Alphabet's SpaceX Stake, Waymo, and the AI Capex Push
Alphabet trades like an ad company but operates like a holding firm. Inside the GOOGL share sit a minority piece of SpaceX, a majority piece of Waymo, and a growing AI infrastructure tab. Today’s notes tie those threads together with a side trip into Ethereum protocol plumbing.
The Alphabet route into private space
Direct exposure to SpaceX is still hard for ordinary investors. The closest publicly traded wrapper, the closed end fund DestinyTech100, often trades far above its net asset value. Premiums of 150% or more have shown up at peak interest. That is a polite way of saying buyers sometimes pay 2.50 dollars to receive 1 dollar of underlying holdings.
Alphabet offers a quieter route. Its early SpaceX position, taken alongside Fidelity in 2015, has compounded as Starlink revenue and Falcon launch cadence scaled. Current estimates place Alphabet’s effective stake at roughly 5 to 6 percent of SpaceX, depending on how later rounds have diluted the slice. That single line of GOOGL exposure is liquid, settled on regulated exchanges, and free of any closed end discount or premium swing.
The same wrapper carries a majority stake in Waymo, the self driving unit. Waymo runs paid robotaxi service in Phoenix, San Francisco, Los Angeles, Austin, and a growing list of cities. Investor presentations have put Waymo’s annualized ride count well into the millions, although none of this shows up cleanly on Alphabet’s segment reporting today.
AI capex and the Missouri build
Alphabet announced a fresh 15 billion dollar AI data center in Missouri. That follows similar size commitments in Texas, Oklahoma, and South Carolina across the last 24 months. Combined with peer spending from Microsoft, Meta, Amazon, and the Stargate partners, US AI capex for 2026 is tracking toward the 350 billion dollar line on the latest dataset estimates.
Power is the binding constraint, not silicon. The Missouri site will pull hundreds of megawatts at full ramp, and hyperscalers continue to sign long term contracts for nuclear restart capacity and modular reactor offtake to feed the load. Grid operators across MISO and PJM territory now treat hyperscaler demand as the largest single source of load forecast risk.
Two numbers are worth holding in mind. First, hyperscaler capex now runs above 40 percent of operating cash flow at several names, a level last seen during the late 1990s build out. Second, depreciation lags spend by roughly 3 to 4 years. That gap means reported earnings still look healthy even as cash leaves the door. The next 12 months will test how patient public market investors stay with that pattern.
Search redesign and the user base
At Google I/O, executives noted that 5 Alphabet apps now sit above 3 billion monthly users each. Search, Chrome, Android, YouTube, and either Gmail or Maps depending on counting methodology all clear the bar. For comparison, Meta’s flagship Facebook app shows roughly 3.1 billion monthly users.
The new Search redesign blends generative answers, video panels, and shopping modules into the standard results page. The risk for Alphabet is direct. AI answers can cannibalize click through to publishers, which over time may erode the very inventory that funds search ads. The opportunity is also direct. Multimodal answers keep users inside Alphabet properties longer per session.
Amazon and the legacy studios share screen time with Alphabet on the video and shopping front. YouTube paid subscribers passed 100 million several quarters ago, and YouTube Shorts revenue has begun to print at scale.
Ethereum protocol work on block access lists
On the other side of the technology stack, Ethereum client teams continue work on the block access list, or BAL. A merged pull request on the Geth client reworks BAL JSON marshalling to align with the execution layer specifications maintained in the EELS repository. A follow up review noted that the correct type for the block access list inside EELS is Bytes, meaning an RLP encoded blob, rather than the structured object form that landed in the first pass.
This sounds like pure plumbing. It is. BAL is part of the work toward stateless and partially stateless block validation. Once it lands cleanly across clients, BAL lets validators verify blocks with smaller state proofs, which lowers hardware needs for new node operators and helps long term decentralization.
The fix has not yet shipped to mainnet. A test net called bal devnet 7 already carries the corrected encoding. The cadence here is slow on purpose, since changes to block format ripple across every client implementation and every block explorer that parses raw block data.
What to watch
Three threads stay live for the rest of the week. First, watch whether Alphabet starts to give cleaner segment level reporting on Waymo and the SpaceX line, since pressure for sum of parts disclosure will only grow as those private positions appreciate. Second, watch power purchase agreements out of the hyperscalers, because the marginal AI build now bids against grid and nuclear timelines, not against chip supply. Third, watch Ethereum client release notes for BAL adoption, since the encoding choice affects every future stateless validator deployment.