06-reference/research

hyperscaler capex merchant vs inhouse silicon decomposition

2026-08-01·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
hyperscaler-capexcustom-siliconcapital-cyclesec-edgarproxy-construction

Hyperscaler capex cannot be split merchant vs in-house from the filings — but a three-legged supplier proxy brackets it to roughly 15-30%

The question

"Can hyperscaler capex disclosures be decomposed into merchant-silicon purchases vs in-house-silicon development spend, and if not, which proxy gets closest?"

Follow-up #2 from [[2026-07-29-cpu-upstream-multiplier-obligated-dollars-thesis]], which closed by saying the CPU question "cannot be resolved without a capex breakdown that separates merchant-silicon purchases from in-house-silicon spend." This brief tests whether that breakdown exists. It does not, for a structural reason worth understanding, and the value here is the proxy ladder plus its error terms.

Answer up front: no. Not at any level, in any of the four filings, in any period. The two categories the question asks about do not even live in the same financial statement: in-house silicon design is R&D operating expense, expensed as incurred and never separately disclosed, while the fabricated in-house part arrives inside a server chassis and is capitalized into the exact same PP&E line as a merchant GPU. No filing separates them, and none is required to.

What we already know (from the vault)

What the web says

All figures below are Tier A — I opened the filing on SEC EDGAR this run and read the text — unless explicitly labelled otherwise. Tier B = verbatim quote surfaced via search but the primary was not raw-read. Tier C = search-summary prose.

1. Total capex is disclosed, cleanly, by everyone. That is where the granularity stops. Calendar Q2 2026: Alphabet $44.9B ("Capital expenditures, which primarily reflected investments in technical infrastructure"); Amazon $54,208M purchases of PP&E; Meta $31.08B ("Capital expenditures, including principal payments on finance leases"); Microsoft $35,802M (derived: FY2026 additions to PP&E of $115,948M minus nine-month $80,146M — my arithmetic, both inputs Tier A). Combined ≈ $166B in one quarter. Microsoft's FY2026 capex of $115,948M is up from $64,551M (FY2025) and $44,477M (FY2024).

2. Asset-category detail exists, and it is a compute-vs-concrete split, not a silicon split. All four disclose a "servers and network equipment" PP&E category. Gross-cost deltas, my arithmetic on Tier A balances:

servers/network Δ total gross PP&E Δ share CIP-inclusive ceiling
MSFT FY25→FY26 +$83,038M +$133,148M 62.4% n/a (no CIP line)
AMZN FY24→FY25 +$59,336M +$140,043M 42.4% 60.3%
META FY24→FY25 +$29,643M +$69,063M 42.9% 77.3%

Alphabet does not use this format but gives the single most useful recurring footnote in the set: "As of December 31, 2025 and June 30, 2026, approximately 60% of technical infrastructure assets were comprised of servers and network equipment. The remaining balance was comprised of data center land and buildings and related assets." The same ~60% appears at Dec-2024. Alphabet is the only one publishing this ratio quarterly; the other three are annual-only, in the 10-K.

3. In-house silicon is literally absent from the text of three of the four filings. Word-frequency check across the Q2 2026 10-Qs and the most recent 10-Ks: "Trainium" appears zero times in any Amazon filing. "MTIA" appears zero times in any Meta filing. "Maia" appears zero times in any Microsoft filing. Alphabet is the sole exception, with 14 "TPU" mentions in its Q2 2026 10-Q — and the reason is not capex disclosure, it is that Alphabet started selling TPU systems externally in Q2 2026: "We have signed a limited number of agreements to supply TPU systems to customers who require or provide on-premises infrastructure for specialized, high-scale workloads. In the second quarter of 2026, we began recognizing revenues from these agreements, with the significant majority to be recognized in 2027." Those TPUs go to inventory and cost of revenues, not PP&E. No dollar figure is given (confirmed: none exists publicly; the only number circulating is a Morgan Stanley 2027 projection, Tier C, not a company disclosure).

4. The finance-lease seam is real but points at real estate, not chips — a negative finding that kills an obvious candidate proxy. Amazon added just $563M of PP&E under finance leases in Q2 2026 ($235M NA/International + $328M AWS) against $63,891M of total net PP&E additions — 0.9%. Amazon's gross finance-lease assets are flat at $55.6B between Dec-2025 and Jun-2026 with $39.8B accumulated amortization, i.e. a legacy, mostly-depreciated book with near-zero new additions. Alphabet's finance-lease PP&E is $7,915M at cost against $304,310M in service (2.6%) and its 10-Q states data-center leasing VIEs are accounted for as finance leases. Meta's finance-lease ROU assets are $8,187M of $233,726M gross (3.5%). The equipment-vs-real-estate distinction the finance-lease line used to carry has inverted: it is now almost purely a data-center-shell instrument.

5. Amazon is the only one with a segment-level capex readout, and it is the sharpest AI-isolation signal in the whole set. Net additions to PP&E by segment, Q2 2026 vs Q2 2025: AWS $48,604M vs $16,043M, North America $12,139M vs $11,272M, International $2,540M vs $2,531M, Corporate $608M vs $915M. AWS is 76.1% of consolidated additions, up from 52.2% a year earlier. Six-month 2026: AWS $90,120M of $118,648M = 76.0%. Retail/logistics capex is flat in absolute dollars; the entire increase is AWS.

6. Microsoft alone discloses a component-scoped dollar figure — and it sits outside capex. From the FY2026 10-K: "the current portion of other receivables related to activities to facilitate the purchase of server components was $27.8 billion and $8.2 billion" (Jun-30-2026 vs 2025), plus "a $22.2 billion increase in cash used in other investing primarily to facilitate the purchase of components," plus "restricted investments pursuant to a supplier agreement were $11.3 billion." This is the finest-grained silicon-adjacent number any hyperscaler publishes. Caution: Microsoft does not disclose whether this is additive to capex or a pass-through that recycles back into PP&E when the assembled server is purchased. Treat it as a gross-up/timing indicator, not as incremental hardware spend.

7. Obligated dollars are disclosed in bulk and decomposed by nothing. Alphabet: "material purchase commitments and other contractual obligations totaling $811.0 billion, of which $200.7 billion was short-term... primarily relate to costs for technical infrastructure and inventory through long-term supply agreements and open purchase orders" (Jun-30-2026). Microsoft: purchase commitments $194,060M ($169,008M due FY2027) + construction commitments $34,566M + leases $443,506M, total contractual obligations $743,821M. Amazon: unconditional purchase obligations $130,065M, leases not yet commenced $137,214M, total commitments $650,034M. Meta: $349.31B non-cancelable commitments ($53.52B due 2026, $81.65B due 2027) + $278.99B of leases not yet commenced, plus "In July 2026, we entered into additional data center leases with lease obligations of approximately $68 billion." Meta guides 2026 capex to $130-145B.

8. Supplier-side disclosure is where the silicon actually becomes visible.

9. TSMC gives platform and node, never a hyperscaler-usable customer split. HPC = 66% of Q2 2026 revenue (Tier A, read from the 6-K investor deck chart) vs 58% for FY2025 (20-F). Node mix Q2 2026: 2nm 3%, 3nm 30%, 5nm 33%, 7nm 11%. Concentration is annual only, in the 20-F, and never named: "our ten largest customers in 2023, 2024 and 2025 accounted for approximately 70%, 76% and 78% of our net revenue... Our largest customer... accounted for 25%, 22% and 19%... Our second largest customer... 11%, 12%, and 17%." CoWoS/advanced-packaging capacity allocation by customer: confirmed not disclosed — every 20-F occurrence is generic service description with no capacity figures.

10. Earnings-call language gives one number, and it is not a capex number. [Tier B] Amazon's Andy Jassy on the Q2 2026 call: the chips business (Graviton + Trainium + Nitro combined) crossed a "$25 billion annual revenue run rate" with "triple-digit growth year over year." That is rental revenue from customers, not what the silicon cost Amazon; it belongs on the demand side, not in a capex decomposition. Microsoft gave ratios only (Maia 200 "30% better performance per dollar," "40% better performance per watt"; Cobalt 200 "racks in over 25 data centers"). Meta's CFO was purely qualitative: "strategic investments in areas like our internal custom silicon effort." Alphabet gave no share figure. None of the four stated what share of AI capacity runs on their own chips.

Correction flag, carried forward deliberately: a widely-circulating attribution of "more than one gigawatt of custom silicon being developed with Broadcom" to Zuckerberg on the Q2 2026 call does not appear in the official Meta transcript, which was read in full (Tier A). Do not use that quote.

Paywalls skipped, not retried: Seeking Alpha, Motley Fool premium, AlphaSense, Bloomberg (transcript services). TrendForce reports are gated.

Convergences and contradictions

Synthesis for RDCO

The decomposition does not exist, and chasing it further is wasted effort. Three of four hyperscalers do not name their own silicon program anywhere in their SEC filings. The fourth names it only because it started selling it. The design spend is in R&D and the fabricated part is inside a server in the same PP&E line as a merchant GPU. Any brief, newsletter, or model that claims a merchant-vs-in-house capex split sourced to "SEC filings" is either using a vendor estimate or fabricating it. That is a useful thing to be able to say with primary-source confidence, and it is worth more than a fake number.

What replaces it: a four-leg proxy stack, ranked by proximity to the actual question.

Leg 1 — in-house numerator: Broadcom Semiconductor Solutions + Marvell data-center revenue. Closest available readout of what hyperscalers pay to have their own designs fabricated. Errors, in descending size: (a) neither breaks XPU out from networking, storage, wireless or broadband in the 10-K/10-Q narrative, so the segment total is a hard ceiling; (b) Broadcom's 42% "customer" is a distributor, so it aggregates end customers and identifies none; (c) Alchip, GUC and Socionext also run hyperscaler ASIC programs and are outside this pair; (d) revenue recognizes on shipment while the hyperscaler capitalizes on deployment, a one-to-two-quarter lead; (e) Broadcom's move to leasing XPUs and full racks moves the asset off the hyperscaler balance sheet entirely, so a growing slice of in-house silicon will never appear in hyperscaler capex at all.

Leg 2 — merchant denominator: NVIDIA's "Hyperscale" market-platform line. This is new and it is the single most valuable disclosure change in the set, because NVIDIA has done the segmentation work the hyperscalers refuse to: it separates hyperscalers from Neoclouds. Errors: direct customers are ODMs, not hyperscalers; NVIDIA does not publish its membership rule for the Hyperscale bucket; fiscal-quarter offset costs about a month of alignment. Add AMD Data Center as a smaller second merchant leg, accepting that AMD publishes no revenue concentration at all.

Leg 3 — scale check: hyperscaler capex × disclosed servers-share. Use Alphabet's ~60% quarterly footnote as the live ratio and the 42-62% annual band from the MSFT/AMZN/META 10-K asset tables as the cross-check. This converts headline capex into compute-hardware dollars so the supplier legs can be sanity-checked against them. It says nothing about merchant vs in-house and must not be asked to.

Leg 4 — AI isolation: AWS share of Amazon net PP&E additions. Only Amazon offers segment capex. At 76% and rising it is the cleanest available separation of AI-infra spend from everything else, and it is quarterly.

Recommended composite, with the arithmetic shown. Taking the most recent verified quarter from each and treating whole segments as ceilings: custom-ASIC vendor revenue ≤ $15,009M (AVGO Semi Solutions) + $1,833M (MRVL data center) = $16.8B, against NVIDIA Hyperscale of $37.9B. Ceiling in-house share = 16.8 / (16.8 + 37.9) = 30.7%. Applying a judgment haircut for the non-XPU content in both supplier segments — call it half of Broadcom semis and two-thirds of Marvell data center, an estimate, not a filed figure — gives (7.5 + 1.2) / (7.5 + 1.2 + 37.9) = 18.7%. So: in-house silicon is roughly 15-30% of the accelerator dollar in mid-2026, measured at the vendor level. Call the error band ±10 percentage points and treat the second derivative as far more reliable than the level. The haircut is the weakest link; it is a guess, and it is labelled as one.

What this stack cannot tell you, stated plainly. It cannot give a per-hyperscaler split, because no supplier names a customer. It cannot separate Alphabet's internal TPU deployment from its new external TPU sales. It cannot see leased XPUs and racks, a channel that is growing by Broadcom's own risk language. It cannot convert dollar share into compute share, and it systematically understates the latter. And it cannot be reconciled to any hyperscaler's reported capex, because the timing, the accrual treatment, and the ODM intermediation all break the chain. It is a directional instrument with a wide level error, and it should never be quoted as a decomposition of capex.

Procedurally, the most valuable output of this run is that it partially closes the empty-anchor gap. anchors/hyperscaler-capex/ is still empty, but the Q2 2026 figures here (~$166B combined, all four, Tier A, with the disclosure language that produced them) are now in the vault and are the first primary hyperscaler capex anchors it holds. The four-leg stack above is computable entirely from free EDGAR filings on the same plumbing [[2026-07-15-capex-financing-layer-as-memory-cycle-phase-marker]] already specified, with no paid data subscription. On the memory-cycle thesis's own anchor #1, the read is unambiguous and no anchor is flipping.

Why this is in the vault

It resolves the blocking open item that [[2026-07-29-cpu-upstream-multiplier-obligated-dollars-thesis]] left on the research backlog with a negative answer plus a usable substitute, and it supplies the first primary-source Q2 2026 capex readings for anchor #1 of [[2026-05-18-memory-cycle-v1.1]] — which the thesis has been running without since inception. It also pre-empts a specific error: it stops any future RDCO artifact from asserting a merchant-vs-in-house capex split "per SEC filings," because no such split is filed by anyone.

Open follow-ups

Related

Sources

Vault

Primary filings — SEC EDGAR (all opened and read this run)

Other