It’s 2:14 in the morning and I’m watching a progress bar. My agent stack is chewing through a batch of 40,000 support tickets, classifying them, summarizing them, and writing draft replies. The job is running on rented GPUs in a data center I will never see. On my second monitor, a cost dashboard ticks up in real time. Every tick is a small payment to the same handful of companies that keep showing up in analyst notes.
That’s the odd thing about building bots for a living. You end up with a ground-level view of an industry you don’t work in. I don’t trade semiconductor stocks. But I pay their bills every night, and the pattern of those bills has started to look suspiciously like the thesis a bunch of Wall Street analysts have been writing for the past year.
What the analysts are saying
Bernstein remains strongly bullish on Nvidia and Broadcom, expecting both to hold market leadership into 2026 on the back of AI and semiconductor demand. Arya, the analyst behind the call, framed the reasoning in a way that stuck with me: a preference for companies with “moats that are quantified by their margin structure.” Beyond those two names, he pointed to four other large-cap semiconductor firms as part of a projected trillion-dollar chip surge in 2026.
The argument for Nvidia gets made bluntly elsewhere too. One analyst discussion described it as an “unassailable monopoly” because the company owns the software ecosystem on top of the hardware lock-in. Broadcom’s case rests on different ground — leadership in networking and custom silicon. Meanwhile AMD’s progress across GPU and CPU architecture supports expectations for real market share gains. Broadcom also showed up on a separate list of AI stocks rated as undervalued as of July 24, 2026, alongside Microsoft, Taiwan Semiconductor, Tencent, and Alibaba.
The part that matches what I see
I want to be careful here. I’m not qualified to tell you whether any of these stocks are priced correctly, and this isn’t investment advice. What I can tell you is which parts of that thesis line up with what happens when you actually ship bots.
The software ecosystem point is the one I feel most directly. When I pick a hosting target for inference, the decision is rarely about raw chip specs. It’s about whether the libraries work. Whether the quantization tooling I already use is supported. Whether the profiler tells me something useful when latency spikes. Whether the Stack Overflow answer from 2023 still applies. That accumulated layer of tools, docs, and tribal knowledge is a real cost to walk away from, and I’ve watched teams quietly abandon cheaper hardware because the switch ate three weeks of engineering time. Analysts call that a moat. I call it the reason my Dockerfile hasn’t changed in a year.
The networking angle is the one I underrated for the longest time. Early on I thought about compute as the whole story — more FLOPs, faster tokens. Then I started running multi-node jobs and learned that interconnect is where the real bottleneck hides. A distributed training run that spends half its time waiting on data movement is a distributed training run you’re paying double for. Broadcom’s position in networking and custom silicon reads differently once you’ve stared at that particular bar chart.
What this means if you build things
The practical takeaway for bot builders isn’t about stock picking. It’s about recognizing that your architecture decisions are being shaped by a small number of very solid supply chains, and planning accordingly.
- Write your inference layer behind an interface. If swapping providers means touching forty files, you’ve given up your only real negotiating position.
- Measure cost per task, not cost per hour. An expensive chip that finishes in a third of the time can be the cheap option. I’ve been wrong on this in both directions.
- Treat interconnect as a first-class concern the moment you go multi-node. Profile data movement before you buy more compute.
- Assume the ecosystem tax is real and budget for it. The cost of leaving a tooling ecosystem is mostly your team’s time, and time doesn’t show up on the invoice.
There’s a version of this story where the analysts are right about market leadership and wrong about margins, or right about 2026 and wrong about 2028. I have no edge on that question. What I do have is a monthly bill that keeps telling me the same thing the research notes do — that a few companies sit at a choke point in how AI actually gets deployed, and that the software stack around the silicon matters at least as much as the silicon.
My progress bar finished around 3:40. The batch cost less than a decent dinner, which still strikes me as absurd. Five years ago that job would have needed a research lab. Now it needs a credit card and a competent understanding of where the bottlenecks live. Those bottlenecks have names, and they show up in quarterly earnings reports.
🕒 Published: