This blog post discusses a trade which will lose you twenty thousand dollars in the median case. Do not do this if you can’t afford to lose twenty thousand dollars. Do not do this if you’re not very comfortable with options. Do not do this out of a feeling that you have no other way of making ‘enough’ money – if you truly find yourself stuck in a ‘permanent underclass’, whatever that means, twenty thousand dollars invested in a booming stock market will be very useful.

I drafted this blog post back in August/September. Stock market returns have been lukewarm since then. I think this trade still makes sense, but I no longer recommend the 14000 strike.

I used to think that options are pretty straightforward, because the payoff function, max{0,S−K}\text{max} \left\{ 0, S - K\right\}, isn’t particularly gnarly. Then I spent two years working in options and was shocked at just how much I didn’t know.

When I heard about this trade idea, I felt like I had the knowledge required to make sense of it and was again surprised at how much I didn’t know when I started modelling it.

If you’ve not spent some years actually placing equity options trades, at least semi-professionally, but you still want to get value from this blog post, I recommend reading the thesis, creating your own model, and then comparing that to my write-up.

Trade

The thesis was presented to me at LessOnline 2026 by Gene Smith (who doesn’t endorse this post, all mistakes are mine). It’s essentially this:

  • RSI is real
  • RSI will massively boost economic growth
  • This will be captured by the SPX index
  • Far OTM SPX calls are not pricing this in
  • Therefore, investing in far out of the money SPX calls is a way to trade RSI

He specifically suggested the Dec’31 20,000 call (at the time of writing, SPX spot sits at around 7,600). I decided this was interesting enough to build an actual model.

Model

Last chance to stop reading and build your own model before you get drawn into my pit of hopeless optimism and financial illiteracy.

I’ll walk you through my squiggle.

The relevant average SPX yearly return was around 11%, vol was somewhere between 10 and 15%1.

mu = log(1.11)
sigma = log(1.1 to 1.15)

Say that with RSI, next year we get 14 months of work done. Actually, I think that frontier models will get done much more than that, but almost everything will get bottlenecked, so we’ll only see 14/12 growth in the first year.

acceleration = 14/12

Now the SPX doesn’t follow an actual lognormal distribution, it follows something that has fatter downwards tails. I’m going to model that by sayin that instead of following the lognormal, the market will crash by 20-40% in about 10% of years2.

pCrash = 0.1
ret(mu, sigma) = mixture([lognormal(mu, sigma), 0.6 to 0.8], [1 - pCrash, pCrash])

Now for the most contentious part. We’re already assuming recursive self-improvement, but I am claiming that the benefits to the economy will accrue recursively as well. I think this is similar to saying that the marginal returns to increasing model strength won’t diminish quickly.

y1 = ret(mu, sigma)
y2 = ret(mu * acceleration, sigma)
y3 = ret(mu * acceleration^2, sigma)
y4 = ret(mu * acceleration^3, sigma)
y5 = ret(mu * acceleration^4, sigma)
y6 = ret(mu * acceleration^5, sigma) ^ 0.45  // Dec expiry means it's about 45% of a year on top of 5 years
cumulativeGrowth = y1 * y2 * y3 * y4 * y5 * y6

The rest is plumbing for the discounted expected option return, assuming I get to keep 76% after tax. I’m assuming a constant rfr of 3% (did I mention I suck at macro?).

ttm = 5.45  // in years
rfr = 1.03
expiry(spot, strike, callPrice) = max(0, spot * cumulativeGrowth - strike) * 100 / rfr ^ ttm - callPrice
netRoi(spot, strike, callPrice) = (
  1 + min(expiry(spot, strike, callPrice), expiry(spot, strike, callPrice) * 0.76) / callPrice
)

Then I ran the real return for a couple of strikes. The 14000 strike had the nicest payoff profile.

I did a lot of tuning the parameters (including putting in a probability that RSE doesn’t happen) and was not able to construct a plausible scenario where this isn’t massively +EV, even ignoring the fact that it’s a good hedge against getting automated soon.

Sizing this

Gene, if I correctly understood his presentation, is yeeting his entire savings in there. Respectfully, this seems insane. I’m guessing he doesn’t care about having money other than in the case of being obscenely wealthy?

For me, this trade is a hedge against my earning power disappearing if RSI happens. Turns out that in those worlds, <5 contracts, once I realize the gain, will pay enough dividend for me to live on comfortably in perpetuity. So that’s how much I’m paying and I’m keeping the rest of my assets broadly invested in the stock market.

What Claude thinks

I also asked Claude. It came up with a different model, but similar results. Pasting its output verbatim. Prompt: Please build a model for the returns of an SPX option, taking the view that RSI will start rapidly improving AI capabilities, starting now. The option I'm interested in is the Dec'31 14000 call. The call price is 210 (the contract multiplier is $100, so this is an effective price of $21000 per contract). The current SPX spot it at 7641. I'm interested in the distribution of percentage ROI, shown as a table of the deciles. Assume I'll have to pay 24% CGT on any proceeds and incorporate that into your model.

Opus 5 High: [a whole bunch of nonesense]

Prompt: Okay, this is a good start [always be nice to your models, maybe the models will be nice to you]. I like the scenario breakdown, but, and maybe you're hiding some of your reasoning, it looks like you just took a guess at the SPX value based on the scenario. Let's go more into detail, and come up with a mechanistic model, and how it would act on the value of SPX over these ~5.5 years to get to values. It also looks to me like your numbers are anchored at the breakeven, whereas I think we should anchor at the current spot and work from there. You might get very different numbers when you do it this way, or you might not, either is fine.

Opus 5 High: [Essentially, "you're absolutely right", followed by the below]

The anchor

Everything now starts from spot. Decompose SPX 7641 = EPS × P/E:

  • S&P 500 revenue/share ≈ $2,000, net margin 14.2% → EPS ≈ $284, trailing P/E ≈ 26.9x
  • Discount rate: 2.0% real risk-free + 2.4% inflation + 4.5% ERP → 8.95% nominal

Feed those into a two-stage DDM (10 years of above-trend growth, 85% payout, then Gordon fade) and solve for the growth rate that reproduces 26.9x. Answer: the market is currently pricing 7.2% nominal / 4.7% real EPS growth for a decade. That is already an AI-optimistic number — well above the ~2% real trend. You are not buying a call on "AI works"; the index has partly priced that. You're buying a call on AI working much harder than 4.7% real for ten years.

Verify my $284 EPS input against actuals — it's the load-bearing number and everything scales off it.

The mechanism

SPX(2031) = RevPS × Margin ÷ Share count × P/E, with four channels the RSI thesis acts through:

  1. Revenue. Per-share revenue grows at nominal GDP plus global/share-count effects. AI raises real output growth.
  2. Margin. The big one. Labor share falls, and index net margin rises from 14.2% toward 19–27%. This is a level shift, not a permanent growth rate — I treat it as such, which is the main discipline the earlier version lacked.
  3. Share count. Compute capex in a takeoff is partly equity-funded. Net issuance flips from −0.6%/yr to +1%/yr, which is a ~9% per-share drag over the horizon.
  4. The multiple. rf_real = −1% + β × (economy-wide real growth), β ~ N(0.6, 0.25). The damped β is deliberate: the textbook Ramsey coefficient (γ ≈ 1.5) would put real rates at 17% under 12% growth, which no historical high-growth episode has ever produced — Japan in the 1960s and China in the 2000s ran 3–5% real rates on 9%+ growth. Financial frictions and precautionary saving break the Euler equation. Post-2031 trend growth is tied to economy growth, and terminal growth is capped at the risk-free rate so the Gordon term can't blow up.

Regimes, now defined by drivers rather than by outcome

Regime p Real rev g Margin 2031 Econ real g → EPS → P/E → SPX
A Diffusion lag 25% 3.0% 15.5% 2.2% 429 23.5x 10,100
B Broad boom 30% 6.0% 19.0% 4.0% 620 27.0x 16,765
C Takeoff, index captures 15% 16.0% 27.0% 13.0% 1,239 19.8x 24,763
D Takeoff, rents escape index 10% 4.5% 13.0% 13.0% 333 18.7x 6,230
E Disruption / bust 12% −2.0% 10.0% 0.5% 205 12.9x 2,682
F Catastrophe 8% — — — — — 0

Regime D is the structurally nasty one and it falls straight out of the mechanism: the economy compounds at 13% real, so the risk-free rate goes to 6.2% and the discount rate to 14.1%, but incumbent margins compress because the rents accrue to private labs, to consumers, or to firms that only join the index after they've already re-rated. Full rate shock, no earnings benefit.

Results

Decile SPX Dec-2031 Gross ROI After 24% CGT
10% 1,510 −100% −100%
20% 4,130 −100% −100%
30% 7,440 −100% −100%
40% 9,360 −100% −100%
50% 11,160 −100% −100%
60% 13,270 −100% −100%
70% 16,020 +805% +612%
80% 20,030 +2,681% +2,038%
90% 27,300 +6,077% +4,619%
95% 35,720 +10,032% +7,625%
99% 65,980 +24,385% +18,533%

P(ITM) = 37.0% (was 48.5%). E[net ROI] = +1,400% (was +4,900%). Median SPX 11,157, i.e. +46% over 5.3 years — below breakeven, which now sits between the 6th and 7th decile rather than between the 5th and 6th.

Remaining weaknesses

Drivers are sampled independently within regimes apart from the rate–growth link. The real correlations run the wrong way for you: high growth attracts competition (margin down), high uncertainty widens the ERP (P/E down), and heavy capex cuts payout (P/E down). Imposing those would shave the upside deciles further. Index-composition turnover is folded into the revenue draw rather than modeled as explicit entry/exit, which understates the "the winner IPOs at 40x and joins the index after the move" problem — though note that Nvidia, Microsoft and Alphabet are already ~35% of the index, so incumbent capture is genuinely plausible, which is what regime C encodes.

Execution

So I did buy this, with Interactive Brokers, with whom I have a margin account.

The bid ask was very wide (I don’t remember exactly, but something like 88/352, for approximately a 220 mid). I put in a limit buy order at mid, then waited a few minutes before raising it by about 0.5 repeatedly until I got filled.

Doing this felt weirdly embarrassing. Because of the low contract size, I knew that it would be very obvious to other market participants that I was amateur retail so I felt quite exposed. The market structure work in my favour though: Since the market is competitive, market makers are still incentivised to give me their best price, even though one-on-one they could easily shake me down. In fact, I might have gotten a better price because the markets are less likely to move against you after your trade if your counterparty is retail.

I was filled at 228.52 (so $12,348 cheaper per contract than a market order!).

Option mids are often far away from any arbitrage-free price curve, so really I should have started below mid, but it worked out.

Interactive brokers performed this trade by putting a short $ position into my account, which I had to close out with £ manually. Closing it out was done via market order, which I received a warning popup and 2 post-factum warning emails about.

Also, because Interactive Brokers uses 15 minutes delayed data and I think they are pricing my position MTM against recent mid, I made an immediate notional profit of $6k: The mid was calculated based off my own bid, which was still showing in the delayed data!

Uncertainties

There is some macro stuff about rates that I just don’t understand.

I am also quite uncertain about how much value will be captured.

The claim that returns in the nn-th year are around μ⋅accelerationn\mu \cdot \text{acceleration}^n feels utterly insane, so even though this is what my analysis implies, that’s where I would expect a lot of weakness

I also worry that my broker, Interactive Brokers, will make me lose all my money even if I get everything right because there is some “don’t lose all my money” checkbox in one of the settings menus that I haven’t found.

Perhaps anything that’s Mythos-tier will not be available commercially, because it’s too dangerous, so even though capabilities will keep improving, the economic impact will be limited.

I executed the trade on August 28th. Now that conversations around an AI pause are picking up, the EV in my model is much lower, though still very positive. It’s hard to be upset when your trade fails due to people attempting not to destroy the world.


  1. Unlike me, the careful reader might notice an error in the squiggle, where I assume a volatility of about 110% instead of 11%. Thankfully, I did enough ~wiggling~ sensitivity analysis that the ROI was still very good after fixing this error. ↩

  2. Yes, I would absolutely rather be stuck in the permanent underclass than try to calibrate a Heston model. ↩