion serviceable AI opportunity from three hyperscalers by FY27. Marvell is the second-source design partner (AWS Trainium, Microsoft Maia, Meta DPU, Google Axion ARM CPU). Custom silicon volume growing 45 percent in 2026 (heygotrade; The Next Platform).

3. Hyperscaler internal silicon. Google TPU v7 'Ironwood' (November 2025) is the most mature internal stack; Anthropic up to 1 million TPU commitment. AWS Trainium2 has 500,000-plus chips deployed. Microsoft Maia is still internal-only; Maia 3 will tape out at Intel Foundry on 18A. Meta MTIA continues with Broadcom. OpenAI is building custom silicon with Broadcom. These programs reduce hyperscaler dependence on NVIDIA but require multi-year ramp.

4. Foundries. TSMC dominates leading-edge logic: N2 in HVM Q4 FY25, N2 capacity ramping to 100,000 wafers per month by end 2026, A16 in 2H 2026. Samsung Foundry lags at advanced nodes for AI silicon. Intel Foundry 18A wins (Microsoft Maia 3 'Griffin'); Google and Apple are in discussion. SMIC is stuck at 7 nanometer (Council on Foreign Relations).

5. HBM and memory. SK Hynix leads (57 to 62 percent HBM share through 2025), reached HBM4 development completion. Micron gained share (21 percent Q2 2025) shipping HBM3E and sampling HBM4 at 11 Gbps. Samsung struggled with HBM3E qualification, lost share from 41 percent (Q2 2024) to 17 percent (Q2 2025), partly recovered to 22 percent (Q3 2025). UBS projects SK Hynix at ~70 percent share of HBM4 for NVIDIA's Rubin platform.

6. Equipment. ASML holds the EUV lithography monopoly: it recognized 48 EUV systems in 2025, 4 of them High-NA, at about 237 million euros of revenue per standard system and 289 million per High-NA system on average (ASML Form 20-F for 2025, a calculation from reported system sales), and Intel's 14A node is being designed to potentially use High-NA. Applied Materials, Lam Research, KLA, and Tokyo Electron supply etch, deposition, and process control.

7. China alternatives (not investable for most international capital). Huawei Ascend 910B and 910C produce more than the H20 on raw compute and energy but lag on HBM bandwidth. SMIC is the captive logic supplier. Stockpiled foreign HBM remains a constraint (SemiAnalysis; CFR).

Cross-cutting framing: every layer is currently benefiting from the same demand wave. The differentiated bets are within layers (who in HBM, which custom-ASIC designer, which foundry node). The macro question of when (not whether) hyperscaler capex normalizes is the single biggest cross-layer risk.

Key metrics to watch

The operational and financial metrics that matter most in this sector. Each one names its source and update cadence.

MetricSourceFrequencyWhy it matters
Hyperscaler combined capex (Microsoft, Amazon, Alphabet, Meta)Quarterly 10-Q filings and earnings calls; consolidated by S&P Global, Statista, CreditSights, FuturumQuarterlyThe single most important demand-side metric for the entire sector. 730 billion of 2026 capex guidance (four companies, midpoints, mixed bases) is the baseline; any guide-down across the four would compress every layer.
NVIDIA Data Center revenue and Compute versus Networking mixNVIDIA quarterly press releases and 10-QQuarterlyDefines the top line of merchant-GPU economics. Q4 FY26 data center 62.3 billion plus networking 11 billion is the current bar. Mix toward networking is the early signal of system-level (not just chip-level) buying.
AMD Data Center segment revenue and Instinct mixAMD quarterly earnings (10-Q and CFO commentary)QuarterlyCleanest read on whether AMD is taking share inside the GPU duopoly. AMD's Q4 FY25 Data Center revenue was 5.4 billion (+39 percent); MI350 versus MI400 commentary indicates the trajectory.
TSMC monthly revenue and advanced-node wafer mixTSMC monthly revenue release and quarterly earningsMonthly revenue, quarterly node mixLeading indicator for the entire foundry-driven supply chain. Q4 FY25 advanced nodes were 77 percent of wafer revenue (N3 at 28 percent, N5 at 35 percent). N2 ramp pace through 2026 is the central foundry-cycle metric.
HBM market share and shipment volume by supplier (Estimate)Counterpoint Research, TrendForce, supplier earnings (SK Hynix, Micron, Samsung)QuarterlySingle biggest near-term constraint on AI chip output. SK Hynix versus Samsung qualification status with NVIDIA, plus Micron share gains, are the variables to watch.
ASML EUV bookings and High-NA EUV deliveriesASML quarterly press releases and capital-markets-day disclosuresQuarterlyLeading indicator for foundry investment intent 12 to 18 months out. The 2025 baseline is 48 EUV systems recognized (44 NXE and 4 EXE; ASML Form 20-F for 2025) and 7.4 billion euros of EUV in Q4 2025 net bookings of 13.2 billion (ASML Q4 2025 results); for 2026 ASML cites low NA EUV capacity of around 65 systems, with a plan to add 30 percent for 2027 (ASML Q2 2026 results, July 15, 2026). Bookings (not just shipments) signal forward demand.
Custom AI ASIC revenue (Broadcom, Marvell)Broadcom and Marvell quarterly press releases and CFO commentary on AI-specific revenueQuarterlyDirect measure of hyperscaler in-house silicon traction. If custom-ASIC revenue grows faster than NVIDIA's data center revenue, the structural shift away from merchant GPU is accelerating.
China AI chip production volume and US export-control changesUS Department of Commerce announcements; CFR, SemiAnalysis, Rhodium Group reportingContinuous (event-driven), with quarterly synthesisGeopolitical wildcard. Each policy turn changes which suppliers can sell into which markets and at what license cost. Huawei's stated plan to double Ascend output in 2026 is one of several signals.

Catalysts and milestones

Known upcoming events that could move the sector. Dated where possible.

  • •NVIDIA Vera Rubin ramp. Vera Rubin began production shipments in Q3 FY2027 (NVIDIA Form 10-Q, filed August 26, 2026), with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and NVIDIA reports certain supply constraints. Uses HBM4 from SK Hynix (projected ~70 percent share). Next marker is the Q3 FY2027 report, with revenue guided to 108.0 billion, plus or minus 2 percent, and no China Data Center compute assumed. Source: NVIDIA Q2 FY2027 press release and Form 10-Q, August 26, 2026; UBS via Astute Group.
  • •AMD MI400 series (432 GB HBM4) launched, and AMD says Helios begins to ramp in the second half of 2026 (AMD Q2 2026 results, August 4, 2026), with Q3 2026 revenue guided to about 13 billion, plus or minus 300 million. AMD targets 20-billion-dollar data center GPU revenue annual run rate over two years (Digitimes). Source: AMD investor day November 2025; AMD Q2 2026 press release, August 4, 2026.
  • •TSMC N2 ramp and the A16 start. TSMC said on its January 15, 2026 call that N2 entered high-volume manufacturing in Q4 2025 with a faster ramp expected in 2026, and that N2P and A16 volume production are scheduled for the second half of 2026; its 2025 Form 20-F says 2026 capital spending goes mainly to 2-nanometer and 3-nanometer capacity and that A16 risk production is expected in 2026. Source: TSMC Q4 2025 earnings call, January 15, 2026; TSMC Form 20-F for 2025, filed April 16, 2026.
  • •ASML High-NA EUV and Intel 14A. ASML recognized 4 EXE (High-NA) systems in 2025, against 2 in 2024, and expects its TWINSCAN EXE platform to start supporting high-volume manufacturing in 2027. Intel's 10-K says Intel 14A is being designed to potentially incorporate an industry-first use of high-NA EUV lithography in high-volume semiconductor logic manufacturing, and Intel said on its July 23, 2026 call that 14A risk production for internal products remains on track for the second half of 2027. Source: ASML Form 20-F for 2025, filed February 25, 2026; Intel Form 10-K for fiscal 2025, filed January 23, 2026; Intel Q2 2026 earnings call, July 23, 2026.
  • •Hyperscaler 2027 capex guidance announcements through 2026 earnings cycle. Confirmation or moderation of the 730 billion 2026 guidance base (four companies, midpoints, mixed bases). Source: Microsoft, Amazon, Alphabet, Meta quarterly earnings.
  • •SK Hynix began HBM4 mass shipments in Q2 2026 and ramps production in the second half (SK hynix 2Q26 results, July 29, 2026). Samsung initiated mass sales of HBM4 for NVIDIA's Vera Rubin platform (Samsung Q1 2026 results, April 30, 2026) and shipped the first HBM4E samples to major customers in Q2 (Samsung Q2 2026 results, July 30, 2026). Next markers are the HBM4 volume split between the two suppliers and HBM4E qualification. Source: SK Hynix newsroom; Samsung Newsroom.
  • •US export-control updates on China AI chips. H20 license posture, allowed memory configurations, and HBM export rules. Source: US Department of Commerce; Rhodium Group analysis.
  • •Intel Foundry 18A external customer announcements. Microsoft Maia 3 'Griffin' is confirmed; Google and Apple in discussion. Source: Tom's Hardware November 2025; Semicon Electronics.
  • •EU Chips Act and US CHIPS Act award disbursements and milestone-based payouts through 2026. Influences where new capacity lands.
  • •Anthropic, OpenAI, and other AI lab compute commitments. Anthropic's stated up-to-1-million-TPU commitment is a single contract that affects Google TPU economics.

What would change the view

Conditions or evidence that would invalidate the thesis or materially shift the risk picture.

  • •Hyperscaler capex pauses or contracts. If Microsoft, Amazon, Alphabet, or Meta materially guides down 2027 capex versus 2026, the supercycle thesis weakens immediately and every layer reprices.
  • •An AI compute breakthrough reduces training and inference cost-per-token by an order of magnitude beyond expected algorithmic gains. Could shift demand from training to inference, change accelerator mix, and pressure NVIDIA pricing power.
  • •Custom ASIC share crosses 50 percent of accelerator TAM. If Broadcom plus Marvell plus internal hyperscaler programs collectively exceed half the dollar value of AI accelerators shipped, the merchant-GPU thesis structurally weakens. Current custom-versus-GPU split is well below 50 percent.
  • •China achieves sub-7 nanometer logic at scale domestically. Would remove a binding constraint on Huawei Ascend and weaken US leverage. Watch SMIC public disclosures and Huawei chip teardowns.
  • •A geopolitical event affecting Taiwan. The single biggest unhedged risk. Watch shipping insurance rates, strait-transit volumes, and US-China policy posture.
  • •HBM oversupply emerges. If Samsung passes qualification and Micron accelerates HBM4 sampling and shipments, HBM ASPs could compress materially, helping accelerator margins but hurting memory-supplier margins.
  • •NVIDIA loses a major hyperscaler design slot. Watch for any major hyperscaler announcement that internal silicon replaces a previously-NVIDIA workload at scale.
  • •ASML loses meaningful 2026 or 2027 EUV bookings. The leading-indicator metric for the equipment cycle. Sustained book-to-bill below 1 would signal an inflection.

What we are not covering

Sub-areas, technologies, or companies we are deliberately excluding from the analysis, and why.

  • •PC and consumer CPUs (Intel client, AMD Ryzen client). Different demand cycle, different competitive dynamics. Covered in adjacent broader semiconductors coverage if and when added.
  • •Smartphone and mobile SoCs (Qualcomm Snapdragon, MediaTek Dimensity, Apple A-series). Different end market with handset-cycle exposure.
  • •Standalone networking silicon below the data-center fabric layer (Cisco, Arista merchant silicon, NIC chips beyond DPUs). The DPU layer (Broadcom, Marvell, NVIDIA networking) is in scope; lower-layer Ethernet silicon is not.
  • •Power semiconductors (silicon carbide and gallium nitride for power conversion). Important for AI data centers but separate market with different drivers.
  • •Analog, mixed-signal, and microcontroller suppliers (Texas Instruments, Analog Devices, Microchip, NXP, STMicro). Industrial and automotive end-market dynamics not relevant to AI compute.
  • •Display drivers and image sensors. Different end markets.
  • •Wafer-level material commodities and consumables (silicon wafers, photoresist, specialty gases). Adjacent but operates as commodity-cycle businesses.
  • •Quantum and neuromorphic computing. Adjacent technology with much smaller current commercial footprint. Covered in separate Quantum sector when launched.

Sources

Primary sources cited in this analysis. Links open in a new tab.

Audit trail

Record of the last review and what changed. Required on every refresh.

Last reviewed: 2026-05-14
Change log
  • 2026-10-29Event-driven correction, not a full refresh (lastReviewedAt unchanged): the ASML EUV key-metric note, the TSMC N2 and A16 catalyst and the ASML High-NA catalyst were sourced to trade press and are now sourced to filings. The key-metric note quoted 60-plus EUV systems planned for 2026 (TechPowerUp), which neither ASML's 2025 Form 20-F nor its 2026 results releases give; it now gives the 48 EUV systems ASML recognized in 2025 (44 NXE and 4 EXE), the 7.4 billion euros of EUV in Q4 2025 net bookings of 13.2 billion, and ASML's 2026 low NA EUV capacity of around 65 systems with a plan to add 30 percent for 2027. The TSMC catalyst quoted about 100,000 N2 wafers a month by the end of 2026 (TrendForce), which neither TSMC's 2025 Form 20-F nor its 6-K earnings releases give; it now records TSMC's January 15, 2026 call statements (N2 in high-volume manufacturing from Q4 2025, N2P and A16 volume production scheduled for the second half of 2026) and the 20-F wording (2026 capital spending mainly on 2-nanometer and 3-nanometer capacity, A16 risk production expected in 2026). The High-NA catalyst quoted 10 High-NA deliveries in 2026 (TrendForce), which ASML does not state in those documents; it now records 4 EXE systems recognized in 2025 against 2 in 2024, ASML's expectation that EXE supports high-volume manufacturing from 2027, Intel's 10-K wording on 14A and high-NA EUV, and Intel's July 23, 2026 call timing of second-half 2027 for 14A risk production. Three TrendForce source entries and the TechPowerUp 60-plus entry were replaced with the filings. Sources: ASML Form 20-F for 2025 (filed February 25, 2026), ASML Q4 2025 results (January 28, 2026), ASML Q2 2026 results (July 15, 2026), TSMC Form 20-F for 2025 (filed April 16, 2026), TSMC Q4 2025 earnings call (January 15, 2026), Intel Form 10-K for fiscal 2025 (filed January 23, 2026), Intel Q2 2026 earnings call (July 23, 2026), accessed 2026-10-29.
  • 2026-10-25Event-driven correction, not a full refresh (lastReviewedAt unchanged): the thesis, driver, risk, key metric, catalyst and open-question text, and the source list, quoted Tom's Hardware totals of 725 billion dollars of 2026 hyperscaler capex and 410 billion in 2025 (up 77 percent, with roughly 75 percent directed at AI). They now quote the hyperscaler capex file: 712.5 billion dollars of 2026 company guidance for Microsoft, Alphabet, Meta, and Amazon (midpoints; Microsoft's and Meta's include finance leases, so it is not like for like with cash capex), against 376 billion of 2025 cash purchases of property and equipment for the same four (424.3 billion with Oracle on its four fiscal quarters to February 2026). The per-company figures are now Microsoft about 175 billion, Alphabet 195 to 205 billion, Meta 130 to 145 billion and Amazon about 200 billion, and the 77 percent growth and 75 percent AI share were dropped because no primary gives them. The full-year fiscal 2026 NVIDIA data center revenue quoted in the thesis and the key metric was 197.3 billion dollars, which no filing gives; the Form 10-K market-platform table shows 193.7 billion (compute 162.4 billion, networking 31.4 billion), which equals the sum of the four quarters, against 115.2 billion in fiscal 2025. Sources: Microsoft FY2026-Q4 call (July 29, 2026), Alphabet Q2 2026 call (July 22, 2026), Meta Q2 2026 release (July 29, 2026), Amazon Q4 2025 release (February 5, 2026), and Microsoft, Alphabet, Meta, Amazon and Oracle 10-K and 10-Q cash-flow statements (SEC XBRL), NVIDIA Form 10-K for fiscal 2026 (filed February 25, 2026), accessed 2026-10-25.
  • 2026-10-22Event-driven correction, not a full refresh (lastReviewedAt unchanged): the NVIDIA Rubin launch catalyst now records that Vera Rubin began production shipments in Q3 FY2027 (10-Q, August 26, 2026) and points to the Q3 FY2027 report (revenue guided to 108.0 billion); the AMD MI400 catalyst records the MI400 launch and the Helios ramp starting in the second half of 2026 (August 4, 2026); the HBM4 catalyst records SK Hynix HBM4 mass shipments from Q2 2026 (July 29, 2026) and Samsung HBM4 mass sales for NVIDIA Vera Rubin (April 30, 2026), and the resolved open question on Samsung HBM4 qualification was removed. Sources: NVIDIA Q2 FY2027 press release and Form 10-Q, AMD Q2 2026 press release, SK hynix 2Q26 results, Samsung Q1 and Q2 2026 results, accessed 2026-10-22.
  • 2026-05-14Initial publication. All eight required SOP components populated using sourced 2025 full-year and 2026 year-to-date data. Primary sources: NVIDIA Q4 FY26 press release (February 2026), TSMC Q4 2025 earnings transcript, ASML Q3 2025 press release and Q1 2026 results, AMD Q4 FY25 earnings, Counterpoint Research and TrendForce HBM share data, Broadcom and Marvell quarterly disclosures, CNBC and Tom's Hardware coverage of hyperscaler capex, NVIDIA newsroom on DRIVE/Drive Thor, Mobileye and other adjacent OEM disclosures, Council on Foreign Relations and SemiAnalysis coverage of China and Huawei, Rhodium Group on export-controls. All sources accessed 2026-05-14.
Unresolved questions
  • •Confirm whether hyperscaler 2027 capex guidance materially exceeds, matches, or steps down from the 2026 guided 730-billion-dollar baseline as Q3 and Q4 2026 earnings disclose forward views.
  • •Intel Foundry 18A yield and customer ramp. Microsoft Maia 3 is the first major external workload; Google and Apple are in discussion. Each conversion (or non-conversion) materially repositions Intel Foundry.
  • •TSMC N2 customer mix beyond Apple. AMD, NVIDIA, and the hyperscaler custom ASICs are the demand engines for advanced foundry capacity; allocation between them determines who can ship at peak performance through 2027.
  • •China AI chip output reality in 2026. US Commerce projected 200,000 in 2025 versus Huawei sources at 300,000 to 400,000. Actual production is opaque. The number affects whether US export controls are working.
  • •Cash flow versus accrual capex. Hyperscalers report capex on different bases (announcements, accrual, cash). Watch free cash flow and debt issuance to confirm announced capex is actually deployed.

Ask Sterling

Register for a premium account to gain access to Sterling AI.

Get Started

Things you can ask Sterling:

Summarize Tesla's latest earnings reportWhy did NVIDIA's margins expand?Compare Apple vs Microsoft's cash flowWhat's driving EV industry growth?
Menu
Favorites
AI Chip Outlook 2026: Drivers, Risks & Catalysts | Sterling