The AI Capex Cycle: From the $800B Buildout to the $1.3T Credit Question — and the Five High-Conviction Positions

The AI capex cycle is the coordinated surge in capital expenditure by hyperscale cloud operators to build the physical infrastructure — data centers, power, accelerators and memory — required to train and serve large AI models. In 2026 the five largest operators — Amazon, Alphabet, Microsoft, Meta and Oracle — have reported or guided a combined $780–800 billion, roughly three times the ~$237 billion deployed in 2024, of which approximately 75% is AI-specific. As of 30 October 2026: Meta from its Q3 release of 21 October, Alphabet from 28 October and Amazon from 29 October — all three left their 2026 guidance unchanged — Microsoft from its FQ4 FY2026 call of 29 July, Oracle from its FY2026 results of 10 June.

Hyperscaler capex tracker — 2026 guidance

Each figure links to the company’s own disclosure · Meta 21 Oct, Alphabet 28 Oct, Amazon 29 Oct 2026 · Microsoft from the Q2 CY2026 cycle · Oracle is a completed FY2026 actual

McKinsey’s research shows global demand for data center capacity could almost triple by 2030, with approximately 70% of that capacity demand driven by AI workloads. The capital split is different: $5.2T of the $6.7T total, or ~78%. Total projected capital expenditure: $6.7 trillion, of which $5.2 trillion is attributable to AI processing loads and $1.5 trillion to traditional IT applications.

The AI capex cycle is defined by the hyperscalers — the companies that build and operate the cloud infrastructure on which AI runs. Amazon, Alphabet, Meta, Microsoft and Oracle collectively spent ~$237 billion on capital expenditure in 2024. The Big-5 estimate for 2025 is $397 billion — the sum of the five companies’ own disclosed actuals, on mixed fiscal bases. Company guidance puts the combined 2026 figure at $780.7–800.7 billion — an ~97–102% increase over 2025 — with approximately 75% of that spend (~$585–600 billion) directed at AI-specific infrastructure: GPUs, servers, data center construction, power systems, and cooling.

Alphabet raised guidance to $195–205B at Q2 (partially due to the Intersect acquisition) and reaffirmed that range unchanged on 28 October; Meta raised its floor to $135–145B, explicitly citing memory price inflation; Microsoft guided to ~$175B CY2026 (lease reclassification, not a cut) ($25B attributed to higher component pricing); Amazon raised to ~$220B. See Exhibit 1 for the full company-level breakdown. Amazon’s 2026 capex of ~$220 billion alone exceeds the combined annual capex of the entire publicly traded US energy sector — and is ~2.5–2.7% of US GDP across the Big-5, against the often-quoted ~5% figure for all AI-related capital formation, which uses a much broader numerator (see Exhibit 6 for the basis).

McKinsey constructed three scenarios ranging from constrained to accelerated demand, shaped by semiconductor supply constraints, enterprise AI adoption rates, efficiency improvements, and regulatory challenges. The base case — $5.2 trillion in AI data center capex — assumes continued growth without runaway acceleration or structural constraints.

Which stocks benefit most from the US data center power gap?

Goldman Sachs expects US data-center power demand to grow 38% in both 2026 and 2027, to 64 GW and 90 GW of year-end capacity (note dated 4 October 2026, via Reuters) — having cut its 2027 figure by 5 GW on community opposition; Morgan Stanley puts the cumulative 2026–28 US shortfall near 57 GW gross and ~32 GW net of “fast-to-power” gas and fuel-cell solutions. On JLL’s published global build rate — roughly 100 GW of new capacity between 2026 and 2030, a 14% CAGR that it says doubles global capacity — supply reaches about 158 GW by 2030 against McKinsey’s 219 GW of demand on a common base.

The gap does not close this decade on either house’s numbers, but its size is far less certain than usually stated: between roughly 20 GW and 60 GW depending on whose 2025 base year you adopt. The investment implication is direct: companies that own existing, entitled, grid-connected data center capacity and energy infrastructure are in a structural scarcity position that cannot be replicated on a 2–3 year horizon. Five positions capture this gap across the supply chain:

EQIX — Equinix

282 data centers across 77 metros globally. North American vacancy ~1.4% (CBRE, H1 2026) gives Equinix structural pricing power on every new lease. Entitled land in super-core markets (Northern Virginia, London, Singapore, Frankfurt) cannot be replicated in under 5 years. The power gap is Equinix’s pricing moat made quantifiable.

CEG — Constellation Energy

Nuclear baseload is the firm carbon-free source best matched to hyperscalers’ 24/7 clean-energy requirements at the scale and timeline they need — not the only one. The US Department of Energy describes geothermal as a steady, high-capacity-factor source suitable for data-center loads, and long-duration storage paired with renewables is advancing. CEG’s case rests on its existing licensed fleet, its locations and its contracts, not on nuclear exclusivity. Hyperscalers cannot build new nuclear — lead time is 10–15 years. Constellation’s existing fleet commands a PPA premium that widens as the power gap expands. The widening power shortfall means more demand for CEG’s contracted output, not less.

VRT — Vertiv Holdings

Every new gigawatt of AI data center capacity requires critical power and thermal management infrastructure. Vertiv’s liquid cooling systems are non-discretionary — AI GPU clusters at 4–8× conventional rack power density (Dell’Oro: ~15 kW/rack today vs 60–120 kW for AI) physically cannot operate without direct-to-chip cooling. Vertiv has not disclosed backlog, orders or book-to-bill since its Q4 2025 release of 11 February 2026, and confirmed on the Q3 2026 call that it does not intend to resume — so the $15.0B backlog and ~2.9× book-to-bill of that release are the last published figures and should not be treated as current. The quantified proxy it does still provide is customer prepayment: current deferred revenue reached $4,120M at 30 September 2026 against $1,815M at 31 December 2025, a 2.3× build in nine months. The Q3 print itself was the strongest evidence: net sales $4,313M, up 61% (55% organic), with full-year organic growth guided to 37–39% after a third consecutive raise. Each gigawatt of gap closed is direct Vertiv revenue.

NVDA — NVIDIA

Every data center that closes the capacity gap must be equipped with AI accelerators. NVIDIA captures an estimated 75–90% of AI accelerator spend — a third-party range, not a disclosed figure; NVIDIA publishes no market share — meaning every gigawatt of new AI data center capacity translates into GPU procurement. NVIDIA now discloses its own revenue per gigawatt of installed AI infrastructure — about $18B/GW for Hopper, $25B/GW for Grace Blackwell and $40B/GW for Vera Rubin (Q2 FY2027 earnings call, 26 August 2026) — so every gigawatt of the gap that does get built is directly measurable NVIDIA revenue. A built gigawatt is measurable revenue for whoever supplies its accelerators — usually NVIDIA, not necessarily NVIDIA. Demand, financing, power, vendor selection and deployment timing are separate conditions, and custom silicon takes a growing share of inference. The gap sizes the opportunity; it does not guarantee the sale.

A.L. Capital Advisory conviction scores: EQIX 23.00/25 · VRT 22.50/25 · NVDA 22.25/25 · CEG 20.50/25. This analysis does not constitute investment advice. See Model Bridge section for full methodology.

The analogy to the late-1990s telecommunications infrastructure bubble is compelling in one dimension — the scale of capital deployment — and misleading in every other. Fiber in the 1990s was built speculatively, with virtually unlimited capacity once laid and zero refresh requirement. Data centers are physically constrained, contractually committed before construction, and subject to accelerated depreciation cycles that naturally absorb any temporary excess. The evidence is visible in vacancy data: North American colocation vacancy has fallen from 9.8% in 2020 to 1.4% by H1 2026 (CBRE, North America Data Center Trends H1 2026, 27 August 2026), while the fiber glut post-2001 saw vacancy exceed 20%.

The key structural difference McKinsey identifies is the cost of carrying excess capacity. Fiber, once laid, is nearly free to maintain. Data centers are the opposite: power, cooling, and maintenance are ongoing high costs regardless of utilization. But crucially, AI accelerators have 3–4 year refresh cycles — meaning any overcapacity is rapidly converted into obsolescence, and new workloads pull spare capacity well before it becomes stranded.

The strongest bear case, stated fairly. By mid-2026 the debate had shifted from whether AI demand is real to the financial mechanics beneath it — and four arguments deserve to be taken seriously. Depreciation: hyperscalers depreciate AI hardware over 5–6 years while its economic life may be 2–3, which Michael Burry and others estimate understates true depreciation by roughly $176 billion across 2026–2028, flattering reported earnings.

Circular financing: interlocking commitments among NVIDIA, OpenAI, Oracle and CoreWeave — vendor equity stakes, take-or-pay compute deals, and debt-funded GPU purchases — can make end-demand look larger and more independent than it is, echoing the Lucent/Nortel vendor-financing dynamics of 1999–2001. Cash-flow strain: PIMCO estimates combined hyperscaler capex will consume ~94% of operating cash flow by 2026–2027, and Moody’s flags ~$662 billion of signed-but-not-commenced data-center leases sitting off balance sheet. Runway: Barclays has flagged a potential hyperscaler cash-runway constraint by late 2026 if the current trajectory holds.

The rebuttal is not that these risks are imaginary — it is that they sit on the buyers’ balance sheets, not the suppliers’. The overbuild question is legitimate; the demand-does-not-exist question is not, with ~1.4% vacancy, sold-out HBM and CoWoS, and 12–18-month GPU lead times. This is precisely why A.L. Capital Advisory’s five High Conviction positions sit on the supply side: NVDA, VRT, EQIX, CEG and MU are paid on delivery, whether or not each hyperscaler’s own AI ROI ultimately clears. The bear case compresses hyperscaler multiples; it does not un-sign the contracts the suppliers are already shipping against.

The AI capex cycle has introduced a structural change absent from prior technology investment waves: the shift from cash-funded to debt-funded infrastructure. Big-5 capex has gone from ~$237 billion in 2024 to a guided $780.7–800.7 billion in 2026 (§01); internal free cash flow cannot scale at that rate. Hyperscalers have collectively turned to debt markets at a scale not seen since the 2001 telecom build-out — but with materially stronger balance sheets underpinning the leverage.

J.P. Morgan Asset Management sizes the requirement at $2.1T of investment-grade bonds by 2030, plus roughly $700B from high yield, loans and securitisation, against $5.5T of AI investment (Jorn Veeneman, 20 August 2026) — materially larger than the ~$1.5T over three years projected in 2025.

Q3 CY2026 — the quarter it stopped looking like timing. Q2 was when self-funding ended; the October 2026 cycle deepened the deficit rather than reversing it. Amazon is now the scale of the thing: trailing free cash flow of −$42.4B against +$14.8B a year earlier, a swing driven by a $94.8B increase in property and equipment. Alphabet’s deficit roughly doubled and now spans the nine months as well as the quarter, Meta turned negative for the first time, and Microsoft — the one name still covering capex out of operations — saw free cash flow fall 27% while operating cash flow rose 29%. Three of the Big-5 are funding the build-out from outside their own operating cash flow, and the combined deficit is no longer a rounding item.

Why this is a timing argument, not a demand argument. The bear case is usually put as “AI capex has no revenue behind it.” The disclosures do not support it, and the October 2026 cycle strengthened the point rather than weakening it. Microsoft’s commercial remaining performance obligation reached $678B, up 84% at Q2 — contracted customer commitments, not forecasts.

Cloud revenue has accelerated in each of the two quarters in which capital expenditure set a record: Azure went +43% at Q2 to +48% in the September quarter, where Microsoft disclosed the dollar figure for the first time at $33.2B; Google Cloud went +82% to +103%, reaching $30.7B, with Alphabet attributing part of it to sales of TPU systems rather than rented capacity. AWS accelerated again, from +37% at Q2 to +39% in the September quarter, reaching $46.0B at a $17.8B segment operating income — its fastest growth in roughly five years. All three of Amazon, Alphabet and Microsoft have described themselves as supply-constrained, which is the opposite of an overbuild. Q3 figures from the Microsoft and Alphabet releases of 28 October and Amazon’s of 29 October 2026.

FactSet’s finding that incremental debt rose from 9% of capex in FY2024 to 32% on a trailing basis by mid-2026 is real, but the same analysis puts total debt to EBITDA at ~1× or below for four of the five. The honest bear case is narrower and harder to dismiss: the depreciation schedule lands in 2027–2028 regardless of how well that backlog converts, and it lands on reported earnings rather than cash. Sources: company filings; FactSet Insight, Kama & Rajput, 23 July 2026.

The implication is dual-edged. Debt-funded capex raises the structural demand floor for suppliers, but rising leverage adds a risk layer to hyperscaler equity itself: if monetisation disappoints, debt service suppresses free cash flow precisely when investor patience is shortest. That is why Microsoft (MSFT) and Alphabet (GOOGL) are rated Selective rather than High Conviction — the infrastructure beneficiaries (NVDA, VRT, EQIX, CEG) capture the upside without carrying the balance sheet risk of the buyers.

Where the debt shift goes next: off the balance sheet entirely. The 2026 answer is not more corporate issuance but a different owner. On 10 August 2026 NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent financing platforms targeting more than $500B of third-party capital — MOUs subject to execution of final agreements, not committed or deployed capital, and as of 21 October 2026 none has converted into a disclosed funded vehicle: no definitive agreement, no per-partner allocation, no named first project and no identified lenders.

The structure was already in use before the announcement: $43.5B across three signed transactions between January and August 2026, with no updated running total published since, beginning with Apollo’s $3.5B capital solution placing institutional capital into ownership of NVIDIA GB200 GPUs under a triple-net lease. That moves the depreciation wall off the hyperscaler income statement and onto an institutional owner — which only holds if compute’s economically useful life is long enough to service the debt written against it. Compute as an Asset Class sets out four frameworks for testing whether it is.

The official sector has moved from commentary to quantification. The Kansas City Fed (5 October 2026) supplies the leverage stack by entity type, and puts AI-firm investment-grade issuance at $330B year-to-date by mid-2026 — ten times the whole of 2023.

The IMF’s October 2026 Global Financial Stability Report quantifies the circularity: 14 disclosed deals since 2023 placing ~$300B of hyperscaler equity and ~$650B of debt into financially weaker AI counterparties in exchange for $1.4T of future revenue commitments. The BIS (14 September 2026) identifies the collateral problem beneath it — tech assets are largely IP, customer relationships and recurring revenue, which are “hard to seize, value or liquidate in standard insolvency.” None of the three forecasts a crisis; all three say the same narrower thing — the financing structure is less transparent than the spending.

The first documented deferral is already on file. Oracle’s 10-Q for the quarter ended 31 August 2026 discloses $288B of data-center lease commitments not yet commenced, on fifteen-to-nineteen-year terms, entirely off balance sheet. The level has been widely reported; the movement has not. In the February 2026 filing those leases were expected to commence between Q4 fiscal 2026 and fiscal 2028; in the August filing the window reads Q2 fiscal 2027 to fiscal 2029 — roughly a year later. What it does not establish: the $288B is an aggregate, so it cannot separate deferrals from newly added commitments — an inference worth drawing, not a confirmed postponement.

In the same quarter it spent $28.5B of capex against $8.5B a year earlier, ran free cash flow negative, and saw remaining performance obligations of $664B grow by just $26B sequentially after $85B the quarter before. On 24 September 2026 it issued a force-majeure notice to Blue Owl’s Stack Infrastructure over the 2.45GW Project Jupiter campus in New Mexico, after the State Land Office twice denied a gas pipeline — though both parties say schedule and commitments are unchanged. Read together, these are the mechanism by which a credit constraint becomes a delivery constraint.

Why are hyperscaler credit default swap (CDS) spreads widening, and what does it mean for the five positions?

Because the buyers are now externally financed, and the credit market has started to price that separately from the equity. J.P. Morgan Asset Management put hyperscaler cash spreads at around 105bp against roughly 80bp for the broader market, having widened about 30bp year-to-date, and noted hyperscaler issuance rising from 2% of total USD investment-grade supply across 2022–24 to an expected 9% in 2026 (Jorn Veeneman, 20 August 2026).

Barclays framed it more starkly: the best hyperscalers trade about ten basis points wide of the index, where a year earlier they traded fifty through it, and at the long end “all the hyperscalers trade on top of the BB index” (28 July 2026).

The single-name derivative market has moved much further, and the tiering is the finding rather than the level. A credit default swap is the cleanest read on how the market prices an issuer’s default risk, and hyperscaler CDS has become the instrument through which AI credit stress is expressed. Press-relayed composites from S&P Global Market Intelligence and LSEG put five-year CDS in mid-to-late October 2026 at roughly 1,080–1,190bp for CoreWeave, ~315–322bp for Oracle (from ~144bp at the start of 2026), and ~100–110bp for Meta and NVIDIA at record wides.

No rating agency acted during the move: S&P still carries CoreWeave at B+, and Oracle’s cut to BBB− came in July 2026, before the widening. The dispersion is wide enough that the market is clearly distinguishing between these balance sheets; Exhibit 13 sets out the tiering and the limits of reading it.

The transmission into positioning runs through who gets paid before the credit question resolves. NVDA, MU and VRT are paid on delivery against purchase orders and, in Micron’s case, take-or-pay agreements with cash deposits. They are exposed to the pace of the build rather than to the buyers’ cost of capital — though not insulated from it: NVIDIA discloses $108.5B of maximum gross guarantee exposure and days sales outstanding extending to 60 from 45, and a buyer whose financing cost rises can defer orders as readily as default on them.

EQIX is paid on signed leases with 80–95% of under-construction capacity pre-committed, which turns tenant credit stress into a re-letting question in a market at 1.4% vacancy. CEG is paid under 15–20-year PPAs with investment-grade counterparties and is the least exposed of the five; the hyperscalers and neoclouds carry the financing risk most directly. The asymmetry is one of timing and channel, not exemption, and it is why this framework rates suppliers High Conviction and buyers Selective. It also sets the falsification condition: if wider spreads start deferring contracted capacity rather than repricing it, the supplier thesis weakens.

Is AI infrastructure now energy-constrained rather than capital-constrained?

Yes — and it is why US data center power demand has become the most-watched number in the build-out. Through 2024 the binding constraint was capital and GPU supply; by 2026 it was reliable power at scale. §01 sets out the interconnect, transformer and permitting lead times behind that. Two 2026 datapoints show the constraint is priced, not merely described.

PJM’s 2028/29 base residual auction cleared at the $325/MW-day price cap for the third consecutive year and still finished 6,831 MW short of its reliability requirement, with PJM attributing a 1,374.5 MW increase in its own forecast “largely due to additional Large Loads.”

And the equipment is spoken for: GE Vernova reported its gas turbine capacity “mostly sold out through ’30”, with more than 125 GW under contract and an Electrification backlog of $41B, up 69%. A capacity market clearing at its cap while short, and a turbine book sold out for four years, are the cleanest available proofs that power is priced as scarce rather than merely described as scarce.

The positioning consequence is developed in §07: companies controlling existing grid-connected capacity (EQIX’s 282 operating data centers) and firm dispatchable power (CEG’s nuclear fleet, the only carbon-free source already licensed and running at this scale) hold a scarcity position that cannot be replicated in under five years — and it widens under an energy-constrained regime.

This analysis is updated each quarter as the hyperscalers report. The Portfolio Fragility Brief carries the revisions — what changed, what was wrong, and what it means for positioning.

In January 2025, OpenAI, SoftBank, and Oracle announced Project Stargate — a $500 billion AI infrastructure programme targeting US deployment, with an initial $100 billion committed within the first four years. Stargate sits in a category McKinsey’s April 2025 work treats only at a high level: sovereign and government-adjacent AI infrastructure, funded at national-strategy scale rather than commercial return logic alone. McKinsey does name government and geopolitical investment as a demand driver, so the right framing is emphasis rather than omission. By mid-2026 the programme was simultaneously advancing and re-sequencing: OpenAI, Oracle and SoftBank added new US sites (including a ~$16 billion campus in Saline, Michigan), while OpenAI walked away from a planned 600 MW Abilene expansion to hold out for next-generation NVIDIA silicon — a reminder that even sovereign-scale demand is now gated by chip-generation timing, not by capital.

Stargate is not an isolated event. Saudi Arabia’s PIF has committed $40B, the UAE runs G42, and the EU’s AI Continent Action Plan targets €200B through 2030 — demand that is contractually committed and largely insensitive to short-term ROI. One correction to the record: OpenAI’s widely quoted “$1.4 trillion of compute commitments” has been superseded by a target of ~$600B of total compute spend by 2030, and $1.4T now more often refers to OpenAI’s valuation. Its revenue run-rate was ~$50B at end-September 2026, not the ~$70B widely reported, because OpenAI books partner sales net.

The structural implication for the AI capex cycle is clear: sovereign programmes are a second demand layer with a different funding logic — national strategy rather than commercial return alone. Two cautions, both previously omitted here. McKinsey’s compute value-chain research does name government and geopolitical investment among its demand drivers, so sovereign demand is not wholly outside the base case; and announced programme values overlap with Oracle’s and the hyperscalers’ own capex, so they cannot be added to a global forecast without double-counting. The defensible conclusion is directional: sovereign demand raises the floor, by an amount this paper does not size. It does not quantify the total demand gap when sovereign programmes are included.

Among the five high-conviction positions, Constellation Energy (CEG) is the most direct Stargate beneficiary: nuclear baseload is the only power source that meets both the carbon-free and uninterruptible specifications that sovereign AI programmes require. Equinix (EQIX) benefits through its hyperscale campus footprint in the geographies where sovereign programmes are concentrating — Virginia, London, Singapore, and the Gulf. NVIDIA (NVDA) benefits from GPU procurement at sovereign scale. Vertiv (VRT) benefits through the thermal management requirements of the dense compute clusters Stargate’s architecture demands.

McKinsey’s analysis maps the $5.2 trillion AI capex envelope across five distinct investor archetypes. Understanding this architecture is essential: the investment case, risk profile, and return dynamics differ fundamentally across archetypes.

The strongest argument against this page’s own thesis is not the bubble case but the efficiency case. The IEA finds that electricity consumed per comparable AI task has fallen by at least an order of magnitude annually in recent years, and NVIDIA’s own figures point the same way — Vera Rubin delivers roughly 30× the throughput per megawatt of Grace Blackwell Ultra at about one-thirty-fifth the cost per token. If efficiency compounds faster than workloads expand, the capacity shortfall closes itself and the scarcity premium embedded in EQIX and CEG compresses. The counter is Jevons: every prior order-of-magnitude fall in compute cost raised total consumption rather than lowering it, and inference demand is currently scaling faster than efficiency is improving. That is an empirical bet, not a certainty, and it is the single assumption most worth monitoring.

The $5.2–$6.7 trillion capex envelope flows through a defined set of public equities. But raw exposure to the AI theme is not sufficient — the archetype, moat, and balance sheet quality of each company determine whether they capture compounding returns or get crushed in the shake-out.

Every dollar of hyperscaler AI infrastructure capex ultimately flows through a semiconductor. The $660–690 billion committed for 2026 does not build itself — it must become physical chips, packaged onto boards, slotted into racks, and cooled. The supply chain is the root constraint on whether the cycle delivers McKinsey’s projected $6.7 trillion of capacity by 2030. Three chokepoints define it: GPU allocation, advanced packaging at TSMC, and the EUV lithography monopoly held by ASML.

NVIDIA GPU Shortage — Why Its Accelerator Market Share Persists Despite Competition

NVIDIA Corporation (NASDAQ: NVDA) enters 2026 in a position with few historical precedents: a near-monopolist in a market growing at 36% annually, constrained not by demand but by its own supply chain.

The Blackwell B200 architecture — 2.5× the inference throughput of the H100 at the same power envelope — carries reported order lead times of supply-constrained through at least end-FY2028 on NVIDIA’s own guidance for hyperscaler allocations as of May 2026. The binding constraint is CoWoS-L (Chip-on-Wafer-on-Substrate with Local) advanced packaging rather than TSMC fab capacity for the die itself: its scale is not publicly known. TSMC has never disclosed a CoWoS capacity figure and reconfirmed that non-disclosure on its Q3 2026 call (15 October 2026); the third-party estimates in circulation diverge by roughly 60% for 2027, so this paper asserts no level. TSMC’s CEO has said only that advanced packaging is “sold out,” and that demand and supply remain “quite big” apart.

NVIDIA’s Vera Rubin architecture (H2 2026 full production) is confirmed, with the first rack already running at Microsoft Azure. Vera Rubin moves to HBM4 at 22 TB/s per GPU — 2.75× Blackwell — and delivers a 10× inference cost reduction, extending the CUDA platform advantage through the 2027–2028 cycle.

The CUDA software moat is as material as the hardware lead. NVIDIA's Compute Unified Device Architecture — the programming model underlying virtually every production AI training workload — has been in continuous development since 2007. The libraries (cuDNN, cuBLAS, TensorRT), the developer tooling, and two decades of academic and commercial code written natively for CUDA constitute a switching cost that Advanced Micro Devices (NASDAQ: AMD) is dismantling only slowly with ROCm. A.L. Capital Advisory estimates enterprise migration from CUDA to ROCm at current pace would require 3–5 years for non-latency-sensitive inference workloads, and meaningfully longer for training.

AMD vs NVIDIA in 2027 — does the OpenAI warrant change the share maths?

The challenger case no longer rests on a chip comparison; it rests on two warrants. Under an agreement first disclosed in October 2025, OpenAI is to deploy 6 GW of AMD Instinct accelerators, with the first 1 GW on MI450 in the second half of 2026 — and only that first gigawatt is binding. Alongside it AMD issued a warrant for up to 160 million shares at $0.01, vesting in tranches tied to deployment volume and to a share-price ladder rising to $600, exercisable to October 2030. In February 2026 AMD issued Meta a near-identical warrant. Together those represent roughly 320 million shares — about 19.6% of shares outstanding — against some 12 GW of prospective commitments. As of the quarter ended 27 June 2026, no warrant shares had vested.

That structure cuts both ways. It converts a share-gain thesis into something closer to a contingent partnership: AMD buys volume with equity, and existing holders pay for the share gain through dilution rather than receiving it as upside. It is also the clearest instance of the vendor-financing pattern running through this cycle — supplier equity or credit extended so the customer can buy the supplier’s product, visible at NVIDIA with SB Energy and at Broadcom, which agreed in October 2026 to lend Anthropic up to $42B against chip residual value. A.L. Capital Advisory keeps AMD at Selective (16.25/25): the deployment is real and MI450 is competitive, but the economics reach shareholders only if volume vests, and the dilution is certain in a way the volume is not. Q3 2026 results are due 3 November 2026 against company guidance of ~$13B.

Intel, Arm and the CPU layer

CPUs matter at the margin: every GPU cluster needs hosts for ingestion, orchestration and inference serving. AMD’s EPYC has taken data-center socket share from Intel for three consecutive years, and Arm is the deeper structural story — its v9 architecture, licensed to Amazon (Graviton), Google (Axion) and NVIDIA (Grace and now Vera), delivers materially better performance-per-watt than x86, and Arm earns a royalty on every chip shipped. NVIDIA disclosed in August 2026 that Grace trailing revenue exceeded $5B and guided CPU revenue to more than double in FY2028. In a cycle where power is the binding constraint, performance-per-watt compounds — but none of the three is a High Conviction position in this framework.

ASML — The Single-Point-of-Failure in Global AI Chip Supply

ASML Holding N.V. (NASDAQ: ASML) manufactures every extreme ultraviolet lithography machine on Earth; there is no second supplier. EUV is required to pattern the sub-7nm transistors in every leading-edge AI chip, tools cost roughly €200M each (High-NA ~€380M), and ASML ships only about 50–60 systems a year on 18–24 month lead times. That makes it the deepest chokepoint in the chain: every new AI fab — TSMC Arizona, Samsung Taylor, Intel Ohio — needs ASML tools, and no competitor has a working EUV machine.

The investment case for ASML is the AI capex cycle expressed through the supply chain’s deepest chokepoint. Every new fab built for AI demand — TSMC Arizona, Samsung Taylor, Intel Ohio — requires ASML EUV machines. The order book extends through 2027 and includes High-NA EUV tools (approximately €380 million per unit) required for sub-2nm nodes. No competitor has a functioning EUV tool; industry analysts estimate 15–20 years and multiple billions to build one from scratch. The Dutch government’s US-pressured restriction of EUV exports to China removes the largest potential demand overhang and creates a China-exclusion premium benefiting Western fabs.

Custom Silicon — Broadcom & Marvell as the ASIC Layer

Custom silicon is one of the most consequential supply-chain developments of the 2026 cycle. Google (TPU v7 “Ironwood” / TPU7x — announced April 2025, preview 24 November 2025, generally available 31 March 2026 per Google Cloud’s TPU release notes), Amazon (Trainium3 on TSMC 3nm, GA December 2025 — 2.5 PFLOPS FP8), Meta (MTIA 400), and Microsoft (Maia 200) are all deploying in-house AI accelerators manufactured at TSMC, explicitly to reduce dependence on NVIDIA and capture gross margin. They cannot design these chips unaided: Broadcom Inc. (NASDAQ: AVGO) and Marvell Technology Inc. (NASDAQ: MRVL) are the two dominant third-party custom ASIC architects for AI workloads.

Broadcom's AI semiconductor revenue reached $10.8 billion in Q2 FY2026 (+143% YoY, reported 3 June 2026), with AI bookings exceeding $30 billion in the quarter alone and Q3 FY2026 AI revenue guided to $16 billion (+200% YoY). Management reiterated a line of sight to more than $100 billion of AI revenue in 2027. Design wins now extend beyond Google, Meta, and Microsoft to OpenAI and Anthropic (confirmed early 2026), putting AVGO across an estimated 60% of hyperscaler custom silicon programmes.

TrendForce shows custom ASIC shipments growing 44.6% in 2026 versus 16.1% for merchant GPUs; Goldman Sachs projects ASIC demand matching GPU demand by 2027, a roughly 50/50 split. Marvell's custom AI silicon is smaller but growing fast (data center now 75% of revenue, +63% YoY in Q1 FY2026), with wins at Amazon (Trainium) and Microsoft; its December 2025 acquisition of Celestial AI ($3.25B) extends it into optical photonic interconnect — a $6B TAM by 2030 as bandwidth demands exceed what copper sustains at 1.6T+ speeds.

Both benefit from a dynamic NVIDIA cannot easily disrupt: hyperscalers have a strategic incentive to fund custom silicon as a CUDA countermeasure regardless of cost premium. The ASIC layer is a smaller threat in training, where CUDA’s flexibility and 19 years of accumulated tooling remain hard to displace — though not no threat: Google documents frontier-model training on TPUs, so custom accelerators are not confined to inference. The sharper shift is in inference share AVGO and MRVL are positioned to capture.

Server Integration Layer — Super Micro Computer & Dell Technologies

Super Micro Computer Inc. (NASDAQ: SMCI) and Dell Technologies Inc. (NYSE: DELL) occupy the final assembly layer, converting raw silicon into deployable rack-scale systems. Super Micro’s liquid-cooled DGX-H100 and MGX platforms are NVIDIA reference designs, and it has been first to market with new NVIDIA server generations three cycles running. Dell’s PowerEdge XE9680 competes for the same enterprise and colocation buyer, with the added advantage of a global direct sales force and ProSupport services.

Both are gated by NVIDIA’s GPU allocation — when B200 supply is constrained, SMCI and DELL backlog builds while gross margins compress on pre-sold orders. The risk is margin: server integration earns 5–8% EBIT at SMCI and 4–6% on AI infrastructure servers at Dell, and price competition intensifies during shortages. The structural opportunity is the move from Phase 1 (Build) to Phase 2 (Deploy), when hyperscalers shift from direct NVIDIA procurement to third-party integrators.

The binding constraint on NVIDIA's Blackwell B200 output in Q1–Q2 2026 is not the GPU die. It is High Bandwidth Memory 3E (HBM3E) — the stacked DRAM that separates a usable AI chip from a theoretical one. A single B200 package carries eight HBM3E stacks, specified by NVIDIA at 192GB nominal; NVIDIA’s shipping HGX and DGX B200 configurations expose 180GB per GPU, 1.44TB across an eight-GPU node. At current shipment volumes, the global HBM market must produce and package more advanced memory in 2026 than in all previous years combined. There are exactly three HBM suppliers on Earth: SK Hynix, Samsung, and Micron Technology Inc. (NASDAQ: MU) — the smallest of the three, and the most investable for Western investors.

HBM3E — The Architecture of Scarcity

High Bandwidth Memory is not DRAM as conventionally understood. DDR5 runs a comparatively narrow parallel interface split into independent subchannels; HBM stacks 8–12 DRAM dies vertically on a very wide interface using through-silicon vias, reaching about 1.2 TB/s per stack against roughly 0.1 TB/s for a DDR5 channel. The difference is interface width and proximity, not serial versus parallel signalling. For AI training — feeding terabytes of weights and activations per second to thousands of CUDA cores — that bandwidth is the ceiling on throughput. Running a 70B parameter training job on HBM3E rather than GDDR6 is roughly 4–5× faster, which at hyperscaler compute costs is tens of millions of dollars per run.

HBM production also requires TSMC (TSM) CoWoS packaging to integrate the stacks with the GPU die — a circular dependency, since GPU and memory compete for the same scarce capacity. None of the three discloses its HBM share and third-party estimates differ by basis (units versus revenue, HBM3E only versus all HBM). The credible range is SK Hynix ~43–55%, Samsung ~25–40%, Micron ~20–24%; this paper adopts no point estimate. NVIDIA has qualified all three for HBM3E, but SK Hynix retains a generation lead: Hynix began HBM3E production in Q3 2024, Micron qualified in Q4 2024, and Samsung’s yield issues delayed qualification into Q1 2026 — leaving Micron a qualified second supplier in a market where the leader cannot meet demand and the third player has quality problems.

Micron Technology — The High-Conviction AI Memory Investment Case

Micron Technology Inc. (NASDAQ: MU) is A.L. Capital Advisory’s fifth High Conviction position, and the one with the widest gap between consensus expectations and structural opportunity. Three independent pillars compound through 2028.

First, the HBM revenue inflection — now confirmed. In fiscal Q3 2026 (reported 24 June 2026), Micron posted record revenue of $41.5 billion (+346% YoY) at a record ~85% non-GAAP gross margin, guided fiscal Q4 to a record ~$50 billion, and signed 16 multi-year take-or-pay Strategic Customer Agreements covering roughly half of future revenue — lifting market capitalisation past $1 trillion.

What was a forward thesis a year ago is a reported inflection: HBM3E is sold out, HBM4 is ramping, and HBM has moved from a negligible FY2024 line to Micron’s primary growth engine. At SK Hynix’s disclosed HBM gross margins (50–55%), that is transformative against a blended profile historically averaging 25–35% across the DRAM/NAND cycle. A single HBM3E 8-Hi stack now carries a blended ASP of approximately $30–40 per GB in 2026 (HBM contract prices raised ~20% amid broad memory “memflation”).

Second, the DRAM pricing cycle. The 2022–2023 oversupply has fully resolved. All three major producers have deliberately constrained bit output growth, redirecting capital expenditure into HBM — capacity that becomes unavailable for standard DRAM. Server DDR5 contract prices are up ~50% quarter-on-quarter (“memflation”) as data center demand — each AI rack carries $100,000–$400,000 of standard DRAM alongside the GPU — grows faster than supply can respond. Micron captures both the volume increase and the ASP uplift.

Third, the NAND recovery. Western investors often model Micron as a pure DRAM company, but NAND flash — SSDs, storage arrays, training data pipelines — was approximately 26% of revenue in FQ4 FY2026 (Micron prepared remarks, 30 September 2026), down from roughly a third historically as the HBM and DRAM mix has shifted. The cycle bottomed in Q3 2023 at prices that pushed every producer, including Western Digital (NASDAQ: WDC), into EBITDA-negative operations.

Enterprise SSD pricing has since recovered 60–80% from the trough as AI training datasets, model checkpoints, and inference caches drive unprecedented enterprise NVMe demand. Western Digital separated its flash business into Sandisk in February 2025 and is now an HDD business, so it is no longer the pure-play expression of a NAND recovery; a flash thesis belongs to Sandisk, which this paper does not cover. Micron is the AI memory story this paper rates.

The Memory Wall — AI’s Hidden Bottleneck

The memory wall is the gap between the growth of model complexity and the growth of memory bandwidth. A B200 carries eight HBM3E stacks; trillion-parameter training runs need more stacks per GPU than the current architecture provides. Every generation — Vera Rubin on HBM4, AMD’s MI400 — therefore requires more HBM, more advanced packaging and more bandwidth. That is the structural escalator beneath the pricing documented above.

Per-rack memory content shows the scale. One NVIDIA DGX H100 system (8× H100) holds 640GB of HBM3 (80GB per GPU), 2TB of DDR5 system DRAM, and 30TB of NVMe SSD — an estimated $80,000–$120,000 of the system cost (A.L. Capital Advisory estimate; SemiAnalysis teardowns imply the lower end), comparable to a single H100 GPU. The per-gigawatt extrapolation this paper previously ran here has been withdrawn. It multiplied a figure for one eight-GPU server by an assumed 10,000 racks per gigawatt, which are not the same unit, and then applied the result to a power shortfall.

Doing it properly requires servers per rack, IT kilowatts per rack, facility PUE, memory content per server and current memory prices, none of which this paper publishes; and a power shortfall is in any case a statement about capacity that does not get built, not a forecast of installed racks or a realisable memory-revenue pool. The defensible observation is narrower: memory is now a first-order share of system cost, and it scales with deployed accelerators rather than with floor space.

Geopolitical Risk — China Memory and Export Controls

Micron's most significant risk is regulatory. China’s Cyberspace Administration banned Micron products from “critical information infrastructure” operators in May 2023, retaliating against US restrictions on Chinese advanced chip imports. China represented approximately 16% of Micron’s FY2023 revenue. Redirecting supply to India, Southeast Asia, and European data center expansion has partially absorbed the impact, but approximately $800M–$1.2B of annualised revenue remains displaced. Resolution — or further escalation — of the US-China technology trade war is the key binary risk in the Micron case.

Chinese domestic memory is a structural headwind, but further from competitive parity than commonly believed. CXMT (ChangXin Memory Technologies), China’s domestic DRAM producer, is shipping DDR4 and early DDR5 at an estimated 25–35% yield against an industry standard of 85–90%. Yangtze Memory (YMTC) reached 232-layer NAND in 2023 but faces ASML EUV import restrictions that will prevent progression to the sub-10nm nodes next-generation HBM requires. China is not a 2026 threat to Micron’s HBM franchise; it is a 2029–2032 risk to commodity NAND share.

Portfolio Construction Framework — Five principles for building the AI infrastructure position without getting burned:

The AI capex cycle is not a theme but a decade-long structural reallocation of capital from consumption to physical infrastructure, at a scale last seen in the electrification of the US economy. The numbers are in the Key Takeaways above and the Data Appendix below; what has changed since the summer is not the scale but the funding. The build is now externally financed, and that introduces a constraint the first two years of this cycle did not have. Across the three phases the value-capture layer rotates — semiconductors and thermal management in Build (2024–26), contracted operators with physical scarcity moats in Deploy (2026–28), all five simultaneously in Compound (2028–30) — which is why a static “AI basket” misses the rotation entirely.

The structural bull case rests on constraints that are physical rather than financial: 18–30 month construction lead times mean supply cannot respond to sentiment; 3–4 year accelerator refresh cycles mean overcapacity converts to obsolescence faster than it strands; and the contracts-first structure means 80–95% of capacity under construction is pre-committed before completion. The honest risks have changed shape since the summer, and the change is the point of this revision. The demand risk is largely unchanged — vacancy at 1.4%, memory sold out into 2027, Micron stating it has no line of sight to when supply and demand rebalance.

What has arrived is a financing and delivery risk: hyperscaler spreads at ~105bp against ~80bp for the market, Oracle’s $288B of not-yet-commenced leases now carrying a commencement window a year later than the February filing showed, at least 47 large US projects delayed or blocked in 2026, and Bernstein judging only ~35% of the announced US pipeline credible. That is why the suppliers are rated on delivery exposure and the buyers on balance sheet — and why the falsification test is no longer “does demand exist” but “does contracted capacity get deferred.”

The monitoring framework is straightforward: vacancy below 3% is healthy; above 6% is the early warning. Capex/revenue ratios declining from 2026 onwards signal the ROI inflection the market is waiting for. Enterprise AI deployment rate — currently in the early phase — is the key leading indicator for whether the McKinsey base case ($6.7T by 2030) is achieved or exceeded. A.L. Capital Advisory updates these readings quarterly.