{"id":676,"date":"2026-08-25T07:47:00","date_gmt":"2026-08-25T14:47:00","guid":{"rendered":"https:\/\/quantcha.com\/news\/?p=676"},"modified":"2026-08-24T20:08:42","modified_gmt":"2026-08-25T03:08:42","slug":"the-ai-compute-trading-tools-im-hoping-youve-always-wanted","status":"publish","type":"post","link":"https:\/\/quantcha.com\/news\/the-ai-compute-trading-tools-im-hoping-youve-always-wanted\/","title":{"rendered":"The AI Compute Trading Tools I\u2019m Hoping You\u2019ve Always Wanted"},"content":{"rendered":"\n

Or you can keep trading GPU contracts like a Neanderthal.<\/em><\/p>\n\n\n\n

A few weeks ago, I walked through the process of using prediction markets to price a stack of AI compute derivatives<\/a>. I had just priced a full derivatives stack\u2014a future, a call, a put, and a perpetual\u2014off nothing but the Kalshi H200 binary ladder. There was no Black-Scholes, no volatility input, no anything beyond a bunch of binary probabilities pulled on July 27. Then I wrote: \u201cAnd the tooling? Well, that\u2019s a tale for another day\u2026\u201d<\/p>\n\n\n\n

Today is that day. Or at least a down payment on that day.<\/p>\n\n\n\n

If you want to skip the deep dive and just start exploring, you can check it out at https:\/\/predictions.qwidgets.com\/compute<\/a>. The platform is free and you don\u2019t even need to register to explore the data.<\/p>\n\n\n\n

Also, please keep three things in mind at this stage:<\/p>\n\n\n\n

    \n
  1. Everything below was captured during drafting on Monday, August 24, 2026. These books roll within days, so read the shapes rather than the decimals.<\/li>\n\n\n\n
  2. All of this is shaped around the context of AI compute, but there\u2019s nothing that ties it to that domain. If there are other scenarios with comparable market coverage then extending support is trivial.<\/li>\n\n\n\n
  3. These markets have thin liquidity and wide spread bands, so a lot of work had to go into handling scenarios where the data would be incoherent (wide spreads causing higher strikes to be priced higher than lower strikes, etc.) so keep an eye out for notes when the data doesn\u2019t make sense.<\/li>\n<\/ol>\n\n\n\n

    Starting at the top<\/h1>\n\n\n\n
    \"\"
    Five GPU underlyings, each with a live realized index from Ornn alongside the level the market prices at even odds.<\/em><\/figcaption><\/figure>\n\n\n\n

    The AI compute landing page provides a quick glance at the supported GPUs along with index data from Ornn and the even-odds level implied by the Kalshi binaries in the nearest term, which is the current week. This even-odds level is where the 50% crossover is estimated based on the underlying threshold market pricing. The final column provides a rough grade of the quality of our even-odds level based on the confidence we have in that value.<\/p>\n\n\n\n

    Working the ladder<\/h1>\n\n\n\n

    A threshold market is a literal question: will an H200-hour settle above $4.50? Its price is the answer. Each column lists the odds of clearing every strike level, which is just a literal read of prices off the chain. It\u2019s so vanilla I\u2019m not even going to bother posting a screenshot of the \u201cChance above\u201d mode. I value your time way too much for that.<\/p>\n\n\n\n

    However, there\u2019s a toggle where you can switch to \u201cChance in band\u201d mode, which is the insight we\u2019re all really looking for. It helps us visualize the distribution the market expects the index to land in on the target date.<\/p>\n\n\n\n

    \"\"
    The same ladder, shaded by the chance of settling inside each band instead of above each strike. Note the September 2026 column divergence.<\/em><\/figcaption><\/figure>\n\n\n\n

    The first thing that I noticed when looking at this view is the September 2026 divergence. Roughly a third of the distribution sits in $3.50\u20134.00 and another third sits in $4.50\u20135.00, with almost nothing in the band between them. Read literally, the market is saying compute either drifts down or jumps up but is unlikely to sit where it is now.<\/p>\n\n\n\n

    That would be a genuinely interesting insight and could be a great trade. Or is it misleading? Stay tuned.<\/p>\n\n\n\n

    Stepping through the tenors<\/h1>\n\n\n\n

    Any single tenor renders as a distribution. The odds of settling above each strike are drawn as a line above and the odds of landing in each band drawn as bars below. The arrows beside the tenor selector step forward and back through every listed settlement, and stepping through them quickly is a great way to see the shape of these distribution expectations evolve over time.<\/p>\n\n\n\n

    \"\"
    September 2026. The two humps are the split from the matrix, and the flat shelf in the line between $4.00 and $4.50 is the tell.<\/em><\/figcaption><\/figure>\n\n\n\n
    \"\"
    August 2027, twelve months out. The distribution has gone nearly flat, and the orange bar is 30.6% of the probability sitting above the highest strike anyone quotes.<\/em><\/figcaption><\/figure>\n\n\n\n

    The near weeks are tight and single-humped. A year out the shape has gone nearly flat, and it\u2019s worth being careful about interpreting what that means. The orange bar is 30.6% of the probability sitting above the highest quoted strike, where there\u2019s no market to read at all. Below it the line is nearly straight, falling about four and a half points per fifty cents from $2.50 all the way to $8.00. The straight line is a uniform distribution, which would put $7.50\u20138.00 on the same footing as $4.00\u20134.50 for a chip trading at $4.27. Every rung is two-sided and quoted around six cents wide, so this isn\u2019t neglect. It just isn\u2019t a traded view either.<\/p>\n\n\n\n

    We\u2019re doing the best we can on the curve<\/h1>\n\n\n\n
    \"\"
    Monthly-average tenors, with the middle 50% of each distribution shaded where the quotes support it.<\/em><\/figcaption><\/figure>\n\n\n\n

    The line is the same story the earlier article told: roughly flat, no dramatic ramp, sitting a little above the realized index. The width is again the real content. Even after discarding the extreme quarter at each end, the middle half of outcomes spans two to three dollars a few months out. This isn\u2019t particularly useful at the moment but should become better as liquidity grows over time.<\/p>\n\n\n\n

    The shading runs out around five months because the quotes get too one-sided to bound a middle. Below the chart the app says which tenor it refused to draw and why. For example, when the top threshold market is \u201cAbove 6.50\u201d and has over 25% probability, there\u2019s no good way to tell exactly where the top 25% point lands.<\/p>\n\n\n\n

    Pricing the stack at the money<\/h1>\n\n\n\n
    \"\"
    One strike, eight instruments. Only the top two are live on Kalshi. Everything else is derived from the same ladder.<\/em><\/figcaption><\/figure>\n\n\n\n

    My earlier article required a bunch of work to produce reference prices for these instruments in the closing list. Now they\u2019re generated and refreshed in real time. The YES and NO contracts are live Kalshi markets. The dated future, the call, and the put are model values derived from the ladder. The perpetuals are marked indicative, which is the app saying it will give you a number but wants you to know how much of it is a modelling choice. The prevailing difference between the confidence levels of \u201cderived\u201d and \u201cindicative\u201d comes down to how much we have to account for the unquoted mass of probability above the top threshold market.<\/p>\n\n\n\n

    Put-call parity still falls out for free: the $4.00 call at $0.4129 less the put at $0.2001 is $0.2128, against a future of $4.21 less a $4.00 strike. Nobody enforced that. It holds because both figures come from the same distribution.<\/p>\n\n\n\n

    \"\"
    The derivation, split into the part that comes from quoted markets and the part that comes from a fitted tail.<\/em><\/figcaption><\/figure>\n\n\n\n

    Every figure is a sum over the quoted rungs. Where an instrument\u2019s payoff runs past the last strike we fit a tail and it shows you how big it is. The dated future shown here is 0.1% estimated tail. The $4.00 call is 1.2%. Both are essentially restatements of quoted markets. A far out-of-the-money call would be mostly tail, and the last column shows how far the number moves under a different tail assumption. This unquoted tail consideration becomes less necessary once we have dated futures trading to produce an independent price to work with.<\/p>\n\n\n\n

    Chasing the split down<\/h1>\n\n\n\n

    Let\u2019s get back to the discussion of the split forecast for September 2026. Every tenor has its own page, so we can drill in for more detail about what\u2019s going on there.<\/p>\n\n\n\n

    \"\"
    The rung table for September 2026. Two adjacent strikes reading the same probability, one of them quoted 41 cents wide.<\/em><\/figcaption><\/figure>\n\n\n\n

    The $4.00 and $4.50 rungs both read 59%, and the $4.50 market is quoted 41 cents wide. The note under the distribution says plainly that two rungs were repaired to make the quotes into a coherent distribution: the raw ladder had $4.50 priced above $4.00, which isn\u2019t allowed. Flattening the inversion to produce plausible prices from within their respective bid\/ask spreads is what emptied the band between them.<\/p>\n\n\n\n

    In other words, the bimodal forecast is not really a prevailing forecast. It\u2019s a 41-cent-wide quote on a thinly traded rung, showing up three views later as a shape that looks like conviction. The platform flagged it twice before I got there: once in the width shown on the rung itself, and again in the note under the distribution saying the rungs had been repaired.<\/p>\n\n\n\n

    \"\"
    The volatility you would have to assume to produce these option premiums.<\/em><\/figcaption><\/figure>\n\n\n\n

    Implied volatility is the one number here that runs backwards. Nothing in this pipeline takes volatility as an input. Since the distribution is read straight off the market, this curve is what you get by inverting the premiums and asking what lognormal volatility would have been required to produce them. That makes it a diagnostic rather than a signal, so don\u2019t read too much into it. A strike showing lower or higher implied volatility isn\u2019t a reason to buy or sell that binary. There is no separate volatility surface here to be rich or cheap against. The curve is just another description of the distribution we were already looking at.<\/p>\n\n\n\n

    Always thinking about perpetuals<\/h1>\n\n\n\n

    The perpetual option was a real pain to price in my earlier article because it needs an assumed anchor tenor. That anchor is a specification rather than a view, and moving it doesn\u2019t refine the price, but rather selects a different contract.<\/p>\n\n\n\n

    \"\"
    A perpetual mark, decomposed into the dated legs behind it, with the funding weight on each. The box at the bottom is the app checking its own coverage.<\/em><\/figcaption><\/figure>\n\n\n\n

    I think this is the clearest expression of what an everlasting option actually is. The mark is a funding-weighted blend of dated legs, and those weights are the contract. The bottom box reports how much of the weight falls past the longest quoted tenor and how many quoted tenors sit inside the anchor, then states whether both are within threshold. It is a coverage check, published.<\/p>\n\n\n\n

    \"\"
    The chain at a one-month anchor, which is one of the settings used in the earlier article.<\/em><\/figcaption><\/figure>\n\n\n\n

    The app defaults to a three-month tenor, but you can drag it to one month to see how everything changes. The $4.50 perpetual put marks $0.2105 and funds at 4.7% a month on the strike. I built that same number by hand off the July 27 pull and got roughly $0.29 and 6% a month. I wouldn\u2019t read much into the gap; the old number came out of a spreadsheet with an assumed tenor and a volatility bridge, while this one is a funding-weighted strip read straight off the ladder. Two different methods landing in the same neighborhood is about as much as that comparison is worth.<\/p>\n\n\n\n

    Turning a view into an optimized position<\/h1>\n\n\n\n

    One of the cool launch features that came with Qwidgets for Prediction Markets was the prediction portfolio optimizer. If you linked a Kalshi account, it allowed you to take a cumulative event (a collection of threshold markets) and define the distribution you wanted to express. It would then model that using a variety of configurable parameters to help you optimize your positions.<\/p>\n\n\n\n

    Suppose I ultimately decided that I wanted to trade the bimodal September 2026 scenario outlined above.<\/p>\n\n\n\n

    \"\"
    Defining a distribution to solve for the optimal binary allocation.<\/em><\/figcaption><\/figure>\n\n\n\n

    The sliders represent my distribution. Let\u2019s say I took that September split at face value and put half my probability in each of the two live bands. The market has them at roughly 32% and 30.5%. Given a $1,000 exposure budget, the optimizer works out which binary legs express that disagreement most efficiently, and the answer is two band trades: long above $3.50 and short above $4.00 for the lower band, long above $4.50 and short above $5.00 for the upper one.<\/p>\n\n\n\n

    It reports a healthy expected return on exposure. It also shows, in the same table, that the $4.50 leg is quoted 0.44 by 0.65 and that the number assumes a fill inside that spread based on how I configured the taker\/maker pricing slider. That is the same 41-cent market from three sections ago, and it is doing most of the work in the headline figure. Which is the honest answer to whether I would put this trade on: not at the ask, and not without working the order.<\/p>\n\n\n\n

    \"\"
    The ticket. Nothing here is a recommendation, and the only tradeable instrument in this entire article is the binary underneath it.<\/em><\/figcaption><\/figure>\n\n\n\n

    Another neat aspect to the portfolio optimizer is that it runs off live data and integrates your actual positions. If pricing changes and you need to rebalance, it makes suggestions with buttons for adjustment trade tickets. Or if you change your view, just move the sliders to express the new one and the changes get teed up for you.<\/p>\n\n\n\n

    Summing it up<\/h1>\n\n\n\n

    My earlier article ended with a warning that still stands: being able to price something does not make it profitable. Nothing above is a claim that the equity options playbook translates to compute, and nothing above is a recommendation.<\/p>\n\n\n\n

    It\u2019s also worth being blunt about what actually exists today. The Kalshi binary ladders are real, live, tradeable markets. The futures, options, perpetuals, and curve are all derived here and not based on external quotes. Multiple exchanges are moving in this space and as compute derivatives are listed I\u2019m very interested to see how real market action compares to what the ladder has been implying all along. I think those instruments will also have a really positive effect on the value of event contracts and may produce the first real hedging scenario that catches on at scale.<\/p>\n\n\n\n

    Please take a look over at https:\/\/predictions.qwidgets.com\/compute<\/a> and let me know what you think. I\u2019m especially interested in whether and how you\u2019d want to see this functionality evolve to support specific trading scenarios.<\/p>\n","protected":false},"excerpt":{"rendered":"

    Or you can keep trading GPU contracts like a Neanderthal. A few weeks ago, I walked through the process of using prediction markets to price a stack of AI compute derivatives. I had just priced a full derivatives stack\u2014a future, a call, a put, and a perpetual\u2014off nothing but the Kalshi H200 binary ladder. There … Continue reading “The AI Compute Trading Tools I\u2019m Hoping You\u2019ve Always Wanted”<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":690,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[8],"tags":[],"_links":{"self":[{"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/posts\/676"}],"collection":[{"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/comments?post=676"}],"version-history":[{"count":1,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/posts\/676\/revisions"}],"predecessor-version":[{"id":691,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/posts\/676\/revisions\/691"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/media\/690"}],"wp:attachment":[{"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/media?parent=676"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/categories?post=676"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/quantcha.com\/news\/wp-json\/wp\/v2\/tags?post=676"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}