Regarded Semi

If you are on the buyside and your PM or CIO insists AI is a bubble you should do everything in your power to find a new seat. How deeply you understand bottlenecks and tokenomics will be the primary driver of alpha over the next five years and you should put yourself in a position where you have the leash to be involved and learning instead of watching from the sidelines. We are at the starting line. The people pointing to memory stocks instead of model IQ tipping points simply lack the IQ or imagination to understand that.

@_vincentpaul_ @brandonjcarl I’m not even attacking him. I’m saying the take is low-IQ. “We’re in a bubble because there will be a breakthrough that reduces compute needs” actually means we’re not in a bubble

@brandonjcarl @regardedsemi I’m sure the guy who created an account in october, 2025 to cover semiconductors has no vested interest here, a perfectly objective view, and no need to resort to personal attacks over facts

There’s a lot of discussion in the markets about whether there’s an AI bubble. Here I walk through a definition of a bubble, why people normally miss bubbles, why AI is likely a bubble, and what the implications are.

@jukan05 I'm polling Edgar SEC every 5 seconds 🤣🤣

Release Leopold Aschenbrenner’s damn 13F already!!

@insane_analyst Wait until Aschenbrenner reveals a $WOLF position in his 13-F tonight. Let’s get extra retarded

When people tell me they are concerned $WOLF recently went bankrupt, my response is simple. Its not a bug, its a feature. Imbrace your inner primate. Market is stupid enough as is just go with it. Its either -50% or +300%. 🎲

Prediction: Leopold Aschenbrenner’s 13F discloses a $WOLF position, sending the short squeeze nuclear.

@HotAisle If it comes with cheap inference I’m sold!!!

@regardedsemi Next time just wire money to my bank account. I will make you feel part of something.

Might as well lock in those cbrs losses.

@HotAisle I just wanted to be a part of something!!

@regardedsemi 🤦‍♂️

This did NOT happen ahead of the dot.com

@insane_analyst HOSTILE TAKEOVER INITIATED

Prepare yourself for the most entertaining activist investor campaign of all time.

@insane_analyst Just add a few more 0’s to the Webull and start a hostile take over

A hedge fund told me I live in Andrew Feldman's head. Ima go full activist investor on him with my peanut $50K stake I buy tomorrow at open.

@insane_analyst Dude you called NVDA to TSEM going back to the rings article like over 6 months ago

One guy called TSEM/NVDA over 6 months ago. Not me. One of the locked accounts I follow.

@PythiaR @DMTCapital When we also remember OpenAI and Anthropic is software so they are both zeros too

@DMTCapital When we also remember CUDA is basically software so $NVDA is a zero too

At what point do $AVGO longs remember that VMWare is actually software which means this thing is a zero?

@OmerCheeema I think some type of InP/rare earth for GPU deal gets done.

Jensen is in China with Trump, does it mean we can expect a news of him being able to sell GPUs?

@AlmaCap114204 Buddy, FDV of WOLF is $6B right now. We’re already there.

$WOLF You are buying $5bn (cost to build) of strategic, irreplaceable Western 200mm SiC capacity for c$3bn EV, with utilisation operating leverage alone, no pricing assumption required, sufficient to deliver a 2-3x re-rating as Mohawk Valley fills. Market should price in the cost to build very quickly, then the capacity buildout as and when orders come in.

@citrini What even is this job in the first place?

This is so unbelievably cool.

@blondesnmoney Obviously PPi is +1.6%. We just had the biggest oil spike in history. It’s a nothing burger headline unless it sticks.

PPI is +1.6% and the market is down 8 points unreal

UNDERSTANDING GaN vs. SiC Lately investors new to semiconductors and new to power electronics have been asking me about "GaN vs. SiC." There is no GaN vs. SiC. Power semiconductors are defined by two variables: voltage blocking capability and switching speed. These trade off against each other, and different applications sit at different points on that curve. SiC is the right material for high-voltage, high-power applications where switching speed matters less than raw blocking capability. Utility-scale solar and very large (larger than 50 MWh) battery systems should often be served by SiC because their value driver is primarily the energy content. This is also why traction motors (like EVs and locomotives) can use SiC –– the computation environment is relatively simplistic so the slow speeds are acceptable. SiC devices today block up to 10 kV and switch cleanly at those levels. Nothing else does. The most interesting area of power development, though, is in projects where the value driver is a mix of energy, power, capacity, and harmonics. Like a data center. Or a grid-supporting 5 MWh BESS. Or V2X devices that link electric cars and trucks up to the grid. This requires deft computational power handling. These are ~all of the interesting next-gen power applications. You serve your local need well (charging the car, powering the data center); but you are also performing ultra high-speed grid support activities. GaN owns the sub-3 kV space, and it owns it decisively. At voltages below 1.2 kV — consumer fast chargers, server power supplies, EV on-board chargers, microinverters — GaN switches at frequencies an order of magnitude higher than SiC, in a fraction of the package size, at a cost that follows a steep downward curve driven by HVM in consumer electronics. The 140-watt USB-C laptop charger uses GaN switching at a million times per second. A comparable 1990s power supply was four times larger and ran hot enough to need a fan. That cost and density curve is not stopping. The framework is simple. If your application is above 3 kV and switching speed is secondary to blocking voltage, use SiC. If your application is below 3 kV and you need speed, density, and cost efficiency, use GaN. Between roughly 1.2 kV and 3 kV is contested territory where the answer depends on the specific thermal, cost, and frequency requirements of the design. The cost trajectories of the two materials reflect their origins. SiC crystal growth happens above 2,000°C at a rate of roughly 100 to 300 microns per hour — a thousand times slower than silicon. Each step up in wafer diameter, from 4-inch to 6-inch to the 8-inch transition now underway, requires reinventing the thermal field and crucible design from scratch. 6-inch SiC substrates fell roughly 30% in price in 2024, driven by Chinese overcapacity, and the move to 8-inch wafers should cut per-die costs another 30 to 40% once yields stabilize — but those gains come hard and slow. GaN's cost curve is of different stuff entirely. Because GaN-on-silicon grows on standard substrates, it inherits the economies of the silicon ecosystem. 150mm GaN-on-Si wafer prices have fallen roughly 40% since 2020. Prices have crossed below $1 per transistor for high-volume GaN devices. The industry is now moving toward 300mm GaN-on-Si, which will yield 2x more chips per wafer than 200mm. That cost curve is being pulled by billions of consumer electronics units and the grid gets to ride it for free. GaN does have a ceiling. At medium-voltage grid applications — like 34.5 kV — GaN cannot block the full line voltage directly. This is where GaN challengers go wrong. They presume that the only architecture possible is a direct, centralized, single-device approach to handling large voltages. Not so. The more interesting path is to keep GaN operating within its native voltage range — below 12 kV — and build the medium-voltage function from an array of GaN-based modules. Each module runs at GaN's switching frequencies, which are high enough that the transformer inside each module shrinks to the size of a paperback. The aggregate system handles medium-voltage grid connection without exposing any single device to voltages it can't block. What you lose in architectural simplicity, you gain in speed, redundancy, cost, and manufacturability — because you're building from parts that already exist in billion-unit volumes on a known transistor cost ramp. Stacking GaN will do something consequential: bring semiconductor switching speeds to grid scale. And thus bringing a cost profile that follows true high-volume manufacturing economics rather than industrial green-metal-cabinet economics. But the putative "competition" between these semiconductors isn't a reflection of reality. There isn't a winner, really. The grid needs both materials deployed where each is strongest.

This is for the innumerate clowns who actually still believe there are GPUs sitting in warehouses somewhere $NVDA : “We are typically seeing 4 or more customers competing for every GPU we bring online” $NBIS earnings call: “We are sold out again in Q1, as we have for several quarters as demand continues to significantly exceed available capacity. The vast majority of capacity coming online over the next several quarters to 12 months is already under contract or earmarked for our AI cloud customers. .. We are typically seeing 4 or more customers competing for every GPU we bring online.”