Is AI Cooked?
Semis are crashing. Iran closed the Strait of Hormuz. New tariffs just landed. Oil ripped 20% in a week. And everyone is asking the same question: is the AI trade over?
We went looking for the answer in the options market. What we found is more interesting than yes or no.
Every GPU has a supply chain behind it. A long one. It passes through chip design, memory fabrication, lithography, foundry packaging, and ends when someone racks the finished server with enough power and cooling to run it. Each link is a company. And each company has an options market that prices its uncertainty.
Line those prices up and they tell a story. (For the tariff, Hormuz, and crypto positioning backdrop, see our companion piece.)
Click any layer above to see the companies inside it. Green means calm. Red means stressed. The numbers are implied volatility, the options market's price of uncertainty. The deeper into the physical supply chain, the louder it gets.
NVIDIA, the company that designs the GPU: 35% IV. Calm. But Micron, which makes the memory soldered onto every GPU? 76%. ASML, the only company that makes the machines to print chips? 70%. Lam Research: 73%. Vertiv, which cools the data centers: 71%.
If AI demand were dying, NVIDIA would be the most volatile name here. It's the calmest. The stress is upstream: memory, tools, foundry, power. The market isn't scared of the GPU. It's scared of everything required to deliver it.
The Chain
Every AI chip takes the same journey:
- A hyperscaler (Microsoft, Google, Amazon, Meta) places GPU orders
- A chip designer (NVIDIA, AMD) creates the design and sends it to a foundry (TSMC)
- The foundry needs lithography tools (ASML), memory (Micron), and equipment (Lam, Applied Materials)
- The finished chip ships to a data center that needs cooling (Vertiv) and electricity (Vistra, Constellation)
Break any link, and everything downstream stops. The options market knows which links are breaking.
Demand is fine. That's not the problem.
Start with the obvious question. Does anyone still want GPUs?
NVIDIA just reported $62.3 billion in quarterly data-center revenue and guided higher. The hyperscalers buying all those GPUs? MSFT at 29% IV. GOOGL at 34%. AMZN at 36%. All calm.
And credit? HYG at 8.2% IV. SMH at 44.7%. That's a 5.5x gap. If AI were actually dying - if this were a bubble popping or a credit crunch - the bond market would be freaking out. It isn't. The stress is in semis and nowhere else.
So demand is healthy, credit is asleep, and the buyers are fine. That narrows it fast. If the market isn't worried about who's buying GPUs, what is it worried about?
NVIDIA: the calmest name on the board
NVIDIA at 34.9% is trading below its own sector ETF. Think about that. The biggest AI company on the planet, in the middle of a semiconductor selloff, and the options market treats it as the safest name in the stack.
Why? NVIDIA doesn't make anything. It designs chips, licenses CUDA, and collects rent. The bottlenecks - memory shortages, packaging constraints, power grid limits - are everyone else's problem. As long as customers keep ordering, NVIDIA gets paid whether the supply chain is smooth or broken.
AMD at 55.8% tells the other side of that story. Same business, 20 points higher vol. AMD loses orders first when budgets tighten. Gets allocation last when TSMC's lines are full. That 20-point gap? That's what CUDA lock-in is worth.
One layer deeper: memory
This is where things get loud. NVIDIA can design all the chips it wants, but each one needs high-bandwidth memory soldered directly onto the package. Only three companies on the planet manufacture it: Samsung, SK Hynix, and Micron.
Micron at 76.1% IV is the most volatile major name in the AI stack. Micron's own management says demand is "constrained by inadequate DRAM and NAND supply." They literally can't make enough.
But 76% IV for a supply-constrained monopoly? That's high even for a bottleneck. The extra vol is coming from somewhere else: memory is the one part of the AI stack that software can eat.
The dual exposure
Micron may be the only node where a bullish story (physical bottleneck, constrained supply) and a bearish story (software could eat your market) coexist in the same options price. That dual exposure explains why it sits at the top of the vol ladder.
The foundry: where demand becomes product
NVIDIA designs the chip. Micron makes the memory. But someone has to actually build the thing. That's TSMC.
At 52% IV, TSMC is stressed but not screaming. The stress is concentrated around one date: April 16, when they report earnings. The implied move is 7.6%. The market is holding its breath.
Why? Because TSMC's CoWoS packaging lines are the tightest bottleneck in the entire AI pipeline. They're spending $52-56 billion in capex to expand them. That earnings call will answer a simple question: is the expansion on track, or not?
Look at the shapes side by side:
TSMC's vol spikes around earnings and drops after. That's a binary event - the market wants to hear from them before deciding. Micron is high everywhere, ongoing stress with no single catalyst to resolve it. NVIDIA is flat and low, no stress expected at any horizon. AMD is elevated across the whole curve, structural uncertainty that isn't tied to one date.
The IV number tells you how much uncertainty. The shape tells you what kind. And right now, the shapes are telling very different stories at each layer of the chain.
The tools that build the tools
One more layer down. TSMC can't manufacture chips without the machines to print them. The most critical machine in the world is ASML's EUV lithography system. ASML is the only company that makes them. There is no alternative.
ASML at 69.7% IV, nearly double NVIDIA's. The company still carries 13.2 billion euros in Q4 bookings and a 38.8 billion euro backlog. But bookings are promises. The options market isn't sure those promises hold.
ASML reports April 15. One day before TSMC. Implied move: 10.1% - the biggest of any name in this analysis. That's the market saying: ASML's order book is the single most important data point for the entire AI capacity expansion.
Lam Research at 73% and Applied Materials at 62% tell the same story. Equipment makers carry the highest sustained IV in the stack because they're the market's forward bet on whether new capacity actually shows up.
Beyond the chip: power and cooling
The supply chain doesn't end when the chip is packaged. A GPU is worthless without somewhere to run it, something to cool it, and enough electricity to keep it alive.
Vertiv at 70.5% IV tells the story. Vertiv makes cooling systems for data centers. Technically an industrial company. But its options trade like an upstream semi name: high vol, stressed, priced for bottleneck risk. The market reclassified it. Cooling isn't a boring utility anymore. It's a binding constraint.
Vistra (61.1% IV) and Constellation Energy (55.4%) carry the power story. But here's a wrinkle: during the recent window, both names actually moved more than their options implied. Realized vol outpaced implied vol. Options on these names may have been cheap relative to actual moves. The power layer might be where supply-chain stress is most underpriced.
The migrating bottleneck
The bottleneck is shifting. It started in fabrication (can we print enough chips?) and is migrating toward deployment (can we plug them in and turn them on?). The IEA projects data-center electricity demand will more than double by 2030. Virginia grid operators have issued capacity warnings through 2028. The physical plant needed to run AI is becoming a constraint in its own right.
What is actually happening: four shocks at once
If you read the entire move as a tariff story, you miss the structure. The options surface is encoding multiple overlapping risk timelines, each hitting different parts of the supply chain on different schedules.
- The Strait of Hormuz raised the floor on energy risk for every Taiwan-based manufacturer
- TurboQuant reshapes the ceiling for memory demand, potentially reducing HBM requirements by 6-8x
- Tariffs added friction across every border crossing in the chip lifecycle
- ASML and TSMC earnings (April 15-16) will convert uncertainty into information
Each one below.
These aren't competing explanations. They're stacked. Different shocks, different layers, different timelines. The term structure reflects all of them. No single narrative explains the full pattern.
Where the options market goes blind
The options market gives strong signal on GPUs, memory, lithography, foundry, cooling, and power. But signal degrades as you move into the deep upstream: silicon wafers, specialty gases, photoresists, and advanced substrates. Most of these companies are listed in Japan or Europe, trade thinly, or both.
Fundamental research fills the gap here.
What would change the picture
The thesis breaks if:
- NVIDIA reprices above the upstream names. That would mean the market is worried about demand, not delivery. Right now NVDA sits 10+ points below its sector ETF. If that inverts, the framework flips.
- ASML and TSMC deliver clean guidance and upstream vol collapses. Clean bookings, clean capex, clean packaging commentary. If the market exhales after those reports, the bottleneck thesis was pricing more fear than reality warranted.
- Credit starts transmitting stress. HYG at 8% IV says the system is solvent. If that moves toward 20%+, the story shifts from "sector delivery problem" to "systemic risk event."
- The Iran conflict escalates or a prolonged Hormuz closure persists into Q3. Taiwan imports 97% of its energy. A sustained disruption doesn't just raise costs - it threatens TSMC's ability to run fabs at full capacity. "Delivery risk" becomes "production halt," and that cascades through the entire stack. A prolonged war also keeps oil above $100, tightening financial conditions globally and eventually feeding into demand destruction the market isn't pricing yet.
- Software efficiency accelerates. If TurboQuant-style compression becomes standard and memory requirements drop 4-8x, the HBM bottleneck thesis unwinds. Micron's 76% IV could be a leading indicator that the market already suspects this.
- AI capex financing cracks. Credit is calm now. But the buildout is funded with real debt.
The signal isn't in any single name. It's in the gradient between them.
So is AI cooked?
No.
The options market is clear on this. Demand is real. Credit is fine. The buyers are calm. NVIDIA is calm. What's not calm is everything underneath: the memory fabs, the lithography machines, the foundry packaging lines, the power grid, the cooling systems. The market isn't pricing an AI collapse. It's pricing a delivery bottleneck.
That's a very different problem. A demand collapse means the trade is over. A delivery bottleneck means the trade is just harder to execute. And if you know where the bottlenecks are - which is exactly what the vol gradient shows you - that's not a reason to panic. It's a map.
The answer
AI isn't cooked. The supply chain is getting stress-tested. The options market is telling you exactly which links are under pressure, which ones are fine, and which ones are about to report earnings that will settle the question one way or the other.
ASML reports April 15. TSMC reports April 16. Watch the vol gradient after those prints. That's when we'll know.
This analysis is educational. Not financial advice. Options trading involves risk of loss. All implied volatility data as of April 4, 2026. Term structure shapes and implied moves are point-in-time snapshots and will change.