CapEx Week
Hyperscalers report Wednesday. $660B in CapEx on the line. Plus: Intel's biggest day since 1987 and Meta + Broadcom go multi-gigawatt.
Good morning. On Friday, Intel had its best trading day since 1987. Has there ever been a better time to be in the server CPU business?
This week, the hyperscalers report! It’s bound to be a wild week.
— Austin & Vik
The most important week of the AI CapEx cycle (so far)
Wednesday after the close: Alphabet, Amazon, Meta, Microsoft. Combined committed 2026 CapEx going in is $660B+, up from ~$400B in 2025. Last quarter, every one of them said roughly the same thing: we’re supply-constrained on AI compute.
Until that sentence changes, the trade is broadly up and to the right. Logic. Memory. Optics. Copper. Substrates. You name it.
What we’re listening for:
CapEx revisions. Everything is downstream of the CapEx guide. Is the number still accelerating, or just growing?
Server CPU language. The CPU:GPU ratio flips thanks to agentic inference (every agent needs a control plane, every tool call needs a host VM). Intel told us the buyers are buying. On Wednesday, the buyers tell us themselves.
Custom silicon. MTIA, Maia, Trainium, TPU. Custom silicon means custom datacenters. Read through to optics and networking.
Power and grid. Any headwinds?
Vik: CPUs are the story here. The more hyperscalers talk about agents in production, the more it reads through to server CPU, memory, and storage demand. Will we see any hints of Arm vs x86? Nah. Right now, I think it’s just “any CPU I can get my hands on is the right CPU”.
Austin: AWS is the one to watch. Bull: Trainium ROI is materializing (a 40% Trn2 to Trn3 price-perf gain is flowing to customers). Bear: Trainium talk is cover for AWS being behind on Nvidia GPU deployment fleet size versus the other CSPs.
Intel had its best day since 1987
INTC closed Friday +24%, the stock’s best single session since 1987. Data Center & AI +22% YoY on “unprecedented demand for silicon”.
Austin: “Data Center & AI” is the segment name, but CPUs pay the bills (accelerators are rounding error). Lucky for Intel, that’s what hyperscalers are buying. Datacenter CPUs, for AI. Name checks out, I guess. Long-term, the real INTC story is still Intel Foundry’s packaging and wafers.
Related: Semi Doped Ep. “Is Intel Finally Back with a $300B market cap?” · Vik’s “TWiC: TPU v8, Intel Rises Up, Marvell-Polariton”.
Meta + Broadcom: four MTIA generations, multi-gigawatt rollout
Meta and Broadcom announced a multi-year partnership to co-develop the next four generations of MTIA chips. Initial commitment is over 1 GW, with a “sustained, multi-gigawatt rollout” to follow. Hock Tan moved from Meta’s board into an advisor role.
Vik: The headlines focus on XPUs and multi-gigawatt deployments, but this is a networking/optics story too. The fine print mentions “advanced Ethernet” and four chip generations in two years. That’s a lot of Tomahawk-enabled networking. Millions of optical lanes, the SerDes and DSPs to drive them, and so on.
Austin: I didn’t know Hock Tan was on Meta’s board!
Related: Austin’s interview with Meta VP Matt Steiner on ads infra, GPUs, and MTIA.
AMD upgraded to Buy at DA Davidson, $375 PT
INTC was the bellwether.
Austin: I’m looking forward to AMD’s earnings on May 5th. If Lisa has a margarita in hand, we’ll know AMD’s CPU sales were EPYC this quarter. Cinco de Mayo!
ON THE POD
Latest on Semi Doped: a masterclass on Google’s TPU v8 networking. Two TPU chips? Old news. Two scale-up networking topologies? Now we’re talking. Workload-specific interconnects, in the wild.
Austin: We get technical on this one, and folks are loving it. Almost 90K views on X! I sent it to my mom. She said, “Keep doing what you’re doing!” Agree, Mom.
Watch on YouTube · Subscribe on Spotify · Subscribe on Apple · Listen here
That’s it for today!