IX
Accelerators & the capex layer
The visible layer: accelerator designers and the hyperscaler capex programs that every other choke feeds.
Accelerators & the capex layer
Who owns it. This is the visible apex of the compute supply chain: the firms that design the accelerator and the hyperscalers whose capex pulls every downstream choke — HBM, packaging, foundry, power, optics. NVIDIA still owns the merchant layer outright: Q1 FY2027 (reported May 20, 2026) data-center revenue of ~$75B, up 92% y/y, with Blackwell "off the charts" and a record $14.8B data-center networking quarter (InfiniBand + Spectrum-X) that makes NVDA its own networking choke too. The second axis is custom silicon: Broadcom posted Q2 FY2026 AI revenue of $10.8B (+143% y/y), guided Q3 AI to ~$16B (+200%), and reaffirmed ~$56B FY2026 / >$100B FY2027 — designing XPUs and networking for Google, Meta, ByteDance, Anthropic (the $10B mystery customer, now Google-TPU racks), and OpenAI (a separate 10GW co-design deal). AMD is the credible #2 merchant GPU: Q1 2026 data-center revenue $5.8B (+57% y/y), with MI450/Helios racks anchored by gigawatt commitments from OpenAI and Meta.
What breaks it. Internalization. The hyperscalers funding the boom are also routing around the merchant layer. Google's seventh-gen Ironwood TPU and previewed 2nm TPU 8t/8i are now winning external workloads — Anthropic committed to up to 1M TPUs and multi-gigawatt next-gen capacity from 2027. AWS Trainium2 runs Project Rainier at ~500k chips, the largest commercial ASIC fleet. Meta (MTIA) and Microsoft (Maia) are building internal parts. Every TPU/Trainium socket is a GPU socket that NVDA/AMD do not fill — but it still feeds HBM, CoWoS and foundry, so the choke moves rather than closes. The second break is policy: the Section 232 25% semiconductor tariff (effective Jan 2026) and the China export regime can reprice the whole stack overnight.
What forces the reprice. Demand is contracted, not hoped-for. The Big Four guided to roughly $700–725B of 2026 capex — Amazon ~$200B, Google $180–190B, Meta $115–135B, Microsoft ~$190B (confirmed on its Apr 29, 2026 fiscal-Q3 call, ~$25B of it component/memory pricing) — and that number is the demand signal every other chokepoint trades against. Each capex print, each gigawatt deployed, and each new XPU customer is a reprice event. The risk that forces a reprice in the other direction is circular financing: NVIDIA's up-to-$100B OpenAI commitment, paid progressively per gigawatt deployed, where OpenAI's CFO conceded "most of the money will go back to Nvidia"; AMD's OpenAI and Meta deals each carry 160M-share warrants (~10% of AMD) tied to deployment and stock-price milestones. Vendor-funded demand can inflate the apparent size of the choke; when financing tightens, the warrants and tranche triggers are where the air comes out first.
Be honest about size and earliness. The merchant revenue is real and enormous — NVDA is running ~$75B/quarter of data-center sales today, AVGO ~$10.8B/quarter of AI, AMD ~$5.8B/quarter — these are not narrative numbers. But three honesty flags matter. First, concentration: a double-digit share of NVDA's recent quarters traces to a handful of mega-customers and to entities NVIDIA itself has invested in, so part of the "demand" is its own capital recycled. Second, the custom-ASIC TAM is front-loaded into guidance and warrants: Broadcom's >$100B FY2027 AI figure and AMD's "tens of billions" from OpenAI are contracted intent, not shipped product, and the MI450/OpenAI ramps are largely a 2H-2026-and-beyond story. Third, the hyperscalers are capex proxies, not pure-plays — AI infrastructure is a fast-growing slice of their P&L, not the whole thing, and the same dollar shows up as merchant revenue, internal silicon, and downstream HBM/power demand at once, so double-counting the choke is easy. Tier by how much of the stack each name actually owns and ships, and treat the financing loop as a structural feature of this cycle, not a footnote.
Who owns the choke
Advanced Micro Devices, Inc.
Q1 2026 data-center revenue of $5.8B (+57% y/y) on EPYC and ramping Instinct GPUs. Signed 6GW deals with OpenAI (Oct 2025) and Meta (Feb 24, 2026), each first 1GW on custom MI450, each carrying a performance warrant for up to 160M AMD shares (~10%) vesting on shipment and stock-price milestones; first-gigawatt shipments begin 2H 2026.
- Bull
- If MI450/Helios ramps from 2H 2026, AMD's $5.8B datacenter quarter plus 6GW OpenAI and Meta deals make it the credible #2 merchant GPU.
- Bear
- The OpenAI/Meta deals are contracted intent with 160M-share warrants tied to deployment, so vendor-funded demand can deflate before product ships at scale.
Amazon.com, Inc.
Guided to ~$200B capex for 2026 (vs ~$132B in 2025). Project Rainier activated with nearly 500,000 Trainium2 chips — the largest commercial hyperscaler ASIC fleet by unit count — with Anthropic scaling toward a >1M-Trainium2 target.
- Bull
- If internal silicon scales, Amazon's ~$200B 2026 capex, ~500k-chip Project Rainier Trainium fleet and ~$74B Anthropic stake give it the deepest vertically-integrated AI position.
- Bear
- AI is a fast-growing slice of a vast retail/cloud P&L, and roughly half of recent 'AI profits' were Anthropic stake markups, not operating revenue.
Broadcom Inc.
Q2 FY2026 (reported June 3, 2026) AI semiconductor revenue of $10.8B (+143% y/y); guided Q3 AI revenue to ~$16B (+200% y/y) and reaffirmed ~$56B FY2026 and >$100B FY2027 AI revenue. In Dec 2025 disclosed its $10B-plus mystery XPU customer was Anthropic (Google-TPU racks); separately co-designing 10GW of accelerators with OpenAI.
- Bull
- If custom-XPU demand converts, Broadcom's $10.8B AI quarter, ~$16B Q3 guide and Anthropic/OpenAI co-design deals validate its reaffirmed >$100B FY2027 AI target.
- Bear
- The >$100B FY2027 figure is contracted intent front-loaded into guidance, not shipped product, and custom ASICs depend on hyperscaler capex holding.
Alphabet Inc.
Alphabet holds roughly 14% of Anthropic in straight equity, contractually hard-capped at 15% per court documents, and committed billions more in 2026. Fortune reported that about half of Alphabet's and Amazon's 2026 'AI profits' came from markups on their Anthropic stakes rather than operating AI revenue.
- Bull
- If Anthropic IPOs near base case, Alphabet's ~14% capped equity stake plus internalizing Ironwood TPU give it both proxy upside and merchant-layer disintermediation.
- Bear
- About half of recent 'AI profits' were Anthropic stake markups, not operating revenue, and the lab stake is a small slice of a vast enterprise value.
Meta Platforms, Inc.
Guided FY2026 capex to $115-135B (later raised toward $125-145B), citing higher component pricing. Signed a 6GW AMD Instinct deal (Feb 24, 2026) with a 160M-share AMD warrant, first deployment on custom MI450 in 2H 2026, while continuing to develop in-house MTIA accelerators.
- Bull
- If its 6GW AMD MI450 deal and in-house MTIA scale, Meta's $115-135B 2026 capex turns it into both a major accelerator buyer and internal-silicon builder.
- Bear
- Meta is a capex proxy, not a pure-play; the AMD deal carries a 160M-share warrant and first deployment is a 2H-2026 story dependent on its own ad P&L.
Microsoft Corporation
Guided to ~$190B capex for calendar 2026 (confirmed on its fiscal-Q3 call, Apr 29, 2026; ~$90B in 2025) — roughly tied with Alphabet at the top of hyperscaler spend. Builds the in-house Maia AI accelerator but remains heavily reliant on NVIDIA GPUs; its deepest AI exposure is the OpenAI relationship, not owned merchant silicon.
- Bull
- If OpenAI lists near base case, Microsoft's ~27% as-converted stake plus ~$190B 2026 capex give it the deepest proxy exposure to the leading private lab.
- Bear
- The OpenAI line is a loss-share, not revenue; the recap gain was one-time accounting, and Microsoft's revenue cut is capped at $38B through 2030.
NVIDIA Corporation
Reported Q1 FY2027 (May 20, 2026) revenue of ~$82B with data-center revenue ~$75B (+92% y/y) on Blackwell ramp, and a record $14.8B data-center networking quarter (InfiniBand + Spectrum-X, +199% y/y). Jensen Huang has framed Blackwell+Vera Rubin as a ~$1T opportunity across 2026-2027.
- Bull
- If Blackwell+Vera Rubin's ~$1T 2026-2027 opportunity converts, NVIDIA's ~$75B datacenter quarter and record $14.8B networking print make it its own merchant and networking choke.
- Bear
- A double-digit share traces to a handful of mega-customers and entities NVIDIA itself invested in, so part of the demand is its own capital recycled.
Catalyst calendar
- 2026-07-01Commerce Section 232 semiconductor report to the President (data-center chip focus)highA scheduled report that could expand or reshape the 25% tariff on advanced AI accelerators, directly repricing the cost basis of the entire accelerator-and-capex layer.
- 2026-07-28Alphabet Q2 2026 earningsmediumSets the next capex guide and Google Cloud/TPU trajectory — the cleanest read on how fast a hyperscaler is internalizing the accelerator choke away from merchant GPUs.
- 2026-08-04AMD Q2 2026 earningsmediumFirst print to test the MI450/Helios ramp narrative and disclosed data-center revenue against the OpenAI and Meta gigawatt commitments that underpin the bull case.
- 2026-08-26NVIDIA Q2 FY2027 earningsmediumThe sector's master demand signal: data-center and networking growth and Vera Rubin commentary set expectations for every downstream chokepoint that NVDA's volume feeds.
- 2026-09-03Broadcom Q3 FY2026 earningsmediumTests the guided ~$16B AI quarter and any new XPU customer disclosure — the key gauge of how fast custom ASICs are taking accelerator share from merchant GPUs.