Where the AV Stack Is Mispricing Its Own Convexity
A long-dated options map of autonomous-vehicle hardware shows the market charging the most for the layer facing the biggest binary, the least for the layer that cannot lose the race, and hedging none of it.
The most useful thing a ride-hailing chief executive said this year was that autonomy is no longer a software problem. On Uber's calls the framing has been consistent: the driving models have been accelerated by large AI, and the genuinely hard part of commercialization is hardware, building robotaxi platforms cheaply and at scale. If that is the bottleneck, the hardware suppliers are where the value accrues. So this note does one thing. It follows the long-dated options money through the autonomous hardware stack to see where convexity is actually being priced, and where that price looks wrong.
The answer is a clean inversion, plus one finding that frames everything around it. The options market charges the steepest premium for the layer with the most binary outcome and the cheapest for the layer that wins no matter which robotaxi operator wins. And out at the multi-year horizon, almost nobody is paying for protection at all.
The thing that frames everything: nobody is hedging
Pull every Jan 2028 contract across the twelve names in this stack and one number stops you. The long-dated put open interest is, for practical purposes, zero, on every single name. The entire autonomous-hardware complex is positioned one direction out to 2028. Roughly 1.7 million long-dated call contracts sit open across the group; the matching downside book is a rounding error.
That is a regime, not a stock signal. It says the multi-year participants in this theme are uniformly long and untroubled by tail risk, which is exactly the condition that makes a theme fragile if the consensus catalyst slips. It does not tell you to be short. It tells you that cheap downside, where it exists, is being given away, and that the crowd's conviction is running well ahead of its caution.
The implied-vol ladder
Ranked by total long-dated call interest and the cost of that convexity, the stack sorts almost perfectly from durable to speculative. The cheaper the implied vol and the deeper and broader the open interest, the more this reads like positioning rather than gambling.
| Name | Layer | LEAPS call OI | Heaviest line | IV |
|---|---|---|---|---|
| NVDA | Compute | 1,277,000 | $300 (92k), $400 (50k), broad ladder | ~44% |
| RIVN | Platform | 234,800 | $40 (58k, 2.7x), narrow | ~75% |
| TSM | Foundry | 104,200 | $350, $300 (both below spot) | ~54% |
| MBLY | Compute / stack | 28,700 | $30 (5.3k, 3.8x) | ~75% |
| QCOM | Compute | 23,500 | $370 (3.4k, 1.85x) | ~68% |
| ON | Power / sensing | 17,900 | spread, mid-tenor | mid |
| OUST | Perception | 8,600 | $70 (1.4k, 1.7x) | ~115% |
| BIDU | Platform | 5,700 | $150 (680, 1.4x) | ~52% |
| AEVA | Perception | 2,700 | $35 (0.9k, 1.7x) | ~120% |
| NXPI | Sensing semi | 900 | near-money, un-positioned | ~53% |
AV hardware stack, Jan 2028 call positioning. OI is total long-dated call open interest across all strikes. Moneyness is the heaviest single strike versus spot.
The compute layer trades near 44 percent implied vol. The lidar layer trades near 120. The market is paying nearly triple the price of convexity on the part of the stack most likely to be designed out, and a discount on the part that supplies every robotaxi regardless of who wins.
Layer one: compute, the bet that cannot lose the race
This is the cleanest risk-adjusted position in the stack, and it is hiding in plain sight. The compute layer holds the heaviest long-dated positioning anywhere in autonomous hardware by a wide margin: roughly 1.28 million Jan 2028 call contracts on the dominant accelerator name, more than five times the next name in the group. The implied vol on that convexity is only about 44 percent, the lowest in the stack.
What makes it convincing is not the headline number but the shape. The open interest is layered across dozens of strikes, from deep in-the-money stock-replacement lines with delta near nine tenths up through the $300 and $400 convexity rungs. That is the footprint of institutional accumulation building a position in size, not a single speculative print. You are buying multi-year exposure to the layer that powers every autonomous stack, training the driving models in the data center and running inference in the car, at half the implied vol the market charges for the speculative names below it. The instrument that matches the thesis is the slightly in-the-money replacement strike, near spot with a delta around one half to two thirds, rather than the far out-of-the-money lottery lines. The honest caveat is that this is a consensus long and the near-term tape is soft. A long-dated call is precisely the instrument for that situation. It lets the thesis breathe through the chop.
The foundry that prints those chips shows a related but distinct pattern. It carries the second-deepest long-dated book in the group at moderate implied vol near 54 percent, but its heaviest strikes sit below spot. That is deep in-the-money stock replacement, a synthetic long, not a convexity bet. Read it as quiet conviction accumulation rather than a leveraged swing. The automotive sensing semiconductor is the opposite extreme: cheap-ish vol near 53 percent and almost no positioning at all, the little there is sitting near the money. That is what a profitable, under-owned quality name looks like in the options market. Nobody is paying up, because nobody is excited. For a patient holder that is a feature.
Layer two: platform, the lottery and the empty room
The platform layer splits into two opposite mispricings. On one side is the loudest single bet in the entire stack: roughly fifty-eight thousand contracts open at a Jan 2028 strike that sits almost three times above spot on the vertically integrated platform name that a major ride-hailing partner is funding through milestone-gated tranches. The whole book is narrow, concentrated in three out-of-the-money strikes, and flow confirms the conviction in both fresh trades and standing skew. But you pay for it. Implied vol near 75 percent means the crowd has already bid up the exact convexity everyone wants, and the position only pays if the robotaxi ramp arrives on schedule. It is the purest expression of the hardware-bottleneck thesis and the most expensive and binary way to own it.
On the other side is the empty room. The Chinese search-and-autonomy platform that the ride-hailing chief executive singled out by name for pairing capable software with genuinely affordable hardware at scale has cheap stock, moderate implied vol near 52 percent, and one of the thinnest long-dated books in the group. The affordable-hardware optionality management praised is barely expressed in listed LEAPS at all. If you want a non-consensus way to play the platform layer, this is the room nobody is standing in. The reason it is empty, thin liquidity and the overhang that comes with the listing, is also the reason it stays cheap.
The asymmetry, stated plainly
When the crowd pays 75 to 120 percent implied vol for the binary names, leaves the standard-setter at 44, and buys essentially no downside anywhere, it is paying up for the outcomes it can imagine and pricing the near-certain one cheaply: whoever wins the robotaxi war, they run on the same compute and, very probably, the same sensors. Durable value tends to accrue where the convexity is cheap, not where it is loud.
Layer three: perception, paying a fortune for a coin flip
The sensor layer is where the inversion is starkest. Both lidar pure-plays carry long-dated implied vol around 115 to 125 percent, the most expensive convexity in the stack by a wide margin, on thin open interest. One of them saw a speculative call frenzy print on a day the stock fell double digits, the textbook signature of gambling rather than accumulation.
That nosebleed pricing is the options market quoting a genuine coin flip. The whole perception layer rides on one unresolved architectural question: sensor fusion versus vision only. If the camera-only camp is right, and at least two serious autonomy programs are betting it is, much of the lidar bill of materials gets designed out and these names fall hard. If fusion wins, which is how most robotaxi fleets are built today, the unit volumes scale with every car deployed. There is no middle outcome, and the implied vol says so. Paying 120 percent for a coin flip is not an opportunity. It is the market doing its job. The only edge here belongs to someone with a differentiated, defensible view on which architecture wins, and that view is a perception-engineering call, not a flow call.
The cadence read
Strip it to one line. The autonomous hardware stack prices its convexity backwards and hedges none of it. The layer with the most durable claim on the value, compute, is the cheapest to own through long-dated calls and is being quietly accumulated in size and breadth. The layer with the most binary claim, perception, is the most expensive, because the market is honestly pricing a fork it cannot resolve. The platform layer offers a crowded, pricey favorite and an un-positioned contrarian, depending on appetite. And across all of it, the near-total absence of long-dated puts says conviction is running ahead of caution.
For a process that wants three independent signals to agree before sizing, the compute layer is the only place where cheap implied vol, deep and broad institutional open interest, and a thesis that survives any operator outcome line up at once. The speculative tiers can still pay off spectacularly. They simply require you to be right about a specific, contested future, and to pay full price for the privilege. The cadence here is not about chasing the loudest options print. It is about buying the convexity the crowd has left cheap on the part of the stack that gets paid no matter who wins, and respecting that when an entire theme stops buying protection, the easy part of the move is usually behind it.