I’m holding all three.
Nvidia, Microsoft, and Meta sit in my portfolio right now, and if you’re building bots in 2025, you’re probably watching these companies too. Not because they’re safe bets or because some analyst said so, but because they’re the infrastructure layer for everything we’re building.
But here’s what changed: I’m not buying more. I’m watching.
Why I’m Not Adding to Nvidia
Nvidia powers the training runs that make our bots possible. Every large language model, every vision system, every multimodal agent—they all run on H100s and now B200s. The company prints money because AI demand is real.
But the recent chatter about an AI bubble isn’t just noise. NPR and multiple analysts are raising flags about circular deals—companies buying AI infrastructure to build products that justify buying more infrastructure. When you’re neck-deep in bot architecture like I am, you see it: some of these enterprise AI deployments are solutions looking for problems.
I’m keeping my Nvidia shares because the compute demand is genuine for those of us actually shipping. But I’m not adding at these valuations. The gap between hype and utility is widening, and Nvidia’s stock price assumes every AI experiment becomes production.
Microsoft: My Steady Hold
Microsoft is different. Azure OpenAI Service is how most of my client projects access GPT-4 and other models. The integration is clean, the enterprise features actually matter, and the pricing is predictable enough to build a business on.
What I like: Microsoft isn’t just selling picks and shovels. They’re using their own tools. Copilot is embedded in Office, GitHub, and Windows. That’s real distribution to real users doing real work. When I build bots for companies, they’re often already in the Microsoft ecosystem.
The risk? They’re heavily exposed to OpenAI’s success. If GPT-5 disappoints or if open models catch up faster than expected, Microsoft’s AI premium evaporates. But for now, they’re the safest of my three positions. I’m holding and might add on any significant dip.
Meta: The Contrarian Play
Meta is my most interesting position. While Nvidia and Microsoft are priced for AI dominance, Meta is still trading like a social media company that happens to do AI.
Here’s what matters for bot builders: Llama 3 is legitimately good. I’ve deployed Llama-based systems for clients who can’t or won’t send data to external APIs. The open-weight approach means I can fine-tune, self-host, and control costs in ways that aren’t possible with closed models.
Meta is spending billions on AI infrastructure—$65 billion in 2025 alone according to recent reports—but they’re giving away the models. The market hasn’t figured out how to value this yet. Are they subsidizing competitors or building the Android of AI?
I’m holding Meta because the downside feels limited and the upside is asymmetric. If open models become the default for production systems, Meta’s investment pays off in ways that aren’t obvious yet. If not, they’re still printing money from Instagram and WhatsApp.
What I’m Actually Doing
I’m not selling any of these positions. But I’m not adding either, except maybe Microsoft on weakness.
The AI infrastructure trade worked brilliantly for two years. Now we’re in a different phase. The companies that win from here aren’t the ones with the most GPUs or the biggest models. They’re the ones that figure out what AI is actually good for.
As someone building bots daily, I see the gap between capability and application. The models are incredible. The use cases are still emerging. That gap is where the bubble concerns come from, and they’re not entirely wrong.
My portfolio reflects this: I own the infrastructure layer because it’s necessary, but I’m not betting the farm on continued exponential growth. The next phase belongs to the builders who figure out what to do with all this compute power.
And that’s exactly where I’m focusing my energy—not on buying more stock, but on shipping bots that actually solve problems.
đź•’ Published: