Who is betting on what in AI, and against whom.
The camps argue about how intelligence works. These companies get paid either way. None of them takes a position on the fault lines, which is why they are not labs in the atlas, and most of the listed exposure to the AI trade runs through them rather than through the camps, which are mostly private. Each row says what the company sells, which camps buy it, and any dated fact that matters, with a source. What it does not say is which one is the better position.
The chips every camp trains on. One company sets the price.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
NASDAQ:NVDA |
GPUs and the CUDA stack for training and inference; also an investor in labs across every camp | every camp | Sep 2026 Agreed to buy Hugging Face for $12.93B, expected to close in the first half of 2027. TechCrunch |
Nvidia | |
AMAMD |
NASDAQ:AMD |
Instinct accelerators; the second source for training and inference GPUs | LLM scaling World models Physics AI | AMD | |
Broadcom |
NASDAQ:AVGO |
Custom accelerators built with the hyperscalers, including Google's TPU, and data-center networking silicon | LLM scaling World models | Broadcom |
Silicon built on the assumption that the transformer is settled enough to harden into hardware.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
Cerebras |
private | Wafer-scale chips sold as fast inference and as a cloud; has filed to go public | LLM scaling | Cerebras | |
Groq |
private | LPU inference chips and the GroqCloud service | LLM scaling | Groq | |
SambaNova |
private | Reconfigurable dataflow chips sold for inference | LLM scaling | SambaNova | |
Etched |
private | A transformer-only ASIC; the purest bet that the architecture is final | LLM scaling | Etched |
Everything above runs through two or three companies that make or equip the fabs.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
TSTSMC |
NYSE:TSM |
Fabricates nearly every accelerator above, including all of Nvidia's | every camp | TSMC | |
ASML |
NASDAQ:ASML |
The only maker of the EUV lithography machines advanced fabs need | every camp | ASML | |
Applied Materials |
NASDAQ:AMAT |
Fab equipment for deposition, etch and inspection | every camp | Applied Materials |
Training is bound by memory bandwidth as much as by compute; three suppliers make HBM.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
SHSK hynix |
KRX:000660 |
The leading supplier of high-bandwidth memory for accelerators | LLM scaling World models | SK hynix | |
Micron |
NASDAQ:MU |
High-bandwidth memory and the DRAM that surrounds it | LLM scaling World models | Micron | |
Samsung Electronics |
KRX:005930 |
HBM and DRAM, and a foundry of its own | LLM scaling World models | Samsung Electronics |
The switches and boxes that turn ten thousand GPUs into one machine.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
Arista Networks |
NYSE:ANET |
Ethernet switching for AI clusters | LLM scaling | Wikipedia | |
Marvell |
NASDAQ:MRVL |
Custom silicon and interconnect for hyperscale data centers | LLM scaling | Marvell | |
SMSuper Micro Computer |
NASDAQ:SMCI |
GPU servers and racks | LLM scaling | Wikipedia | |
Dell Technologies |
NYSE:DELL |
AI servers for enterprises and neoclouds | LLM scaling | Dell Technologies |
The three clouds rent compute to every camp and run their own labs.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
NASDAQ:MSFT |
Rents compute to OpenAI and others; runs Microsoft AI's own models | every camp | Microsoft Azure | ||
NASDAQ:AMZN |
Trainium chips and the cloud Anthropic trains on; runs Amazon's own AGI lab | every camp | Amazon Web Services | ||
NASDAQ:GOOGL |
TPUs rented to outside labs, and the compute behind Google DeepMind | every camp | Google Cloud | ||
Oracle |
NYSE:ORCL |
Cloud capacity leased at scale to model labs; the most indebted of the clouds | LLM scaling | Jul 2026 Interest expense doubled and free cash flow was negative $4.7B for the first quarter, per the credit-market reporting on the capital page. Tech Times |
Oracle |
GPU landlords built for AI demand. The most leveraged part of the trade, and where the credit-market doubts sit.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
CoreWeave |
NASDAQ:CRWV |
GPU cloud built on Nvidia hardware and debt secured by it | LLM scaling World models | Jul 2026 Five-year credit default swaps reached about 855 basis points in July 2026. 24/7 Wall St. |
CoreWeave |
Nebius |
NASDAQ:NBIS |
GPU cloud spun out of Yandex's non-Russian assets | LLM scaling World models | Nebius | |
Nscale |
private | GPU cloud that signs dedicated compute deals with labs | LLM scaling Physical AI | Sep 2026 Committed $3.5B of compute to Figure, with intent to exceed $6B, and took an undisclosed equity stake. Tech Times |
Nscale |
Data centers need firm power; nuclear operators and gas plants signed the contracts.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
Constellation Energy |
NASDAQ:CEG |
The largest US nuclear fleet, selling firm power to data centers | LLM scaling | Constellation Energy | |
Vistra |
NYSE:VST |
Nuclear and gas generation contracted to data-center growth | LLM scaling | Vistra | |
Talen Energy |
NASDAQ:TLN |
Nuclear power sold directly to a data-center campus | LLM scaling | Talen Energy | |
NENextEra Energy |
NYSE:NEE |
The largest US renewable developer, building for data-center load | LLM scaling | NextEra Energy | |
Oklo |
NYSE:OKLO |
Small modular reactors pitched at data centers; no plant operating yet | LLM scaling | Oklo | |
NuScale Power |
NYSE:SMR |
Small modular reactor designs; no plant operating yet | LLM scaling | NuScale Power |
The buildings, sold as leases.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
Equinix |
NASDAQ:EQIX |
Colocation data centers leased to clouds and labs | LLM scaling | Equinix | |
Digital Realty |
NYSE:DLR |
Data-center real estate leased at hyperscale | LLM scaling | Digital Realty |
The incumbents camp 3's learned simulators are trying to replace or plug into.
| Company | Ticker | Sells | Sells to | Dated facts | Sources |
|---|---|---|---|---|---|
SYSynopsys |
NASDAQ:SNPS |
Chip design software and, through Ansys, the physics simulation camp 3 wants to learn | Physics AI | Synopsys |
Missing a supplier that sells to a camp? Send it, with a source, through the corrections form. Companies that merely use AI are out of scope. The bearish side of the same trade is on the Capital page.