Our house view coming into this week: the binding constraint on AI compute is moving from leading-edge silicon toward grid energization.

Power equipment, interconnection queues, and site readiness now look at least as important as GPU allocation in determining how quickly AI capex becomes usable compute.

This week tests that view across the stack.

Four hyperscalers report Wednesday. Samsung, Apple, Quanta, and Western Digital follow Thursday. Eaton closes the week Friday.

The capex backdrop is the strongest of the cycle. AMZN guides to $200B for 2026. GOOGL guides to $175 to 185B. META guides to $115 to 135B. That trio alone implies $490 to 520B before adding Microsoft, which is already running above $35B per quarter.

The market no longer needs proof that AI capex is real.

The harder question is where the constraint has moved.

Is it leading-edge wafers? CoWoS-L capacity? HBM allocation? Grid interconnection? Transformers? Site readiness?

Our bias: the constraint is moving toward the physical layer of AI deployment. Not whether the dollars are being spent, but how quickly those dollars can become energized compute.

This week tells us whether we’re right.

If Q1 lands above implied run-rate across all four hyperscalers, advanced packaging tightens through mid-2027, memory allocation stays firm, and power backlogs stretch further.

If even one hyperscaler hedges H2, the chain separates. The trade narrows to names tied to spenders that do not blink.

SK Hynix already gave us the leader’s print: HBM revenue more than doubled YoY, DRAM ASP rose 20% QoQ, and 2026 capacity is effectively pre-sold.

Samsung tells us whether that is a cycle or a leader.

WDC and Samsung NAND give the storage cross-check.

Both outcomes are investable. They point to different winners.

PWR backlog conversion, ETN book-to-bill, and hyperscaler site-readiness language converge on one question:

How fast does announced AI capex become energized compute?

If the power names print tight, NVDA allocation timing stops looking like the only constraint and starts looking like one symptom of a larger deployment problem.

Q1 has seasonality. Hyperscalers can pull or push spend around quarter ends. FX can move several billion dollars of optical capex year-over-year for Korean and Taiwan-exposed names. Memory ASPs can move on inventory rebalancing rather than demand.

Positioning also matters. The buy-side is long the AI infrastructure basket and selectively short the consumer chain. The magnitude of beats relative to whisper numbers may matter more than the beats themselves.

Three policy variables sit underneath the prints. Korean memory export licensing remains the wildcard for Samsung and SK Hynix. China handset tariff treatment shapes the QCOM and AAPL reads. US grid permitting reform determines whether ETN and PWR backlogs convert to revenue in 2027 or 2028.

AMZN, GOOGL, META, and MSFT

Hyperscaler capex does not stay inside the capex line.

It flows into TSMC wafers, CoWoS slots, HBM allocation, data center power equipment, retimers, optics, and switch silicon. Q1 prints set the cadence for a large share of the AI infrastructure stack.

Threshold: Q1 capex ≥ $50B · Trainium 3 mid-2026 reaffirmed

Amazon’s $200B FY2026 capex guide implies roughly $50B per quarter.

A print at or above pace, paired with reaffirmed Trainium 3 timing, supports tighter TSMC CoWoS-L through mid-2027, extends ALAB PCIe 6 retimer visibility, and confirms AVGO’s 4nm Trainium production curve.

Break case: Q1 below $45B with H2-loaded language.

That weakens the front-loaded thesis, raises the odds of NVDA GB300 backfill, pauses ALAB inventory builds around new-rack qualification, and slows ETN switchgear backlog growth.

Threshold: Q1 capex $44 to 46B · named H1 2026 Ironwood ramp

Alphabet’s $175 to 185B FY guide implies roughly $44 to 46B per quarter.

The headline number matters. Ironwood timing matters more.

A named H1 2026 ramp would lift confidence that Google’s TPU program is absorbing a larger share of AI silicon demand. That would support AVGO 3nm, TSMC CoWoS-L, and CRDO optical DSP demand for inter-pod links.

Break case: silence on Ironwood plus Q1 below $40B.

That implies NVDA picks up the inference demand Ironwood was supposed to absorb. It also takes pressure off SK Hynix HBM3E lead times relative to a clean TPU ramp.

Threshold: Q1 capex ≥ $29B · FY guide reaffirmed · MTIA or Broadcom partner-chip milestone

Meta has the largest sequential step-up of the group.

Q4 capex was $22.1B. The FY guide implies $29 to 34B per quarter.

A print at the new run-rate matters for VRT cooling backlog, Prometheus and Hyperion site buildout, and ANET Cassini orders through H2.

Site-readiness commentary may matter more than the capex number.

If Prometheus or Hyperion timing slips, the issue is probably not GPU demand. It is physical deployment. That readthrough moves toward ETN switchgear and PWR grid construction.

Threshold: capex ≥ $35B · Maia 200 named FY2027 fleet share > 10%

The first test: was last quarter’s $37.5B a peak or a new floor?

Sustained capex above $35B with FY language reaffirmed extends TSMC packaging commitments through 2027 and keeps ETN data center switchgear backlog past 18 months.

The second test is Maia 200.

A named fleet-share target above 10% is one of the cleaner signs that Microsoft’s custom silicon effort has moved from strategy to budget allocation.

That would be positive for AVGO and TSMC 4/5nm loading, with a modest negative read for NVDA.

KLAC
Quarterly advanced packaging revenue at or above $275M, against $950M in CY2025, with CY2026 growth raised above mid-teens, would confirm CoWoS-L tightening.

Readthroughs: SK Hynix and Micron HBM packaging.

EQIX
The question is whether Q4’s record $474M of bookings, with 60% AI mix, are converting to MRR.

MRR growth above the 10% Q4 pace confirms the pipeline. Bookings stacking without conversion points to fill-rate or activation delays.

ASX
LEAP advanced packaging revenue is measured against the $3.2B 2026 guide, doubled from $1.6B in 2025.

ASX likely needs roughly $600M in Q1 to stay on pace. A print below $500M with softer FY language would suggest CoWoS overflow allocation is not broadening as much as the supply chain implies.

QCOM
Watch whether Chinese OEM handset inventory drawdowns from DRAM-to-HBM reallocation are stabilizing or extending.

This is the cleanest indirect read on whether memory tightness has reached the consumer chain.

Threshold: named HBM3E qualification progress · HBM revenue > 20% of DRAM · HBM4 sampling on schedule

Samsung is the most important memory print of the week.

SK Hynix already showed us a tight HBM market. Samsung tells us whether tightness is broadening or leader-concentrated.

If qualification progress and the 20% HBM share both land, Samsung starts re-rating as a credible co-leader into HBM4.

If neither lands, the cycle is real but the scarcity premium stays at SK Hynix and Samsung’s discount holds.

Both outcomes are investable. They point to different winners.

Threshold: iPhone revenue ≥ $46B · Mac growth · Arizona volume commentary

Apple is the largest single consumer of TSMC N3-family wafers.

Strong iPhone revenue plus Mac growth keeps N3 utilization tight, which constrains the same leading-edge pool needed by NVDA Grace Blackwell, AMD MI-series, and QCOM Snapdragon through late 2026.

iPhone weakness or China deceleration loosens N3 at the margin, a small tailwind to the AI accelerator queue.

Arizona fab volumes are the extra tell. A meaningful shift could ease Taiwan N3 pressure.

Threshold: backlog ≥ $44B · NiSource 3 GW project on schedule

Quanta is the cleanest public read on US grid construction.

Q4 backlog was $44B, up from $39.2B in Q3.

Backlog above $44B with accelerating large-load starts confirms power as a binding constraint on AI cluster energization at Microsoft and Amazon sites.

Flat or declining backlog suggests one of two things: delivery is catching up, or hyperscalers are routing around the public utilities.

Either way, announced capex is not the same as energized compute.

Threshold: gross margin ≥ 46% · HAMR qualification on schedule

Q2 gross margin was 46.1%.

Sustained expansion supports the view that NAND supply discipline and UltraSMR mix are structural rather than cyclical.

HAMR slipping past H1 CY2026 would cede next-generation share to STX.

Pair this with Samsung NAND and SK Hynix’s recent print. Three converging margin reads tell us whether memory tightness is broadening across DRAM, NAND, and storage simultaneously.

Threshold: Electrical Americas backlog ≥ $13.2B · book-to-bill ≥ 1.2x

This is the most direct public read on whether data center electrical equipment is still tight.

Q4 2025 backlog was $13.2B, up 31% YoY, with 1.2x book-to-bill and data center orders growing triple digits.

If backlog and book-to-bill hold, grid interconnection and electrical equipment stay tight through 2027…

…that would keep gating NVDA cluster energization at MSFT, AMZN, and META sites while keeping VRT and PWR backlog growth firm.

If backlog eases, it would suggest either Eaton’s added manufacturing capacity is starting to catch up, or customer order growth is slowing. Either outcome would soften the megawatt thesis.

This is the most important print of the week because it tests our house view directly.

The week strengthens the trade if all hold:

The week weakens the trade if any land:

A few tells from the precall cards that go beyond headline numbers:

If memory pricing holds, power backlogs keep growing, and hyperscaler capex lands ahead of run-rate, the AI infrastructure trade gets its strongest single confirmation window of 2026.

More importantly, the megawatt thesis tightens.

If hyperscalers hedge, Samsung lags, and ETN or PWR normalize, the story narrows.

AI spend is still real. The winners get fewer.

The market has spent two years asking who gets the GPU.

This week, we find out who gets the megawatt.

We’ll publish post-call assessments after each print on the Tessara terminal, including verdict scoring, scenario matching, and updated readthroughs to downstream names.

Full precall cards with confirm/break thresholds, supply-chain readthroughs, and call checklists are live for private beta users → Apply here

We’ll also make the full earnings call transcripts publicly available within the first few hours to after each call. You can access them here:

tessara.chainofthought.xyz/earnings/transcripts

— Tessara