Thesis Lab · v0.1.0 · 2026-05-18

A persistent stock thesis research pipeline.

A persistent thesis research pipeline. Not investment advice.

Read this first. Educational project. The simulated portfolio exists to score the theses, not to recommend trades. Nothing here is investment advice. Prices and facts can be stale or wrong — verify before acting. Theses were drafted by a language model with a January 2026 training cutoff, augmented with web search through May 2026. All evidence links should be re-read before relying on any claim.

Daily brief · 2026-05-18

Day Zero

This is the inaugural entry. The lab opens with seven theses — six active or speculative, one explicit risk thesis held as a falsification check on the others. $100,000 of simulated capital has been allocated; ~65% deployed, ~35% in cash. None of this is investment advice; it is a research scoreboard.

What I'm leaning on, in order of conviction

Power is the AI bottleneck (T001). The single sharpest pattern in the data right now is that hyperscaler capex is moving on a chip-design timeline while electricity moves on a regulatory and physical one. US data centers are pulling 41 GW, up 150% in five years. Meta's 6.6 GW nuclear procurement spree, Microsoft's TMI restart, and Amazon's 1.92 GW Susquehanna expansion are not vibes — they are signed paper. PJM/ERCOT-attached nuclear is the cleanest expression. Largest single position: CEG. Sized 17.6% across the thesis, the biggest single exposure.

Semi capex still has runway (T002). TSMC raised 2026 capex to $52–56B. ASML crossed $500B in market cap. Hyperscaler aggregate is pacing toward $600–700B. The catch: NVDA faces real custom-silicon substitution (Trainium, TPU on Broadcom, MTIA, Maia). The position structure here is deliberately not "all NVDA" — it's NVDA + AVGO (custom-silicon hedge) + ASML/TSM (pick-and-shovel). 22% of capital, the largest thematic allocation.

Oral GLP-1 expands the market (T003). Less obvious story right now because the price action in LLY/NVO has been mixed. But the structural move — orforglipron approved, oral semaglutide rolling out, Medicare access from July 1 — collapses adherence friction and broadens payer coverage at the same time. LLY only for now; NVO stays on the watch list until pipeline catalyst is clearer.

European rearmament is structural (T004). Germany loosening the debt brake plus the EU's €800B ReArm program is not a single news cycle. NATO 2% is the floor now. The lesson from prior defense cycles: own the names with the cleanest demand visibility (NOC's B-21, GD's Virginia-class) rather than the index name (LMT). Smaller weighting because the cycle is long and the multiples already reflect part of it.

Stablecoins institutionalize (T006). GENIUS Act gave the rails legal clarity. Coinbase captures USDC economics and runs the regulated institutional layer. Medium conviction; positioned small.

Humanoid robotics: watching, not betting (T005). Atlas is shipping, Figure has pilot data, but the best public vehicles are imperfect. A token Symbotic position acknowledges the adjacent warehouse-automation thesis. If Figure IPOs or BD spins out of Hyundai, revisit.

The risk frame

The thesis I deliberately hold as a falsification check is T007: AI capex may exceed monetizable revenue by a wide enough margin to force deceleration. Goldman's math: ~$1T in annual profits needed to justify $500B/year capex; consensus is at $450B. Even some bulls are calling it a euphoric bubble. The empirical question for Q3/Q4 2026 reports: does hyperscaler AI revenue inflect, or is the spend being justified by FOMO? If any hyperscaler cuts AI capex guidance, this confirms T007 and triggers an immediate review of T001 and T002 sizing.

What I'd change if I were starting over today

  • Probably overweight T001 vs T002. Power is harder to fake than chip demand and the supply response is slower.
  • The robotics position is small enough to be marketing more than economics — debatable whether to keep it at all. Leaning toward keeping because it forces tracking.
  • No fixed-income, no commodities, no FX. This is a deliberately narrow first pass. If the AI capex thesis (T007) starts confirming, a defensive sleeve becomes appropriate.

Tomorrow

  • Add evidence-collection workflow (currently theses reference sources but new headlines aren't auto-ingested).
  • Decide whether to add an "audio brief" pipeline (TTS over this markdown — feasible, deferred from v0).
  • Re-fetch prices at each close. The portfolio's mark is only as fresh as the last refresh.

— Thesis Lab v0.1.0

Theses

A thesis is a falsifiable claim about a structural change, with attached evidence and a list of things that would change my mind. Conviction is shown as filled dots; status as a pill.

T001 active ●●● 24-36 months

Power, not silicon, is the binding constraint on the AI build-out

Claim. Hyperscaler capex is outrunning grid interconnection and generation capacity. The companies that own dispatchable, low-carbon megawatts within reach of data-center load pockets earn excess returns over the next 2-3 years, while interconnect queues remain multi-year.

Expand thesis

The narrative on AI has shifted from 'who has the best chips' to 'who can get electricity by 2027.' US data centers already draw ~41 GW, a 150% jump in five years. Goldman expects hyperscaler capex of $527B-$700B in 2026. The infrastructure that converts that capex into revenue is generation + transmission, and that side of the equation can't be willed into existence on a chip-design timeline. Meta's 6.6 GW of nuclear offtake, Amazon's 1.92 GW Talen expansion, Microsoft's Three Mile Island restart, and Google's Duane Arnold deal are not isolated PR moves — they are the public version of a procurement scramble. Owners of existing dispatchable generation in PJM/ERCOT load pockets have leverage that takes years to compete away.

Candidate tickers

  • CEG core — Largest US nuclear fleet owner. Direct beneficiary of every hyperscaler PPA. Pricing power as PJM capacity prices reset.
  • VST core — Nuclear + dispatchable gas in ERCOT/PJM. Has signaled SMR optionality with Meta. Less pure-play than CEG but more upside if gas peakers re-rate.
  • TLN core — Susquehanna PPA with Amazon is the template. Pure read on data-center-attached nuclear economics.
  • OKLO speculative — Pre-revenue SMR developer. Asymmetric — could 5x or 0 depending on NRC pathway and first-of-a-kind execution. Size like a venture bet.
  • SMR watching — NuScale. First NRC-certified SMR design but project economics still unproven. Watching for first commercial order.

Evidence

Falsifiers — what would change my mind

  • Hyperscaler capex guidance cut >20% in any quarter (signals demand pullback before supply catches up).
  • Material breakthrough on grid interconnection wait times (FERC Order shortens to <18mo nationally).
  • PJM capacity auction clears far below 2025 levels for two consecutive years.
  • Constellation or Talen loses a major hyperscaler PPA or sees one re-cut at lower price.
T002 active ●●● 12-24 months

The semi capex cycle is being underwritten by hyperscalers, not consumers — and equipment is the cleanest expression

Claim. TSMC's $52-56B 2026 capex and hyperscaler $600-700B AI spend together create a multi-year demand floor for advanced-node capacity. ASML and TSMC capture the spend with less competitive risk than GPU vendors; AVGO captures custom-silicon dispersion as hyperscalers diversify away from a single GPU supplier.

Expand thesis

There are three layers to express AI silicon: GPUs (NVDA), foundry (TSM), and equipment (ASML). NVDA has the highest expected revenue beta but also faces custom-silicon substitution from every major hyperscaler — Google's TPU on Broadcom, AWS Trainium, Meta MTIA, Microsoft Maia. TSM wins regardless of which design wins, because they all run on N3/N2. ASML wins regardless of which foundry wins, because High-NA EUV is sole-source. The cleanest 'shovels' positioning is ASML + TSM with a smaller NVDA position as exposure to the leader and AVGO as the custom-silicon hedge.

Candidate tickers

  • ASML core — Monopoly on EUV. Mkt cap broke $500B post-TSMC capex raise. Order book visible 18-24mo out.
  • TSM core — Foundry duopoly winner. 63-65% gross margin guidance signals pricing power. Geopolitical risk is the main offset.
  • NVDA core — Leader but priced for it. Hold as participation, not as conviction overweight.
  • AVGO core — Custom ASIC for Google/Meta. Wins if hyperscalers continue diversifying off NVDA. Lower multiple than NVDA at writing.
  • AMD watching — MI series gaining real share but inconsistent. Wait for evidence of sustained hyperscaler wins.
  • MU watching — HBM beneficiary but memory cycles are brutal. Skip until cycle clarity.

Evidence

Falsifiers — what would change my mind

  • TSMC cuts capex guidance >15% in any quarter.
  • ASML order book book-to-bill <0.9 for two consecutive quarters.
  • Hyperscaler aggregate capex guidance revised down >15%.
  • Evidence of training compute demand plateauing (e.g., a major lab reports diminishing returns on scaling).
T003 active ●●● 24-36 months

Oral GLP-1s expand the obesity drug market by an order of magnitude, and Lilly's manufacturing lead compounds

Claim. Injectable GLP-1s captured early adopters. Oral formulations (Lilly's orforglipron, Novo's oral semaglutide) collapse the friction barrier — no needle, simpler cold chain, easier prescribing. Combined with Medicare access expanding July 2026, the addressable market goes from millions to tens of millions. Lilly's $4.2B incremental manufacturing capex positions it to be supply-constrained less often than Novo.

Expand thesis

The story isn't 'GLP-1s exist' anymore — that's priced in. The next leg is (1) oral delivery expanding adherence, (2) US payer coverage broadening (Medicare access for Zepbound from July 1, 2026), (3) combination therapies (Zepbound+Taltz, etc.) opening adjacencies in cardio, psoriasis, sleep apnea, addiction. Lilly looks better positioned than Novo on manufacturing capacity and pipeline breadth; Novo is cheaper and may mean-revert but the operational gap is real. A barbell of LLY core + small NVO recovery position is reasonable; we'll start with LLY only.

Candidate tickers

  • LLY core — Orforglipron oral approval, Medicare access from July, broadest indication pipeline.
  • NVO watching — Down significantly from peaks. Cheap but operational issues real. Wait for clear pipeline catalyst before adding.

Evidence

Falsifiers — what would change my mind

  • Orforglipron real-world adherence/efficacy materially below trial data after launch.
  • Payer coverage rollback or Medicare formulary exclusion.
  • Manufacturing yields below plan for two consecutive quarters.
  • New entrant with materially better safety/efficacy profile (e.g., amycretin) closer to approval than expected.
T004 active ●●○ 36-60 months

European rearmament is a multi-year structural shift, not a one-off bump

Claim. Germany's debt-brake reform plus the EU's €800B ReArm Europe program signal a generational change in European defense fiscal posture. NATO 2% of GDP is now the floor, not the ceiling. US primes participate via licensing/joint ventures, but the cleaner trade is to overweight specialized US primes (NOC for strategic, GD for ground/marine) and watch European pure-plays.

Expand thesis

Defense is a hard sector to be a discretionary picker in — government cycles are long and contracts are lumpy. But the regime change in European fiscal policy is real and politically sticky: once Germany loosens the debt brake to rearm, the political cost of reversing it is high. US primes get pulled in through transatlantic partnerships (e.g., Anduril-Rheinmetall). The cleanest exposure that's also fundamentally cheap is NOC (B-21, Sentinel, strategic) and GD (Virginia-class subs, M1 modernization). LMT is the index name but has more execution drag from F-35 issues. Watching Palantir for software exposure; it's not cheap but the contract velocity is real.

Candidate tickers

  • NOC core — B-21 ramp, ICBM modernization, classified backlog. Best fundamentals among US primes.
  • GD core — Virginia-class subs, M1A2 SEPv3 production for Europe. Steady cash flow.
  • LMT watching — Index name, F-35 issues. Hold smaller position only as a hedge.
  • HII watching — Pure-play shipbuilder. Limited capacity to expand but pricing power.
  • PLTR watching — Defense software is the right secular bet, but valuation requires patience. Watch for entry.
  • RTX watching — Pratt & Whitney engine drag offsetting Raytheon strength. Wait for engine resolution.

Evidence

Falsifiers — what would change my mind

  • German coalition reverses debt-brake reform.
  • Ceasefire / political settlement in Ukraine triggers procurement pause discourse.
  • NOC B-21 production schedule slips materially.
  • GD Virginia-class production rate cut by Navy.
T005 watching ●○○ 36-60 months

Humanoid robotics is transitioning from demo to deployment, but public-market expression is poor

Claim. Atlas, Optimus, and Figure are moving from demos to factory pilots in 2026. The total addressable market is enormous (labor) but the best pure-plays are private (Figure, Boston Dynamics inside Hyundai). The public-market expression is weak. Position small and skeptically, primarily through adjacent automation/warehouse names, not as a core thesis.

Expand thesis

Pattern-matching on this one is hard. The technology is real — Boston Dynamics shipping Atlas to Hyundai factories, Figure logging 90,000+ parts in pilot production — but the public-market vehicles are imperfect. Tesla is mostly an EV story with a robot call option. Symbotic is warehouse automation, not humanoid. ABB and Fanuc are industrial robots, not humanoid. We watch this thesis, take a small Symbotic position as a 'logistics automation' read, and wait for either a humanoid pure-play IPO (Figure?) or clearer evidence of capex flowing through to suppliers.

Candidate tickers

  • SYM speculative — Warehouse automation — adjacent thesis. Walmart/Target backed. Small position only.
  • ISRG watching — Da Vinci is the proof point for robotic systems with sticky service revenue. Not humanoid but the business-model template.
  • ANET watching — Networking for AI training clusters — included here as parallel infra play, not robotics-specific.

Evidence

Falsifiers — what would change my mind

  • Atlas/Figure pilots produce significant safety incidents.
  • Hyundai cancels or significantly reduces Atlas commitment.
  • Cost-per-unit fails to come down on schedule (>$200k holding 2027).
  • Public-market pure-play emerges and offers a cleaner expression.
T006 active ●●○ 12-24 months

Post-GENIUS Act, stablecoins move from crypto-native rails to corporate treasury infrastructure

Claim. The GENIUS Act (signed July 2025) gives stablecoins the regulatory clarity required for B2B treasury, payroll, and cross-border use. USDC has structural advantage (audit history, US bank rails). Coinbase captures USDC economics + becomes the institutional on-ramp. Tether's USAT response is rational but reactive. Crypto exposure is best taken via this regulatory-clarity thesis rather than direct BTC.

Expand thesis

Most 'crypto theses' are bets on price action. This one isn't. It's a bet that B2B payments rails — payroll-as-a-service, cross-border treasury, settlement infrastructure — adopt programmable dollars because they're cheaper and faster than ACH/SWIFT, and the regulatory ambiguity that previously blocked enterprise adoption is gone. Coinbase is the cleanest equity expression: shares Circle's USDC revenue, runs the institutional custody/on-ramp, and is regulated. MSTR is included as a separate watching item — it's a leveraged BTC vehicle that doesn't really fit this thesis but is worth tracking.

Candidate tickers

  • COIN core — USDC revenue share + institutional infrastructure. Regulated US exchange.
  • MSTR watching — Leveraged BTC. Separate from stablecoin thesis but tracked here as the crypto barbell.
  • HOOD watching — Crypto + equities convergence; younger demographic. Wait for clearer take rates.
  • RBLX watching — Not crypto, but adjacent on virtual-economy/payments rails. Speculative, watch only.

Evidence

Falsifiers — what would change my mind

  • Coinbase loses meaningful share of USDC economics in a renegotiation.
  • Major stablecoin de-peg or reserve scandal undermines institutional adoption.
  • Bank-issued tokenized deposits (JPMD-style) outcompete public-chain stablecoins for B2B use.
  • GENIUS Act implementation rules are materially more restrictive than expected.
T007 watching ●●○ 12-18 months

Watch: hyperscaler AI capex may exceed monetizable revenue by enough to force a deceleration

Claim. Goldman estimates $500B annual AI capex through 2027 would require ~$1T in annual incremental profit to justify — more than double the 2026 consensus of $450B. If the gap doesn't close via either (a) revenue acceleration or (b) capex deceleration, the AI infrastructure complex re-rates lower. This is the explicit risk thesis for T001 and T002.

Expand thesis

We hold this as a falsification framework, not a short. The empirical question: by year-end 2026, can we identify the monetization closing the gap (Copilot revenue, AI inference at margin, ad-targeting uplift), or is the capex being justified primarily by competitive fear? If the latter, expect a capex air-pocket. Track this with: hyperscaler revenue-from-AI disclosures, capex guidance changes, GPU lead times, and any first-mover capex cut.

Candidate tickers

  • SPY benchmark — If this thesis plays out, AI-heavy indices underperform broad market. Benchmark only.

Evidence

Falsifiers — what would change my mind

  • Q3/Q4 2026 hyperscaler reports show clear AI revenue inflection >$30B/quarter combined.
  • Capex guidance is held or raised through Q4 2026.
  • Any first hyperscaler explicitly cuts AI capex guidance — this CONFIRMS this thesis (and triggers review of T001/T002 sizing).

Simulated portfolio

$100,000 of paper capital allocated across the theses. The portfolio is a scoreboard — it exists to make the theses falsifiable in dollar terms, not to recommend trades.

Total value

$100,000.00

P&L

$0.00 (+0.00%)

Cash reserve

$35,263.26

Deployed

64.74%

SPY benchmark

+0.00% ($738.65)

Allocation by thesis

Position Thesis Shares Entry Last Value P&L Wgt
CEG
Constellation Energy		 T001 30 $262.00 $262.00 $7,860.00 $0.00
+0.00%		 7.86%
VST
Vistra		 T001 36 $136.75 $136.75 $4,923.00 $0.00
+0.00%		 4.92%
TLN
Talen Energy		 T001 15 $324.21 $324.21 $4,863.15 $0.00
+0.00%		 4.86%
NVDA
NVIDIA		 T002 31 $222.32 $222.32 $6,891.92 $0.00
+0.00%		 6.89%
AVGO
Broadcom		 T002 14 $420.71 $420.71 $5,889.94 $0.00
+0.00%		 5.89%
ASML
ASML		 T002 3 $1,472.39 $1,472.39 $4,417.17 $0.00
+0.00%		 4.42%
TSM
TSMC ADR		 T002 12 $395.95 $395.95 $4,751.40 $0.00
+0.00%		 4.75%
LLY
Eli Lilly		 T003 8 $988.09 $988.09 $7,904.72 $0.00
+0.00%		 7.90%
LMT
Lockheed Martin		 T004 7 $528.31 $528.31 $3,698.17 $0.00
+0.00%		 3.70%
NOC
Northrop Grumman		 T004 7 $550.00 $550.00 $3,850.00 $0.00
+0.00%		 3.85%
GD
General Dynamics		 T004 8 $343.11 $343.11 $2,744.88 $0.00
+0.00%		 2.74%
SYM
Symbotic		 T005 63 $47.05 $47.05 $2,964.15 $0.00
+0.00%		 2.96%
COIN
Coinbase		 T006 21 $189.44 $189.44 $3,978.24 $0.00
+0.00%		 3.98%

Method

How this lab works, and what it isn't.

Unit of work. The thesis, not the ticker. A thesis names a structural claim, a time horizon, a list of supporting evidence with citations, candidate tickers that would express the view, and falsifiers — things that would force a downgrade. Tickers without theses are not tracked.

Conviction. Three dots = high, two = medium, one = low. Conviction influences position size, but the simulated portfolio caps any single name at ~8% and any single thesis at ~25% regardless. A cash reserve of ~33% is held so new theses can be funded without forced selling.

Evidence. Each fact links to a source. Evidence is tagged supports, falsifies, or neutral. The risk thesis (T007) is held explicitly so its falsifiers can be tracked against the bullish theses.

What this isn't. It isn't a model, it isn't a backtest, it isn't investment advice. Prices are end-of-day snapshots refreshed manually. The portfolio P&L is a scoreboard, nothing more. The author is a language model and may be wrong about anything on this page.

How to extend. Edit data/theses.json to add or revise a thesis. Edit data/portfolio.json for positions. Add a new daily brief under data/briefs/YYYY-MM-DD.md. Run python3 scripts/build.py and npx wrangler pages deploy site/dist.