In today's capital markets, if you haven't made it into the $4 trillion club, you don't really get to walk tall. The club currently has exactly two diamond members: NVIDIA ($4.77T) and a freshly re-enthroned Alphabet ($4.01T).
In this war for the AI crown, some are betting on the brute-force beauty of GPUs. My own conviction: this war is ultimately won by the company that owns the whole stack. Here's the full case.
1. Google's "Swiss Army Knife" Playbook: It's Not Just Search Anymore
Stop thinking of Google as a search engine that serves ads. Alphabet's current lineup is basically the AI Avengers:
- Gemini — frontier large language model
- DeepMind — the research lab shipping production-grade breakthroughs
- Waymo — already several lengths ahead of Tesla in autonomous driving
- YouTube — the world's largest streaming video platform
And they hold a hidden trump card: SpaceX equity. Reports suggest the SpaceX IPO could deliver Google an estimated $100 billion return on its stake. NVIDIA's pockets don't contain any Musk rockets.
The most enviable number: Google Cloud's backlog stands at $240 billion, up 55% QoQ. That "too many orders to handle" feeling is what people call a happy problem.
2. Broadcom: The Low-Profile "Top Arms Dealer"
If Google is the frontline general charging into battle, Broadcom (AVGO) is the quartermaster in the rear — handing out weapons, paving the roads, and laying the rail tracks while they're at it. The Google-Broadcom TPU partnership now spans over a decade — less a business relationship, more a childhood friendship in compute.
Broadcom's XPU business is on autopilot:
- Producing TPUs on 2nm for Google this year
- Landed Meta, ByteDance, OpenAI, and Anthropic along the way
- Apple may preach "privacy first" in public, but behind the scenes seems dependent on Broadcom's custom silicon and Google's Gemini
One line summary of Broadcom: the network is the computer. While everyone else scrambles for GPUs, Broadcom's industry-leading Tomahawk switches and SerDes technology solve the fatal data bottleneck inside AI clusters.
3. The 2027 Thesis: From "Training" to "Inference" — the Great Power Transfer
The market is currently obsessed with the NVIDIA-dominated training market. My bold prediction: by the end of 2027, Alphabet overtakes NVIDIA as the world's most valuable company.
The reasoning is straightforward:
- Training is just the beginning; inference is the real final frontier. Training is one-time CapEx; inference is eternal OpEx. A customer who trains a model can turn the GPUs off. But every single user query to that model is real revenue on the inference side.
- Google's platform monetization potential dwarfs a single-point platform built around GPUs + CUDA. Search, Ads, YouTube, Cloud, Workspace — every surface is an inference demand generator.
- The Google-Broadcom TPU is proving its dominance on cost-efficient compute. For inference workloads, the cost/watt/dollar optimum isn't an H100 — it's a TPU.
4. Beyond the Humor: What Are the Real Risks?
The path to $5 trillion has potholes. The biggest risk is not technology — it's gas.
- High-purity helium: If TSMC's helium supply chain snaps, it doesn't matter whether you're shipping XPUs or GPUs. Production stops.
- Middle East geopolitics: Oil spiking to $200/barrel crushes valuation multiples across the board.
- Antitrust: Google's search antitrust case is still live. A worst-case breakup of Chrome or the default-search agreement would hit Alphabet's earnings materially.
5. Conclusion: Strong Buy, Don't Get Off
- Alphabet (GOOG): Strong Buy. Target: cloud revenue share hits 20% by 2027, with revenue growth sustained above 15%.
- Broadcom (AVGO): Strong Buy. 12-month price target: $550. Rationale: it is the irreplaceable "clot-buster" in the AI compute architecture, with projected $70B free cash flow by 2027.
The underlying bets: the power transfer from training to inference, platform companies outcompounding single-point companies on multiples, and TPU's cost advantage over GPUs in inference workloads.
FAQ
When will Alphabet overtake NVIDIA?
Based on the current market cap gap (NVIDIA $4.77T vs Alphabet $4.01T, roughly a 19% gap) and the relative growth rates of Cloud and AI inference, my target window is end of 2027. Key catalysts: Google Cloud's $240B backlog converting to revenue, Waymo's commercialization at scale, and equity crystallization from the SpaceX IPO.
Why is Broadcom (AVGO) worth buying?
Broadcom is the "network-layer hegemon" of AI infrastructure — it doesn't make GPUs, but AI clusters literally cannot run without its switches. Its XPU business covers nearly every major AI lab outside the NVIDIA camp (Google, Meta, ByteDance, OpenAI, Anthropic). It's the classic "sell picks and shovels" play.
What does the training-vs-inference distinction mean for investors?
Training is one-time CapEx — once a model is trained, the hardware can sit idle. Inference is perpetual OpEx — every user interaction generates inference demand. Investment implication: the inference market scales linearly with user base and has a far higher ceiling than training. Whoever is cheaper, faster, and more reliable in inference wins the next decade.
What is the biggest risk variable?
Supply chain (especially TSMC helium) > geopolitics (Middle East oil) > antitrust (Google search breakup). The first two are short-term tail risks that can blow up suddenly; the third is a slow-moving variable.
Risk Disclaimer
This article reflects the author's personal views and does not constitute investment advice. All investments carry risk, including potential loss of principal. Tech stock valuations are subject to macro liquidity, regulatory policy, supply chain, and geopolitical factors — volatility is material. Readers should conduct their own due diligence and consult qualified financial advisors before making investment decisions.