06-reference

technically google ai full stack bet

2026-09-15·reference·source: Technically·by Paolo Perrone
google-cloudhyperscaler-strategyai-platform-layerdistributionharness-engineering

Why this is in the vault

A guest-post framework (four layers: chips, models, platform, distribution) for reading Google's, Microsoft's, and Amazon's competing full-stack AI bets — useful shorthand for where RDCO's own bets sit in that stack. Note on completeness: this issue is a paid-subscriber preview; only the TL;DR and the framework setup rendered (confirmed against the web version via WebFetch, which returned the same preview-gated content). The layer-by-layer breakdown of each company is behind the paywall and not captured here.

The core argument

Perrone frames Google, Microsoft, and Amazon's AI strategies with a coffee-company analogy: each wants to grow the beans (chips), roast them (models), run the cafés (the platform developers build and govern agents on), and own the busiest streets (distribution — the channels that put agents in front of users). A few years ago each company rented most of the stack and owned a layer or two; now all three are building the whole chain.

The visible-in-preview positioning: Google's lead is chips (TPUs, "a decade head start") and models (Gemini 3, plus its Anthropic funding/hosting stake). Google's gap is distribution — Microsoft walks agents straight into the Office apps already sitting on 90% of Fortune 500 desktops, while Google's Workspace/Gmail reach is narrower and Amazon barely reaches non-developer users at all. The named trigger event: at Cloud Next, Google replaced Vertex AI (its AI platform since 2023) with the Gemini Enterprise Agent Platform — Thomas Kurian's framing on stage was "competitors hand you the pieces, Google hands you the platform." Perrone's twist: Google has the most to lose in this race, since for Microsoft and Amazon AI is incremental revenue, while for Google "every agent answer is a search that doesn't happen" — the AI layer cannibalizes the core business in a way it doesn't for its two rivals.

The issue also carries a house plug for Technically's own live workshop (Sung Won Chung, Oct 1) — self-promotional, not a third-party sponsor, and unrelated to the article's content.

Mapping against Ray Data Co

This four-layer read (chips → models → platform → distribution) is a clean external frame for where RDCO's own L5 bet actually sits: per [[project_l5_north_star_strategic_direction]], RDCO's bets are downstream of agent capability, not chip or model ownership — RDCO has never competed on the bottom two layers and doesn't intend to. The piece's implicit thesis that chips and models are becoming table stakes while platform + distribution is where the real fight is happening echoes the harness-engineering framing already anchored in the vault ([[2026-07-21-technically-harness-engineering]]: model quality commoditizes, the harness is the moat) — Google, Microsoft, and Amazon fighting over the "platform you build and govern agents on" layer is the enterprise-scale version of the same claim. Concretely relevant to phData/OI: Google's Gemini Enterprise Agent Platform replacing Vertex AI is exactly the kind of hyperscaler platform move that will reshape what "the platform layer" means for enterprise data/AI consulting engagements — worth a watch-item for how OI's Snowflake-centric positioning interacts with a Google land-grab on the same layer. The TPU "decade head start" claim is also a data point for the chip-fab/memory capital-cycle investing thesis ([[project_investing_markov_capital_cycle]]) — Google is the one hyperscaler with in-house silicon at scale, a structurally different position than Microsoft/Amazon's more merchant-silicon-dependent stacks.

Related