The Compute Constraint Just Moved. Capital Is Rotating down the Stack.

Four Layers, Two Moats

Coverage of Friday centered on Kimi K3 as an existential threat to Cadence and Synopsys. Both stocks confirmed the read with sharp drops. That framing is accurate as far as it goes. It also stops one step short of the question that matters for capital allocators: if the design-tool moat is under siege, where does the displaced capital find a new moat?

The answer, in this tape, is not “somewhere else in software.” It is one floor down. Compute architecture. Advanced packaging. High-bandwidth memory. Interconnect. Power. The physical constraints an intelligent model can automate around but cannot dissolve.

Read the compute stack as four layers.

The model layer sits on top. Frontier models command premium pricing today, but distillation, open-source pressure, and the Kimi K3 demonstration argue that model-level pricing power narrows over the next 18 months.

The design-tool layer is next. Cadence trades at 41 times forward earnings, Synopsys at 24. Those multiples price durability. Friday’s tape argued the market no longer fully believes the durability holds.

The silicon layer, which most desks still treat as one bucket, is actually two. General-purpose GPU pricing is under attack from custom silicon: Google’s TPU program, Amazon’s Trainium, Meta’s MTIA, Microsoft’s Maia. That is one part. The other part is the physical enablers of compute: TSMC’s advanced packaging capacity, HBM memory from Micron and SK Hynix, optical and photonic interconnect. Those cannot be downloaded. Their moat is physical.

The infrastructure layer sits underneath everything. Data centers, transmission, cooling, power. This layer cannot be commoditized because you cannot download a substation.

Two moats hold. Physical enablers and infrastructure. Everything above them is now negotiable.

The tape has been marking the rotation for weeks and no one has said it plainly. Intel is up 160% year-to-date on advanced-node buildout. TSMC raised 2026 capex to a record. HBM supply is oversold through 2028. Data-center power demand is forecast to double toward 100 gigawatts by decade end, per the Department of Energy, which is why utilities with compute exposure have outperformed the broad utility sector.

Meanwhile the bond market stays calm. The 10-year at 4.52% and easing is not the print of a market pricing an overheat or a macro crack. It is the print of a market that has decided the capex cycle is real, financeable, and non-inflationary at the aggregate. Our view: the trade is no longer “own the winner.” It is “own the constraint.” The constraint is now physical.

Alphabet reports after Wednesday’s close. Tesla the same evening. Intel Thursday. What matters is not the earnings beat. It is the direction of the capex guide and the composition of the spend. Alphabet’s cloud capex has been running above $18 billion quarterly. Any acceleration, particularly on TPU cadence or custom silicon disclosure, confirms that hyperscalers are deepening their ownership of the compute stack.

Worth watching: the composition of the guide, not the headline number.

For an investor thinking six to eighteen months out, the operating question is which layer earns its capital cost as compute capability improves, and which does not.

The physical-enabler layer appears to command pricing power that widens as compute demand accelerates. This is packaging capacity, HBM and next-generation memory, optical and photonic interconnect, cooling, and datacenter-grade power. General-purpose GPU tier faces margin pressure from custom silicon adoption at every hyperscaler. Design-tool tier faces margin pressure from automation. Pure-model tier faces margin pressure from open-source distillation.

The invalidation is specific. If hyperscaler capex guides materially lower this week, or if HBM and CoWoS backlogs begin shortening in the next two quarters, the physical-scarcity moat weakens and the thesis pauses. Absent that, capital rotation down the stack is the base case.