A Field Guide to the Biggest Earnings of the Year

After the close on Wednesday, April 29, Alphabet, Microsoft, Amazon and Meta will all update investors on 2026 AI capex. Together they are guiding a spending cycle worth roughly $710 billion this year alone. In about one hour of after-hours trading, the market will likely reprice the entire AI supply chain. Here is the setup.

How big is this?

Hyperscaler capex has exploded. It moved from roughly $240B in 2024, to about $430B in 2025, and is projected to reach roughly $710B in 2026. That would mean three straight years of 60%+ growth, one of the most aggressive capex cycles software and internet investors have ever seen. The key guides so far: Alphabet at $175 to $185B, Amazon around $200B, Meta at $115 to $135B, and Microsoft signaling FY26 capex growth above FY25's level. Add Oracle's Stargate spending and the neoclouds, and the total climbs meaningfully higher still.

The system: where the money goes

The hyperscalers sit at the top of the AI infrastructure stack. Their cash flow turns into orders across the whole supply chain: chip designers like Broadcom and Marvell, GPUs from NVIDIA and AMD, foundry capacity at TSMC, equipment from ASML, AMAT, LRCX and KLAC, HBM memory from SK Hynix, Samsung and Micron, networking and optics from Arista, Coherent, Lumentum and Fabrinet, power and cooling from Vertiv, Eaton, Schneider and GE Vernova, and eventually long-term power agreements.

The bottlenecks

The biggest bottleneck today is power. Goldman Sachs expects data-center power demand to rise 165% by 2030, while the U.S. grid still faces an estimated $720B capex gap; the utility system has not committed anywhere near enough spending to close it. Behind the power problem is an equipment bottleneck: data centers need hardware that moves electricity from the substation into the building, and power transformers have roughly 128-week lead times, generator step-up units closer to 144 weeks, and switchgear is the only major category under one year, at about 44 weeks. This does not clear quickly.

On the chip side the picture is more mixed. HBM memory is still the tightest constraint: SK Hynix, Samsung and Micron have all said 2026 capacity is fully booked, and TrendForce does not expect HBM4 to ramp in volume until Q2 2026. TSMC's CoWoS advanced packaging is also tight; CEO C.C. Wei has said capacity is sold out through 2025 and into 2026, with NVIDIA taking roughly 60% of available wafers. That matters because HBM and CoWoS are both needed for leading-edge accelerators, including Blackwell, Rubin, MI400, and Google's TPU Ironwood; you cannot easily swap around these bottlenecks. The one easier spot is merchant GPU availability: H100 lead times have fallen to 8 to 12 weeks, down from 16-plus through much of 2024, and H200 lead times sit around 12 to 24 weeks for Tier-2 customers. The hyperscalers pulled forward a lot of demand, but the Tier-2 enterprise market has loosened meaningfully.

On Wednesday, the most important signal may not be EPS. The better trade signal will be whether management says these constraints are easing or getting worse.

Revenue exposure: who gets paid by AI?

Not every "AI stock" has the same AI exposure. Some companies get only 5% of revenue from AI; others are almost pure-play, and the higher the exposure, the more sensitive the stock should be on Wednesday night. Arista has roughly 42% combined exposure to Microsoft and Meta. Broadcom's largest AI customer is Google TPU, with Meta MTIA second. Marvell's largest AI customer is Amazon Trainium, at more than 30% of AI revenue. Vertiv has about 50% hyperscaler exposure across the four names and a $15B backlog, up 109% year over year.

One major subplot is custom silicon. Each hyperscaler now has its own ASIC program: Broadcom designs Google's TPU v7, also known as Ironwood, and Meta's MTIA; Marvell designs AWS Trainium; Microsoft's Maia 200 is now deployed in Iowa. Across the four hyperscalers, custom silicon could absorb roughly $185B of 2026 AI capex, about 37% of AI silicon spend, up from less than 15% in 2023. The overall 2026 AI silicon market could be roughly $500B: merchant accelerators like NVIDIA Blackwell/Rubin and AMD MI350/MI400 taking about $250B, or half; custom ASICs about $185B, or 37%; and networking and interconnect about $65B, or 13%, possibly the fastest-growing slice.

If Sundar Pichai gives a TPU customer count, Amy Hood talks about Maia's inference share, Andy Jassy highlights Trainium dollar growth, or Mark Zuckerberg cites MTIA Gen-4 adoption, the market may read it as negative for NVIDIA's hyperscaler share.

Priors: what's priced in already

The best clue going into Wednesday is what the supply chain has already reported over the past month. TSMC's April 16 Q1 report made the demand picture clear: net income rose 58% year over year, the company raised its FY26 growth guide above 30%, and HPC now makes up 61% of revenue. That kind of mix shift only happens when AI orders are actually shipping, and management's "extremely robust" language was a direct validation of the capex cycle.

Going into Wednesday, investors are focused on four debates. For Microsoft, whether Azure growth justifies the AI spending: in Q2 FY26 Microsoft lost roughly $357B of market cap in one session after Azure decelerated, and the Street may now treat an Azure beat plus another capex increase as neutral or even negative. The key numbers are Azure constant-currency growth and Amy Hood's comments on supply constraints: is demand strong enough, is capacity catching up, is the spending starting to pay off?

For Google, whether Sundar Pichai finally puts numbers around TPU. At Cloud Next on April 22, Google made Ironwood generally available and previewed the eighth-generation TPU, and Anthropic has confirmed demand for more than 1M TPUs. The market knows the technology is real; what it still needs is a revenue figure or customer count, evidence that TPUs are becoming a real external cloud business rather than a technical advantage with limited disclosure.

For Amazon, the debate is AWS margin. Consensus sits around 35.7%, but sell-side estimates range from 30.9% to 40.0%, the widest spread of the group. The bear case is that AI infrastructure depreciation pushes AWS margins into the low 30s. The bull case is that Anthropic-on-Trainium, tied to a $100B, 5GW commitment, plus a $70B+ advertising run-rate with high incremental margins, protects the profit story. Watch AWS operating margin, Trainium dollar growth, and how much of Anthropic's workload is actually deployed.

For Meta, the debate is whether investors punish AI capex the way they punished Microsoft. Meta jumped 10% after its Q4 print but has since fallen roughly 14% from the peak. The key issue is the capex range: holding near the $115B low end would likely be viewed as relief, while a move to $140B or higher could trigger the multiple-compression trade. Reality Labs losses are the secondary signal.

Across all four names, the headline numbers may not matter much. Revenue and EPS are modeled closely by the time results hit the tape. Unless there is a huge beat or miss, traders will focus on the fragile parts of the story: cloud growth, capex ranges, custom silicon, and supply constraints. A clean headline beat with weak Azure growth or higher capex can still sell off; a modest revenue miss with a tighter capex range or real TPU disclosure can still rally.

Past prints and the ripple map

Looking at how each stock traded the session after its recent prints, three patterns stand out. Meta has been the most volatile, with next-session reactions running from minus 9% to plus 11%. Microsoft's last two prints were both red despite headline beats, including the Q2 FY26 move that erased roughly $357B of market cap. Amazon's Q3 2025 reacceleration print was the cleanest beat, with AWS growing 20%, its fastest pace since 2022.

The tightest movers around hyperscaler prints are the leveraged plays: CRWV carries a +1.19 beta to Microsoft and +0.67 to Google, NBIS +0.90 to Microsoft, SMCI +0.97 to Microsoft, ALAB +1.05 to Microsoft and +0.95 to Google, and CLS +0.77 to Microsoft and +0.95 to Google. These names amplify whatever the hyperscalers do. The middle group is the core supply chain: NVIDIA at +0.54 to Microsoft, Broadcom at +0.56, Arista at +0.73, along with AMD, Marvell, the foundry and wafer-fab equipment names, Micron, and most optics; they should move most with whichever hyperscaler tells the cleanest story. The loosest movers are the power and industrial names, CEG, VST, TLN, Eaton, GE Vernova, AAON, which have their own drivers, power purchase agreements, regulated returns, capacity-auction prices. Hyperscaler earnings matter to them, but they are not the whole story.

The night unfolds

All four companies report at the 4:00 PM ET close. Alphabet's call starts at 4:30 PM. Then the real chaos begins at 5:30 PM, when Microsoft, Amazon and Meta all start their calls at the same time.

The key point is that Wednesday is not really a simple directional bet on AI. The supply chain has already validated demand for the first half of 2026: TSMC raised, ASML raised, NVIDIA guided to $78B for Q1, and all three HBM vendors are sold out. What the market does not know is what 2027 looks like. That is what these calls can clarify, and traders will listen for whether 2027 capex is still growing, starting to plateau, or showing early signs of slowing. That framing will set positioning into Q2.

We see three scenarios worth planning for. The melt-up case, roughly 30% probability: at least two hyperscalers show cloud reacceleration while keeping capex flat or only modestly higher, with backlog or RPO acceleration making the setup cleaner. This validates AI spending without raising new funding concerns: GOOG could rally 6 to 9% on cloud strength plus any TPU disclosure, AVGO 5 to 8% if TPU and MTIA are validated, ANET 5 to 8% on concentration relief, CRWV and NBIS 8 to 12% as leveraged neocloud beneficiaries, VRT and CEG 3 to 6%, MRVL 6 to 10% on constructive AWS commentary, and QQQ could squeeze to new highs.

The base case, roughly 45% probability: all four beat revenue by 1 to 3%, capex is mostly unchanged, cloud growth is in line, no major narrative surprise. Moves are muted: GOOG up 1 to 3%, META flat to up 2%, MSFT between minus 2% and plus 2%, AMZN up 1 to 3%, QQQ close to unchanged. The trade is rotation rather than direction; the tactical pair is long AVGO against short NVDA, or the reverse, depending on how the custom-silicon mix-shift narrative lands.

The collapse case, roughly 25% probability: capex is raised again without matching cloud acceleration, reviving the "AI capex hostage" concern. MSFT could fall 7 to 10% on a third straight Azure disappointment, META 8 to 12% if capex moves to $140B or higher, ANET 10 to 15% on its Microsoft and Meta concentration, CRWV and NBIS 10 to 15% as the leveraged neocloud trades unwind, with NVDA, AMAT, LRCX, AVGO and MRVL down 5 to 8%. QQQ could retest the February lows, defensives catch a bid, and the VIX rises 25%.

Whichever scenario plays out, the best opportunities Wednesday probably are not simple long-or-short AI bets. The supply chain has already front-run the demand validation. The real trades are in the second derivatives: names with concentrated hyperscaler revenue exposure, custom silicon winners and losers, and mismatches between how the hyperscalers trade and how their suppliers react.