The processors that actually run AI models — the reason every other sector on this map exists.
Sorted by one-month move. “Vs target” compares the current price with the mean analyst price target — a figure from analysts, not from us.
| Company | Price · todayThe latest closing price, and how far it moved from the previous close in dollars and percent. | 1 MonthPrice change since the last close on or before the same date one month ago — a calendar month, not 21 trading days. | 1 YearPrice change since the last close on or before the same date one year ago. | Vs targetHow far the price sits below the average of analysts’ published 12-month price targets. Negative means the price has already risen past that average.Analysts’ figures, not ours. Not a forecast we make or endorse. |
|---|---|---|---|---|
| AMDAdvanced Micro | $559.82+$14.73 (+2.70%) | ▲ 15.6% | ▲ 255% | ▲ 10%$365 … $1250 |
| MRVLMarvell | $244.25+$3.49 (+1.45%) | ▲ 13.1% | ▲ 230% | ▲ 18%$210 … $400 |
| INTCIntel | $108.60-$0.20 (-0.18%) | ▲ 12.3% | ▲ 255% | ▲ 7%$75 … $200 |
| QCOMQualcomm | $177.72-$10.99 (-5.82%) | ▲ 11.5% | ▲ 8% | ▲ 9%$100 … $400 |
| AVGOBroadcom | $357.61+$10.31 (+2.97%) | ▼ 5.9% | ▲ 4% | ▲ 49%$216 … $715 |
As of close, Sep 18, 2026 · prices from Yahoo Finance
Data: daily close · Reporting, not recommendations · Not financial advice.
AI compute is the layer everything else on this map exists to serve. These are the processors that actually run AI models — general-purpose GPUs that can train almost anything, and custom chips designed by a single buyer for a single job.
The economics are unusual. A handful of designers capture most of the value, but none of them own factories; they rent capacity from foundries and depend on packaging and memory suppliers to finish the part. That makes this sector a demand signal for the six or seven sectors sitting behind it.
What moves it: hyperscaler capex announcements, each new chip generation, and any hint that a large buyer is shifting volume between suppliers or toward its own custom silicon.
The Weekly Pulse covers AI Compute (GPUs & Custom Chips) alongside every other sector — one email, Sunday, free.
A GPU is general-purpose — it can train or run almost any model, which is why it dominates when workloads are still changing. A custom chip (often called an ASIC) is built for one company's specific workload. It is cheaper per unit of work once the workload is stable, but useless if the job changes.
Building a leading-edge fab costs tens of billions and takes years. Most designers went "fabless" decades ago, handing manufacturing to dedicated foundries so they could focus capital on design. It is why a designer's output is capped by someone else's capacity.
Historically it has been advanced packaging and high-bandwidth memory rather than the processor itself. A finished AI accelerator needs memory stacked onto the logic die, and that packaging capacity has repeatedly been the binding constraint.