Thaisan Tonthat | Writing

Which model is the better stockpicker?

Motivation:

I was curious to learn more about the physical components that power AI, so I spent 30 minutes asking ChatGPT about what goes into a data center, how they are built, etc. I dug into the components and potential bottlenecks from GPUs to electrical to cooling to power and down to the turbine blades.

The thought occurred to me: If we believe that there will be a datacenter boom, what companies stand to benefit? What if we compared ChatGPT and Claude on stockpicking performance? I am not a professional public markets investor and this is mostly for fun, but it gives me the opportunity to have skin in the game, learn more about the datacenter space, and play around wrangling multiple tools.

Process and results:

I asked ChatGPT to come up with a big list of companies in the datacenter supply chain. After a lengthy dialogue with ChatGPT, I asked it to summarize what we discussed, including company names and bottlenecks.

Using Codex and starting with that CSV list generated by the chat thread, I enriched the data by grabbing financial metrics and analyst recommendations from Robinhood available at https://robinhood.com/us/en/stocks/[ticker]/

Most stocks have a tally of how many analysts think the stock is a Buy, Hold, or Sell. Using the formula:

1 x %Buy + 3 x %Hold + 5 x %Sell

I calculated an analyst score as a single number (lower is better). I asked Codex to grab the data for each stock, compute the analyst score, and put it into a sortable HTML table. Codex one-shotted it after doing some planning with me.

Note: Because I knew this would be a one-off project, I was ok with taking a snapshot of data on May 1, 2026. I arbitrarily picked the scoring formula to try out, but it would be trivially easy to use a different score such as Target Price.

I had to do some manual cleanup when some values came back as N/A. These stocks are not available on Robinhood so I either removed them or corrected them with the available equivalents.


TSMC	TSM	Foundry	Chip manufacturing and packaging	$397.32	32.92	2.06T	97.9%	2.1%	0%	104
Celestica	CLS	Manufacturing	Electronics manufacturing	$419.96	49.50	48.18B	96%	4%	0%	108
Broadcom	AVGO	Networking	Switch silicon and custom AI chips	$420.61	81.43	1.99T	94.4%	5.6%	0%	111
Micron Technology	MU	Memory	DRAM and HBM	$541.13	24.42	611.64B	92%	8%	0%	116
Comfort Systems USA	FIX	Infrastructure	HVAC contractor	$1,873.15	53.12	65.72B	90%	10%	0%	120
Cheniere Energy	LNG	Fuel	LNG export terminals	$270.06	11.40	56.75B	88.9%	11.1%	0%	122
Hitachi	HTHIY	Grid	Transformers and HVDC (ADR)	$30.57	26.56	138.67B	88.2%	11.8%	0%	124
Energy Transfer	ET	Fuel	Midstream pipelines	$19.96	16.69	68.62B	86.4%	13.6%	0%	127
POSCO Holdings	PKX	Materials	Steel and electrical steel	$78.68	51.90	24.47B	82.6%	17.4%	0%	135
Howmet Aerospace	HWM	Industrial	Turbine castings	$239.56	65.52	95.98B	81.5%	18.5%	0%	137
Chesapeake Energy	EXE	Fuel	Natural gas	$99.26	7.61	23.95B	80.6%	19.4%	0%	139
Applied Materials	AMAT	Equipment	Semiconductor tools	$388.98	40.41	308.75B	80.6%	19.4%	0%	139
Marvell Technology	MRVL	Networking	Networking and custom silicon	$164.61	53.79	144.24B	82.2%	15.6%	2.2%	140
Advanced Micro Devices	AMD	Compute	CPUs and GPUs	$360.60	133.78	587.82B	78.6%	21.4%	0%	143
Tokyo Electron	TOELY	Equipment	Semiconductor tools (ADR)	$153.32	38.61	143.51B	78.3%	21.7%	0%	143
ASML	ASML	Equipment	EUV lithography	$1,429.79	47.94	552.40B	82.2%	13.3%	4.4%	144
Lumentum	LITE	Networking	Optical components	$951.68	275.91	67.80B	76.9%	23.1%	0%	146
DuPont	DD	Materials	Specialty materials	$46.24	-24.48	18.96B	75%	25%	0%	150
Equinix	EQIX	Data Center	Colocation	$1,085.03	74.88	107.06B	78.1%	18.8%	3.1%	150
Seagate Technology	STX	Storage	HDD	$727.89	63.94	163.00B	78.6%	17.9%	3.6%	150
Vertiv	VRT	Grid/Data Center	UPS systems and cooling	$329.02	82.46	126.05B	78.6%	17.9%	3.6%	150
Western Digital	WDC	Storage	HDD and NAND	$428.95	25.94	146.41B	78.6%	17.9%	3.6%	150
GE Vernova	GEV	Power	Gas turbines and power generation equipment	$1,064.50	31.57	285.59B	75%	22.5%	2.5%	155
Lam Research	LRCX	Equipment	Semiconductor etch tools	$255.88	48.67	321.02B	75%	22.2%	2.8%	156
Antero Resources	AR	Fuel	Natural gas	$38.89	12.72	12.04B	72%	28%	0%	156
EQT Corporation	EQT	Fuel	Natural gas production	$58.67	11.39	36.71B	71%	29%	0%	158
Freeport-McMoRan	FCX	Materials	Copper production	$56.62	30.66	81.28B	76.9%	15.4%	7.7%	162
Digital Realty	DLR	Data Center	Data centers	$200.70	53.08	70.03B	70.6%	26.5%	2.9%	165
Eaton	ETN	Grid	Electrical distribution equipment	$425.55	41.45	165.06B	69.7%	27.3%	3%	167
Siemens	SIEGY	Power	Turbines and grid infrastructure (ADR)	$148.74	26.00	232.09B	73.1%	19.2%	7.7%	169
Williams Companies	WMB	Fuel	Pipelines	$75.54	35.75	92.39B	74.1%	14.8%	11.1%	174
Corning	GLW	Networking	Fiber optic cables	$158.02	78.67	135.93B	65%	30%	5%	180
Ciena	CIEN	Networking	Fiber networking	$538.22	335.04	75.66B	63.6%	31.8%	4.5%	182
Carrier Global	CARR	Cooling	HVAC systems	$67.41	44.13	56.50B	58.6%	41.4%	0%	183
Chevron	CVX	Fuel	Oil and gas	$190.55	29.17	380.36B	65.5%	27.6%	6.9%	183
Kinder Morgan	KMI	Fuel	Natural gas pipelines	$32.43	22.17	72.38B	50%	46.2%	3.8%	208
ArcelorMittal	MT	Materials	Steel	$56.85	13.81	44.51B	55%	35%	10%	210
Trane Technologies	TT	Cooling	HVAC systems	$486.48	38.10	107.71B	48.3%	44.8%	6.9%	217
3M	MMM	Materials	Industrial materials	$142.50	28.27	74.33B	45%	50%	5%	220
Thyssenkrupp AG	TKAMY	Materials	Industrial materials (OTC)	$11.89	40.05	7.40B	50%	40%	10%	220
Bloom Energy	BE	Power	Fuel cells	$290.52	-5024.11	82.70B	46.7%	43.3%	10%	227
ExxonMobil	XOM	Fuel	Oil and gas	$152.75	23.03	635.16B	43.3%	50%	6.7%	227
Shell	SHEL	Materials	Energy	$88.98	14.94	248.66B	40%	56.7%	3.3%	227
Intel	INTC	Compute	CPUs and foundry	$99.50	-150.85	500.94B	31.4%	58.8%	9.8%	257
Range Resources	RRC	Fuel	Natural gas	$42.63	11.49	10.04B	25%	64.3%	10.7%	271
Cleveland-Cliffs	CLF	Materials	Electrical steel (GOES)	$10.47	-4.36	6.00B	18.8%	68.8%	12.5%	288
Huntsman	HUN	Materials	Chemicals	$14.63	-9.03	2.54B	17.6%	58.8%	23.5%	312
Southern Copper	SCCO	Materials	Copper	$171.18	28.79	141.41B	10%	40%	50%	380
Glencore	GLNCY	Materials	Mining (ADR)	$15.22	252.29	89.25B	N/A	N/A	N/A	N/A
Mitsubishi Heavy Industries	MHVIY	Power	Large gas turbines (ADR)	$29.32	52.24	98.92B	N/A	N/A	N/A	N/A
MTU Aero Engines	MTUAY	Industrial	Aircraft engines (ADR)	$170.32	16.03	18.36B	N/A	N/A	N/A	N/A
Park-Ohio Holdings	PKOH	Industrial	Industrial manufacturing	$28.54	16.31	411.94M	N/A	N/A	N/A	N/A

Now given this large list of potential stocks, I asked the models to construct a portfolio. I asked both ChatGPT and Claude: What 10-12 stocks would it pick, how would it weight them, and why? I then asked ChatGPT to create HTML for these 3 tables with columns [Name, Weight]. For the Analyst's Recommendation table, take the top 12 and create a weighting based on relative scores. ChatGPT easily figured out how to do this.

Portfolio Comparison

Given the option of picking 10-12 names, Claude picked 11. I specified 12 for the third table, to match ChatGPT's recommendation

ChatGPT

Company Name	Weight
Broadcom	15%
Lam Research	12%
Applied Materials	11%
Eaton	10%
Vertiv	10%
Micron Technology	9%
GE Vernova	8%
Marvell Technology	7%
Celestica	6%
Trane Technologies	5%
Freeport-McMoRan	4%
Corning	3%

Claude

Company Name	Weight
GE Vernova	12%
Freeport-McMoRan	11%
ASML	11%
Equinix	11%
Eaton	10%
Vertiv	9%
TSMC	9%
Broadcom	8%
Prysmian	6%
Marvell Technology	7%
Celestica	6%

Analyst Rating Score

Company Name	Weight
TSMC	9.81%
Celestica	9.44%
Broadcom	9.19%
Micron Technology	8.79%
Comfort Systems USA	8.50%
Cheniere Energy	8.36%
Hitachi	8.22%
Energy Transfer	8.03%
POSCO Holdings	7.56%
Howmet Aerospace	7.45%
Chesapeake Energy	7.34%
Applied Materials	7.34%

Subjectively, I preferred Claude's picks based on its explanation, but it's tbd which portfolio does best in practice.

Robinhood conveniently allows you to buy fractional shares as well as set up different accounts. I have set up a Claude Portfolio account and a Chat Portfolio account. Let's see how they perform over the next months/years.

Learnings on Process:

Using Codex for this project made scraping and formatting data very easy. Being able to switch between data formats seamlessly was incredible. I could easily provide text, CSV, screenshots, HTML, structured data in different formats, and the models would handle it.
ChatGPT balked when asked to do large analyses. Asking it to enrich the data the way I did with Codex would not have yielded good results
Building in this ad hoc way creates an overall workflow that is easy but not repeatable. It required jumping between tools, having a generator script, then editing the output of the generator, etc. Working like this leaves artifacts all over the place: multiple chat threads, locally saved CSVs, a small generator script, HTML outputs. For one-off analyses this is not really a problem, but it does diverge from how one might think about structuring a project in a pre-AI era. The tradeoffs of creating a repeatable system weren't worth it in this case since it was part research curiosity, part writing exercise, and part code. Only in this era does disposable tooling become economically rational.
HTML is a great output format for these analyses. It offers maximal flexibility since there is already a robust system for formatting and interaction that models can reliably manipulate.
The twists and turns are a part of the process. Had I known exactly what I wanted at the beginning, this would have been a very quick task. Instead, I added ideas iteratively as I learned and built. I view this as a creative benefit.
The downside is, because it's so easy to take the next step, you may end up doing things suboptimally, or going down a rabbit hole. For example, instead of scraping Robinhood, which is brittle, I could have used the yFinance Python package. I did not even think about this because I had it in my mind that I was going to open these position in Robinhood. Similarly, starting with a ChatGPT-generated list of stock names and then picking 12 from there is very likely suboptimal because it assumes that ChatGPT picked the right pool of names to begin with. It would be easy to try another approach like analyzing Dwarkesh / SemiAnalysis podcast content, for example.

What AI unlocks:

The set of tools available today makes doing arbitrary analysis easy and fun. For example, it would be trivial to rate stocks based on "coolest company logo" or CEO's astrological sign (pick a different example if you don't like my cute ones). Exploring these types of questions used to take too much time to justify.

A more serious example might be: allowing a middle-school teacher to see how students do in class based on where they sit. Before, that analysis would require technical skill and more time than a teacher would care to allocate. LLMs allow the teacher to become an amateur sociological researcher by collapsing the costs.

Building in this exploratory fashion leads to open-ended creativity and I'm excited to do more of it. Not knowing precisely where you're going becomes more of a benefit than detriment as the capabilities available shift the work further from planning to action. Historically, ambiguity was expensive, but today the cost of trying new ideas approaches zero. I expect that going forward, much of research, learning, and product development will take the form of broad inquiry driven by curiosity and play.

Update 6/10/26: Claude Fable 5 was released this week and people are already showing incredible examples of one-shotted analysis presentations, like a McKinsey style report on a brand new topic, or an interactive webpage showing travel times to all places on a map. This will further unlock the free exploration of topics, although I remain convinced that doing the process partly by hand results in a better creative process by forcing you to think, make new connections, and drive the direction of the research.