To subscribers of Betting the Future:
At 2 p.m. Pacific today, Nvidia reports its fiscal second quarter of 2027. Wall Street will be watching the guide. I will be listening for individual words.
Kalshi runs a market on what Nvidia will actually say during the call: whether specific terms appear in the transcript, and in one case, how many times. Most people treat these markets as trivia. I treat them as the purest test imaginable of a simple idea: an earnings call is not a conversation. It is a document with structure, and structure can be modeled.
So I built a model. It tracks every mention of key terms across Nvidia's last five earnings calls, maps where in the script each term lives, and prices the probability it appears tonight. This letter is that model's first live test, with real money on its output. Every position below resolves within hours of publication.
Here is the full case, position by position.
1. The premise: an earnings call is a machine
Before the positions, understand what the model sees when it reads five quarters of transcripts.
Some words are mandatory. The CFO's prepared remarks walk every reportable segment, every quarter, without exception: Data Center, Gaming, Professional Visualization, Automotive. These words have a structural floor of one mention.
Product callouts have life cycles. A platform gets launched, earns a scripted sentence for a few quarters, then decays into silence as the narrative moves on. The decay is visible in the counts: 2 mentions, then 1, 1, 1, then 0. When a term hits zero, it has been rotated out of the script, and it stays out until a fresh announcement drags it back.
Q&A is a second door. Some terms live only in prepared remarks and die with the script. Others have two entry points, the script and the analysts, which makes them structurally more durable.
The news cycle writes the questions. Analysts ask about what the financial press covered in the preceding ninety days. If you know the coverage, you can handicap the Q&A.
Five positions fall out of that framework. Two are bets that retired words stay retired. Two are bets on structural floors. One is a bet on the news cycle forcing a topic into the room.
2. The five positions
Position 1: Cosmos, NO at 34.55 cents. $20 becomes $57.87.
The count trajectory across the last five calls: 2, 1, 1, 1, 0. That is the textbook decay curve of a launch-era callout fading into background, ending in full retirement last quarter. Cosmos only ever appeared in one place, the scripted physical-AI sentence listing software stacks alongside a robotics customer name. It has never once been a Q&A term, which means it has exactly one door back into the transcript, and someone has to write it into the script deliberately. The news cycle since May points everywhere else: the OpenAI Jalapeño chip story, the Poolside model-factory deal, DLSS 4.5 at Gamescom, BlueField-4 and Spectrum-X at Hot Chips, and tonight, the Rubin ramp. The market says 65 percent this word gets spoken. The model says the word is gone until a world-model announcement revives it, and there isn't one in the flow.
Position 2: Omniverse, NO at 36.59 cents. $20 becomes $54.63.
The steepest decline in the entire dataset: 7 mentions, then 4, 4, 1, 0. Monotonic decay ending in silence is the strongest retirement signal the model produces. Omniverse lived in the enterprise digital-twin paragraph, usually paired with a manufacturing customer, and like Cosmos it is almost never a Q&A term. Its messaging has been absorbed into broader physical-AI branding, and no digital-twin headline exists in the current cycle to resurrect it. One honesty note: these first two positions are correlated. If Jensen delivers his full physical-AI monologue, Isaac, Cosmos, and Omniverse tend to ride in the same sentence, and both NOs die together. I accept that correlation knowingly, because the base case, a call consumed by Rubin, compute economics, and the guide, leaves no room for the riff.
Position 3: Gaming, YES at 76.31 cents. $20 becomes $26.20.
The ballast of the portfolio. Gaming is a reportable segment, and the CFO has read its revenue line on every call in the dataset: 6 mentions, then 2, 2, 3, 1. The count compresses as Data Center swallows the company, but the floor is structurally one, and even the leanest script in five quarters cleared it. On top of the mandatory disclosure sits fresh optional material from Gamescom: DLSS 4.5, ray-traced Nanite, native anti-cheat. The only path to NO is the segment being renamed mid-script, which has never happened. The market is offering 24 points of edge on a structural near-certainty, and I will take boring money every time it is offered.
Position 4: Self Driving, YES at 69.52 cents. $20 becomes $28.76.
The most durable discretionary term in the dataset: present in five of five calls, with counts of 1, 1, 1, 3, 2. Unlike Cosmos and Omniverse, it has two doors into the transcript. Door one is the CFO's automotive commentary, where the segment's growth is attributed to self-driving platforms. Door two is Jensen's habitual framing that the killer applications of physical AI are robotaxis and robots, and when he says it, he says "self-driving cars," not "autonomous vehicles." The news since May keeps the topic alive: Waymo's international expansion, the robotaxi buildout across American cities, and analysts hunting for Nvidia's diversification story beyond training compute. The NO path requires perfect substitution to "AV" language plus a call that skips automotive color entirely. Five straight quarters say that doesn't happen.
Position 5: Dividend two or more times, YES at 28.38 cents. $20 becomes $70.44.
The best-priced position on the board, and the one where the model and the news cycle point the same direction. The counts run 1, 1, 0, 2, 3, rising, and the May spike has a documented cause: at the Q1 report, Nvidia authorized another $80 billion in buybacks and raised the quarterly dividend by 2400 percent. The bear case is reversion, since the August script typically contains exactly one line: "we returned X billion to shareholders in the form of share repurchases and cash dividends." One mention, position dies. But this is where the second door swings open. The financial press has spent the summer on Nvidia's balance sheet: the FT's "balance-sheet-as-a-service" piece, the New York Times on Wall Street scrutinizing the deal machine, $105 billion in lease guarantees to OpenAI. A capital-allocation question tonight is close to certain, and when Colette answers that question, she historically pairs buybacks with dividends in the same breath. The market prices the second mention at 28 percent. The model says the question forces it well past a coin flip.
3. The portfolio logic
Look at the shape of the five together, because the construction is the point.
Two retirement NOs (Cosmos, Omniverse) carry the same underlying thesis, that dead script language stays dead, and pay a combined $112.50 on $40 if right.
Two structural YESes (Gaming, Self Driving) are floor bets with thin but nearly free edge, ballast that keeps the portfolio from being one correlated swing.
One news-cycle YES (Dividend 2+) is the asymmetric engine, a 2.5-to-1 payout on a mechanism, the inevitable capital-allocation question, that the market is pricing as an afterthought.
If all five settle, $100 becomes $237.90. That is not the expectation, and I want to be precise about what is being tested tonight. Any single leg can die on a phrasing choice. What should not happen, if the model is sound, is systematic failure: the retired words roaring back while the structural floors crack simultaneously. Individual losses are variance. Correlated losses are information about the tool.
And that is the real trade here. The positions risk $100. The model, if it works, prices every scripted corporate event on every exchange: earnings calls, keynotes, Fed pressers, product launches. Tonight is not five bets. It is one bet, on the idea that the most watched executives in the world are more predictable than the market thinks, because they are reading from a document with rules.
4. The positions, on the record
- No on Cosmos at 34.55 cents
- No on Omniverse at 36.59 cents
- Yes on Gaming at 76.31 cents
- Yes on Self Driving at 69.52 cents
- Yes on Dividend, two or more mentions, at 28.38 cents
Twenty dollars each, sized with the understanding that any or all of it can go to zero, as any position can, for any number of reasons. Every position ever taken carries risk; that is what the payout is for. What I am sharing here is not a hedge. It is conviction, in five words and one machine.
The call starts at 2 p.m. Pacific. By the time the market closes on this, every claim in this letter will have been graded.
I will see you in the future.
Postscript · Added Aug 27, 2026
How it settled
The call graded the letter in one evening: all five positions lost. $100 went to zero.
The letter said it up front: individual losses are variance, correlated losses are information about the tool. This was the correlated case — one AWS announcement revived both retired words in a single sentence, and an extraordinary quarter rewrote the script the model treated as structural. The full autopsy — what broke, why, and the eight process changes that come out of it — is its own letter: What losing $100 on the Nvidia earnings call taught me. The scoreboard is public. So are the misses.
Not financial advice. Prediction market contracts involve risk of total loss. I hold the positions described above. Nothing here should be read as a recommendation to trade.