TL;DR — Past the daily headline noise, Alphabet’s AI advantage comes down to unit economics. Custom TPUs push its inference cost curve below what GPU buyers can match, long-term power deals secure the input that limits everyone, and Genie and glasses are options on top. That is compounding at the infrastructure level.
Watching Big Tech’s daily price swings is a good way to develop an ulcer. Zoom out past the end-of-the-world headlines and the Alphabet story gets simpler. It is not magic. It is the ordinary machinery of compound returns and market position, applied to a new product cycle.
In the cold logic of a competitive market, the winner over time is the producer with the lowest unit cost. In AI, the unit is a query, a token or a generated frame. Alphabet is using its custom silicon to drive an inference cost curve that is increasingly hard for competitors to follow.
Owning the Chip Is How Alphabet Wins on Unit Cost
Most AI companies rent intelligence from someone else’s hardware. They buy GPUs, mostly from Nvidia, and pay Nvidia’s margin on every one. Alphabet designs its own accelerator, the Tensor Processing Unit, and has done so for about a decade.
The seventh generation, Ironwood, reached general availability in November 2025. Outside demand is a useful test of whether the chip is competitive, and it passed one: Anthropic said in October 2025 it plans to use up to one million TPUs for its Claude models. We break down the hardware in detail in our TPU v6 and v7 analysis.
The economics compound in three ways:
- Lower cost per query, because Alphabet captures the margin a chip vendor would otherwise keep;
- Faster iteration, because model and chip teams sit inside the same company and can tune one for the other;
- A second revenue line, because the same TPUs can be rented to Google Cloud customers.
In the cold logic of capitalism, the winner is the one with the lowest unit costs.
Electricity Is the One AI Input Nobody Can Print
While the public argues about whether AI writes bad poetry, Google is signing power contracts. The relevant partner is NextEra Energy. In October 2025, Google signed a 25-year agreement tied to restarting NextEra’s Duane Arnold nuclear plant in Iowa. In December, the two companies agreed to co-develop multiple gigawatt-scale data center campuses.
This is not about brochures. It is about securing the literal electrons that keep servers running when every hyperscaler wants the same grid capacity. In the 19th century, you owned the rails. In the 21st, you own the compute and the power, and Alphabet is assembling both.
Scale explains why this matters. NextEra has said it expects to add at least 15 gigawatts of new generation for data centers by 2035. Being an anchor customer for that pipeline is a cost advantage of its own, because power secured early is cheaper and more reliable than power bought late.
Genie and Glasses Are Options on Top of the Core Business
There is plenty of noise about Project Genie disrupting video game stocks. Google began rolling it out on January 29, 2026, to its top AI Ultra subscribers in the US. The tool generates interactive 3D worlds from a prompt. Making games is slow and expensive, and Alphabet wants to make parts of it fast and cheap.
Traditional studios are nervous because the moat of having hundreds of developers drawing trees for years looks thinner. Still, Genie today produces short sessions, not finished games. Treat it as an option, not a revenue line.
The market did not wait for that nuance. On January 30, Bloomberg reported Unity down 24%, Roblox down 13% and Take-Two down 7.9%. That is classic market evolution: capital moves before the revenue does.
The same goes for AI smart glasses. Google Glass was a commercial failure a decade ago. In December 2025, Google said new Gemini-powered glasses built with Samsung, Warby Parker and Gentle Monster would arrive in 2026. If Alphabet owns the interface on your face, it worries less about who makes the phone in your pocket.
Update: The Unit-Cost Argument Showed Up in the Numbers
Update, October 2026. On the February 4 call, Sundar Pichai said Google lowered Gemini serving unit costs by 78% over 2025 through model optimization, efficiency and utilization gains. That is the cost curve described above, stated by management as a single figure.
The operating results followed. In the second quarter of 2026, Google Cloud revenue grew 82% to $24.8 billion and its operating margin reached 35.6%, up from 20.7% a year earlier. The capital bill grew too: 2026 capex guidance now stands at $195 billion to $205 billion, and quarterly free cash flow turned negative at minus $5.9 billion.
Shareholders also absorbed something the original thesis did not anticipate. In June 2026, Alphabet sold Class A and Class C stock plus mandatory convertible preferred shares for $49.6 billion in net proceeds to fund AI infrastructure. Our guide to Alphabet’s share classes explains how those securities sit alongside existing GOOGL and GOOG holdings.
What to watch next: whether management updates the serving-cost figure, Cloud operating margin as new capacity comes online, the pace of capex against operating cash flow, and any further equity issuance. The capitalist engine is running at full throttle. Holding the shares is a bet that unit costs keep falling faster than the capital bill rises.
Sources: Alphabet earnings releases (Q3 2025, Q2 2026) and filings on SEC EDGAR; reports and webcasts at Alphabet Investor Relations; Sundar Pichai’s Q4 2025 earnings remarks. Disclosure: the author owns Alphabet shares (GOOGL/GOOG). This is not investment advice, so do your own math. Our full disclosure is on the About page.
