GOOGL, not NVDA, is framed as the AI winner because it owns the full stack: models, TPU chips, data centers, networks, Search, Chrome, and YouTube. That vertical integration lets it cut Gemini serving costs 78% last year and another 30% since Gemini 3 launched last November, while still scaling usage to 3.2 quadrillion tokens a month and 900M+ app users. But the earnings math is strained: Q2 free cash flow was -$5.9B, capex was $45B in one quarter, 2026 capex was raised to $200B, and $811B of future purchase agreements now stretch to 2054. The thesis is that GOOGL can absorb the spend because it monetizes every layer and is now selling TPUs directly, while MSFT, AMZN, NVDA, AMD, AVGO, TSMC, INTC, SSNLF, MU, CRWV, NBIS, Airen, ANET, COHR, and LITE each own only pieces of the stack.
Training, fine-tuning, and inference are separated into three cost layers: training burns cash on more data, chips, time, and parameters; fine-tuning uses feedback, guardrails, and distillation to adapt open-source models like Llama, DeepSeek, or Chimney K3; inference happens on every prompt and is where AI spend compounds with usage.
OpenAI and Anthropic are exposed to training costs, while inference gets more expensive as an AI product succeeds, because customers control prompt length, prompt complexity, prompting frequency, and output format.
Ground News analyzes 60,000+ articles a day; in its Elon Musk/Odyssey example, left-leaning coverage outnumbered right-leaning coverage almost 3-to-1, only half the sources had high factuality, and blind-spot feeds showed which stories each side ignored.
GOOGL owns the whole AI stack: Chrome and Gemini, TPUs, its own data centers, fiber networks, private undersea cables, and YouTube; its TPUs have been running in Google data centers since 2015.
GOOGL’s Gemini models are optimized for TPU chips and Google-owned servers, and the company can swap models, chips, or cooling without supplier approval because it controls the stack end to end.
On July 3, GOOGL marked 10 years since TPUs launched and said it split eighth-generation TPUs into 8T for training and 8I for inference.
Google said those optimizations cut Gemini serving costs 78% last year and cut core AI response costs another 30% since Gemini 3 launched last November.
Gemini now powers 13 products and services with over 1B users each; AI Overviews reaches 2.5B people monthly, AI mode passed 1B monthly active users a year after launch, and the Gemini app hit 900M monthly active users and is more than doubling year over year.
Google said more than 8.5M developers build on its models every month, Gmail blocks 10M spam emails every minute, and it processes more than 3.2 quadrillion tokens a month, more than 300x two years ago.
Alphabet reported Q2 on July 22, seven weeks after the July 3 presentation: revenue was $120B, up 24% y/y; net income was $112B, but $99B came from stock gains, mostly ANTH and SPACE; excluding paper gains, EPS was about $2.85, below expectations.
Alphabet’s operating income rose 30%, not 300%; Google Cloud revenue hit $24.8B, up 82% y/y, with growth accelerating from 48% to 63% to 82% y/y and operating income tripling from $2.8B to $8.8B, lifting margin from about 21% to 36%.
Google Cloud backlog rose from $462B to $514B, above half a trillion dollars, and Google Cloud now generates product revenue from direct TPU sales, which the notes say is the first quarter Alphabet recognized revenue from selling TPUs directly.
For a decade TPUs were only rentable through Google Cloud, but now they are being installed directly in other companies’ data centers, with product revenue expected to ramp over the next few quarters.
Alphabet spent $45B in Q2 on land, buildings, chips, and cooling, doubled 2025’s same quarter, raised 2026 capex from a $185B midpoint to $200B, and said 2027 spending will be much higher.
Alphabet signed $811B of future purchase agreements for chips, equipment, data centers, and electricity, with $200B due in the next year; three months earlier the total was $332B, and some contracts run to 2054.
Free cash flow was -$5.9B in Q2 after $24.6B two quarters ago and still above $10B last quarter; Alphabet has started issuing shares, selling long-term bonds, stopped buybacks for the first time since 2017, doubled long-term debt in six months, and raised its interest bill about 5x.
The stock fell 7% the day after earnings, even as the case is that GOOGL can justify the spend because it is the only company said to own every layer of the AI stack and can monetize the browser window as well as the models.
MSFT has Maia chips, Azure data centers, Windows and Office, but still uses OpenAI models for its heaviest workloads; AMZN uses Anthropic’s models; NVDA, AMD, and AVGO design chips; TSMC, INTC, SSNLF, and MU build them; CRWV, NBIS, and Airen rent out data centers; ANET, COHR, and LITE build networks.