Combined hyperscaler AI capex for 2026 is now tracking above $600 billion, according to updated guidance from Alphabet, Meta, and Microsoft on their April 29 earnings calls. Alphabet raised its 2026 range to $180 billion to $190 billion, Meta to $125 billion to $145 billion, and Microsoft signaled roughly $190 billion in total annual spend with a Q4 capex bar above $40 billion. AWS quarterly revenue hit $37.6 billion at +28% year over year, its strongest growth in 15 quarters, and the implicit capex behind that print sits in the same neighborhood as the other three.
The capex curve has not bent. What has changed is the tone from public market investors, who rewarded Alphabet's number with a 7% after-hours pop and punished Meta's with a 6%+ drop. The market is no longer scoring capex by absolute size. It is scoring it by visible revenue attribution, and Google Cloud's $462 billion backlog gave investors the proof point Meta couldn't produce.
The per-hyperscaler breakdown for 2026
Stack the four guidance numbers side by side and the shape of the year becomes obvious:
Alphabet: $180B to $190B in 2026, with CFO Anat Ashkenazi telling analysts 2027 capex will "significantly increase" versus 2026. Google Cloud Q1 revenue was $20B at +63% YoY, with backlog at $462B, roughly doubled quarter over quarter.
Meta: $125B to $145B, up from a previous $115B to $135B range. The ad business is carrying the spend alone, which is why a soft ROI answer on the call moved the stock more than the capex line itself did.
Microsoft: roughly $190B for the year, Q4 above $40B, of which CEO Satya Nadella attributed about $25B to higher component pricing. Microsoft expects to remain capacity constrained through 2026.
Amazon: AWS at $37.6B for the quarter, +28% YoY, with AI infrastructure deals around OpenAI, Anthropic, Meta, and Nvidia driving the back half of the year. AWS does not break capex out the way the others do, but the underlying trajectory matches.
Two thirds of Microsoft's spend goes to GPUs and CPUs. The other third is land, power, network, and shell construction. That ratio is roughly stable across the four, and it is the ratio that determines who actually gets capacity in 2027.
Why Alphabet's number landed and Meta's didn't
Investors didn't punish capex. They punished capex without an attribution story. Alphabet walked onto the call with a $462B cloud backlog, Gemini Enterprise paid MAUs up 40% quarter over quarter, named deals with Bosch, Mars, and Merck, and revenue from products built on its GenAI models up nearly 800% year over year. Sundar Pichai also disclosed that Alphabet had doubled the number of $100M to $1B deals year over year and signed multiple $1B-plus deals.
Meta's call had none of that texture. Mark Zuckerberg's response when asked about ROI, that the formula is to "build experiences that can get to billions of people and focus on monetizing them once you get to scale," reads as a 2014 playbook applied to a 2026 capex number. Meta is also the least diversified Magnificent 7 stock, almost entirely ad-revenue driven, so the spread between capex pace and ad-revenue pace shows up faster in the share price than it does for the others.
Microsoft sat in the middle. Azure plus other cloud services grew 40%, 21 points of which CFO Amy Hood attributed to AI, and Hood also noted that AI margins are already better than cloud margins were at a similar stage. That kept the stock flat instead of red, but it did not produce a Google-style pop.
Power, land, and chip supply are the actual ceiling
The $600B headline buries the supply story. Nadella's $25B component-price callout, and Microsoft's explicit "capacity constrained through 2026" language, both point at the same bottleneck. The constraints are not capital. They are GPUs (Nvidia allocation plus Blackwell yield), grid interconnects (the multi-year queue for new substations in Northern Virginia, Phoenix, and Dublin), and shell construction (18 to 24 months on a new build, longer for liquid-cooled retrofits).
Vendor concentration around Nvidia is being actively diversified. Google's TPU v5p and v6 cadence, AWS Trainium2 deployment, and Microsoft's Maia ramp are not science projects anymore. They are line items in the same capex bucket. None of them displaces Nvidia in 2026, but the order of magnitude of internal-silicon spend is now large enough that Nvidia's pricing power in 2027 negotiations will be measurably softer than it was in 2025.
Smaller clouds are the squeezed cohort. CoreWeave, Crusoe, Nebius, and the regional sovereign clouds are competing for the same Blackwell allocation, the same substation capacity, and the same construction crews as the four hyperscalers, with one or two orders of magnitude less balance sheet to win the bidding war. Frontier training at the 100k-GPU-plus tier is consolidating into the four. Inference and fine-tuning workloads are where the smaller clouds will keep winning, and where enterprise buyers should keep them on the shortlist.
What we are telling clients to lock before Q3 2026
For our enterprise clients sizing 2027 GPU budgets, the read on a $600B capex year is not "prices will fall." It is "capacity will stay tight at the top of the rack and loosen at the bottom." H100 and H200 on-demand pricing on the major clouds has already drifted down 15 to 25% from the 2024 peak. Blackwell (B200, GB200) is the inverse: list pricing is firm, reservation windows for 2026 H2 are largely closed at AWS and Azure, and we are seeing 12 to 18 month wait quotes for net-new GB200 NVL72 racks from anyone other than Google Cloud.
Our procurement guidance for CTOs and VPs of Engineering with material 2027 inference spend is concrete: lock multi-year reservations on the GPU class you actually need, not the one the vendor wants to sell you. For most production inference workloads, that is H200 or MI300X on a 24 to 36 month reserved commit, not Blackwell at list. Pricing power sits with the buyer on the prior generation through at least Q2 2027. Pricing power sits with the seller on Blackwell through 2027, full stop.
On the build versus buy split, we are not telling anyone to build their own data center on the back of this capex print. We are telling them to negotiate the multi-cloud egress and reserved-instance terms now, while three of the four hyperscalers need to show analysts a cloud-backlog number on the Q2 call. The bargaining position for a sub-$50M annual cloud commit is the strongest it has been in 18 months, specifically because Meta's reaction Tuesday told every CFO in the cohort that backlog dollars matter more than capex dollars.
The Q2 print is the next real signal
The next test is the Q2 2026 earnings cycle in late July. If Alphabet's cloud backlog grows again from $462B and Microsoft narrows the gap between Azure-AI growth and total Azure growth, the $600B+ number holds and 2027 guidance ratchets up further on those calls. If Meta cannot put a concrete revenue attribution number on the $125B to $145B range by July, expect either a capex trim into the back half or a continued multiple compression that forces the trim by Q3. The signal to watch is not the capex line. It is the backlog line and the named-deal disclosure underneath it. If the four stop disclosing backlog growth, the capex story has broken.
